Probabilistic Wind Power Forecasting With Missing Data Tolerance: An End-to-End Nonparametric Approach

Missing data occurs due to sensor failures, communication issues, or temporary gaps in the measurement process, which may significantly reduce the performance of the wind power forecasting system. Therefore, an end-to-end nonparametric approach is proposed for probabilistic wind power forecasting (WPF) incorporating missing data imputation. The proposed method comprises both end-to-end training and online application procedures. In the end-to-end training phase, a deep learning-based nonparametric forecast model undergoes iterative processes involving imputation for missing data and model training with revised loss. In the online application phase, the trained forecast model is deployed online to provide multi-step-ahead probabilistic WPF through continuously imputating online observations. Compared with other state-of-the-art benchmarks, the advantages of the proposal include: 1) the proposed method is nonparametric, i.e., no hypotheses on distribution types are needed, and 2) the proposed end-to-end training process will automatically regulate the imputed value from the deep learning-based forecast model for higher probabilistic forecast performance, thereby mitigating the negative impact of missing observations. This approach leverages the advantages of the nonparametric method and the deep learning-based end-to-end structure. As a result, the proposed approach showcases an outstanding approximation capability for the future probability distribution of nonstationary wind power while simultaneously addressing missing values. Experiments validate that the proposed end-to-end nonparametric approach is more effective in mitigating the negative impact of data missingness on forecast performance compared to other representative two-phase methods integrated with standalone missing data imputation steps. Additionally, it outperforms its parametric end-to-end counterpart across various missing rate scenarios, especially in multi-step-ahead probabilistic forecasting.