The output power of photovoltaic (PV) plants is primarily influenced by solar radiation, which is closely related to their geographical location. As a result, PV output exhibits significant spatiotemporal correlation characteristics. To address the inherent volatility and complexity in PV power forecasting, an ultra-short-term PV power output prediction method based on multi-dimensional IoT data collection and Attention-BiLSTM is proposed. This paper first extracts the correlation between geographical locations and weather conditions across multiple photovoltaic stations. Then, through an attention mechanism, we compute a weight matrix. This attention matrix is subsequently fed into a BiLSTM for training iterations, further uncovering and leveraging the interrelationships among multiple stations in both geospatial and photovoltaic power generation data. By combining the historical data of 10 adjacent PV power stations collected by IoT with weather forecast data, a new attention matrix is generated, enabling the model to focus on spatial environmental variations at the same time point through deep learning. Compared with existing methods, the proposed model shows better accuracy in forecasting random environmental disturbances. It improves the forecasting robustness and accuracy under random disturbances such as cloud and wind speed changes, and opens up new ways for efficient utilization of the PV energy.
Zhenyu et al. (Wed,) studied this question.