Thermal performance analysis of gas turbine power plant using soft computing techniques: a review

Abstract

Gas turbines are pivotal in electricity generation and industry, prized for their efficiency and flexibility in meeting diverse power needs. Optimizing their thermal efficiency is essential for improving energy output and sustainability. Soft computing methods such as neural networks, genetic algorithms, and fuzzy logic offer potent tools for this optimization due to their ability to handle the turbines' nonlinear and dynamic characteristics. These techniques facilitate a deeper understanding of the intricate interplay among various parameters affecting thermal performance, thereby enabling the development of intelligent and adaptive turbine systems. By leveraging soft computing, researchers can enhance gas turbine designs to align with modern energy and environmental objectives. This review emphasizes the application of soft computing approaches in analysing and improving gas turbine thermal performance. Such advancements are instrumental in achieving higher energy efficiency, reducing greenhouse gas emissions, and promoting a sustainable energy landscape. Ultimately, integrating soft computing into gas turbine operations promises to advance both technical capabilities and environmental stewardship in the detection of a more resilient and efficient energy infrastructure.

KEYWORDS:

• Gas turbine optimisation
• thermal performance analysis
• fuel and emission control
• waste heat recovery
• turbine fault diagnosis
• soft computing techniques

1. Introduction

A gas turbine power plant is a sophisticated and highly efficient energy generation facility that harnesses the principles of thermodynamics to convert fuel into electricity. Also known as a combustion turbine plant, this technology relies on the combustion of a gaseous or liquid fuel to produce a high-speed rotating shaft that, in turn, drives an electric generator. Gas turbine power plants play a pivotal role in meeting the ever-growing global demand for electricity due to their versatility, quick start-up times, and relatively low environmental impact. These plants are widely utilised in various applications, ranging from large-scale electricity generation for grid supply to smaller, decentralised power systems for industrial, commercial, and military purposes. The combined gas turbine power plant represents a cutting-edge and highly efficient approach to electricity generation, seamlessly integrating gas turbine technology with supplementary components to enhance overall thermal performance (Gandhi & Kathirvel, 2021; Latif et al., 2020). Gas turbines use the Brayton cycle, which consists of four major components: the compressor, combustor, turbine, and exhaust. The compressor compresses the entering air, which is then combined with fuel in the combustor and burnt under continuous pressure. The high-temperature, high-pressure gas expands through the turbine, creating shaft work that powers the compressor and external loads. Finally, the exhaust gases are released into the environment. A gas turbine's functioning is determined by numerous fundamental equations that describe the thermodynamic processes involved (Kabengele et al., 2022; Khani et al., 2023). The basic governing equation for the performance of a gas turbine is expressed through the First Law of Thermodynamics. It relates the energy input, energy output, and changes in the internal energy of the working fluid, which is shown in equation (1). (1) ${\dot{Q}}_{\mathrm{in}}-{\dot{W}}_{\mathrm{out}}=\dot{m}\left(h_{\mathrm{out}}-h_{\mathrm{in}}\right)$(1) Where, ${\dot{Q}}_{\mathrm{in}}$ is the rate of heat input to the system, ${\dot{W}}_{\mathrm{out}}$ is the rate of work production by the turbine, $\dot{m}$ is the mass flow rate of the working fluid, $h_{\mathrm{out}}$ is the specific enthalpy of the fluid entering the turbine. The primary concept behind gas turbine functioning is the burning of gaseous or liquid fuels to produce a high-speed rotating shaft that powers an electric generator. Gas turbines can operate on a variety of fuels, including duel-fuel setups, biogas, natural gas, and hydrogen. To optimise performance and economy, each fuel type requires specific design and operational considerations. Duel-fuel systems, for instance, provide flexibility by enabling various fuel sources to be used, increasing resilience and adaptability in reaction to market swings or environmental requirements. Biogas, obtained from organic waste, offers an environmentally friendly alternative, but one with distinct combustion characteristics that necessitate specialised control measures. Natural gas, which is plentiful and relatively clean compared to other fossil fuels, remains a popular choice for gas turbine applications. Hydrogen, with its potential for zero-emission power production, is gaining popularity as a fuel alternative, despite technological obstacles relating to storage, handling, and combustion characteristics (de Castro-Cros et al., 2021; Mohamed & Khalil, 2020; Nallagownden et al., 2020).

This innovative power generation system leverages the strengths of gas turbines, known for their high power density and rapid response capabilities, and combines them with complementary technologies such as steam turbines or heat recovery systems. In the contemporary energy landscape, continuous improvement in the performance of gas turbine power plants is indispensable given the critical role this plays in electricity generation and industrial processes. Gas turbines are favoured for their efficiency, flexibility, and ability to respond rapidly to varying power demands. However, optimising their thermal performance is crucial to enhancing energy output and minimising environmental impact. This necessitates the adoption of innovative methodologies to address the inherent complexities in their operation. To enhance their performance and efficiency, researchers have increasingly turned to soft computing techniques (Hai et al., 2023; Rajagopalan et al., 2022).

1.1. Soft computing and its applications in gas turbines

A branch of computer science called soft computing seeks to create intelligent machines that mimic human decision-making. The combination of several approaches incorporating the ideas of approximation, uncertainty, partial truth, etc. is known as soft computing. Soft computing approaches are distinct from analytical methods due to their capacity for experience-based learning. Evolutionary computing, neural networks, fuzzy computing, machine learning, swarm intelligence, and a few other approaches are the six basic categories of soft computing techniques. Figure 1 presents a branches of soft computing approaches (Falcone et al., 2020).

Figure 1. Branches of soft computing techniques.

Soft computing techniques play a pivotal role in optimising turbine design, offering innovative solutions to enhance performance and efficiency. Adaptive Neuro-Fuzzy Inference Systems (ANFIS) provide a flexible framework to model complex relationships between design parameters, allowing for efficient design space exploration (Assareh et al., 2023). Genetic Algorithms (GA) and Particle Swarm Optimisation (PSO) excel in searching for optimal design configurations by mimicking natural evolution and swarm intelligence (Modu et al., 2023). These techniques enable the optimisation of turbine geometry, blade profiles, and operating conditions, taking into account multi-objective criteria such as efficiency, reliability, and environmental impact. Additionally, the application of machine learning algorithms, such as Artificial Neural Networks (ANN), aids in predicting turbine performance based on various design parameters, facilitating a data-driven optimisation approach (Moradi & Seyedtabaii, 2022). Integrating these soft computing techniques not only accelerates the design process but also contributes to the development of more robust and adaptable turbine systems, addressing the intricacies and uncertainties associated with real-world operating conditions. Employing methods such as fuzzy logic, artificial neural networks, and genetic algorithms enables the development of adaptive and intelligent models that can effectively handle the complex relationships inherent in fuel systems and emission control (Ahmed, Rehan, et al., 2022; Prabakar, 2022).

Genetic algorithms offer an evolutionary approach to searching through a solution space, exploring optimal configurations. By integrating these soft computing techniques, gas turbine filter house optimisation can benefit from enhanced adaptability, accuracy, and efficiency, ensuring optimal performance under various operating conditions and contributing to the overall efficiency of the gas turbine power plant. The adaptive methods, such as artificial neural networks, genetic algorithms, and fuzzy logic systems, offer intelligent solutions for efficiently recovering waste heat and dynamically adjusting parameters based on real-time conditions. In fault detection and diagnosis control, soft computing techniques excel in handling the inherent complexity and non-linearity of gas turbine systems, ensuring accurate and robust diagnostic capabilities (Al-Awad, 2020; Mohan et al., 2023; Reddy et al., 2023). Traditional methods often fall short in dealing with the intricate, non-linear dynamics inherent in these systems. The drive for sustainable energy solutions and the ever-increasing demand for electricity necessitates the continuous enhancement of gas turbine power plants. Using soft computing approaches, optimise design parameters and operating tactics, resulting in increased efficiency, dependability, and environmental friendliness. This research aims to explore the potential of these innovative approaches to revolutionise gas turbine thermal performance analysis and contribute to the advancement of clean and efficient energy generation. The main objectives of this review approach are given below:

• Various soft computing techniques for the optimisation of Gas turbine design, fuel control and emission control are analyzed with their significance and limitations considering the future necessity of improving the capability of these techniques.

• The gas turbine filter house optimisation, fault detection and diagnosis and waste heat recovery system control using soft computing methods are analyzed with their significance and limitations.

The above-mentioned objectives are covered and various soft computing-based control and optimisation strategies have been discussed in this review. The organisation of this study is as follows section 2 takes five directions about this review and its objective then summarises its ideologies, section 3 provides a comparison study, Section 4 summarises the whole review and, section 5 concludes this study and Section 6 includes future scope.

2. Literature survey

In this section, the review has been provided discussing various important factors to be considered for the thermal performance improvement of combined gas turbines using soft computing techniques. The directions for reviewing the soft computing techniques for the optimisation of combined cycle power generation units are shown in Figure 2.

Figure 2. Directions for reviewing the soft computing techniques used for the thermal performance analysis of gas turbine.

This review has been made in five different phases of soft computing-based control and optimisation for, gas turbine design, fuel control and emission control, gas turbine filter house optimisation, fault detection and diagnosis and waste heat recovery system.

2.1 Review on turbine design optimisation using soft computing techniques

Chao Deng et al. (2020) introduced an adaptive neuro-fuzzy inference system (ANFIS) to predict ideal gas-turbine operating parameters. In-depth introductions were given to the fundamental formulations of various operating conditions for gas-turbine configurations. The best gas-turbine configuration was chosen after consideration of the effects of various parameters. The adopted ANFIS model had three outputs, namely fuel consumption, power output, and thermal efficiency, as well as five inputs, including isentropic turbine efficiency (Teff), isentropic compressor efficiency (Ceff), ambient temperature (Ta), pressure ratio (rp), and turbine inlet temperature (TIT). The anticipated outcomes showed that the suggested model established the operational circumstances that had a significant impact on the gas turbine's performance. But, this approach failed to provide any provisions for the control and monitoring of fuel injection in the combustion chamber.

