Solar irradiance estimation using a hybrid deep learning model with metaheuristic algorithms

Solar irradiance forecasting is important for precise scheduling and an effective solar energy system. However, its unpredictability and dynamic nature make forecasting extremely challenging. This study proposes a dualstream hybrid Bidirectional Long Short-Term MemoryConvolutional Neural Networks (BiLSTM-CNN) model for multi-time ahead solar irradiance forecasting. Two stochastic metaheuristic algorithms, Particle Swarm Optimization (PSO) and the Genetic Algorithm (GA), are used to optimize the hyperparameters of the dual-stream hybrid model. The performance of the optimized model is then evaluated on a real dataset obtained from Alice Springs, and the results are compared to single and hybrid models. Analysis of forecast results show that the GA-tuned dual stream model outperforms single and hybrid models in 3-step and 6-step ahead forecasting. The GA-optimized BiLSTM-CNN achieved the best scores on all five evaluation metrics in both 3-step and 6-step ahead forecasting. Hence, the proposed model can be useful for forecasting solar irradiance in different weather conditions.