Short-term Prediction Of Wind Power Based On Temporal Convolutional Neural Networks

Wind power forecasting technology plays a key role in grid production scheduling,operational safety and grid connection needs.In the context of rising global CO2 content,severe environmental damage,and energy shortages,focusing on clean energy,especially wind energy,is of far-reaching significance for adjusting the energy structure and improving climate change.The most critical aspect of effective forecasting of wind power is the handling of the complex influencing factors of wind energy and the resolution of uncertainties caused by the fluctuations of the wind itself.In order to further understand the change law of wind power,reduce the impact of wind power fluctuations on the economic operation of the grid,and ensure the safe operation of wind power after the grid is connected,this paper focuses on short-term wind power forecasting.The main work done from this will be shown as follows:One is the establishment of the data set.All stages of wind power data from collection to storage may be affected by external factors,resulting in some data no longer conforming to the trend of the wind speed-wind power theoretical curve,such data will definitely have an adverse effect on the prediction results,in order to improve the prediction accuracy data cleaning is bound to be carried out.The correlation coefficient method is used to analyze the correlation between each input variable and the measured power data.In order to simplify the data cleaning variables as much as possible,the wind speed variable that has the greatest impact on the power data is selected as the data cleaning object.On this basis,the type of outliers corresponding to the data in this paper is determined to be left and right outlier data.The box plot method is used to identify and remove the outliers in this type of data.In this process,in order to identify as many outliers as possible,a grouping method is used to perform box plot operations.The second is to use three classic single prediction models BPNN,extreme learning machine ELM and Elman neural network to predict the "healthy" data set.The results show that the overall prediction trend of each model in each quarter is consistent,but the accuracy is not high.The overall performance is: the same prediction model has different prediction effects in different seasons,and different prediction models in the same season have different prediction effects,and the generalization performance is poor.Third,based on the shortcomings of a single prediction model,an improved empirical mode decomposition(ICEEMDAN,Improved Complementary Ensemble Empirical Mode Decomposition with Adaptive Noise)and temporal convolutional neural network(Temporal Convolutional Neural Networks,TCNN)short-term wind power prediction research method.ICEEMDAN is used to decompose the wind power power sequence of each quarter to obtain the corresponding IMF component and a trend component;Generally,the higher the frequency,the less information the component contains.Therefore,we calculate the number of maximum points of each component to divide the high frequency and low frequency,and calculate the correlation coefficient between the high frequency component and the original power data.The components greater than a certain threshold are regarded as non-noise components and retained,and the remaining components are eliminated as noise components,and the non-noise components are reconstructed and superimposed together with the numerical weather forecast data as the input variables of the combined model.Finally,through TCNN,each input variable is trained to obtain the final predicted value.In order to evaluate the pros and cons of the proposed model,the results of each model are compared and analyzed.The experimental results show that the combined model has the highest prediction accuracy,the smallest error evaluation indicators in each quarter,and good generalization performance.