Feature Optimization for Short-Term Solar Power Forecasting using Bidirectional LSTM Networks

Solar energy offers the advantages of low cost, abundant availability, and environmental friendliness. However, energy conversion output fluctuates significantly, depending on weather and environmental conditions. Key influencing factors include solar irradiation, ambient temperature, relative humidity, photovoltaic (PV) surface temperature, and wind speed. This study investigates which of these parameters have the greatest impact on solar power output forecasting accuracy. The research examines feature optimization for short-term solar power forecasting using Bidirectional Long Short-Term Memory (BiLSTM) networks. Four feature selection methods—Pearson Correlation Test, Spearman Correlation Test, Recursive Feature Elimination (RFE), and Mutual Information (MI)—were applied to identify the most significant predictors. Results indicate that solar irradiance and PV surface temperature are the most critical predictors. The BiLSTM model, implemented with the Adam optimizer and the optimal feature subset, demonstrated superior performance, achieving a Mean Absolute Error (MAE) of 0.002766153 and a Root Mean Square Error (RMSE) of 0.008086442. A comparison of optimizers revealed that Adam yielded the best results, closely followed by RMSprop, whilst Stochastic Gradient Descent (SGD) exhibited lower performance. The optimized BiLSTM model substantially outperformed the baseline Autoregressive Integrated Moving Average (ARIMA) model, with 86.36 times lower MAE and 34.91 times lower RMSE. These findings demonstrate the effectiveness of integrating feature optimization with deep learning methods in handling the complexity and non-linearity inherent in PV power output data. Consequently, it improves the precision of short-term solar power forecasting, which is crucial for efficient grid integration and management.