One-Step-Ahead Solar Radiation Forecasting Using Bi-LSTM and GRU Models in Semi-Desert Climates

Abstract

The increasing demand for energy and the imperative to reduce greenhouse gas emissions have heightened the need for renewable energy sources. Hence, there has been a notable surge in research efforts focused on advancing solar energy forecasting. The aim of this study is to forecast solar radiation using Deep Neural Network models, including Bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Unit (GRU), based on historical solar radiation values recorded at half-hour intervals in a semi-desert climate over a two-year period starting in January 2020. To assess and compare the accuracy of the two Deep Neural Networks (DNNs), various evaluation metrics were used, including Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Maximum Error, and R-squared (R²). Solar radiation predictions were carried out based on the values recorded over the previous twenty-four hours at half-hour intervals. The results obtained indicate that both forecasting models achieve exceptional accuracy in one-step-ahead solar radiation prediction, with correlation coefficients exceeding 97.2%. This highlights the strong potential of Gated Recurrent Unit (GRU) and Bidirectional Long Short-Term Memory (BiLSTM) models for optimizing and estimating solar radiation. Consequently, these models can be effectively used to estimate solar energy production, thereby enhancing the control and efficiency of solar energy systems. However, the study also highlights challenges in forecasting under partial cloud cover, where sudden fluctuations in solar radiation adversely affect prediction accuracy. Addressing these complexities could further improve the robustness of forecasting models and enhance their practical applicability in semi-desert climates.