Multiple Decision Expert Systems for Performance Analysis of a Boiler System

Abstract

Artificial neural networks (ANN)-based multiple decision expert systems (MDES) were developed for assessing the performance of a boiler system. Different configurations of ANN were used with different decision combination methods, including a neural combiner, to propose the model. The model was developed using the plant data collected over a period of five months to predict steam temperature, pressure, and mass flow rate, using feed water pressure, feed water temperature, conveyor speed, and incinerator exit temperature as the input parameters. The predictive capability of the model is evaluated in terms of both correlation coefficient (R) and mean absolute percentage error (MAPE). The results observed from this work demonstrate that neural combiner and ANN-based MDES can efficiently predict the data on steam properties consistently, and that the model can serve as an efficient tool for monitoring boiler behavior under real-time conditions. Superiority of the proposed model over others under various scenarios is also demonstrated.

INTRODUCTION

Precarious energy crises around the world warrant the exploration of new and renewable energy sources. As the world’s tenth largest energy consumer, The Republic of Korea imports 97% of its total required energy (MOE 2013). Therefore, waste-to-energy (WTE) conversion is drawing significant attention in the recent times. Among the options for WTE, steam, generated via heat energy recovery from waste incinerators, can be utilized by many process and allied industries, wherein boilers play a significant role. The performance of a boiler with respect to steam production, both in terms of quality and quantity, is important especially when some process industries are highly dependent on a particular plant to meet their steam requirement for different processes. However, it is often cumbersome to predict the performance of industrial boilers during prolonged period of operation owing to the continuous change in the boiler and fuel characteristics (Behera et al. 2014). Nevertheless, physical modeling and simulation of a boiler is a complicated phenomenon (Adam and Marchetti 1999; Bhambre, Mitra, and Gaitunde 2007; Kwan and Anderson 1970; Lu 1999; Wu et al. 2011). For real-time simulation, monitoring, and prediction of boiler behavior, artificial neural network (ANN)-based models have emerged to be very useful, especially for the prediction of fresh steam properties from a newly installed boiler, for which no physical model exists (Kljajić, Gvozdenac, and Vukmirović 2012; Smrekar et al. 2009; Wu et al. 2011). The most widely used neural network for forecasting/prediction purposes is the multilayer perceptron (MLP) (Maier and Dandy 2001) using the backpropagation (BP) algorithm (Rummelhart, Hinton, and Williams 1986). In MLPs, the neurons (processing units) are connected to each other in a layered configuration, containing an input layer (receives information from external sources), a hidden layer (receives information from the input layer and processes them), and an output layer (receives processed information and signals the desired output; Figure 1).

FIGURE 1 Generic architecture of multiple decision system.

The capability of soft computing methods such as ANNs, fuzzy logic, and expert systems to solve engineering problems including energy systems is well documented (Afgan et al. 1996; Khoshjavan, Rezai, and Heidary 2011; Lalot and Lecoeuche 2003; Monedero et al. 2012; Qin et al. 2012; Syn et al. 2011). It is, however, experimentally verified across various application domains that different decision makers trained under the same environment might provide contradictory decisions for different inputs. This motivated us to design a decision system by combining multiple decision expert systems (MDES) (Du et al. 2012; Kittler 1998; Kuncheva 2004; Kusiak, Li, and Zhang 2010; Polikar 2006; Tang et al. 2012). The basic idea is not to rely on a single decision system, rather on a consensus decision obtained by combining the individual performances with an anticipation of improved result. The objective of MDES is to exploit the complementary discriminatory ability of different expert systems for achieving improved learning and better generalization. In general, MDES has two major goals: (1) increasing the decision accuracy, and (2) increasing the robustness of the system in terms of generalization ability on unknown inputs. Combination of multiple decisions has been treated as one of the most exciting advancements in pattern recognition over the last decade. Improvement in performance due to combining decision systems is reviewed and presented by several researchers in various applications (Kittler 1998; Kuncheva 2004; Polikar 2006). Several methods for combining multiple models with applications to different domains have been described in the literature (Kuncheva 2004; Polikar 2006; Kusiak, Li, and Zhang 2010; Tang et al. 2012). For example, a neural network ensemble with five MLPs have been described to develop a predictive model for energy management applications, and Monte Carlo simulation was used to investigate uncertainty propagation of the model built by using weather forecast data (Kusiak, Li, and Zhang 2010). Based on the formulated model and weather forecasting data, future steam consumption was estimated. A hybrid ensemble learning paradigm integrating empirical mode decomposition and least squares support vector regression was proposed for nuclear energy consumption forecasting (Tang et al. 2012). This model was formulated specifically to address difficulties in modeling nuclear energy consumption, which inherently has high volatility, complexity, and irregularity. Recent literature survey describes the design, implementation, and application of many of the existing ensemble based systems in decision-making processes (Du et al. 2012; El-Melegy and Ahmed 2007; Ghosh et al. 2010; Kusiak, Li, and Zhang 2010; Polikar 2006; Tang et al. 2012).

