A Two-Stage Ensemble Learning Framework for Short-Term Nonparametric Probabilistic Photovoltaic Power Forecasting Integrating Multi-Source Numerical Weather Predictions

Utilizing numerical weather prediction (NWP) for short-term solar power probabilistic forecasting (SSPPF) has proven to be an effective approach to enhance the integration and utilization of solar energy. However, most existing SSPPF studies only employ single-source NWP, overlooking the potential benefits of multi-source NWPs for improving the accuracy and robustness of probabilistic forecasting. This paper proposes an improved two-stage ensemble learning forecasting framework (ITS-ELFF) for SSPPF. ITS-ELFF utilizes multi-source NWPs as key exogenous covariates to generate multi-step quantile forecasts. In the first stage, a set of robust and diverse base learners provides the initial quantile forecasts. In the second stage, a meta-learner integrates the quantile forecasts from all base learners to generate the final predictive quantiles. To explore the valuable information in multi-source NWPs, we introduce skip connections in the ITS-ELFF and incorporate a feature attention mechanism to dynamically select critical features. Furthermore, a multi-head temporal attention mechanism is incorporated into the meta-learner to effectively capture hidden long-term temporal dependencies within the solar power time series. A non-crossing quantile generation strategy is proposed to ensure the monotonicity of the predictive quantiles. Finally, a two-stage training strategy is developed for ITS-ELFF to avoid overfitting. Numerical results on public real-world data verify the effectiveness of the proposed ITS-ELFF in SSPPF.