Research On Short-term Wind Power Prediction Model Based On Deep Learning

With the economic development of society and the increase of population,people’s demand for energy is increasing,and the reserves of fossil resources such as coal,natural gas,and oil are getting smaller and smaller.The current energy structure is mainly based on traditional primary energy,supplemented by new energy.However,due to the non-renewable nature of traditional primary energy sources and serious pollution of the air,energy reform is imperative.Among the many clean energy sources,wind energy is an accessible and non-polluting renewable resource,and the large-scale construction of wind farms also provides conditions for the use of wind energy.However,due to the characteristics of fluctuating,intermittent and uncontrollable wind resources,large-scale wind power integration brings great challenges to the stable operation of the power system,so accurate wind power prediction has become a key technology to promote wind power consumption.Combining feature selection,signal decomposition and deep learning,this paper proposes a short-term wind power prediction method from the aspects of neural network and model optimization.The specific research contents are as follows:(1)On the basis of analyzing the effect of various factors on wind power,the feature selection method and the basic theory of wind power prediction are studied.Firstly,the basic theory of LSTM,neural network and differential autoregressive moving average model(ARIMA)is described.The wind power data set used in this paper is defined,and its Pearson correlation coefficient is calculated through experiments.Finally,recursive feature elimination method is used to analyze the key factors in the data set.(2)For complex data such as wind power with multiple influencing factors,it needs to be decomposed into different components to improve the accuracy and reliability of prediction.Therefore,this chapter studies and compared signal decomposition techniques such as empirical mode decomposition,ensemble empirical mode decomposition,adaptive noise-complete ensemble empirical mode decomposition and variational mode decomposition.It is found that the variational mode decomposition technique can significantly reduce the influence of mode aliasing and end effect,and is a more effective method for wind power sequential signal decomposition.Further,the method of variational mode wind power sequence decomposition is studied in depth.The energy difference principle is used to optimize the value of decomposition mode number k,and the optimal decomposition mode number is determined.The problem that decomposition mode number k does not have adaptability in variational mode decomposition is solved,and the influence of wind power sequence volatility is reduced.(3)In order to solve the problem that ARIMA model can only predict the stable time series data and has poor performance in predicting the noise and outliers of non-stationary series,optimization methods such as stochastic gradient descent,adaptive moment estimation and root-mean-square propagation are introduced,and the advantages of adaptive moment estimation in solving the optimization problems of large-scale data and parameters as well as non-stationary and non-convex optimization problems are studied.Through the comparison of different optimization algorithms and the experimental comparison with the traditional long and short term memory network model,the results show that the proposed method has a good performance for wind power prediction,and the adaptive moment estimation method for ARIMA model optimization effect is more significant than other gradient descent algorithms.(4)This paper studies Nag-DNN,an algorithm model that integrates Nesterov Accelerated Gradient(NAG)and Deep Neural Network(DNN),which is used to solve the high-precision short-term wind power prediction problem.The time series prediction model can only analyze stable data and deep learning is difficult to meet the requirements of real-time and efficiency.Firstly,the recursive feature elimination method is used to select key factors of wind power data,and then the variational mode decomposition technique is used to separate different frequency components contained in the data,and then the deep neural network model optimized by gradient descent algorithm is used to predict the decomposed data.Finally,different models are compared.The experimental results show that the proposed NAG-DNN model has better prediction performance when solving the data prediction problem of wind power with high volatility,and it also proves once again that the gradient descent optimization algorithm has excellent enhancement effect on the neural network.