Although many efforts aimed at MDES have become popular, the performance of the MDES is highly dependent on the individual decision-making system and the combination technique (Kuncheva 2004). The use of multiple decisions led to increase in computational complexity with significant increase in the performance. To overcome these constraints, a trade-off between the decision accuracy and the cost of computation needs to be carefully considered in real-world applications. Normally, a set of classifiers is used in the design of MDES for combining the output through some linear or nonlinear method. The process of combination can be anything, starting from simple voting or linear combination to adaptive learning. Recently, two different MDES models were developed for energy management applications (Kusiak, Li, and Zhang 2010; Tang et al. 2012). However, these models do not provide consistently better results in different environments. This means that, for various individual decision systems, based on the datasets collected from diverse applications, and use of different combination strategies, the corresponding MDES model provides dissimilar outputs.

Our aim in the present study is to use the whole feature space as input to MDES, combine all the outputs with different combination methods, including the adaptive combination method called neural combiner, to justify the potentiality of the use of MDES for parameter estimation of a boiler. Irrespective of the different datasets, the motivation to use an adaptive neural combiner to get a consistently better result has been drawn from various applications. We have considered MDES consisting of four ANNs (e.g., see Kusiak, Li, and Zhang 2010), all of which receive the same input data. The network structures of all ANNs are the same but are trained with four different learning algorithms, such as Levenberg–Marquardt optimization BP (Wilamowski and Yu 2010), resilient BP algorithm (Riedmiller and Braun 1993), extended Kalman filter (EKF)(Singhal and Wu 1989), and extreme learning machine (ELM)(Huang, Zhu, and Siew 2006). The decisions provided by the individual systems are combined with ten different combination techniques (Kuncheva 2004), such as voting, aggregation reasoning rules such as maximum, minimum, product, sum, and mean, probabilistic product, decision template, Dempster-Shafer, and neural network.

The inspiration drawn from the aforementioned literatures, and the scope of using real plant data from a newly installed refuse plastic fuel (RPF)-fired boiler helped us to formulate the research objective of this work, given as follows; (1) to create MDES by using different learning algorithm-based ANNs and a neural combiner that would predict the pressure, temperature, and mass flow rate of steam generated from the boiler, (2) to test the developed model with real plant data that was not presented to the trained decision system, and (3) to emphasize the application of MDES for real-time control of boiler parameters and (4) to address the potential advantages of MDES for predicting the boiler performance compared to the individual learning-method-based ANN. Finally, the article demonstrates the advantages of MDES with neural network-based individual decision-making systems and the decision combination that consistently provided the best output for the boiler performance analysis.

MATERIALS AND METHODS

Boiler Operation

The boiler under discussion is a water-tube type established in an incineration plant fired with RPF located within the Ulsan National Industrial Complexes, South Korea. It produces fresh steam with temperature and pressure in the range of 180 °C–190 °C and 12 kgf/cm², respectively. A part of the steam is supplied to a nearby paper mill, and the remaining is used to meet its own energy requirement.