Qi Wang et al. (2021) presented an innovative idea of transfer learning – for the transfer of information from a large-scale, low-fidelity dataset to a small-scale, high-fidelity dataset. A pre-trained network called a Conditional Generative Adversarial Neural Network was created in order to regress the surface pressure distributions via transfer learning. A detailed comparison was conducted between an independent model and two models that were translated from datasets with varying levels of accuracy. The outcomes shown that the suggested strategy successfully decreased the modelling expense while forecasting the surface pressure distributions with a low error rate. Better generalisation performance of the model transferred from the higher-fidelity dataset resulted in a 40.2% and 9-fold reduction in root mean square error and modelling cost, respectively. However, aligning feature representations between the source (large-scale, low-fidelity) and target (small-scale, high-fidelity) datasets, impacts the model's ability to transfer knowledge accurately and leading to suboptimal performance in certain cases.

Zuming Liu and Karimi (2020) created two comprehensive HDMR and ANN models to capture GT part-load and full-load performance. It was discovered that all surrogate models accurately represented GT's overall part-load and full-load performance using private operational data. The quality of the holistic ANN models was comparatively higher than that of the other surrogate models. Additionally, the best forecasts for the maximum power generating capacity were provided by ANN-3 and HDMR-3. However, the lack of interoperability makes it difficult to extract meaningful insights into the physical relationships and mechanisms governing gas turbine performance.

Zuming Liu and He (2020) provided an exergo-economic optimisation for a combination system consisting of a GTMHR and an ORC, with the ORC recovering the GTMHR waste thermal energy. The system performance is optimised by utilising a simulation-based optimisation method that determines the design parameters as well as the components and their respective compositions for the ORC working fluid at the same time. Parametric research is carried out to investigate the effects of major design factors on system performance. Furthermore, the helium temperature and pressure ratio are discovered to have the greatest influences on system performance. The integration of complicated systems such as the GTMHR and ORC may provide maintenance and operational logistics problems.

Yu-Zhi Chen et al. (2021) suggested a unique performance diagnostic approach that 13 divides engine diagnosis into a number of steps to remove the ‘smearing effect’ and 14 reduce matrix dimensions in the iterative diagnostic algorithm. In order to evaluate the accuracy and computational performance of the suggested method, an engine performance model of a 15 triple-shaft gas turbine was constructed and evaluated against commercial software. With a smaller set of 21 measurements, the newly discovered approach delivers an accurate diagnosis. However, it is critical to determine whether the method is scalable to various engine sizes and complexities.

Homam Nikpey Somehsaraei et al. (2020) explored an interdisciplinary approach to replace the laborious and prone to mistake manual technique by evaluating a machine-learning-based methodology for autodetecting outliers from actual data. The raw data came from tests conducted in Norway on a 100-kW micro gas turbine test rig. Density-Based Spatial Clustering of Applications with Noise (DBSCAN), a component of the suggested approach, was used to identify and remove outliers. Artificial neural networks (ANNs) were developed using the filtered information as a baseline to forecast the system's typical performance for monitoring applications. The filtering approach given is fast and dependable, reducing the amount of time and resources needed for data processing, according to the results. However, this technique requires fine-tuning and will impact the algorithm's accuracy and robustness across diverse datasets or operating conditions.

Alirahmi et al. (2021) investigated, and optimised a unique hybrid system made up of compressed air energy storage (CAES), organic Rankine cycle (ORC), and solid oxide fuel cell (SOFC). To determine the best system performance and design, the grey wolf multi-objective optimisation (MOGWO) technique is used in this situation. To do this, multi-objective optimisation is applied to a trained neural network that was fed into the MOGWO algorithm as a fitted function. The recommended method's main advantage was that it saves time. The thermodynamic performance of the suggested system was examined at three different times – full-time, charging, and discharging – from the perspectives of energy, exhaustion, economy, and environment (4E). However, the optimisation technique assumes that the system's behavior is stationary or does not change significantly over time which is not valid in dynamic conditions.

Ahmed, Alvi et al. (2022) proposed a novel soft computing optimisation method for complicated non-convex machine dynamic economic dispatch problem (DEDP) with many constraints. This approach used sequential quadratic programming (SQP) to fine-tune the pre-optimised run of the genetic algorithm (GA) as the first optimiser. In spite of the valve point loading effect (VPLE) and multiple fuelling option (MFO) constraints, the GA-SQP simulation analysis performed well by acquiring less computational cost and finite time of execution, while providing optimal generation of powers according to the targeted power demand and load. However, the sequential quadratic programming (SQP) method is susceptible to getting stuck in local optima which converges to a suboptimal solution.

Yinghao Zhao and Foong (2022) predicted the electrical power (PE) output of Combined cycle power plants (CCPPs) using new soft computing techniques. To this goal, the suggested hybrid is made by combining an artificial neural network (ANN) with a metaheuristic method known as the electrostatic discharge algorithm (ESDA). To examine the impact of hybridisation, its performance is contrasted with many conventionally trained ANNs. Through a 4 x9x 1 network, the PE is anticipated by taking into account the impact of ambient temperature, exhaust vacuum, atmospheric pressure, and relative humidity. However, the CCPPs exhibit complex behaviours under diverse circumstances that are not fully captured by the selected input features.

Rodríguez et al. (2022) introduced an adaptive neuro-fuzzy inference system (ANFIS) to simulate a Combined Cycle Gas Turbine (CCGT) to identify and recognise its operational variables. The pressure ratio (compression ratio) employed in the gas cycle, the pressure of the steam extraction used to warm feed water, and the heat lost to the outside in the steam turbines were the three input factors whose effects were examined with regard to the combined cycle's thermal efficiency. It was demonstrated that the gas cycle's pressure ratio had the most impact on efficiency. The steam extraction ideal pressure of 62.38% was found, which corresponds to the cycle's maximum efficiency. However, the heat loss was not accurately measured which led to the variation in the pressure ratio calculation.

Aquize et al. (2023) proposed a systematic approach to constructing and developing black box models using NARX neural networks (Nonlinear Autoregressive Networks) to establish a reliable and resilient GT model. The proposal outlines 9 clear phases from selecting GT factors to generating a successful NARX model. This study demonstrated that the suggested NARX model creation technique accurately identified and predicted GT output characteristics based on system input changes. The approach accurately anticipated the behaviour of similar gas turbine systems, as it is not affected by the type of turbine. Furthermore, the method's success is strongly dependent on precise and extensive data collecting, which is time-consuming and not always viable in different situations or for distinct turbine types.

In this Table 1, various design techniques have been explored for predicting and optimising the performance of gas turbines and combined cycle power plants (CCPPs). The studies encompass a range of methodologies, including adaptive neuro-fuzzy inference systems (ANFIS), transfer learning, high-dimensional model representation (HDMR), exergo-economic optimisation, electrostatic discharge algorithm (ESDA) coupled with artificial neural networks (ANN), and machine learning-based outlier detection. Notably, these techniques aim to address different aspects of performance prediction, optimisation, and diagnostics in the context of gas turbines and CCPPs. However, several limitations are identified across the studies, such as the inability to account for control and monitoring of specific parameters, challenges in aligning feature representations in transfer learning, lack of interoperability in surrogate models, and issues related to scalability and accuracy in diagnostic approaches. Additionally, assumptions of system behaviour being stationary and susceptibility to local optima are recognised in optimisation studies. Hence, there is a need for more comprehensive and standardised approaches, particularly in handling diverse operational conditions and improving the accuracy and efficiency of prediction and optimisation models for gas turbines and CCPPs. Top of Form.

Table 1. Review on gas turbine design optimisation using soft computing.

Ref. No.	Technique	Significance of technique	Limitation
Deng et al. (2020)	Adaptive Neuro-Fuzzy Inference System (ANFIS)	Predicting gas turbine operating parameters	Lack of provisions for control and monitoring of fuel injection in the combustion chamber
Wang et al. (2021)	Transfer Learning (Conditional Generative Adversarial Neural Network)	Reducing modelling expense and improving generalisation	Suboptimal performance due to challenges in aligning feature representations between source and target datasets
Liu and Karimi (2020)	High-Dimensional Model Representation (HDMR), Artificial Neural Network (ANN)	Capturing gas turbine part-load and full-load performance	Lack of interoperability for extracting meaningful insights into the physical relationships governing gas turbine performance
Liu and He (2020)	Exergo-Economic Optimisation, Simulation-Based Optimisation	Optimising a combination system of GTMHR and ORC	Integration of complex systems may pose maintenance and operational logistics problems
Chen et al. (2021)	Performance Diagnostic Approach	Dividing engine diagnosis into steps to reduce ‘smearing effect’	Scalability concerns regarding applicability to various engine sizes and complexities
Nikpey Somehsaraei et al. (2020)	Machine Learning (Density-Based Spatial Clustering of Applications with Noise – DBSCAN, Artificial Neural Networks – ANNs)	Autodetecting outliers and forecasting typical performance	Requires fine-tuning and impacts algorithm accuracy and robustness across diverse datasets or operating conditions
Alirahmi et al. (2021)	Hybrid System Optimisation (CAES, ORC, SOFC), Grey Wolf Multi-Objective Optimisation	Optimising a unique hybrid system's performance and design	Assumption of system behaviour being stationary may not be valid in dynamic conditions
Ahmed, Alvi, et al. (2022)	Soft Computing Optimisation (Genetic Algorithm – GA, Sequential Quadratic Programming – SQP)	Solving a complicated non-convex machine dynamic economic dispatch problem	Susceptibility to local optima and convergence to suboptimal solutions for the sequential quadratic programming (SQP) method
Zhao and Foong (2022)	Hybrid (Artificial Neural Network – ANN, Electrostatic Discharge Algorithm – ESDA)	Predicting electrical power output of CCPPs	Incomplete capture of complex behaviours of CCPPs under diverse circumstances with the selected input features
Rodríguez et al. (2022)	Adaptive Neuro-Fuzzy Inference System (ANFIS)	Simulating Combined Cycle Gas Turbine (CCGT) operational variables	Inaccurate measurement of heat loss leading to variation in pressure ratio calculation
Aquize et al. (2023)	NARX neural networks	NARX model creation technique accurately identified and predicted GT output characteristics.	The method's success is strongly dependent on precise and extensive data collecting.

