A multiscale network with mixed features and extended regional weather forecasts for predicting short-term photovoltaic power

Photovoltaic (PV) power is regarded as one of the most critical renewable energy sources for mitigating climate change. The generation process of PV power is significantly influenced by meteorological and geographic factors, resulting in intermittent and fluctuating variations. Accurate short-term PV power prediction is essential for optimizing the utilization of PV resources in grid integration. In this paper, we presented a multiscale network with mixed features and extended regional weather forecasts for predicting short-term photovoltaic power. To unravel complex temporal patterns, multiscale modeling is employed to learn temporal patterns from both local and global perspectives, with features mixed in temporal and variant dimensions, respectively. Additionally, the original model is improved with specially designed modules to manage multiple input data sources. Building on this, the effectiveness of incorporating regional meteorological forecasts for PV power prediction is evaluated. Based on the observed PV power data from five PV stations of China, comparative experiments show that the proposed model outperforms all baseline models in most cases, as measured by R2 and RMSE. This model achieves optimal results with an R2 of 0.706 when incorporating the future weather parameters. Furthermore, it shows improvements of at least 0.007, 0.018, 0.027, and 1.491 in MAE, MSE, RMSE, and SAMPE, respectively, compared to other models. The results also indicate that this model achieves the lowest RSME values on sunny and rainy days. This improvement in predicting short-term photovoltaic power has the potential to enhance grid stability and further promote the development of renewable energy.