Data Acquisition and Preprocessing

Time-series data at 1 min intervals was provided by the plant manager for a period of five months. The data set consisted of the following parameters: steam temperature, steam pressure, mass flow rate, feed water pressure, feed water temperature, conveyor speed. and incinerator exit temperature. The data was filtered according to the procedure described elsewhere (De et al. 2007; Romeo and Gareta 2006; Rummelhart, Hinton, and Williams 1986). After the data filtering, the plant data was averaged at every 12 h and mixed together, which resulted in 215 data points for all the parameters used in this study. This amount of data was easier to handle, because the training process did not require longer time, which consequently resulted in developing a reliable and more representative MDEC model. The data was normalized and scaled to the range of 0 to 1 using Equation 1, in order to suit the transfer function in the hidden (sigmoid) and output (linear) layers of the ANN used in MDES.

(1)

where is the normalized value, and X_min and X_max are the minimum and maximum values of X, respectively.

Neural Combiner-Based MDES Development

As described earlier, the basic motivation in the design of MDES is to overcome some (known or unknown) limitations of individual approaches, which are often difficult to find for a given task at hand, while different decision makers are designed for the same task. For example, the use of the same base decision-maker trained on different training sets, or on different feature, can show evidence of complementary strengths and weaknesses. Moreover, the problem of decision making becomes a little complex process at the cost of improved performance, particularly when the expert system is designed in a multistage process.

Two major steps exist in the design of MDES, such as construction of the individual decision makers and the design of the combination rule (Figure 1). The basic aim of the combination rule is to exploit the complementary decisions made by the individual systems. In general, most of the MDESs follow a parallel design structure, wherein the systems are trained independently so that all of them can make a decision on the input samples simultaneously. The outputs of these systems are finally combined for a consensus decision using some rules. Researchers have adopted different strategies to obtain complementary decisions from the systems. This strategy is broadly categorized into three levels (Figure 2): decision system level, feature level, and data level (Kuncheva 2004).

FIGURE 2 Designing MDES in different levels of operations.

Operation on training data (i.e., in the data level), is performed by training the individual system (or maybe the same base system) with different training sets. This method is usually preferred for unstable systems such as decision trees and neural networks. With the manipulation of input features (i.e., at the feature level), different feature sets for the same or different individual decision-making systems are used. At the decision system level, various contemporary and efficient decision-making systems are considered for the MDES design.

MDES with ANN

Among the different approaches for designing MDES, we have considered a simple and efficient approach in which the input sample data, with its original number of features/parameters are used for the next step of processing (Figure 1). The proposed MDES is modeled using four ANNs with different learning algorithms, and the decisions obtained from these networks are finally combined using a conventional BP neural network.

The input and output parameters in the modeling of an individual ANN are usually selected on the basis of the objective of the modeling and the availability of reliable data (Rummelhart, Hinton, and Williams 1986). In order to develop the ANN model for the RPF-fired boiler, the output parameters viz., steam flow rate, (ṁ_s), steam temperature (t_s), and steam pressure (p_s) were decided on the basis of real-life requirements. The selection of input parameters (belt conveyor speed (s_c), incinerator exit temperature (t_ie), feed water pressure (p_fw), and feed water temperature (t_fw)) was made (Figure 3) on the basis of prior experience and system knowledge. The mass flow rate of fuel, feed water pressure, and feed water temperature have been selected previously as the input parameters for prediction of steam properties (De et al. 2007). However, considering the practical utility of the ANN model, a more directly controllable parameter, viz. belt conveyor speed, was included as a surrogate input parameter against RPF mass flow rate. Besides, the data received from the plant on the RPF mass flow rate were averaged calculated values. The basic statistics of the variables for the training and test matrix are shown in Table 1. These parameters are then fed in parallel to four ANN-based decision-making systems. The ANN is trained with four different learning methods: Levenberg-Marquardt optimization BP, resilient BP algorithm, EKF, and ELM. A brief introduction and operation of a conventional ANN is given in the following.

