Research On Hybrid Model Of Short-term Wind Speed Forecast For Wind Farm Based On Machine Learning

Since modern times,the global economy has continued to develop,and non-renewable resources with extremely limited reserves have been consumed in large quantities.At the same time,wind energy is widely used as a renewable energy source with high economic efficiency and no pollution.The installed capacity of wind power generation has grown rapidly,and the proportion of wind power in the power grid has also risen sharply.The characteristics of the system have a great influence on the stability and dispatch of the power grid.Therefore,it is extremely necessary to propose an effective wind speed prediction method.The superior wind speed prediction model is one of the effective methods to deal with the impact of wind energy on the power grid.In recent years,many researchers have developed numerous models of wind speed prediction for mid-and long-term wind speed prediction,which has improved the accuracy of wind speed prediction.However,in the field of short-term wind speed prediction,the effect of the classic wind speed prediction model is not very ideal and cannot meet the actual requirements.Therefore,in view of the non-linear and random characteristics of short-term wind speed series,this master's thesis selects wind speed data from three different wind farms(sampling intervals are 1 hour,totaling 800 data per dataset),and launches a short-term wind speed prediction study:(1)An improved Gray Wolf Optimization(IGWO)algorithm is proposed to improve the global search ability of the grey wolf optimization algorithm in the early stage and accelerate the late convergence speed for the problem that the Gray Wolf Optimization(GWO)is easily limited to the local optimal solution in the process of parameter optimization.(2)Aiming at the disadvantage that the predictive effect of the Back Propagation network(BP)is greatly affected by the initial weights and bias,the improved grey wolf optimization algorithm is selected to optimize the parameters of BP network,which significantly improves the prediction ability of the network and the prediction accuracy of BP network.(3)Aiming at the characteristics of unstable and linear wind speed series,intermittent,etc.,it is proposed to use the variational mode decomposition algorithm(VMD)to preprocess the data.The K value of the variational mode has a greater impact on the decomposition effect.The complete empirical mode decomposition of adaptive white noise(CEEMDAN)decomposes the data,and uses the mode number obtained by the CEEMDAN decomposition algorithm as the K value of the variational mode to improve the effect of VMD decomposition;After using the improved gray wolf optimization algorithm,the optimized BP neural network predicts all the components decomposed by VMD and performs equal weight accumulation to obtain the prediction result.The experimental results show that the prediction effect of this model is better than that of the wavelet and feed-forward neural networks models in each performance index.(4)Based on the machine learning model of(3),it is proposed to use deep learning models for wind speed prediction,and LSTM as a classic deep learning model,due to its unique forgetting gate.The structure has great advantages in the field of wind speed prediction.However,the number of network layers and the number of neurons of the LSTM need to be obtained by trial and error.This master's thesis proposes to use the improved GWO(IGWO)algorithm to optimize the structural parameters of the LSTM,thereby improving the predictive ability of the network model.Through the simulation verification of the data at the sampling interval of 1h,it is proved that the IGWO-LSTM model is better than the models of SVR,ANN,LSTM.(5)Aiming at the problem that the sequence complexity after VMD decomposition differs greatly,the model prediction of which component to choose after decomposition is determined by the sample entropy value.Here,a combined prediction model(VMD-IGWO-LSTM-BP-SVR)is used to predict the sequence used after decomposition using the combined model of SVR,LSTM and BP network optimized by the improved Grey Wolf optimization algorithm.The penalty factor and regularization parameter(IGWO-SVR)for optimizing SVR using IGWO are proposed.Then,IGWO-SVR is used to predict the component of low sample entropy value,and IGWO-BP is used to predict the score of the general sample entropy value.IGWO-LSTM network predicts complex components.Three sets of comparative experiments were set up using three datasets.The results show that VMD-IGWO-LSTM-BP-SVR can significantly improve the prediction accuracy and follow-up performance compared with the traditional machine learning model,other decomposition models(EMD,EEMD and CEEMDAN)and 11 comparison models of a single model in VMD-IGWO-LSTM-BP-SVR.