ATTENTION-ENHANCED LSTM-BASED SOLAR POWER MANAGEMENT FORECASTING: A COMPARATIVE ANALYSIS WITH RECURRENT NEURAL NETWORK ARCHITECTURES

Abstract

The prediction of solar power production plays an important role in the proper functioning and efficient management of renewable energy systems. Nevertheless, the solar power generation varies continuously with changes in weather conditions, which make the process of forecasting difficult. The purpose of this paper is to evaluate the accuracy of forecasting with the application of Long Short-Term Memory (LSTM) and Bidirectional Long Short-Term Memory (Bi-LSTM) models combined with the attention mechanism for solar power prediction. The efficiency of the models is measured by means of the coefficient of determination (R²), Mean Squared Error (MSE), and Mean Absolute Error (MAE). The research results show that the best results are obtained with the Attention + LSTM model, with the highest mean values of R² of 0.9555, MSE of 0.4205, and MAE value is 0.4521. The mean R² for the Attention + Bi-LSTM model is 0.9468 and the conventional Bi-LSTM model shows only slightly better results in terms of mean R² value of 0.9473. Therefore, the implementation of attention does not necessarily improve the performance of LSTM-based architectures. The overall results of the comparative analysis reveal that the efficiency of attention is based on the particular recurrent architecture chosen. Of all the considered models, the one that delivers the best prediction results is the Attention + LSTM model, which is capable of identifying the temporal patterns for solar power generation.