TABLE 1 Basic Statistics of the Data Used for ANN Model Development (a) Training and (b) Test Data

(a)	Basic Statistics of the Training Data
Variable	N	Mean	Std Dev	Minimum	Maximum
Feed water pressure (kgf/cm^2)	151	13.96	3.94	5.46	16.79
Feed water temperature (°C)	151	44.64	18.31	14.5	78.83
Conveyor speed (rpm)	151	18.32	3.07	11.22	29.6
Incinerator exit temperature (°C)	151	912.14	42.80	780	1,015.65
Steam pressure (kgf/cm^2)	151	9.84	0.75	3.88	10.8
Steam temperature (°C)	151	181.56	4.04	142.58	185.65
Steam flow rate (ton/h)	151	7.46	1.20	3.1	10.33
(b)	Basic Statistics of the Testing Data
Variable	N	Mean	Std Dev	Minimum	Maximum
Feed water pressure (kgf/cm^2)	64	14.5539	3.401	5.95	16.61
Feed water temperature (°C)	64	42.1561	18.3442	15.08	78.98
Conveyor speed (rpm)	64	18.268	2.8089	12.19	25.26
Incinerator exit temperature (°C)	64	919.0517	39.2881	812.5	1,015.90
Steam pressure (kgf/cm^2)	64	9.8725	0.5948	6.94	10.76
Steam temperature (°C)	64	181.9273	2.0954	176	185.66
Steam flow rate (ton/h)	64	7.5758	0.9918	5.52	9.33

FIGURE 3 Schematic of a multilayer perceptron (4-15-3) neural network used for predicting the boiler behavior.

The basic motivation in the design of an ANN is to mimic the behavior and reasoning processes of the biological nervous system. Intuitively, the model gives a hope of accomplishing humanlike performance artificially by capturing the key ingredients responsible for the remarkable capabilities of the human nervous system. Making a collective decision is the inherent aspect of an ANN because of its interrelated connections among the processing units called neurons. In addition, the key characteristics of ANNs is that they are data driven, adaptive, fault tolerant, robust, and play important roles in a decision-making process. Furthermore, to discover the essential regularities in the problem space, the ANN fits very well to model complex nonlinear boundaries. In a conventional neural decision model, the number of nodes in the input-layer is equal to the number of parameters (features) present in the input data pattern and the number of output-layer nodes is equal to the number of output parameters (Figure 3). Thus, the key information of different input parameters will automatically be encoded in the connecting links during training, and the nonlinear decision boundaries are generated to estimate the output parameters by making collective decisions with regard to the processing units’ functionalities.

Several attempts have been made to use ANNs for parameter estimation of different boilers (De et al. 2007; Kljajić, Gvozdenac, and Vukmirović 2012; Rummelhart, Hinton, and Williams 1986). In the present work, we have tried to explore the advantages of ANNs with different learning methods and finally use them to design an efficient MDES, which finally provides the best result compared to the individual ANN in estimating the performance of the RPF-fired boiler for a better analysis and diagnosis. The values of estimated parameters are obtained from these four ANNs and fed to the second step of processing, where the values are combined with a metalearner or combiner using a conventional BP-based neural network (Figure 1). A brief description of the neural network-based decision fusion system is made in the following section.

Neural Network-Based Decision Fusion

Just as no system exists for providing optimum decision in general, there exists no “best” combiner for all problems. This motivates designing a decision combiner that performs consistently well for most of the problems and is not dependent on the dataset and the selection of individual decision-making systems. Among the different combining methods, “majority voting” is the most popular one, which interprets the decision as a vote for one of the decisions provided. However, for any specific task, each of these combiners could attain a different degree of success, but none of them is perfect and consistently performs well. Fusion decision varies with input data, type of individual decision-maker, and combination methods (Kuncheva 2004; Polikar 2006).

