Spatial-Temporal Genetic-based Attention Networks for Short-Term Photovoltaic Power Forecasting

Photovoltaic (PV) output power is significantly random and fluctuating due to its sensitivity to meteorological factors, making PV power forecasting a big challenge. Accurate short-term PV power forecasting plays a crucial role for the stable operation and maintenance management of PV systems. To achieve this target, the paper proposes a novel Spatial-Temporal Genetic-based Attention Networks (STGANet), which consists of a spatial-temporal module (STM) and a genetic-based attention module (GAM). STM serves to predict the missing solar irradiance to support the generation forecast, and contains a graph convolutional neural network to learn the spatial and temporal dependencies between historical meteorological data, while using dilated convolution as the non-linear part to simplify the network structure. The GAM efficiently explores for potential relationships in input features with attentional mechanism and uses genetic-based operation and LSTM which takes forecasting error as reference to find global optimal solutions and to avoid getting trapped in local optimal solutions. The model is verified through comparative experiment with several benchmark models using a real-world historical meteorological dataset and a power generation dataset of PV plants in southeastern China. The results have illustrated that the proposed model can provide better prediction performance in PV systems.