A Power Prediction Enhancement Method for Multi-Photovoltaic Stations Based on Two-Layer Spatio-Temporal Anomaly Detection and Dynamic Weight Integration Learning

New energy bases have multiple types of resources such as wind power, photovoltaic(PV), hydropower, thermal power, energy storage, etc., and the scale of the stations is large; data anomalies and lack of power prediction accuracy constrain their power generation and grid-connected operation. In this paper, a method based on two-layer spatio-temporal anomaly detection and dynamic weight integration learning is proposed to improve the power prediction accuracy for multiple photovoltaic stations. Firstly, a two-layer spatio-temporal calibration method combining the $3 \sigma$ criterion and the Pearson’s correlation coefficient is used to improve the quality and accuracy of the collected data; then, a dynamic weight integration method combining the physical model, the statistical model and the machine learning model is proposed. Experiments show that the average prediction accuracy gain is improved up to $42.7 \%$ by the method proposed in this paper, which provides a technical means to improve the accuracy of power prediction for PV station clusters in new energy bases.