Broadly, the methods of combining decisions can be categorized as fixed and adaptive. The maximum, majority and average rules (Kuncheva 2004) are among the approaches that can be categorized as fixed combining methods. These combining methods do not acquire any information from the input data. On the other hand, the adaptive combining method uses the approach to possess parameters acquired via training. Such approaches are Demspter–Shafer, weighed average, behavioral knowledge space, and fuzzy integral approaches, fuzzy templates. Another way of categorizing these different classifier-combining techniques is based on the method of mapping between the input and output of the fusion module. This mapping may be linear or nonlinear. The feature-based approach, stacked generalization, or rank-based methods, which involve a more complex mapping of the input, also use a nonlinear mapping in the combining method. Decision fusion using a neural network is a trained, nonlinear scheme. The present study uses a neural network-based combination method that provides the final decision adaptively and is independent of input data and individual decision systems.

The neural network-based decision combining method proposed in this study falls under the adaptive combination category. The elements of the individual network’s output becomes the input to a feed forward MLP (Figure 3), which acts as the combiner. The number of input nodes of the neural network is equal to the product of the number of individual decision expert systems used in the MDES and the number of output parameters to be estimated in the dataset. The number of output nodes of the neural network is equal to the number of output parameters present in the dataset. As a rule of thumb, the number of hidden nodes in the neural network is the square root of the product of the number of input nodes and the number of output nodes.

Working Principles of the ANN Models

Selection of Training Data

ANN detects the statistical interrelationship between a set of input and a set of output parameters. It can learn by training and cannot be more accurate than the initial training data. Thus, the proper selection of training data, from the available raw plant data is very important for the accurate prediction of the developed ANN-based MDES model. For the individual systems, four three-layered ANN models were developed to predict steam temperature, steam pressure, and mass flow rate, using feed water pressure, feed water temperature, conveyor speed, and incinerator exit temperature as the input parameters. The outputs of these systems are then used for the input to the ANN-based combiner to get the final decision output. The 215 data points were divided into different sets of training and testing sets, e.g., 80% (N_Tr: 170) of the data points were used for training the network, whereas the remaining 20% (N_Te: 45) were used for testing the developed models. The sets of training and testing data were picked up randomly and tested for the performance of the model. The testing dataset of the boiler was kept aside during the training process and was used only for validation purposes.

Network Training, Testing, and Error Evaluation

Four ANNs have been used in this study for the MDESs that are trained with different learning algorithms, such as Levenberg–Marquardt optimization BP, resilient BP algorithm, EKF, and ELM, each of which has its own advantages and disadvantages. However, these algorithms have been selected because of their superiority over other existing algorithms. Anew, they have also been observed to perform fairly well when used individually. The basis of operations for all these four networks is the same although the learning methodologies are different; i.e., the input data are processed through the network, layer by layer, and finally the parameters are updated. This is an iterative process, which continues until an acceptable level of errors (difference between desired and estimated) is obtained. Each time the network processes, the whole set of data is called an epoch. After training, the ANN can accept new “unknown” data as input and predict the corresponding output based on its training. The performance of the training and test dataset was evaluated in terms of the correlation coefficient (R), and the mean absolute percentage error (MAPE) was used as an error-estimating index to evaluate the accuracy/predictive capability of the model (Kljajić, Gvozdenac, and Vukmirović 2012). It can be noted that higher R and lower MAPE values are desirable.

(2)

(3)

where x_i is the target value, y_i is the predicted value at time i, and n is the number of test or training samples. For the modeling work, Neural Network Tool box, Version 7.0 of MATLAB was used.

RESULTS AND DISCUSSION

Selection of Parameters for the Best Network Topology

In the training stage of three individual ANN decision systems and one ANN-based combiner, successive iterations consisting of sequences of learning and verification processes were repeated until the accuracy of the networks reached the assumed model of the MDES. For the individual systems, four ANNs with different training algorithms; such as Levenberg–Marquardt (LM) optimization BP, resilient BP, EKF and ELM have been used. In the decision combination process, ELM-based ANN is used because of its superiority over the other three (Table 2) to obtain the final output. The training process was repeated for different sets of internal network parameters: training count (T_c) or epoch, number of neurons in the hidden layer (N_h), learning rate (η), and momentum term (α) of a BP neural network. A detailed study on the effect of internal parameters on the performance and procedures involved in selecting the best network topology of different training algorithms-based ANNs has been described elsewhere (Hornik, Stinchcombe, and White 1989). For the model development, the dataset was divided into two parts: (1) the first part was taken for the estimation of model parameters (training data), and the second part was taken for testing the performance (test data). Three different sizes of data were taken for training, viz., 60%, 70%, and 80% and 40%, 30%, and 20%, respectively, were considered as test data. Selection of the training data is random in nature and an equal percent of data is collected from each class.

