Ultra-short-term photovoltaic power hybrid interval forecasting model based on the “decomposition-training-prediction-reconstruction” framework

Accurate forecast of photovoltaic power generation is fundamental to ensuring safe dispatch in power systems and serves as a crucial means to mitigate technical and economic risks faced by electricity market participants. This paper constructs a framework for ultra-short-term photovoltaic power forecasting. First, the Extreme Gradient Boosting algorithm is employed for key feature selection and decomposition of raw photovoltaic power sequences. Second, the Informer- Encoder Forest model is utilized to forecast short-term photovoltaic power, enhancing the model's generalization capability and prediction accuracy. Finally, nonparametric estimation methods are applied for photovoltaic power interval forecasting, providing richer decision-making reference information for photovoltaic power generation enterprises and dispatch agencies, thereby ensuring the stable operation of grid companies. To validate the practicality and reliability of this forecasting framework, empirical analysis was conducted using actual data from a Chinese PV power plant. Results showed that the Extreme Gradient Boosting - Wavelet Transform -Informer- Encoder Forest method achieved an R2 improvement of 26.4% and 21.25% compared to the benchmark models Recurrent Neural Network and Long Short-Term Memory, respectively, effectively demonstrating the efficacy of the proposed hybrid forecasting model. Furthermore, based on coverage width criteria, Kernel Density Estimation outperformed other interval forecasting methods by delivering more reliable prediction intervals. By forecasting future trends of individual photovoltaic power power generation components, this approach generates ultra-short-term photovoltaic power power predictions, offering significant guidance for the vigorous and sustainable development of China's photovoltaic industry.