Combined Model Of Short-Term Wind Speed Prediction For Wind Farms Based On Deep Learning

With the development of economy and society,wind-based new energy power generation has been increasingly developed.Considering the intermittency and randomness of wind energy,the wind speed forecasting for wind farms has significant meaning for the wind power development.At present,many professional scholars pay more attention to the deep learning-based wind speed prediction.Compared with the traditional methods,deep learning has better prediction accuracy,but this method still has some defects.1)The learning ability of a single prediction model is limited by randomness,which leads to poor generalization ability.2)The deep learning is a black-box model,and the spatial information and meteorological information are difficult to be integrated into the DNN.Thus,regional wind energy resources cannot be dynamically described,and it afflicts the wind power generation by the wind farms for the grid connection.The main research content and innovation of this paper are listed as follows:For the time-varying problem of the wind speed time series within the small-scaled wind fields,the ARIMA-NARX model is proposed.The model used the minimum information criterion to analyze the correlation of wind speed data,and extracted the linear trend of wind speed based on the autoregressive integrated moving average(ARIMA)model,and used the output of the ARIMA model as the input of the NARX neural network variable.Our results showed that the ARIMA-NARX model is better than the single model.Through these experiments,an independent small-scale area,multi-reference point,wind speed evaluation model was established.Then,a comparative experiment with NAR,NARX,and other models was carried out.The indicators of MAE,MAPE,and RMSE of the ARIMA-NARX combined model all dropped significantly.Our experiments demonstrated that the combined model can effectively solve the time-varying data problem on a small scale.To solve the problem that the environmental parameters of wind speed in the mesoscale wind field are varied and related,and that their features are difficult to be extracted,we proposed a Multi-TCN-LSTM model.In this model,TCN first quickly extracts the spatiotemporal features of wind speed,and then the LSTM model performs high-level feature information abstraction on the features of the output results of the received TCN,and then pools it through the attention,and finally sends it to the fully connected layer of the LSTM model.Wind speed forecasting experiments have verified that the model can cooperate with TCN environmental parameter extraction and LSTM prediction to improve the accuracy of wind speed prediction and to achieve the results of accurate short-term wind speed prediction for wind farms in the medium-scale range.In view of the wind speed prediction of large-scale wind farms,it is necessary to consider the spatial-temporal correlation,geographic location,and other factors.By applying the temporal-spatial variation theory,our report analyzed the influence of time,space,and temporal-spatial distance on wind speed changes,and proposed a new method that combines temporal-spatial kriging with the spatiotemporal wind speed prediction method which is combined with the spatiotemporal variability theory and realized the short-term spatiotemporal wind speed prediction of large-scale wind farms.It is verified through experiments that the model can improve the accuracy of wind speed prediction through the integration of the spatial geographic model and the time series' model to reach a large scale.The results of accurate short-term wind speed prediction for large-scale wind farms have been achieved.