TABLE 2 Performance Comparison of Individual Decision Expert Systems

Percentage of Training/Test Samples (15 Hidden Neurons), 1000 Epochs
	60/40		70/30		80/20
ANN with learning algorithm	R	MAPE	R	MAPE	R	MAPE
Levenberg–Marquardt BP	0.5256	17.3862	0.8061	10.8106	0.8984	9.5936
Resilient BP	0.8089	7.8396	0.8504	7.5364	0.9112	6.9853
EKF	0.8524	7.3913	0.8886	7.0724	0.9213	6.7914
ELM	0.8913	7.0318	0.9189	6.7811	0.9498	6.5304

Note: R and MAPE values are estimated from test data.

The best network architectures were based on the R value (Elias et al. 2006) and were achieved by a vigorous trial-and-error approach (Silva et al. 2008), keeping some training parameters constant and slowly moving the other parameters over a wide range of values. For example, as a first step of the modeling procedure for the BP network, studies were initiated to identify the influence of both T_c and N_h for obtaining the suitable network topology. During model development for the boiler, T_c was varied from 500 to 1000 by keeping other network parameters such as η and α at some constant value (0.75). The best values of η and α for the developed model were thus determined by keeping some training parameters constant and by slowly moving these parameters from 0.1 to 0.9. It was observed that (1) increasing the number of neurons (N_h) from 6 to 18 increased the R values, (2) increasing the training count (T_c) increased the R values, (3) increasing the learning rate (η) from 0.01 to 0.9 decreased the R value significantly, and (4) increasing the momentum term (α) from 0.1 to 0.9 did not significantly increase the R value. The simulation results showed that the network architecture 4–15–3 (i.e., four input parameters, 15 neurons in one hidden layer and three output parameters) resulted in high R value during training, and this MLP had the following best values of network parameters: N_h = 18; T_c = 1000; η = 0.01, and α = 0.1.

Performance Analysis

The ANN-based MDES model for the boiler was specifically formulated to predict the values of output parameters more accurately using the values of the selected input parameters collected from the plant, during real-time operation. Successively, to verify this, the trained model was provided with a separate set of test data containing real measured values of input parameters, which was not used earlier during training. By comparing the prediction of the model with the actual/target measured values of output parameters, the expected performance of this trained model was assessed for its practical application. The results were evaluated in terms of the most commonly employed error measurement indexes: R and MAPE.

The comparative performances of the predictions of individual ANN models with different training algorithms are depicted in Table 2. The results in Table 1 were evaluated on the test data for three sets of training and test percentages. It was observed from Table (1) that, individually, the ELM-based ANN performed better compared to the rest with high R and low MAPE values for all three percentages of training and test sets. For example, with 80% training and 20% test set, the ELM-based model provided R and MAPE values of 0.9498 and 6.5304, respectively, compared to 0.8984 and 9.5936 with LM BP, 0.9112 and 6.9853 with resilient BP, and 0.9213 and 6.7914 with EKF-based models.

TABLE 3 Performance Comparison of MDES in Terms of MAPE with Different Combiners

	Percentages of Training/Test Samples
Decision combination method	60/40	70/30	80/20
Majority voting	6.7214	6.4322	6.2813
Maximum	6.8032	6.3356	6.0984
Minimum	6.6214	6.4201	6.127
Mean	6.8833	6.1674	5.7665
Decision template	6.7702	6.0012	5.8133
Dempster–Shafer	6.3779	6.2189	6.0218
Neural networks	5.8933	5.2467	5.1369

Note: MAPE values are estimated from test data.

