Proactive failure warning for wind power forecast models based on volatility indicators analysis

Abstract

With the promotion of low-carbon models, the proportion of wind power energy has significantly increased. Accurate wind power forecasting is of great significance for the scheduling of power systems. Previous studies often focused on improving forecasting accuracy, neglecting the issue of model failure (abnormally large forecast error occurring). However, forecasting model failure brings significant misleading information to the scheduling of the power system. To address this issue, this paper firstly analyzes the error distribution of prediction models under various neural networks (CNN, CNN-GRU, DNN, ConvLSTM, ELM, GBDT, AR, TREE and XGBoost) and various loss function (MAE, MSE, MLSE and Log-cosh) combinations. Subsequently, based on the Backward cloud generator and Pearson correlation analysis, the paper confirmed that the forecasting errors mainly come from the volatility of the wind power sequence itself, rather than the types and structures of the models. Finally, the paper uses variation and variance to assist Bi-LSTM in 1-h-ahead early warning of forecasting model failure under two kinds of thresholds, and has achieved excellent reliability and accuracy. The warned models show obvious failure situations, both in terms of single error peaks and cumulative errors.