Research On Wind Power Prediction Based On DNN And Random Forest Feature Information Selection

Wind power has strong volatility,randomness,and intermittency.The exact problem of wind power is a research hotspot and difficulty in the field of new energy power generation.At present,the commonly used wind power forecasting methods have certain deficiencies,which are reflected in the forecasting process.Most forecasting models use all historical data of the wind farm as model input,failing to make important feature information(wind power,wind speed,temperature,pressure,etc.)Sorting and screening of sex,thereby reducing the prediction accuracy and effect.In addition,the relatively excellent heuristic prediction method represented by BP-ANN(Back Propagation-Artificial Neural Networks)also has the problem of insufficient hidden layers,which cannot capture the relationship between model input and output in depth.In response to these problems,this paper introduces DNN(Deep Neural Networks)and RF(Random Forest)models,and conducts research in the following three aspects.In order to analyze the influencing factors in the process of wind power forecasting,the correlation between wind power and weather factors is analyzed from the aspects of wind speed,wind direction,temperature,air density,etc.,so as to establish a set of input weather variables for the forecasting model.Finally,in order to measure different forecasting methods For the forecast error,four commonly used wind power forecasting evaluation indicators are selected:(1)average absolute error(2)average absolute percentage error(3)mean square error(4)root mean square error.A DNN-based wind power prediction method is proposed,which inputs the training set data to the DNN network,uses RBM(Restricted Boltzmann Machine,Restricted Boltzmann Machine)to pre-train each hidden layer,and uses it after training PSO(Particle Swarm Optimization,particle swarm optimization algorithm)fine-tunes the network weights and threshold parameters to obtain a DNN network with excellent parameters and is used for wind power forecasting.To evaluate the forecasting effect,there are several representative wind power forecasting methods(BP-ANN,SVM,ARMA),the results show that the DNN prediction effect is significantly better than the other three comparison algorithms,the prediction accuracy is high,the prediction effect is goodIn order to improve the prediction accuracy of DNN,a wind power prediction method based on RF-DNN is proposed.It uses historical samples of wind power characteristic information to train a random forest model,find the prediction importance index PI corresponding to all characteristic information,and then select The optimal feature information set Q,and the test data set consisting of the historical value of the set Q is input as the DNN model,and the wind power value to be predicted is obtained by calculation,and compared with the DNN prediction model without feature information selection,to filter feature information Effectively improve the prediction accuracy of the DNN model.