Prediction of the output of the tri-generation concentrated solar power system based on artificial neural network

Solar-powered systems are complex and affected by numerous parameters such as climate conditions, shading, and panel orientation. Artificial neural networks can effectively handle this complexity and capture nonlinear connections between inputs and outputs. This study used historical experimental data of a solar tri-generation system utilizing solar energy through a parabolic trough matrix to feed an ANNs model and enhance the forecasting for best performance and energy generation from these solar energy systems through more precise and effective control. Elman, NARX, and feedforward models are used to predict the output of the trigeneration system. The three models using LevenbergMarquardt (trainlm) as a training algorithm are designed and tested using MATLAB neural network module to solve this problem and also to test the ability of each ANN to predict the output of the solar power block such as energy rate storage tank Q, power of steam engine P
seng
, energy rate of distillation Q
distilled
, distilled water rate m
ds
, and solar field output Q ptc. The performance of the proposed model has been conducted using the three global statistics: coefficient of determination (R
²
), root mean squared error (RMSE), and mean bias error (MBE). ANN network with neuron numbers (3, 10, 8) were designed and tested within the MATLAB. One hundred fifty sample sets were obtained from previously historical experimental data and are used as the input of the ANN network.Furthermore, a total of 70% of this data is used for training, 15% is used for validation, and 15% is used for testing. The tangent sigmoid function was applied for the hidden layer, and the linear transfer function was used in the output layer. In this study, after many trials and errors, the number of hidden layers was selected to be ten layers. It was found that the Feedforward model provided the best performance, i.e., the lowest RMSE and highest R
²
, for the training period and validation period. Also, Elman's forecasting capability is poor compared to other training algorithms.