An Improved Wind Power Forecasting Framework Based on Seasonal Feature Selection and Temporal Convolutional Network

Large-scale wind power forecasting with high accuracy and reliability can proficiently decrease the unfavorable impacts of wind power uncertainty. However, the existing wind power forecasting performances are greatly influenced by intermittent changes of environmental and meteorological factors, which brings much challenge for wind power integration. To solve the above problem, based on a dynamic selection of seasonal features, an improved Seasonal Feature Selection Temporal Convolutional Network (SFS-TCN) model is proposed to forecast ultra-short-term wind power forecasting in this paper. Firstly, to simplify the model input and improve the forecasting efficiency, the correlations between multiple environmental variables and wind power under different seasonal characteristics are studied respectively. Secondly, a wind power forecasting framework based on temporal convolutional network is developed. The seasonal characteristic variables are selected as model input to forecast the wind power 10 minutes ahead. Finally, the forecasting accuracy of the proposed method is compared with the original TCN model by using the actual dataset of a case wind farm. The case study results show that compared to the original TCN model, the proposed improved SFS-TCN model has better forecasting performance, of which the forecasting errors decrease by 16.7% (MSE), 22.5% (MAE) and 11.5% (RMSE) respectively. At the same time, compared to benchmark models including multi-layer perceptron (MLP) and long short-term memory (LSTM), forecasting errors of the proposed model decrease by 28.6%, 16.7% (MSE); 39.2%, 36.7% (MAE) and 20.7%, 13.8% (RMSE) respectively.