Improving the Short-Term Photovoltaic Power Forecasting Accuracy Based on Deep Learning Hybrid Model

ABSTRACT: The increasing penetration of renewable resources in smart energy systems has increased the need for accurate forecasting of photovoltaic (PV) power to reduce uncertainty and improve operational planning. Rapid changes in atmospheric conditions, especially cloud movement, cause severe fluctuations and pose a significant challenge to PV power forecasting in the short term. In this paper, an integrated Deep Learning-based framework for short-term PV power forecasting is proposed that models the relationships among sky conditions, solar irradiance, and PV power. First, the cloud movement is forecasted based on the Optical Flow algorithm using sky images. Then, the forecasted images are fed into a ResNet50 network to estimate the solar irradiance. Finally, the forecasted solar irradiance, meteorological parameters, and historical PV power data are fed into the CNN-BiGRU-Attention model, whose hyperparameters are optimized using Bayesian Optimization to forecast the PV power for the next hour. The performance of the proposed model was evaluated under different atmospheric conditions and compared with other benchmark models. The results showed that the proposed model achieves high forecasting accuracy, obtaining a R² of 0.987 and RMSE, MAE, and MAPE error values of 0.231 Kw, 0.147 Kw, and 8.69%, respectively, and outperforming benchmark models.