2.2 Review on fuel system optimisation and emission control

Amirhossein Hasanzadeh et al. (2022) introduced a suitable solid oxide fuel cell scale combined with a GT and a thorough comparative analysis between 10 MW GT and hybrid SOFC-GT systems was carried out using energy, exergy, exergo-economic, and environmental (4E) criteria. For the comparative study, three distinct scenarios – baseline operation, parametric behaviour, and ideal conditions – were taken into account. The systems were modelled using an appropriate artificial neural network (ANN), and the ANNs that were produced were then fed into the Grey Wolf optimiser. A multi-objective optimisation issue with energy efficiency, CO2 emission, and unit cost of the product as the objectives was solved using the Grey Wolf Optimiser. However, it has a more intricate setup that makes it more challenging to manage properly.

Shiyu Yang et al. (2023) introduced an integrated optimisation approach to concurrently optimise building electrification technologies and DER design and control that spans many timescales. To replace and approximate the computationally intensive control optimisation, a unique deep learning-based model for predicting building operational performance was created. This contributed to the solution of the difficult, multi-timescale integrated design and control optimisation issue that is computationally intractable. The experimental findings showed that the suggested framework was successful in reducing carbon emissions at a reasonable cost when applied to a residential structure. However, the limitation of this method is its potential lack of generalizability to diverse building types and locations due to reliance on a deep learning model trained on specific residential structures and geographical regions.

Zuming Liu and He (2020) provided an exergo-economic optimisation for a combination system consisting of a GTMHR and an ORC, with the ORC recovering the GTMHR waste thermal energy. The system performance is optimised by utilising a simulation-based optimisation method that determines the design parameters as well as the components and their respective compositions for the ORC working fluid at the same time. Parametric research is carried out to investigate the effects of major design factors on system performance. Furthermore, the helium temperature and pressure ratio are discovered to have the greatest influences on system performance. The integration of complicated systems such as the GTMHR and ORC may provide maintenance and operational logistics problems.

Jo-Han et al. (2019) focused on the emissions characteristics of syngas combustion in a gas turbine, as well as the associated cost to reduce emissions using the selective catalytic reduction (SCR) method and gas mixes. The syngas was made up of four gases: H2, CO, CO2, and CH4. The syngas mixture is burned at an equivalency ratio (ER) of 0.4-0.9. Using Design of Experiments (DOE) optimisation approaches for simultaneous NOx-CO reduction, the emission indices of the NOx and CO pollutants for the best H2-rich syngas (ER = 0.5) were found to be 0.0189 and 0.0028 g/kWh lower than that of the best CO-rich syngas (ER = 0.5), respectively. The combustion of syngas in gas turbines may lead to heightened corrosion and erosion of turbine components, jeopardising overall durability and lifespan due to the presence of impurities and contaminants in the syngas.

Wang, Feng, et al. (2022) proposed a novel precooling fuel to address the issues of excessive carbon emission and inadequate heat sink of traditional hydrocarbon fuel, as well as the problem of excessive aircraft volume caused by the low density of traditional liquid hydrogen. Three models were created to assess the performance of engines. According to the data, ammonia had a larger heat sink than other fuels in the same volume. Taking into account this higher heat sink, the dual-fuel precooling system that combines ammonia and n-decane would optimise engine performance. Moreover, for the multi-objective optimisation of engine performance and carbon emission, the ideal fuel ratio is 2.78 and the ideal pressure ratio is 12.49. However, ensuring safety and regulatory compliance, would be crucial for the successful real-world application of the proposed dual-fuel precooling system.

Tao Hai et al. (2022) used effective ejector-based organic cycles to harness the chemical energy in biomass to produce electricity, heat, and cold. The gasification process, an externally fuelled gas turbine that acted as the system's upper cycle and prime mover, and an organic flash cycle with an ejector were the three separate subsystems that made up the system. The steam Rankine cycle, whose condenser is the evaporator of the ejector-based organic cycle, was then heated by the hot gases that escape. Multi-criteria optimisation, parametric research, and 3E analysis were carried out using ANN and TOPSIS methodologies. The results show that the properties of biomass have a quantifiable effect on system performance. However, the inherent variability in biomass composition and quality poses challenges in accurately quantifying its impact on the performance of the effective ejector-based organic cycles.

Xezonakis and Ntantis (2023) examined how an artificial neural network may be used to increase the accuracy of real measurement data in a thermal power plant model. Algebraic expressions were processed using neural networks, and they are trained using a MATLAB-implemented Feed-Forward Back Propagation technique. Three distinct algorithms were used in the applied training example in a thermal power plant in Paracha: the Levenberg-Marquadt, the Scaled Conjugate Gradient, and the Bayesian Regularisation. These algorithms take into account fewer samples in order to produce more accurate results. However, this algorithm faces generalizability issues when applied to combined cycle power plants.

Wang, Ma, et al. (2022) proposed a technical, financial, and environmental examination of a hybrid system based on solid oxide fuel cells (SOFCs) that includes a proton exchange membrane electrolyzer, biomass, and gas turbine. Based on an improved version of Aquila Optimiser (DAO), a multi-objective optimisation approach has been applied to optimise the total cost of the product and the energy effectiveness. The primary goal of utilising the produced version is to raise the original Aquila optimiser's accuracy and precision. Next, the energy-economic efficiency and energy-perturbative efficacy of the system are verified. However, the optimal configuration generated by the Aquila Optimiser is highly sensitive to the input parameters and operating conditions.

Abdalla et al. (2023) created models based on optimal clustering data to integrate fuzzy subtractive clustering (FSC) and accelerated particle swarm optimisation (APSO) to perform the behaviour of CHP-PV (Combined Heating and Power – Integrated with Photovoltaic). The purpose of the APSO method was to use a proportional–integral (PI) controller to adjust the parameters of the data clustering-based FSC. The primary objective of the article is to minimise the overall energy and fuel consumption while maintaining the cooling load requirement. The suggested model interacted with the energy produced by PV systems and gas turbine generators (GTGs). Additionally, it used the partial load condition to split the cooling burden based on the external weather. However, the emission contents of the combined power plant were not analyzed in this optimisation.

Yingqi Xia et al. (2024) The proposed two-stage deep learning structure and three representative deep learning networks (Recurrent Neural Network, Long Short-Term Memory network, and Gated Recurrent Unit network) are applied to build electricity-carbon models for carbon dioxide emission estimation using real-time data from a Chinese power plant. The experimental findings show that models from three different deep learning networks exhibit a significant improvement in estimation accuracy when using the two-stage structure as opposed to conventional models. The two-stage structure lowers the mean squared error (MSE) for models spanning these three networks at the ideal window size. However, the model's effectiveness relies on continuous and accurate monitoring of electricity consumption and key equipment data.

Yousif et al. (2024) provided a hybrid model that combined the Feed Forward Neural Network (FFNN) model and the PSO algorithm to estimate gas emissions from natural gas power plants. The FFNN predicted gas turbine NOx and CO emissions, whereas the PSO optimised FFNN weights to improve prediction accuracy. The PSO employed a novel random number selection technique that included the K-Nearest Neighbour (KNN) algorithm to decrease prediction errors. Neighbour Component Analysis (NCA) selected factors that were most strongly connected with CO and NOx emissions. The hybrid model produced good prediction accuracy, especially when the PSO parameter selection was optimised using seed random generators. However, the emission contents of the combined power plant were not analyzed in this optimisation.

Faqih et al. (2023) presented a semi-supervised method for predicting the operational range of Dry-Low Emission (DLE) gas turbines. The prediction model was created by combining XGBoost and K-Means algorithms with actual DLE gas turbine data. 15 parameters, including operating and emission concentrations, were analyzed. The XGBoost model estimated turbine combustion temperature, NOx, and CO emissions. The expected output was given into the K-Means model, which predicted the operational regions. As a consequence, the model's capacity to react to dynamic and unpredictable circumstances in real-time operation restricts its accuracy in forecasting the operating range.

The literature survey from Table 2 presents a wide array of techniques employed in the optimisation and analysis of energy systems. These techniques include Artificial Neural Networks (ANN), Grey Wolf Optimiser, deep learning-based models, simulation-based optimisation, Design of Experiments (DOE) optimisation approaches, multi-objective optimisation, and fuzzy clustering methods. The significance of these studies lies in their contributions to addressing various energy-related challenges, such as the optimisation of hybrid systems, reduction of carbon emissions, and effective utilisation of biomass. However, common limitations across the studies include intricate system setups, potential challenges in generalizability, and sensitivity of optimal configurations to input parameters. Challenges also arise from the integration of complicated systems, such as potential operational and maintenance logistics problems. There is a need for new approaches that focus on improving the robustness and adaptability of models, considering real-world complexities, and addressing the challenges associated with sensitivity and generalizability. The quest for innovation in methodologies is crucial to advancing the optimisation and sustainability of energy systems in diverse applications.

Table 2. Review on fuel system optimisation and emission control.

