A Robust Photovoltaic Power Forecasting Method Based on Multimodal Learning Using Satellite Images and Time Series
Kai Wang, Shuo Shan, Weijing Dou, Haikun Wei, Kanjian Zhang
Abstract
Ultra-short-term photovoltaic (PV) power forecasting holds significant importance in enhancing grid stability. Most PV power forecasting methods based on satellite images rely on pixel-level predictions, which are inefficient and redundant. Meanwhile, current deep-learning approaches struggle to establish correlations between large-scale cloud features and PV generation patterns. In this paper, an end-to-end model based on multimodal learning is proposed for directly obtaining multi-step PV power forecasts from satellite images and time series. To capture cloud dynamics and features within the region of interest (RoI), ConvLSTM-RICNN is utilized to encode satellite images. To mitigate the impact of noise and missing data in PV power, a robust fusion approach named DCCA-LF is introduced. This approach integrates deep canonical correlation analysis (DCCA) into late fusion (LF) to strengthen cross-modal feature alignment. The proposed model is verified using publicly available data from BP Solar in Alice Springs and Himawari-8, from January 1, 2020, to October 8, 2022. Comparison with current state-of-the-art research shows that the suggested model achieves the best RMSE and MAE with minimal complexity across all cloud conditions. Moreover, the proposed approach is robust to noise and missing data.