Research On Wind Power Forecasting Based On Deep Learning

Operations such as dispatching,backup,and energy storage arrangements are based on the forecast value of wind power in power grids with wind power.Underestimation of wind power will result in abandonment of wind and overestimation will result in imbalance between supply and demand of the power grid,which will seriously threaten the security of the power grid.It is of great significance for providing accurate wind power forecast and error evaluation values to restrain the impact of wind power on the grid.Wind power has complex characteristics of volatility and uncertainty,and sudden rises and drops occur from time to time.Wind power forecast and error evaluation are a technical problem.Deep learning technology can learn the internal characteristics of complex systems and discover potential laws that were not easy to discover by previous technical means.Therefore,deep learning networks are used to conduct research on wind power forecast,ramp forecast,and error evaluation in this paper.The main completions are as follows:1)In order to suppress the random component to long term memory of neural,a wind power short term forecast method based on Improved Long Short Term Memory(ILSTM)network structure is proposed,combined with the characteristics of wind power.Firstly,Variational Mode Decomposition(VMD)method is adopted to decompose wind power signal to the long-term component,the fluctuation component and the random component,which are used as the input of forecast model.Then a parameter was defined and added to the memory cell to suppress the random component to long term memory of neural.To provide a pass for the current random component,the output gate was cancelled accordingly.As a result,the learning for the real patterns of wind power is strengthened,avoiding over-fitting and improve the accuracy of wind power forecast.2)To improve the accuracy of wind power ramp forecast,a method based on convolutional neural network(CNN)is proposed.Wind power ramp event rarely occurs and has complex characteristics,and it’s difficult for forecast model to effectively learn from small number of ramp event samples.So CNN is used to extract features in wind power time series.And then forecast model is built,and AM is added to the model to weight outputs of forecast model to strengthen the learning of wind power features,thus improving the accuracy of ramp forecast.3)Aiming at the problem that the system error of wind power forecast exists in the entire time series and requires a larger receptive field to correlate all relevant data,a method for evaluating the error interval of wind power forecast based on the Temporal Convolution Network(TCN)is proposed.Wind power error is related to wind power characteristics and forecast methods.There are system errors that exist in the entire time series and stronger randomness and volatility.There is a longer receptive field with TCN and can remember the associated wind power forecast error data in a longer period of time.Unlike Recurrent Neural Networks,TCN does not share weights in different time periods,so there is basically no problem of gradient explosion and disappearance.Use TCN to establish a forecast error evaluation model and add residual connections and skip connections to improve the structure.So that the accuracy of error evaluation is improved.There are 27 figures,11 tables and 116 references in this thesis.