Wind power interval forecasting based on RF-VMD-LSTM and non-parametric estimation

Wind power has a high level of technological maturity and is ideal for large-scale applications as a renewable energy source. The large-scale grid connection will pose problems to the safe and steady functioning of the grid due to the fluctuation and intermittency of wind power output. For secure delivery of the energy system and to avoid the economic risks associated with wind power grid connection, precise and reliable short-term wind power forecasting is crucial. This work offers a combination forecasting model including Random Forest, Variational Modal Decomposition, and Long Short Term Memory that utilizes the notions of choosing features and decomposition-reconstruction to improve forecasting accuracy. Firstly, the feature selection is performed by RF to calculate the feature importance of each environmental factor. Secondly, to diminish the non-smoothness, VMD decomposes the wind power time series. Thirdly, to acquire the wind power point forecasting results, the decomposed components of VMD from the chosen environmental factor series are input into the LSTM model. The computed results are then compared with those from the BPNN and SVM models. Finally, the absolute error of point prediction is used for kernel density estimation. The experimental findings reveal that the RF-VMDLSTM built in this work has an excellent degree of accuracy and proves the effectiveness of the model. For the interval prediction results, the PICP (prediction interval coverage probability) reaches 94.53% with a 90% confidence interval, which proves the validity of interval prediction and contributes to the decisionmakers making relevant decisions.