Solar and Wind Energy Generation Forecasting Using Deep Learning Time-Series Models

Abstract

Since solar and wind energy are becoming more and more integrated into electrical grids, accurate forecasting of renewable energy generation is essential for modern power systems. However, accurate prediction is difficult due to the highly sporadic and nonlinear nature of these energy sources, which are driven by atmospheric and environmental variability. The majority of current research concentrates on predicting solar or wind energy separately, frequently employing various preprocessing techniques, models, and assessment approaches, which restricts fair comparison and useful scalability. In order to predict solar and wind energy generation within a single methodological pipeline, this study suggests a unified deep learning-based forecasting framework that uses a hybrid Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) model. The suggested framework ensures consistency and cross dataset comparability by applying the same data preprocessing procedures, sequence generation strategies, model architecture, and evaluation metrics to both energy sources. To efficiently capture historical dependencies, time-series data are normalized using Min Max scaling and converted into fixed-length temporal sequences. The hybrid LSTM–GRU architecture is made to take advantage of the complementary capabilities of both recurrent networks, with GRU layers effectively capturing shortterm fluctuations and LSTM layers modeling long-term temporal patterns. To evaluate the effect of dynamic temporal weighting on forecasting performance, adaptive attentionbased variations are also investigated. Root Mean Square Error (RMSE), Mean Absolute Error (MAE), accuracy obtained from confusion matrix analysis, and visual inspection using actual-versus-predicted plots are used to assess model performance. According to experimental results, the Hybrid LSTM–GRU model without adaptive attention outperforms standalone LSTM, GRU, and attention enhanced variants, achieving the highest forecasting accuracy of 92.61% for solar energy and 93.90% for wind energy. These results validate the ability of a single hybrid deep learning framework to efficiently generalize across various renewable energy sources. The suggested method supports better energy planning, grid stability, and decision-making in smart power systems by providing a scalable and workable solution for integrated renewable energy forecasting.