Ultra-Short-Term Power Forecasting for Distributed PV based on Multi-Source Remote Sensing Information and Adaptive Feature Extraction

The lack of meteorological data for distributed photovoltaic (DPV) makes it impossible to incorporate meteorological information sources highly related to PV output into power forecasting model, which limits the further improvement of forecasting accuracy. Therefore, this study employs satellite cloud image data, including Cloud Fraction Rate (CFR), Cloud Top Height (CTH), and Surface Solar Irradiance (SSI), for research purposes. The multi-source satellite remote sensing information describes the distribution of solar irradiance from different angles, which provides a powerful data support for the forecasting of PV power generation. However, existing forecasting methods that rely on satellite cloud images face a problem where convolutional autoencoder (CAE) used to extract features from these images may inadvertently discard some critical attributes, which can negatively impact the final power forecasting results. Therefore, this study proposes a novel ultra-short-term power forecasting method for regional DPV systems using multi-source remote sensing information and adaptive feature extraction techniques. The process begins with preprocessing both satellite cloud image data and PV power generation data. Building upon this foundation, the proposed approach harnesses the spatial feature capturing capabilities of a Convolutional Neural Network-Long Short-Term Memory Network (CNN-LSTM) network to identify and extract the correlation between the multi-source remote sensing information and the PV power output, thus compensating for the dearth of meteorological information in distributed PV power forecasting and enhancing its ultra-short- term forecasting accuracy. Finally, the effectiveness of the proposed methodology is validated through practical application to DPV power generation data.