Ref. No.	Techniques Used	Significance	Limitations
Hasanzadeh et al. (2022)	ANN, Grey Wolf Optimiser	Comparative analysis between GT and hybrid SOFC-GT systems, Multi-objective optimisation	Intricate setup, potentially challenging to manage properly
Yang et al. (2023)	Deep learning-based model, Optimisation approaches	Integrated optimisation of building electrification technologies and DER design, Reduction of carbon emissions at a reasonable cost	Potential lack of generalizability to diverse building types and locations
Liu and He (2020)	Simulation-based optimisation, Parametric research	Exergo-economic optimisation for a combination system of GTMHR and ORC, Investigation of major design factors	Integration of complicated systems may provide maintenance and operational logistics problems
Jo-Han et al. (2019)	Design of Experiments (DOE) optimisation approaches	Emissions characteristics of syngas combustion in a gas turbine, Simultaneous NOx-CO reduction	Syngas combustion may lead to heightened corrosion and erosion of turbine components
Wang, Feng, et al. (2022)	Multi-objective optimisation, Three models	Dual-fuel precooling system for excessive carbon emission and inadequate heat sink, Ideal fuel ratio and pressure ratio	Safety and regulatory compliance crucial for real-world application
Hai et al. (2022)	ANN, TOPSIS, Multi-criteria optimisation, Parametric research	Ejector-based organic cycles for biomass utilisation, Quantifiable effect of biomass properties on system performance	Inherent variability in biomass composition and quality poses challenges in accurately quantifying its impact on performance
Xezonakis and Ntantis (2023)	Artificial Neural Network (ANN), Feed-Forward Back Propagation	Increase accuracy of real measurement data in a thermal power plant model, Three distinct algorithms used for training	Generalizability issues when applied to combined cycle power plants
Wang, Ma, et al. (2022)	Multi-objective optimisation, Improved Aquila Optimiser (DAO)	Technical, financial, and environmental examination of a hybrid system based on SOFCs, Optimal configuration highly sensitive to input parameters	Dependence on real-time data availability and quality for effective optimisation
Abdalla et al. (2023)	Fuzzy subtractive clustering (FSC), Accelerated Particle Swarm Optimisation (APSO)	Optimal clustering for CHP-PV behaviour, Minimisation of overall energy and fuel consumption	The emission contents of the combined power plant were not analyzed in this optimisation
Xia et al. (2024)	Two-stage deep learning structure, Recurrent Neural Network (RNN), LSTM, GRU	Electricity-carbon models for CO2 emission estimation, Significant improvement in accuracy with the two-stage structure	The model's effectiveness relies on continuous and accurate monitoring of electricity consumption and key equipment data
Yousif et al. (2024)	Feed Forward Neural Network (FFNN) model and the PSO algorithm	FFNN predicted gas turbine NOx and CO emissions.	The emission contents of the combined power plant were not analyzed in this optimisation.
Faqih et al. (2023)	XGBoost and K-Means algorithms	XGBoost model estimated turbine combustion temperature, NOx, and CO emissions.	The model's capacity to react to dynamic and unpredictable circumstances in real-time operation restricts its accuracy.

2.3 Review on filter house optimisation

Sabah Ahmed Abdul-Wahab et al. (2020), proposed a prognostic approach based on a MISO (multiple inputs and single output) fuzzy logic approach for modelling and simulation of clogging of the gas turbine filter house. In this artificial intelligence-based methodology, nine real-time process variables (ambient temperature, humidity, ambient pressure, GT-produced load, inlet guide vane position, airflow rate, wind speed, wind direction, and PM10 dust concentration) were fuzzified using a graphical user interface. The findings showed that the proposed fuzzy logic model, with a very high determination coefficient of 0.974, produced very small deviations and demonstrated superior predictive performance over the traditional multiple regression methodology. However, additional experimental data is also needed to be provided for the validity of the implemented deep learning strategy.

M. Gul et al. (2020), proposed simple optimisation methods of grey Taguchi and ANN to enhance gas turbine performance. By using the Grey-Taguchi method to optimise various levels of input process parameters, the industrial gas turbine's horsepower, SFC, and heat release were to be increased. Since air inlet temperature is the main factor influencing the machine's performance characteristics, an evaporative cooling system was used to regulate it, which was a necessary modification for a hot climate. ANOVA analysis also showed that ‘air-inlet-temperature’ is the dominant process parameter and ‘type of air-inlet-filter’ is the least effective one, with 71.17% and 1.40% impacts on the gas turbine's output parameters, respectively. The aforementioned ideal configuration had enhanced gas turbine performance in terms of higher thermal efficiency, increased horsepower, decreased heat rate, and low SFC of natural gas. However, the cleaning and maintenance approaches for the air inlet filter for the gas turbine were not discussed in this technique.

Sabah Ahmed Effiom et al. (2023), proposed a prognostic approach based on a MISO (multiple inputs and single output) fuzzy logic approach for modelling and simulation of clogging of the gas turbine filter house. In this artificial intelligence-based methodology, nine real-time process variables (ambient temperature, humidity, ambient pressure, GT-produced load, inlet guide vane position, airflow rate, wind speed, wind direction, and PM10 dust concentration) were fuzzified using a graphical user interface. The findings showed that the proposed fuzzy logic model, with a very high determination coefficient of 0.974, produced very small deviations and demonstrated superior predictive performance over the traditional multiple regression methodology. However, additional experimental data is also needed to be provided for the validity of the implemented deep learning strategy.

Zhengbo Zou et al. (2020) developed DRL training environments by approximating real-world HVAC operations using long-short-term-memory (LSTM) networks, with the goal of achieving optimum control over Air Handling Units (AHUs). For the best possible control over the AHUs, the framework was also made to incorporate cutting-edge DRL techniques, such as deep deterministic policy gradient. Three AHUs served as assessment testbeds, each having two years’ worth of building automation system (BAS) data. Using the BAS data from the first year, the LSTM-based DRL training environments were constructed. These settings produced an average mean square error of 0.0015 for normalised AHU parameters. However, n order to achieve a more complete optimum control policy for the AHUs, the indoor air quality indicators need to be added to the present incentive function and control action design.

Ahmad Abubaker et al. (2021) proposed a unique integration of cascaded solar heat exchangers and a combined cycle power plant The CCPP was utilised to remedy the GT's deficiencies at the same time. PTCs (parabolic trough collectors) were used to warm the air at the combustion chamber input. The collectors were then employed to power an absorption inlet-air cooling cycle, which will regulate the air temperature at the compressor's inlet. Following that, a linear-regression LR-based optimisation was undertaken to develop extremely accurate polynomial equations for predicting the system's performance. The integration introduces additional components such as the parabolic trough collectors and the absorption cooling cycle, increasing the overall complexity of the system.

Tunckaya (2021) used the live process data of Erdemir BF#2 in Turkey, which was collected chronologically, to design a unique top gas pressure tracking system. According to the recommendations of the plant maintenance team, eight process parameters were taken into consideration as input parameters. Soft computing techniques, artificial neural networks, and adaptive neuro-fuzzy inference systems were used, and autoregressive integrated moving average, a statistical regression tool, was also used for comparison. The coefficient of determination (R2), mean absolute percentage error, and root mean squared error parameters are used in performance and success ratio analysis. However, this study does not explicitly address the impact of external factors, such as environmental conditions or variations in raw material properties, which can significantly influence the performance of blast furnace processes.

Deng et al. (2020) predicted the ideal gas-turbine operating parameters using multiconfiguration gas-turbines and the adaptive neuro-fuzzy inference system (ANFIS). A detailed introduction was given to the basic formulations of gas-turbine setups with different operating circumstances. The best gas-turbine arrangement was chosen by analysing the impact of various factors. The selected ANFIS model includes three outputs: fuel consumption, power output, and thermal efficiency, and five inputs: ambient temperature, pressure ratio, ambient temperature, isentropic turbine efficiency, and turbine inlet temperature (TIT). Using the ANFIS model, real reported data from Iraq's Baiji Gas-Turbines as well as simulated data were used. The expected model findings are, however, significantly influenced by the pressure ratio, ambient temperature, and isentropic turbine efficiency.

Nand Gopal and Dilbagh Panchal (2023) proposed an innovative fuzzy modelling technique based on the Jaya Lambda-Tau Optimisation (JLTO) approach to tabulate several BU Boiler Unit reliability indicators. Concepts from the Fuzzy Set (FS) theory were used to take the ambiguity and uncertainty of the data that was gathered into account. Furthermore, Failure Mode and Effect Analysis (FMEA) was carried out utilising qualitative data gathered from expert opinion to increase the boiler unit's availability. The Fuzzy-Complex Proportional Assessment (FCOPRAS) decision-making technique was linked with the FMEA approach to highlight the riskier failure reasons contributing to a drop in system availability and to overcome the inadequacies of the standard FMEA approach-based conclusions. The ranking results were further compared using the Fuzzy Combinative Distance-based Assessment (FCODAS) technique to enable precise decision-making. The availability of the boiler unit decreases as the uncertainty level rises.

Huiyong et al. (2023) proposed a simulated annealing-grasshopper optimisation algorithmtion algorithm support vector machine to establish the overall simulation model of the air circulation system of the aircraft and to conduct fault injection analysis. By introducing the support vector machine to classify the results of the system and applying of grasshopper algorithm to optimise the support vector machine with methods such as simulated annealing and position migration, the optimal parameter values were obtained. The results indicate that the simulation system can effectively simulate the temperature changes of the aircraft in various operating states. However, the incorporation of a computational fluid dynamic approach would help improve the analysis.

Wood (2020) proposed a transparent open box (TOB) machine-learning approach that evaluates the electrical power output (PE) of a combined cycle gas turbine (CCGT) to offer precise PE forecasts. Compared to published predictions for the dataset from fifteen correlation-based machine learning methods, the PE predictions obtained by using the TOB optimised data matching methodology are more accurate. By employing a tuning subset of less than 150 (∼1.5%) data records, TOB was able to attain this high accuracy. Testing the optimised solutions against every dataset record in 15 runs distributed over five shuffled datasets verifies its correctness. The four independent variables in the dataset contain a few severe outliers, which have a detrimental effect on machine-learning algorithms’ ability to make accurate predictions. However, a deeper investigation into the specific characteristics of the TOB method is needed to improve the forecasting accuracy.

