A feature transformation and extraction approach-based artificial neural network for an improved production prediction of grid-connected solar photovoltaic systems

ABSTRACT

The grid operators face challenges due to the fluctuations in the energy production from Photovoltaic (PV) power plants. Therefore, reliable prediction models are critical as they aid grid operators in optimizing operational planning and lowering operating costs. One of the most challenging components of production prediction accuracy is determining a proper set of features that affect the models’ prediction accuracy. In this work, an accurate data-driven prediction model has been developed based on new timestamps and weather condition features that affect the PV system. Specifically, standard PV physics-based models have been employed for feature transformation and extraction from the traditional available historical features. The new features are then utilized to build an Artificial Neural Network (ANN), whose production predictions are compared to those obtained using the old features and other features extracted using the principal component analysis (PCA). The suggested approach’s efficacy is demonstrated using a 264 kWp PV system in Jordan. The proposed approach has shown superior with performance gain in accuracy reaches up to 21% and 64% compared to the models that employ old and PCA features, respectively. Additionally, the proposed approach reduces the computational cost by 20% and 7% compared to old and PCA models, respectively.

KEYWORDS:

• Photovoltaics
• feature transformation and extraction
• prediction
• artificial neural network
• principal component analysis

List of acronyms and notations

CSP	=	Concentrated Solar Power
PV	=	Photovoltaic
ASU	=	Applied Science Private University
NWPs	=	Numerical Weather Predictions
ML	=	Machine Learning
FS	=	Feature Selection
PCA	=	Principal Component Analysis
RF	=	Random Forest
YJT	=	Yeo-Johnson Transformation
MLR	=	Multiple Linear Regression
ESN	=	Echo State Network
CNN	=	Convolutional Neural Network
GRU	=	Gated Recurrent Unit
DL	=	Deep Learning
KM	=	K-Means
EEMD	=	Ensemble Empirical Mode Decomposition
MFA	=	Modified Firefly Algorithm
NN	=	Neural Network
SVR	=	Support Vector Regression
WPD	=	Wavelet Packet Decomposition
LSTM	=	Long Short-Term Memory
ANN	=	Artificial Neural Network
WT	=	Wavelet Transform
BC	=	Bias Compensation
GBM	=	Gradient Boosting Machine
PC	=	Physics-Constrained
SVM	=	Support Vector Machine
BP	=	Back-Propagation
$W$	=	Historical timestamps and weather data
$W^{\mathrm{\prime }}$	=	New timestamps and weather data
$W^{PCA}$	=	Tranformed $W$ data
$\vec{P}$	=	Historical PV production data
$X$	=	Overall available input-output dataset
$X^{\mathrm{\prime }}$	=	New overall input-output dataset
$X^{PCA}$	=	PCA overall input-output dataset
$Y$	=	Number of years
$\mathrm{\Delta }t$	=	Timestep of measurements
$\overrightarrow{hr}$	=	Hour counter
$hr$	=	Hour counter index
$\vec{d}$	=	Day counter
$d$	=	Day counter index
$\overrightarrow{GHI}$	=	Global solar irradiance on a horizontal surface
$\vec{T}$	=	Ambient temperature
$\overrightarrow{WS}$	=	Wind speed
$\vec{S}$	=	Solar time
$\overrightarrow{S_a}$	=	Solar angel
$\overrightarrow{G_t}$	=	Global solar irradiance on a tilted surface
$HRA$	=	Hour Angle
$LST$	=	Local Solar Time
$\delta$	=	Declination angle
$\alpha$	=	Elevation or altitude angle
$\phi$	=	Latitude of the location
$\zeta$	=	Zenith angle
$\gamma$	=	Solar azimuth
$\theta$	=	Incidence angle
$Z_S$	=	Surface azimuth angle
$R_B$	=	Beam irradiance tilt factor
$G_B$	=	Beam irradiance on a horizontal surface
$G_D$	=	Diffuse irradiance on a horizontal surface
$\beta$	=	Tilt angle of the surface
${\rho }_G$	=	Ground albedo
$F$	=	Number of overall features/ANN nodes
$F^{\ast }$	=	Number of selected PCA features/ANN nodes
$H$	=	Number of ANN hidden neurons
$h$	=	Index of the hidden neuron, $h$=1,…,$H$
$\overrightarrow{X_j^{\mathrm{\prime }}}$	=	Generic training pattern
$j$	=	Generic training pattern index
$P_j$	=	The $j$-th actual production
${\hat{P}}_j$	=	The $j$-th predicted production
$\bar{P}$	=	Average actual production
${\vec{W}}_h,{\vec{\beta }}_h$	=	Internal ANN parameters (weights)
$b_h,b_o$	=	Internal ANN parameters (biases)
$C$	=	PV plant capacity
$f_1()$	=	Hidden neuron activation function
$f_2()$	=	Output neuron activation function
Rprop	=	Resilient Backpropagation
OSS	=	One Step Secant
LM	=	Levenberg-Marquardt
$MSE$	=	Mean Square Error
$h^{opt}$	=	Optimum number of hidden nodes
$PC$	=	Principal Component
$X_{train}$	=	Old training input-output dataset
$X_{train}^{\mathrm{\prime }}$	=	New training input-output dataset
$X_{valid}$	=	Old validation input-output dataset
$X_{valid}^{\mathrm{\prime }}$	=	New validation input-output dataset
$X_{test}$	=	Old test input-output dataset
$X_{test}^{\mathrm{\prime }}$	=	New test input-output dataset
$N_{train}$	=	Number of training patterns
$N_{valid}$	=	Number of validation patterns
$N_{test}$	=	Number of test patterns
$RMSE$	=	Root Mean Square Error
$MAE$	=	Mean Absolute Error
$WMAE$	=	Weighted $MAE$
$NMAE$	=	Normalized $MAE$
$R^2$	=	Coefficient of Determination
$Combined$	=	Overall combined performance metric
$Metric$	=	A performance metric
$P{G}_{Metric}$	=	Performance gain of a metric $metric$
$Metri{c}_{Benchmark}$	=	Benchmark’s performance metric
$Metri{c}_{Proposed}$	=	Proposed’s performance metric

Acknowledgments

The authors would like to acknowledge the Renewable Energy Center at the ASU for sharing with us the solar PV system data.

Disclosure statement

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

Data Availability

The data that support the findings of this study are available from the authors upon reasonable request and with permission of the Renewable Energy Center at ASU.

Additional information

Funding

This research received no external funding.