In a comparative analysis with different decision-combiners-based MDES, results are shown in Table 2 for the MAPE values with three different sets of training and test samples, such as 60% and 40%, 70% and 30%, and 80% and 20%, respectively. A similar trend of observation was also observed with respect to R values, and thus, we have not included these values in Table 3. Note that the MAPE values in the table are based on the test samples. The performance comparison between Tables 2 and 3 revealed that any of the MDES models performed better than different individual models. Interestingly, among the MDES with six decision combiners (except the proposed neural combiner), the observation varies for different combiners and none of the MDES (Table 3) provided consistently better results for all three training and test samples. For example, with 60% training set, the Dempster–Shafer-based model performed better, whereas with 70% and 80% training sets, decision template and mean combiners, respectively, performed better. Moreover, the improvement of MDES is dependent on different randomly selected training and test samples. Hence, it is difficult to choose a combiner for a particular dataset with a particular training set. However, with the proposed neural network combiner-based MDES, the accuracies were improved and found to be consistently superior for all three training sets compared to the results obtained with any of the six discussed combiners-based MDES.

The prediction capability of the proposed MDES was also verified by comparing the actual and predicted values of three output parameters of the boiler using four different models for 30% testing samples (Figures 4, 5 and 6). The models that were considered for comparison are (1) MDES with neural-combiner, (2) MDES with Dempster–-Shafer combiner and 3) MDES with mean combiner. It was observed that the predicted values from the neural combiner-based MDES were more aligned with the targeted values compared to the other two models.

FIGURE 4 Targeted and predicted values of MDES models for steam pressure.

FIGURE 5 Targeted and predicted values of MDES models for steam temperature.

FIGURE 6 Targeted and predicted values of MDES models for steam flow rate.

It can be seen from Figures 4, 5 and 6 that the variations in the actual and predicted profiles of steam flow rate is comparatively large when compared to the steam pressure and temperature. This variation can be attributed to the large variation of feed water pressure (5.95 kgf/cm² and 16.61 kgf/cm²; Table 1). The fluctuation in the steam flow rate during the period of data collection could be attributed to the less homogeneous nature of RPF injected to the incinerator, which could have resulted in varying heat generation. Thus, the fluctuation in the input and output parameters would have caused an impact in the learning/generalization pattern of the neural networks during the performance prediction. The sudden drop in the measured steam temperature, pressure, and flow rates on data point 36 is due to a transient condition during the boiler operation. Despite this unavoidable transient-state condition, the results of the ANN-based MDES were excellent for predicting the steam temperature and pressure, with a slight variation in the steam flow rate prediction. Furthermore, the errors in prediction were due to the sudden fluctuation of the measured data presented to the MDES during training and testing. Nevertheless, the parts with higher fluctuation were predicted with quite good accuracy. It is, in general, observed that operation of a boiler is more or less stable at a nominal load, and the errors during ANN prediction can be the lowest (Smrekar et al. 2009). However, the prediction error increases with the load on the boiler, and under such circumstances ANN-based prediction does not provide satisfactory results. At the same time, deviations during prediction can also be expected from real boiler behavior operation due to the lack of a representative dataset (Romeo and Gareta 2006). This problem has been sorted out to a certain extent using MDES-based prediction, which is clearly observed from the results presented in this study. This implies that the degree of complexity of the data for the MDES-based network was not high and that, it can sufficiently map the relation between the given set of inputs and outputs.

Performance Analysis in Terms of Computational Time for Both Individual and MDES with Different Combiners

In this section, we have made a comparative performance analysis with computational time (C__time; the sum of training and testing times), as required by different individual and MDES with seven combiner-based models. The results for 70% training and 30% testing sets are depicted in Figure 7. All the simulations were done in a MATLAB environment. It was observed from Figure 7(a) that the ANN model with EKF- and RP-learning-based models took high and low time, respectively, compared to the rest. Interestingly, the ELM-learning-based model took little more time than the RP-learning-based model, and at the same time performed the best among the four (Table 2). In a comparison analysis of MDES with different combiner-based models, as expected the proposed NN-combiner-based MDES took more time compared to others. However, at the cost of these times, it yielded the best and most consistent prediction results toward the performance analysis of the boiler system irrespective of different training and testing sets, whereas other combiner-based MDES provided varying results with less C__time.