Yang et al. (2021) proposed a novel method of controlling the fuzzy PID controller settings by hybridising the improved particle swarm optimisation algorithm with the cuckoo search algorithm (HIPSO_CS). First, by linearly reducing both the particle count and the inertia weight value, the classical particle swarm optimisation technique was enhanced. Second, to increase the variety of particles, the particle swarm optimisation method incorporated the local random walk approach of the Cuckoo algorithm. By contrasting it with conventional optimisation methods, the proposed HIPSO_CS algorithm's quick iteration speed and excellent convergence accuracy were confirmed. This enhances the controller's resilience and dynamic performance. yet, it also lengthens computation times and makes simulation analysis and experimentation more challenging.

The literature survey from Table 3 encompasses a diverse set of techniques employed in the optimisation and predictive modelling of various energy systems. These techniques include MISO fuzzy logic, Grey-Taguchi, artificial neural networks (ANN), deep reinforcement learning (DRL) with long-short-term-memory (LSTM) networks, linear regression (LR), and Jaya Lambda-Tau Optimisation (JLTO) coupled with Failure Mode and Effect Analysis (FMEA). The significance of these studies lies in their ability to enhance the performance, reliability, and predictive accuracy of energy systems, addressing challenges such as clogging in gas turbine filter houses, optimal control over air handling units, and improvement of gas turbine efficiency. However, common limitations include the need for additional experimental data, a lack of discussion on maintenance approaches, potential impacts of external factors, and challenges in generalizability. These limitations underscore the necessity for a new approach that integrates more comprehensive datasets, considers a wider range of external factors, and focuses on improving the robustness and adaptability of the models to dynamic operating conditions. The quest for new approaches is driven by the aim to further optimise energy systems, enhance reliability, and overcome the identified limitations for more effective real-world applications.

Table 3. Review on filter house optimisation techniques.

Reference	Techniques Used	Significance	Limitations
Abdul-Wahab et al. (2020)	MISO Fuzzy Logic	Proposes a prognostic approach for gas turbine filter house clogging using MISO fuzzy logic, achieving high determination coefficient and superior predictive performance.	Requires additional experimental data for the validity of the implemented deep learning strategy.
Gul et al. (2020)	Grey-Taguchi, ANN	Uses Grey-Taguchi and ANN for gas turbine performance optimisation, emphasising the dominant role of air-inlet temperature.	Cleaning and maintenance approaches for the air inlet filter are not discussed.
Effiom et al. (2023)	MISO Fuzzy Logic	Similar to (Xezonakis & Ntantis, 2023), introduces a prognostic approach for gas turbine filter house clogging using MISO fuzzy logic, demonstrating superior predictive performance.	Calls for additional experimental data to validate the implemented deep learning strategy.
Zou et al. (2020)	DRL, LSTM, Deep Deterministic Policy Gradient	Develops DRL training environments for optimal control over Air Handling Units using LSTM networks, showcasing accurate predictions.	Needs to incorporate indoor air quality indicators into the incentive function for a more complete optimum control policy.
Abubaker et al. (2021)	Linear Regression (LR)	Integrates cascaded solar heat exchangers with a combined cycle power plant, employing LR for accurate polynomial equations, addressing GT deficiencies.	Introduces additional components, increasing overall system complexity.
Tunckaya (2021)	Soft Computing, ANN, ANFIS	Designs a top gas pressure tracking system using live process data from Erdemir BF#2, employing soft computing techniques for accurate predictions.	Does not explicitly address the impact of external factors, potentially limiting the model's applicability to dynamic operating environments.
Deng et al. (2020)	Multiconfiguration Gas Turbines, ANFIS	Predicts ideal gas-turbine operating parameters using multiconfiguration gas turbines and ANFIS, highlighting the influence of pressure ratio, ambient temperature, and isentropic turbine efficiency.	Model findings are significantly influenced by certain parameters, potentially limiting generalizability.
Gopal and Panchal (2023)	Jaya Lambda-Tau Optimisation (JLTO), FMEA	Introduces a fuzzy modelling technique based on JLTO for tabulating BU Boiler Unit reliability indicators, enhancing availability through FMEA and FCOPRAS decision-making.	The availability of the boiler unit decreases as the uncertainty level rises.
Huiyong et al. (2023)	Simulated Annealing, Grasshopper Optimisation	Utilizes a simulated annealing-grasshopper optimisation algorithm for fault injection analysis in the aircraft air circulation system.	The incorporation of computational fluid dynamics could enhance the analysis.
Wood (2020)	Transparent Open Box (TOB) Machine Learning	Proposes a TOB machine-learning approach for precise PE forecasts in a CCGT, demonstrating higher accuracy compared to other methods.	Severe outliers in the dataset may affect the accuracy of machine-learning algorithms; a deeper investigation into TOB characteristics is needed for improvement.
Yang et al. (2021)	Hybridising the improved particle swarm optimisation algorithm with the cuckoo search algorithm	Adaptively adjust the parameters of the PID controller.	It lengthens computation times and makes simulation analysis and experimentation more challenging.

2.4 Review on waste heat recovery system control in gas turbine

Kasra Mohammadi et al. (2020) suggested a novel triple power cycle in which waste heat from a gas 12 turbine cycle is used sequentially to drive a supercritical carbon dioxide (s-CO2) recompression cycle 13 and a recuperative organic Rankine cycle (ORC). To analyze the performance of the 15 proposed cycles under different operating conditions, a thorough thermoeconomic 14 model was built and implemented in MATLAB. The particle swarm 16 optimisation (PSO) algorithm was used to minimise the levelised cost of electricity 17 (LCOE) and identify the optimum cycle design conditions. Rapid fluctuations in demand or other disturbances might have an impact on the system's transient response, thereby affecting its overall performance.

Kamarulhelmy Talib et al. (2018), introduced a performance prediction model for a naturally aspirated, spark-ignited engine with a waste heat recovery system (WHRM) that works with a steam turbine. An artificial neural network (ANN) is used in the simulation method to forecast the power generated by this WHRM. The simulation was run using an automated neural network (ANN) that used multilayer perceptron, a feed-forward neural network architecture with unidirectional full connections between successive layers, and iterative techniques to train the data using the Broyden-Fletcher-Goldfarb-Shanno algorithm. The experimental results showed that when using ANN, power generated from this WHRM was predicted with good accuracy and less error with training and testing. However, the data were separated data that were not used previously to train the ANN so, better neural network training is needed to reduce the errors further.

Hongqiang Ma et al. (2022) suggested a new flue gas waste heat recovery system and control scheme based on phase-change heat transfer theory and intelligent control technology. It was made up of air heat transfer systems (AHTS) and flue gas heat transfer systems (FGHTS). The AHTS blower frequency regulates both the system pressure and the temperature of the flue gas exit. A dynamic control approach based on enhanced particle swarm optimisation (PSO) is suggested for FGHTS. When compared to other inertia weight strategies, the inertia weight strategy with the inverted S-shaped function (IWS-ISF) can improve the capacity to search globally in the early stages and locally in the latter stages. However, the fitness function and inertia weight have an impact on the PSO algorithm's convergence and optimisation capacity.

Alirahmi et al. (2023) presented a unique but useful technique for dragonfly optimisation that concurrently considered four competing goals. The suggested technique was used in a hybrid system that combined waste heat recovery units with a solid oxide fuel cell (SOFC). To reduce the computation time, a function-fitting neural network is created to integrate the system's thermodynamic model with the dragonfly method. The optimisation results show that, in comparison to the design condition, the optimal parameters produce a large amount of power, have a higher energy efficiency, and have lower product prices and CO2 emissions. The sensitivity study shows that fuel utilisation factor and current density have a considerable impact on performance indicators, however, turbine inlet temperatures of power cycles are ineffectual.

Tari et al. (2023) considered a 150 m² off-grid flat in a five-story building for waste heat recovery analysis. Since the building was situated in a remote place, all of the needs of the tenants were met without the need of gas and power from the grid. Reverse osmosis (RO) desalination systems are used to produce water. The HVAC system is designed to simultaneously provide the lowest CO2 emissions and the highest level of thermal comfort for the occupants. the investigation and comparison of two distinct sources – solar collector and waste heat from diesel generators – for the building's domestic hot water supply. The PV panel installed on the roof is taken into account for powering the building's energy needs. However, each apartment has a limited amount of rooftop space, and the diesel generator was regarded as the building's backup power source which was not sufficient to meet the demand.

Manente and Costa (2020) developed a new s-CO₂ layouts to maximise heat extraction from the heat source while maintaining the high thermal efficiency that was intrinsic to the system. The most promising ones have half heating, dual expansion, and dual recuperation. A systematic technique based on the superimposition of simple thermodynamic cycles was used to conceptually develop these unique s-CO₂ arrangements. The single flow split with dual expansion (also known as cascade), partial heating, and dual recuperated cycles were broken down into basic Brayton cycles to determine the components needed to achieve high performance in the use of waste heat sources. However, this complex layout induces additional initial costs when applied to a real-time scenario.

Jamali and Noorpoor (2019) introduced a revolutionary multi-generation energy system that based on renewables. A double-effect absorption chiller, an organic Rankine cycle, an ejector refrigeration cycle, a proton exchange membrane electrolyzer, an amine-based CO2 collection system, and a heater form the foundation of the system. A biomass combustor, photovoltaic thermal solar panels, and waste heat recovery from an Iranian cement company in Abyek power the proposed integrated system. This creative energy system arrangement removes CO2 from the cement plant's flue gas while also producing electricity, cooling, heating, and hydrogen in the summer and winter. Thermoeconomic, energetic, and energetic analyses were performed on the system. However, this multiple integration leads to operational instability and a degradation in renewable energy utilisation.

