Novel hybrid BiLSTM-BiGRU and CNN-LSTM architectures for enhanced solar irradiance forecasting in semi-arid climates

Currently, the world is transitioning towards renewable energy sources, and accurate forecasting of solar irradiance has become increasingly essential for effective energy generation, grid integration, and power system management. This study presents a holistic hybrid framework that combines traditional machine learning and cutting-edge deep learning architectures can enhance solar irradiance forecasting accuracy and reliability. This study introduces novel hybrid architectures and presents the first detailed evaluation of Global Horizontal Irradiance (GHI) prediction models, addressing a crucial void in renewable energy research. The approach includes traditional machine learning algorithms (Support Vector Regressor, Decision Tree Regressor, and Random Forest Regressor) and novel hybrid deep learning architectures (BiLSTM-BiGRU and CNN-LSTM). The novel cascaded BiLSTM-BiGRU architecture uses complete bidirectional processing. The CNN-LSTM framework effectively combines spatial and temporal features. These advanced architectures achieved a 30 % reduction in computational complexity compared to equivalent standalone deep learning approaches, enhanced robustness to missing or noisy data, and controlled seasonal variations. Using five years of meteorological data from Quetta, Pakistan, the proposed hybrid models demonstrated superior performance with the BiLSTM-BiGRU achieving R2 = 0.9726 and RMSE = 0.0465, representing a 15 % improvement over traditional methods. The CNN-LSTM also showed excellent results (R2 = 0.969, RMSE = 0.5896), significantly outperforming conventional approaches like SVR (R2 = 0.8621). These hybrid architectures provide enhanced accuracy and computational efficiency for solar irradiance forecasting in complex meteorological conditions.