A Modified Neural Network for Predicting the Solar Photovoltaic Power Generation Using Weather and Operational Parameter

Solar PV systems often face challenges feeding power into the local grid due to weather dependence. Although solar PV generation is variable, it is predictable and can help maintain grid stability. In this study, a modified neural network is developed to forecast power generation for a 500-kW solar farm under Thailand’s climatic conditions. Year-round operational data from the solar PV plant are used to train the forecasting model, and over 15% of the period is reserved for power generation prediction and validation against the actual power profile. Keras provides an effective interface, with TensorFlow as the backend engine, which is well-suited for high-computation processes. For training and testing, a batch size of 32 and 50 epochs is used as standard parameters, helping avoid overfitting and improving computational efficiency. It was found that during 75% of the sunshine period, the solar PV system generated 50% of its nominal DC capacity, indicating efficient operation. A 0.22 kW average difference between the forecasting model and the actual power profile indicates 99.86% accuracy over the testing period. The difference between actual and predicted power ranged from 2.88 kW to -4.67 kW, and the corresponding MAE, MSE, and RMSE were 0.87, 1.32, and 1.15, respectively. Furthermore, the developed ANN-based forecasting model is highly recommended for commercial use to avoid penalties from the grid authority and enhance grid stability.