Enhancing a temporal fusion transformer using GRU-LSTM encoder-decoder for effective solar generation forecasting

Abstract

Efficient renewable energy management requires precise solar power forecasting. This study enhances prediction performance by integrating a Gated Recurrent Unit (GRU) – Long Short-Term Memory (LSTM) encoder-decoder architecture within a Temporal Fusion Transformer (TFT), enabling more effective modeling of complex temporal dependencies in solar generation data compared to traditional models. The novel contribution lies in the synergy between GRU’s ability to handle vanishing gradients and LSTM’s capability of maintaining long-term dependencies, resulting in improved forecasting accuracy. Additionally, we incorporate relevant meteorological data as supplementary inputs to refine the model's predictive precision. The results using the UNISOLAR and Solcast weather data reveal that our GRU-LSTM encoder-decoder within TFT (GRU-LSTM) model consistently outperforms the standard LSTM encoder-decoder within TFT (LSTM) and GRU encoder-decoder within TFT (GRU) models, achieving superior accuracy across both short-term and long-term forecasting tasks. This GRU-LSTM model exhibits significantly lower Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE), particularly during periods of high solar output. In short-term forecasting, the model achieved an MAE of 2.687, MSE of 15.603, and RMSE of 3.950 for Campus 1, and an MAE of 0.509, MSE of 0.585, and RMSE of 0.978 for Campus 2. Long-term results followed a similar trend, reinforcing the model’s ability to identify underlying patterns in solar generation data. These findings validate the effectiveness of the proposed GRU-LSTM encoder-decoder within TFT (GRU-LSTM) model for robust and exact solar power forecasting.