High-resolution Photovoltaic Power Prediction based on Dynamic Time Constants Residual Recurrent Neural Network

Renewable energy production often utilizes statistical models and regression analyses for forecasting on a daily or hourly basis. However, these conventional approaches may fall short of fulfilling the high-precision demands of minute-level scheduling critical for future integration of energy generation, distribution, storage, and consumption systems. Addressing this limitation, this paper introduces an innovative approach with the Dual-channel Residual Recurrent Neural Network that incorporates Dynamic Time Constants (DTC-RRNets). This model uses five-minute interval public photovoltaic data from the Desert Knowledge Australia Solar Center (DKASC). Our experiments show that DTC-RRNets markedly enhance forecasting accuracy, achieving a 23% improvement in SS-RMSE compared to the persistence model, with MAE of $\mathbf{1 1. 5 8}$ and an R-square value of 0.985. Notably, this model also significantly mitigates the time lag issues commonly associated with overfitting historical power data in recurrent neural networks, enhancing both the precision and reliability of renewable energy forecasting. This breakthrough shows that DTC-RRNets could be essential in advancing the capabilities of renewable energy management systems.