Hani et al. (2022) analyzed a combined power system to recover energy from a regenerative supercritical Brayton cycle. An Engineering Equation Solver (EES) simulation algorithm was constructed through the thermodynamic study of the system. Conventional exergy analysis, on the other hand, may be used to identify areas with greater irreversibility rates but cannot forecast the portion of the exergy destruction rate that can be disregarded in various components. Advanced exergy analysis, which takes into consideration technical limitations with thermodynamic principles, can offer useful information about many elements of exergy destruction rate in each component. However, the turbine has the primary exergy destruction rate in the regenerative supercritical Brayton cycle, according to conventional exergy analysis.

Dubey and Mishra (2021) The stack flow unit of the combined gas turbine and steam turbine (GT-ST) plant are connected with the LiBr-H2O-based, environmentally friendly, and low energy consumption vapour absorption refrigeration system (VARS). The predicted VARS can generate an effective cooling effect if the stack flow temperature is kept between 200 and 250 °C. The primary findings of this study have been reached about component energy losses, combined plant energy efficiency, and total plant energy efficiency. The GT system's combustion chamber (CC) and the ST system's exhaust flow have been shown to have the highest irreversibility at 66.8% and 13.4%, respectively. The entire performance is influenced by the efficiency of the GT plant. However, the heat loss factor that represents the most energy deterioration was not effectively addressed when a cycle was coupled.

Pati et al. (2021) introduced a new approach to energy management using multiple-stage evaporators (MSE). The best choice of unknown steady-state process parameters, such as liquid flow rates and vapour temperatures, yields the highest efficiency of MSE. A significant improvement in energy efficiency has been attained by the integration of multiple energy reduction strategies (ERSs). A series of nonlinear mathematical models for different ERSs are developed and converted into optimisation problems for energy optimisation. The energy efficiency is assessed by computing the best process parameters using three nature-inspired algorithms: GA, DE, and PSO. However, the optimisation needs to be further improved to handle the non-linearities effectively.

Turja et al. (2024) focused on the comparison and optimisation of three supercritical carbon dioxide (sCO2) Rankine cycles for gas turbine waste heat recovery. The study begins with a parametric analysis to determine the important effects of key factors such as turbine intake temperature, condenser inlet temperature, and pinch point temperature on the thermal performance of advanced sCO2 power cycles. A multi-objective optimisation technique was used to determine the optimal cycle design. This method used a Genetic Algorithm in conjunction with machine learning regression models (Random Forest, XGBoost, Artificial Neural Network, Ridge Regression, and K-Nearest Neighbours) to forecast cycle performance using a dataset derived from cycle simulations. The TOPSIS approach made it easier to make decisions on which cycle arrangement was best. Integrating diverse cycle topologies with multistage turbines and compressors increases the complexity and possibility of system inefficiencies.

Kareem et al. (2023) presented a Taji gas turbine power plant integrated Rankine and organic cycles to generate more energy and lower environmental emissions through waste heat recovery from its present plant. The innovative triple trigeneration cycle was explored from a thermodynamic and economic perspective. The whole cycle's processes were parametrically studied and optimised with thermodynamic principles and EES software. The combustion chamber loses the greatest exergy, whereas pumps destroy the least amount relative to other portions of the cycle. However, the rise in condenser temperature reduces the First- and Second-Law efficiencies while increasing the system's overall cost rate.

The reviewed literature in Table 4 employs various waste heat recovery techniques such as thermoeconomic modelling, particle swarm optimisation (PSO), artificial neural networks (ANN), phase-change heat transfer theory, dragonfly optimisation, and nature-inspired algorithms like genetic algorithm (GA), differential evolution (DE), and particle swarm optimisation (PSO) for waste heat recovery systems and energy management. The challenges and problems addressed in these studies include the optimisation of levelized cost of electricity (LCOE), accurate prediction of power generation, control scheme design, system efficiency, and addressing issues related to operational instability, degradation in renewable energy utilisation, and handling nonlinearities in optimisation. The necessity for a new approach arises from the need to enhance overall system performance, improve global and local search capabilities in control systems, address complex trade-offs between system complexity and efficiency, and further optimise energy management strategies to handle nonlinearities effectively.

Table 4. Review on waste heat recovery system control in gas turbine.

Reference	Techniques Used	Significance	Limitations
Mohammadi et al. (2020)	Thermoeconomic Model, Particle Swarm Optimisation (PSO)	Introduces a novel triple power cycle with waste heat recovery, optimised using PSO to minimise the levelized cost of electricity (LCOE).	Rapid fluctuations in demand or disturbances might impact transient response and overall performance.
Talib et al. (2018)	Artificial Neural Network (ANN)	Develops a performance prediction model for a spark-ignited engine with waste heat recovery, using ANN for accurate power forecasting.	Separated data for training and testing the ANN requires better neural network training to reduce errors further.
Ma et al. (2022)	Phase-change heat Transfer, Particle Swarm Optimisation	Proposes a flue gas waste heat recovery system with control based on phase-change heat transfer and PSO, aiming for improved global and local search capabilities.	The fitness function and inertia weight impact PSO algorithm's convergence and optimisation capacity.
Alirahmi et al. (2023)	Dragonfly Optimisation, Neural Network	Presents a dragonfly optimisation technique for a hybrid system with waste heat recovery and solid oxide fuel cell, optimising for power, energy efficiency, cost, and emissions.	A sensitivity study shows that certain parameters have a considerable impact on performance, while turbine inlet temperatures are ineffectual.
Tari et al. (2023)	Reverse Osmosis, HVAC System, PV Panel	Analyzes waste heat recovery in an off-grid flat using reverse osmosis and HVAC systems, considering solar and waste heat for hot water supply and powering energy needs.	Limited rooftop space for PV panels and reliance on a diesel generator as a backup power source may constrain meeting demand.
Manente and Costa (2020)	s-CO2 Layouts, Thermodynamic Cycles	Develops unique s-CO2 layouts maximising heat extraction, using a systematic technique based on the superimposition of simple thermodynamic cycles.	The complex layout introduces additional initial cost when applied to real-time scenarios.
Jamali and Noorpoor (2019)	Multi-Generation Energy System, Thermoeconomic Analysis	Introduces a multi-generation energy system based on renewables, utilising various technologies, but acknowledges operational instability and degradation in renewable utilisation.	Multiple integrations in the system may lead to operational instability and reduced renewable energy utilisation.
Hani et al. (2022)	Supercritical Brayton Cycle, Advanced Exergy Analysis	Analyzes a combined power system with a regenerative supercritical Brayton cycle, comparing conventional and advanced exergy analyses.	Conventional exergy analysis identifies the turbine with the primary exergy destruction rate, and advanced analysis considers technical limitations but cannot forecast disregarded exergy.
Dubey and Mishra (2021)	Vapour Absorption Refrigeration System, Combined GT-ST	Connects a vapour absorption refrigeration system with a combined gas turbine and steam turbine plant, emphasising component energy losses and overall plant efficiency.	The heat loss factor representing the most energy deterioration is not effectively addressed when the cycle is coupled.
Pati et al. (2021)	Multiple-Stage Evaporators, Genetic Algorithm (GA), etc.	Introduces multiple-stage evaporators (MSE) for energy management, optimising energy efficiency using GA, DE, and PSO.	Optimisation needs improvement to handle nonlinearities effectively and achieve better results.
Turja et al. (2024)	Multi-objective optimisation technique	TOPSIS approach made it easier to make decisions on which cycle arrangement was best.	Integrating diverse cycle topologies with multistage turbines and compressors increases the complexity and possibility of system inefficiencies.
Kareem et al. (2023)	Thermodynamic and exergoeconomic assessment	Lower environmental emissions and	Multiple integrations in the system may lead to operational instability and reduced renewable energy utilisation.

2.5 Review on fault detection and diagnostics

Choayb Djeddi et al. (2021), proposed an adaptive neural-fuzzy inference system to provide solutions for monitoring and predicting the degradation based on an approach of faults detection and isolation of the components in the gas turbine system. With this method, the faults and defects affecting this rotating machine can be located and immediately diagnosed. As a result, the architecture of the adaptive neuro-fuzzy inference system's module, which is linked to a decision support system to increase turbine system availability, is used to plan diagnostic actions to maintain the turbine system's availability while ensuring a compromise between safety and operating costs. To achieve maximum availability and prevent any breakdowns or failures in the turbine's components, corrective measures are taken while the machine is in operation or, in the event of serious faults, unplanned operational shutdowns of the turbine. However, this approach failed to focus on the thermal characteristics that determine the ability and performance of the gas turbine.

Nayeri et al. (2022) designed an ensemble-based hierarchical classifier for the foundation of a Fault Detection and Isolation (FDI) system to identify and isolate twelve common turbine defective situations from a healthy scenario. Based on eight researched classifiers’ best-practice confusion patterns – which range from traditional to cutting-edge techniques – the hierarchical classifier was created. The proposed hierarchical classifier improved the accuracy of healthy scenario classification by 10% while discriminating five defective scenarios with almost perfect accuracy. When dealing with unlabelled data, in particular, a suitable concept of confidence rate helped assess the classifier's conclusions in addition to the classifier itself. However, these Ensemble-based classifiers, including hierarchical ones, heavily rely on the quality and representativeness of the training data.

Montazeri-Gh and Yazdani (2020) proposed the FDI system, which consists of a bank of IT2FLSs that have been taught to identify and analyze the state of an industrial gas turbine under various operating situations. Train and test data are created for this purpose by adding mechanical fault signs to the mathematical model of the gas turbine. After that, the Fuzzy Rule Base is created using Interval Type-2 Fuzzy C-Means (IT2FCM) clustering, and the IT2FLS parameters are optimised using a metaheuristic approach. When compared to typical Type-1 Fuzzy Logic Systems, Interval Type-2 Fuzzy Logic Systems have a higher computational complexity.

