Short-term Wind Farm Power Forecasting Based On Temporal Convolutional Neural Networks

In recent years,wind power has developed rapidly in China.However,the uncertainties arising from the increasing scale of wind power have significantly impacted the scheduling and control of power systems.Accurate short-term wind power forecasting models can help people making more reasonable dispatch plans,improving power quality and reducing operating costs.To improve the efficiency of wind power utilization,this thesis analyses the uncertainty of wind power.To describe the time series of a single wind farm,this thesis first develops Temporal Convolutional Networks(TCN)with a causal dilation convolution layer.The point prediction model based on TCN effectively extract timing characteristics from historical power data through the dilation convolution.Secondly the residual block connection in the model ensures the convergence speed of the forecasting model.In order to extract the time series features of the input sequence more accurately,the thesis adopts Multiple Train and Test Splits(MTTS)method for data pre-processing.In the case,the proposed model’s forecasting error is compared to a conventional recurrent neural network algorithm across all four seasons.The results show that the error of the proposed model is lower than that of other forecasting models regardless of the season.For example,in spring,the forecasting model based on TCN shows a maximum error reduction of 23.86%and a minimum error reduction of 20.88% compared to the baseline model.Next,this thesis analyzes the operational characteristics of a wind farm in Australia.The volatility,stochasticity and probability distribution of wind power are quantified and analyzed at different time scales.A quantile regression interval forecasting model based on the improved TCN algorithm is introduced: On the one hand,the improved model achieves effective dimensionality reduction by sharing the weights of the causal expansion convolution operation at each time step.On the other hand,the incorporated gating mechanism avoids excessive decay or explosion of information,which makes the improved model’s nonlinear fitting ability further improved.At the same time,the absolute value quantile regression is used as the loss function,and the skill score evaluation index is adopted as the overall evaluation index.The interval prediction results at 80%,90% and 95% confidence levels all indicate that the improved prediction model further improves the prediction accuracy of wind power.Finally,based on the modeling of single wind farm prediction,this thesis analyzes the modeling of wind power in multi-wind farm state.Ten wind farms in the Australian wind farm region are selected,and the Pearson correlation coefficient is used to analyze the spatial relationships between the wind farms.And the CCF and ACF coefficients are used to analyze the correlation and autocorrelation of the power output time series between the wind farms.One wind farm is selected as the primary focus,with other highly correlated wind farms in both time and space dimensions selected for analysis: based on the improved algorithm,the forecasting analysis is performed for wind farms considering the historical power data of the related wind farms.The results show that the reliability and sensitivity of the forecasting algorithm are improved after considering the power factor of the relevant wind farms.