Assessing Predictor Influence in LSTM Models for Enhanced Solar Energy Forecasting

Accurate forecasting of solar energy production is highly important for an adequate integration of renewable energy into the power grid. This study explores the importance of various predictors for enhancing the accuracy of solar energy forecasting using two distinct Long Short-Term Memory models: one that relies solely on historical power production data and past exogenous weather data, and another that additionally incorporates future exogenous weather data. Various settings are tried to improve the results and relatively similar parameters are used by both, yielding more accurate results for the model that also uses information from the future. The results also demonstrate that predictor importance varies significantly between the two models. These findings suggest that different models benefit from distinct predictor combinations and incorporating exogenous data can significantly enhance forecasting performance.