Tsoutsanis et al. (2023) introduced a unique approach to gas turbine performance diagnostics using artificial neural networks (ANN) in conjunction with a dynamic engine model. A series of artificial neural networks (ANNs) was trained to identify engine deterioration during transitory situations, when all of the engine's components were undergoing degradation. An engine model of a two-shaft gas turbine was constructed in MATLAB/Simulink. The carried out case studies took into account different scenarios of deterioration. The suggested approach has the benefit of efficiently handling both fixed and time-evolving deterioration. Moreover, the model bridged the gap in instances when there is insufficient data to train ANNs by simulating a multitude of scenarios. However, the effect of component interactions was not effectively considered for simulation.

Zhao et al. (2023) presented a unique hierarchical diagnostic technique to get insight into the physical systems, automatically generate hierarchies, and recommended classification structures, instead of changing and using current ML methods for regression or classification. The suggested hierarchical diagnostic approach was assessed using a high-bypass, two-spool turbofan engine from the NASA model. The results of NASA's blind test case indicated that Kappa With a coefficient of 0.693, the suggested hierarchical diagnostic approach outperforms the other diagnostic techniques in the publicly available literature by at least 0.008 points. However, considering variations in sensor configurations, fault modes, and operational conditions that are specific to various types of aircraft engines is essential.

Montazeri-Gh and Nekoonam (2022) proposed a bank of online sequential extreme learning machines (OSELMs) for a component failure diagnosis system, which can be updated gradually with an infinite number of additional training samples. This system was designed to work with an industrial gas turbine with two shafts that produced electricity. Additionally, utilising a hybrid technique based on the variable length genetic algorithm and extreme learning machine (VLGA-ELM), an ideal set of measures was chosen as the system's input. By examining the confusion matrix of each system, the effectiveness of the OSELM-based diagnostic system was contrasted with several batch learning methods. However, this approach introduces computational complexity, particularly during the optimisation process.

Cheng et al. (2023) introduced a surrogate technique to achieve real-time gas route defect diagnosis of GTs under transient operating circumstances, which improved the diagnostic speed. The component level model (CLM) was constructed and validated first. The required physical model and the artificial neural network (ANN) were then combined to create the surrogate model. Nearly all of the CLM simulation findings can be replicated by the built surrogate model under all operating situations. Lastly, the unscented Kalman filter (UKF) and surrogate model were combined in the real-time defect diagnostic system. However, the surrogate model's diagnostic accuracy is compromised if it encounters operating conditions or gas route defects that significantly differ from those present in the training data.

Talebi et al. (2022) suggested an ANN-based diagnostics system to identify and isolate problems in a variety of part loads and deterioration in the face of uncertainty. A Micro Gas Turbine off-design model was created, and uncertainties were taken into account to create an extensive training database. To comprehend the nonlinear link between data and the health condition of components, an artificial neural network was used. To reduce the total number of measures needed, several sets of measurements are evaluated. Power, shaft speed, exhaust temperature, compressor discharge pressure, and temperature were needed to show the proper fault isolation utilising these factors. However, diagnostic performance is highly sensitive to the variety of degradation severities but insensitive to the load variety included in the database.

Yu-Zhi Chen et al. (2022) provided an altered version of the sequential diagnostic technique to present a fresh way to measure component degradation, such as fouling and erosion. The suggested approach might be used to diagnose aero engine faults in real time under both dynamic and steady-state settings. When the engine under examination runs on hydrogen, the economic effects of engine deterioration on fuel costs and payload revenue were also assessed. When compared to a benchmark diagnostic approach, the suggested method showed 15% more accuracy, which can be extremely beneficial given the expense associated with deterioration. However, the fuel quality also needs to be analyzed to fully capture the component degradation.

Dan Liu et al. (2022) developed a new fault diagnosis framework that combined clustering-based downsampling with a deep Siamese self-attention network (CBU-DSSAN) to decrease the number of typical training samples through clustering and improve the capability of fault feature extraction through the multi-head self-attention mechanism. To balance the normal and fault classes, first clustering-based down sampling was limited to the normal samples, and the cluster centres were combined with the fault samples to form the training data set. Second, the original data set was mapped by the Siamese network into an embedded feature space where normal samples and fault samples from distinct classes are separated by great distances. However, the clustering-based down-sampling used in this approach introduces an extra layer of complexity in terms of data pre-processing.

Sarwar et al. (2024) offered a hybrid model for IGT engine fault diagnosis and detection (FDD). As a preliminary stage in pre-processing, the multi-sensor monitoring data was initially combined using principal component analysis (PCA). To improve accuracy and optimise the use of gas turbine information, the PCA technique extracted insights from raw data and optimised the combination of several conditions monitoring datasets. Later, the fused multiple sensors monitoring data was subjected to the ANN-based FDD approach. This research additionally incorporated a comparative description of pattern classification evaluations for supervised and unsupervised ANN learning approaches, such as multilayer perceptron and self-organising maps. Using real-time data from actual operating settings, the suggested model was verified and tested to help achieve early FDD with the least amount of error. The work is not as applicable to real-world industrial applications because it depends too much on labelled sensor data for training, which makes it impractical in situations where labelled data is limited.

Montazeri-Gh et al. (2021) provided a unique method for diagnosing gas turbine faults that was used to simultaneously find, isolate, and diagnose gas route problems. This method was based on combining a residual compensation extreme learning machine (RCELM) and a developing neural gas network to learn the fault characteristic mappings of gas turbine components. The health parameter vector was first estimated by training a bank of RCELMs. Next, the GNG network was employed as a tool to learn the topology of the maps such that each network neuron represents a probable gas turbine health state and a certain degree of degradation. Finally, by connecting the RCELM network to the GNG model, the defect diagnostic procedure was finished. The primary drawback is the compressor's decreased fault detection competency as a result of common fault patterns, which makes precise identification more difficult.

The provided reviews in Table 5 cover a diverse range of diagnostic techniques and frameworks for various types of engines, including gas turbines and aero engines. These include adaptive neural-fuzzy inference systems, ensemble-based hierarchical classifiers, FDI systems with IT2FLSs, artificial neural networks with dynamic engine models, unique hierarchical diagnostic approaches, banks of online sequential extreme learning machines, surrogate techniques, and ANN-based diagnostic systems. While these approaches demonstrate advancements in fault diagnosis, common limitations include computational complexity, sensitivity to operating conditions, and challenges related to the quality and representativeness of training data. The studies emphasise the importance of innovative models and hybrid approaches for enhanced diagnostic accuracy, with a focus on addressing these shared limitations to ensure robust performance in real-world applications.

Table 5. Review on fault detection and diagnostics in gas turbine.

Reference	Techniques Used	Significance	Limitations
Djeddi et al. (2021)	Adaptive neural-fuzzy inference system	Locating and diagnosing faults in gas turbines, Decision support system for planning diagnostic actions	Lack of focus on thermal characteristics
Nayeri et al. (2022)	Ensemble-based hierarchical classifier	Improved fault detection and isolation, 10% increased accuracy in healthy scenario classification	Heavy reliance on quality and representativeness of training data
Montazeri-Gh and Yazdani (2020)	FDI system with IT2FLSs	Identification and analysis of gas turbine states, IT2FCM clustering for rule base, Metaheuristic optimisation	Higher computational complexity compared to Type-1 Fuzzy Logic Systems
Tsoutsanis et al. (2023)	ANN with dynamic engine model	Efficient handling of fixed and time-evolving deterioration, Simulation of scenarios with insufficient data	Ineffective consideration of component interactions in simulation
Zhao et al. (2023)	Hierarchical diagnostic technique	Automatic generation of hierarchies, Outperforms other diagnostic techniques, Kappa coefficient of 0.693	Essential consideration of variations in sensor configurations, fault modes, and operational conditions
Montazeri-Gh and Nekoonam (2022)	Online sequential extreme learning machines (OSELM)	Gradual updates, Hybrid technique (VLGA-ELM), Contrast with batch learning methods	Introduces computational complexity, especially during optimisation
Cheng et al. (2023)	Surrogate technique with CLM and ANN	Real-time gas route defect diagnosis, Improved diagnostic speed	Diagnostic accuracy compromised with significantly different operating conditions or defects
Talebi et al. (2022)	ANN-based diagnostics system	Identification and isolation in various part loads and deterioration, Consideration of uncertainties	Sensitivity to degradation severities, Insensitivity to load variety in the database
Chen et al. (2022)	Altered sequential diagnostic technique	Real-time aero engine fault diagnosis, Economic assessment	Requires fuel quality analysis for comprehensive component degradation understanding
Liu et al. (2022)	CBU-DSSAN framework	Clustering-based down sampling, Multi-head self-attention mechanism	Introduction of complexity in data pre-processing due to clustering-based down sampling
Sarwar et al. (2024)	ANN-based FDD	To improve accuracy and optimise the use of gas turbine information.	The work is not as applicable to real-world industrial applications because it depends too much on labelled sensor data for training
Montazeri-Gh et al. (2021)	Residual compensation extreme learning machine	A quick and dependable diagnostic system that handles fault isolation, identification, and detection operations together is made possible due to this strategy.	Reduced fault diagnosis proficiency for the compressor, due to similar fault patterns making accurate identification more challenging.

3. Comparison of performance of various soft-computing techniques for gas turbine plant optimisation

Various soft computing methods used in the existing studies for thermal performance improvement have been compared in this section with respect to different parameters such as waste heat recovery efficiency, fault detection accuracy, optimisation response time and emission control performance.

Figure 3, compares the turbine design optimisation time for different soft compouting techniques such as ANFIS, CGAN, HDMR-ANN, ANN-DBSCAN and GWO (Liu & He, 2020; Liu & Karimi, 2020; Reddy et al., 2023). From this comparison, it is understood that the ANFIS has the least time for optimisation of 12 sec and GWO takes more time about 25 sec for optimising the design parameters. The CGAN takes an optimisation time of 21 sec which is more that of HDMR-ANN and ANN-DBSCAN approaches.

Figure 3. Comparison of optimisation response time.

