A Temporal Convolution Memory Network Model for Short-Term Wind Speed Forecasting

Wind speed forecasting (WSF) is the basis for realizing wind farm power forecast, but the high randomness of wind speed makes this problem of very difficult solution. In this paper, a deep architecture called Temporal Convolutional Memory Network (TCMN) is put forward for short-term WSF. The presented data-driven solution combines monitoring data of wind farm wind towers with the forecasted wind speed from numerical weather prediction (NWP). TCMN effectively consists of a long short-term memory network (LSTM) and a temporal convolutional network (TCN). Initially, the monitoring data and the NWP forecast data are composed of input variables in time-series. Then, the input variables first enter the TCN network to complete the initial feature extraction. Finally, LSTM is applied to capture the overall time series wind speed characteristics of TCN layer output to obtain the short-term WSF. This architecture is trained in an end-to-end manner, combining the advantages of wind velocity time sequences extrapolation forecasting and numerical forecasting. The proposed method is tested in an actual wind farm and the results show that it outperforms the current typical methods.