Enhanced Forecasting of PV Power Output Using LSTM Integrated with LASSO-RFR Hybrid Models in Extreme Weather Conditions

In light of the recent climate variations observed in regions like Siberia, it is crucial to enhance the precision of predicting photovoltaic power generation to optimize the utilization of energy resources. This research introduces a blended model utilizing Long Short-Term Memory (LSTM), Least Absolute Shrinkage and Selection Operator (LASSO), and Random Forest Regression (RFR) to improve PV power forecasting accuracy during severe Siberian weather conditions. The model leverages LSTM's ability to capture time-related patterns effectively and combines it with LASSO's feature selection power and RFR's strong regression capabilities to handle the high-dimensional characteristics of the data. By analyzing datasets spanning from January 1, 2021, to January 23, 2024, the integrated LSTM-LASSO-RFR model surpassed traditional techniques by delivering reduced Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and a notable coefficient of determination (R²). The model performed well with an MAE of 0.0651, an RMSE of 0.0934, and an R2 value of 0.7837, showing its excellent predictive abilities and capacity to incorporate important components for predicting future trends accurately over time. These results indicate that the LSTM-LASSO-RFR model fits the data well and surpasses other models in all performance measures, presenting possibilities for enhancing the accuracy of forecasting solar power output influenced by weather conditions.