The emission control performance of various soft computing techniques such as ANN, APSO, RNN, LSTM and GRU (Ahmed, Alvi, et al., 2022; Hai et al., 2022; Wang, Feng, et al., 2022) have been compared and presented in Figure 4. Among all these techniques, the APSO has a higher performance of 88% in emission control and next to this, the GRU has an emission control efficiency of 84%. Other techniques have relatively low performances and the ANN is found to have a lower performance of 74%.

Figure 4. Comparison of Emission control performance.

The waste heat recovering efficiency of different soft computing approaches (Abubaker et al., 2021; Talib et al., 2018; Tunckaya, 2021) have been compared in figure 5. The PSO has a higher waste heat recovery efficiency of 48% compared to other techniques and the IES-ISF has a lower efficiency of 29% in recovering the waste heat from gas turbine exhaust.

Figure 5. Comparison of Waste heat recovery efficiency.

Figure 6 depicts the comparison of fault detection accuracy of different soft computing techniques such as ANFIS (Alirahmi et al., 2023), OSELM (Dubey & Mishra, 2021), ANN-UKF (Pati et al., 2021) and CBU-DSSCAN (Djeddi et al., 2021). From this comparison, it is found that the ANFIS has a higher fault detection accuracy of 97% and the CBU-DSSCAN falls next to this with an accuracy of 95%. The OSELM has a lower fault detection accuracy of 86%.

Figure 6. Comparison of Fault detection accuracy.

Overall, the performance of various soft computing techniques in the optimisation and control of the thermal performance of gas turbines have been compared. From these comparison graphs, it is found that the particle swarm optimisation has a higher waste heat recovery efficiency of 48%, the ANFIS approach has a higher fault detection accuracy of 97% also it has a lower optimisation response time of 12 sec and finally, in the case of emission control performance, the APSO performs better compared to other approaches with an efficiency of 88%.

3.1 Co-occurrence analysis

The co-occurrence analysis of keywords in the supplied research work sheds light on the important themes and their links inside the thermal performance study of gas turbine power plants utilising soft computing methods. By studying keyword frequency and co-occurrence patterns, they discover important areas of concentration as well as the interdependence of various components of the research. As a result, a bibliometric analysis of the keyword co-occurrence analysis was studied using the program VOSviewer. Co-occurrence analysis of gas turbine optimisation nature keywords was used to better visualise the research subject. Figure 7 depicts the findings of the hierarchical cluster analysis in terms of repetitions. The size of each node (circle) indicates the number of papers containing the linked phrase. These topics are interrelated, as represented by the lines linking them, implying that optimising gas turbine performance necessitates taking into account the interdependence and trade-offs between these many factors.

Figure 7. Analysis of co-occurrence by keywords.

4. Result and summary

Various soft computing techniques for the improvement of the thermal performance of gas turbine power plants have been reviewed in this paper with five different directions such as turbine design, fuel system and emission control, filter section optimisation, waste heat recovery system and fault detection and the summary is as follows:

• Various design techniques, including ANFIS, transfer learning, HDMR, exergo-economic optimisation, ESDA with ANN, and machine learning-based outlier detection, have been explored for gas turbines and CCPPs. While addressing aspects of performance prediction and optimisation, studies reveal limitations such as neglect of control parameters, challenges in transfer learning representation alignment, and scalability issues in diagnostic approaches. Optimisation studies note assumptions of system stationarity and susceptibility to local optima. A call for more comprehensive and standardised approaches is emphasised to enhance accuracy and efficiency in diverse operational conditions for gas turbines and CCPPs.

• The diverse optimisation techniques in energy systems, including ANN, Grey Wolf Optimiser, deep learning, simulation-based optimisation, DOE, multi-objective optimisation, and fuzzy clustering. These approaches contribute significantly to addressing challenges like hybrid system optimisation, carbon emission reduction, and biomass utilisation. Common limitations include complex system setups, potential generalizability issues, and sensitivity to input parameters. The integration of intricate systems poses operational and maintenance logistics challenges, emphasising the need for innovative approaches to enhance model robustness, adaptability, and real-world applicability in advancing energy system optimisation and sustainability.

• The various optimisation techniques for energy systems, including MISO fuzzy logic, Grey-Taguchi, ANN, DRL with LSTM networks, LR, and JLTO with FMEA. These methods address issues like clogging in gas turbine filter houses and optimal control of air handling units, improving system reliability and efficiency. Despite their significance, common limitations include the need for more experimental data, lack of maintenance discussions, and challenges in generalizability. To overcome these, a new approach is sought, emphasising comprehensive datasets, consideration of external factors, and enhanced model adaptability for dynamic operating conditions in real-world applications.

• The literature reviews diverse waste heat recovery techniques, such as thermoeconomic modelling, PSO, ANN, phase-change heat transfer, dragonfly optimisation, and nature-inspired algorithms like GA, DE, and PSO. These methods tackle challenges like optimising LCOE, power generation prediction, control scheme design, and addressing issues of operational instability and degradation in renewable energy utilisation. A new approach is deemed necessary to enhance overall system performance, improve control system search capabilities, balance system complexity and efficiency trade-offs, and optimise energy management strategies to effectively handle nonlinearities.

• The reviews explore diverse diagnostic techniques for engines, including adaptive neural-fuzzy systems, ensemble-based classifiers, FDI with IT2FLSs, dynamic engine model-based neural networks, and unique hierarchical approaches. Despite advancements, common limitations include computational complexity, sensitivity to operating conditions, and challenges in training data quality. Studies emphasise the need for innovative and hybrid models to enhance diagnostic accuracy, with a focus on addressing shared limitations for robust real-world performance.

4.1. Challenges in applying soft computing or metaheuristics to optimise gas turbines in practice

• Implementing advanced soft computing techniques frequently necessitates significant computational power and resources, which can prove prohibitively expensive. The high computational needs result in lengthier processing times and the requirement for specialised technology, which can't be accessible in all operational contexts.

• The initial investment required for developing, implementing, and maintaining advanced optimisation techniques can be substantial. Organisations may be hesitant to allocate funds without a clear and immediate return on investment, especially in industries with tight budgets and stringent cost-control measures. This financial barrier can delay or prevent the adoption of innovative soft computing solutions.

5. Future scope

• The exploration of various design techniques for gas turbines and combined cycle power plants (CCPPs) reveals notable advancements but also highlights limitations in neglecting control parameters, transfer learning challenges, and scalability issues in diagnostics. To address these, a promising future scope lies in leveraging soft computing techniques and integrating advanced automatic control theory methods. Integrating adaptive neuro-fuzzy inference systems (ANFIS), machine learning-based outlier detection, and enhanced transfer learning frameworks provides a more comprehensive and standardised approach to the design and performance optimisation of gas turbines. Additionally, implementing Model Predictive Control (MPC) can optimise GT performance by predicting future behaviour and adjusting control inputs accordingly.

• In the fuel and emission control methods, common limitations such as complex system setups and sensitivity to input parameters underscore the potential for future advancements through the integration of soft computing techniques. Incorporating a Multi-Layer Perceptron MLP-based neuro-fuzzy approach in the fuel control system to select the proper air–fuel ratio in the combustion chamber and selective catalytic reduction for NOx emission control with adaptive optimisation can enhance model robustness, adaptability, and address generalizability issues.

• Optimisation strategies for gas turbine filter systems have proven helpful in resolving issues such as clogging and optimum control, resulting in increased system dependability and efficiency. Future studies might concentrate on multi-input single-output (MISO)-based adaptive and real-time filter house cleaning strategies, as well as a prediction model approach for filter house pressure drop estimation. This can help to overcome data limits, streamline comprehensive maintenance plans, and improve the generalizability of optimisation methodologies. Furthermore, using automated feedback control systems can enhance these cleaning procedures, assuring constant performance under changing situations.

• The techniques on waste heat recovery need to further enhance overall system performance and tackle complexities. Integrating genetic algorithms (GA), and a fuzzy-based model for heat transfer characteristics optimisation with a bio-inspired algorithm can improve control system search capabilities, balance trade-offs between system complexity and efficiency, and optimise energy management strategies for effectively handling nonlinearities. Furthermore, applying automatic control theory techniques, such as PID control, adaptive control, and robust control, can enhance the precision and efficiency of heat recovery systems.

• The machine learning-based fault diagnostic has persistent limitations in computational complexity, sensitivity to operating conditions, and challenges in training data quality and thus calls for future developments in soft computing. Leveraging hybrid models that integrate policy gradient-based fuzzy logic, gated recurrent unit or Bi-LSTM-based outlier detection and correction approach can enhance diagnostic accuracy and robustness. Moreover, incorporating automatic control techniques such as state estimation and fault-tolerant control can provide real-time adaptive responses to detected faults, thereby improving the resilience and reliability of gas turbine operations.

6. Conclusion

In conclusion, the reviewed literature underscores the significant advancements achieved through the application of soft computing techniques in enhancing the thermal performance of gas turbine power plants. Techniques such as neural networks, genetic algorithms, fuzzy logic, ANFIS, and machine learning have been instrumental in addressing critical areas, including design optimisation, fuel system optimisation and emission control, filter house optimisation, fault detection and diagnosis, and waste-heavy recovery control. While these studies demonstrate significant progress in addressing specific challenges, such as control parameter neglect and scalability issues, they also reveal common limitations such as computational complexity, sensitivity to operating conditions, and issues with data quality. A comparison of the performance of various soft computing techniques in terms of detection accuracy, waste heat recovery efficiency, emission control performance, and optimisation response for thermal performance analysis of gas turbine power systems has been provided. Future research should focus on developing more adaptive, robust, and real-time optimisation techniques that can handle the nonlinear and dynamic nature of gas turbine operations. The review highlights how important soft computing is to moving the next generation of gas turbine power plants into alignment with modern energy and environmental goals by increasing efficiency, reducing emissions, and improving flexibility. All things considered, the combination of several soft computing techniques provides incisive information and a way to develop intelligent, flexible, and sustainable gas turbine systems, which will ultimately lead to a more reliable and effective energy infrastructure.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The authors declare that no data availability.