Improved Temporal Convolutional Network Based Ultra-Short-Term Photovoltaic Power Prediction

Accurate ultra-short-term power predictions are important for eliminating fluctuations in new energy power generation systems. To improve the accuracy of ultra-short-term photovoltaic (PV) power prediction, this paper proposes an ultra-short-term PV power prediction method based on improved temporal convolutional neural (TCN) network and feature modeling. First, the Spearman coefficient is applied to filter existing meteorological features while simultaneously combining solar illumination and the three-dimensional modeling of PV panels to identify the key factors affecting PV power generation and to mine astronomical features that affect PV power prediction. Second, the high correlation between astronomical features and PV power prediction is analyzed based on the correlation coefficient, which theoretically proves the feasibility and necessity of astronomical features. Third, an improved TCN network model is proposed for the algorithm. Multiple experiments indicate that compared with the existing prediction models, the proposed forecasting method is superior, the accuracy of PV power prediction over the next 4 h in the absence of meteorological conditions improves by 20.5%.