FIGURE 7 Computational time taken by (a) individual and (a) different (a) combiner-based models for 70% training and 30% testing set.

Application of ANN-Based MDES for Real Time Performance Analysis of Boiler System

The developed ANN-based MDES for the boiler was trained by a very short computation time (0.018 s–0.025 s for one run), which resulted in predictions of high accuracy. Although the operational time for the individual learning network is insignificantly lower than the MDES, the prediction capability of any of the MDES is 8% to 9% better, and the proposed MDES is nearly 16% to 17% more accurate than the best individual model (Tables 2 and 3). This justifies the significance of MDES as a decision-making tool for evaluating the performance of the boiler. The proposed ANN-based MDES integrated in a simple user interface can enable the estimation of the quality of fresh steam properties on change of input parameters. Thus, the model can be used for online prediction of fresh steam properties and can also alert an operator about possible error in the system. The model developed might yield more promising results when the online control system is integrated with adequate application-specific optimization programs. Nevertheless, ANN-based MDES, when combined with more prominent applications such as genetic algorithms, fuzzy logic based-control systems, and fuzzy neural optimization software would allow better prediction and control of the boiler performance. This work might enable researchers to extend and intensify research in applying neural networks for optimizing state variables in a boiler for enhanced steam production.

Research Contribution

It has been proved that during the inception period of the MLP NN, MLP can approximate any function with a certain level of precision using two hidden layers of threshold nodes (Haykin 2009). Subsequently, however, it has been shown that an MLP with a single hidden layer and threshold nodes can also approximate any function with a specified precision (Haykin 2009). Interestingly, none of the findings provided any likely method of building and training MLP, and thus, the significance was observed only theoretically. In spite of this, MLP has been applied in various domains for solving highly complex problems (e.g., nonlinear function mapping), and it successfully gained attention for its human-like intelligence and discriminative power. The NN provides an alternative information processing approach to biological neural networks. Its major characteristics are adaptivity (adjusts to a change in environment/new data/information), speed (via massive parallelism), fault tolerance (to missing, confusing, and/or noisy data), ruggedness (to failure of nodes/links), and optimality (with regard to error rates in classification).

This motivated us to use the merits of NN in such a way that their shortcomings can be diminished in the resultant network using the concept of MDES. As described in the previous section, MDES has its own drawback in the selection of individual decision-making systems and the decision combination rules. Our proposed MDES has aimed to address these issues and used four individual NNs trained with different learning algorithms and an adaptive/trainable combination strategy, wherein the task of determining the data-dependent/decision, system-dependent-based combination rules can be avoided. Here, a NN was used as a combiner that takes the decisions of the individual systems as input to the network and provides respective outputs with consistent improvement in the performance, irrespective of data and decision-making systems. The efficacy of the proposed model for analyzing the behavior of RPF-fired boiler have been justified.

CONCLUSION

In the present study, we have proposed a decision-making model MDES based on a neural network combiner in which the outputs of a set of neural networks are fused adaptively. The usefulness of the model is demonstrated by showing its consistency in performance over other existing combiners and individual decision systems for various training and test sets. A set of four base neural systems with different training algorithms were considered, and the outputs of these systems were combined in a multiple classifier paradigm with different combiners (existing and proposed neural combiner), following which the performance analysis was evaluated. The proposed model was observed to work well and the improvement was consistent. However, the results were different for different existing combiners, depend highly on the input dataset.

The developed neural combiner-based MDES for the boiler was found to have very good prediction accuracy. As it can predict the steam properties (temperature, pressure, and mass flow rate) with good accuracy, the model can be used for evaluating the performance of the boiler through an online monitoring system. With the developed model, it will be easier to control, monitor, and predict the three output variables. As further research, it might be interesting to compare the performance of the proposed MDES model with other learnable combiners such as support vector machine and Fisher discriminant analysis (the two successful classifiers).

Notes

Color versions of one or more figures in the article can be found online at www.tandfonline.com/uaai