Research On Short-term Power Load Forecasting Based On Data Mining And Machine Learning

High-precision load forecasting is of great significance for adjusting power generation plans and rationalizing energy use,which is one of the prerequisites for safe and stable operation of power systems.Since the power load itself is highly random and there are many factors affecting the short-term load,the traditional forecasting methods cannot meet the current requirements of the power system for load forecasting accuracy.In view of this,this paper combines clustering methods and machine learning models for research.In order to fully exploit the electric load data,a similar day set construction method based on MF optimized clustering and mahalanobis distance is proposed.On the one hand,for the problem that the fuzzy C-mean clustering(FCM)cannot automatically determine the number of clusters and the mean drift algorithm clustering effect is greatly affected by the bandwidth parameters,a MF optimized clustering method is proposed by combining the two.On the other hand,combining the results of MF optimized clustering,the set of similar days is constructed by calculating the mahalanobis distance between the daily load to be predicted and similar loads.The effectiveness of the proposed method is proved by experimental comparison.To improve the short-term load prediction accuracy,a CNN-GRU based short-term prediction model is built by combining convolutional neural network(CNN)and gated recurrent unit(GRU)under similar days.the CNN has a better ability to extract the nonlinear features of the input vector,and the GRU can capture the time-series relationship in the electric load data,and the hybrid model can fully utilize the advantages of both models.A grid search method is used to search for the best parameters of the CNN-GRU model for model training and prediction,and several models are built at the same time.The prediction errors of each model are compared with the example analysis of the power load data in a place in East China,and the higher prediction accuracy of the proposed model is verified.In order to further improve the prediction effect of the model,a short-term prediction model based on GWO-TCN-GRU is proposed by improving the CNN-GRU model using temporal convolutional network(TCN)and gray wolf optimization algorithm(GWO).The feature extraction ability of TCN is stronger than that of CNN,and the optimization of hybrid model parameters using GWO can solve the problem of traditional parameter search method The GWO-TCN-GRU model extracts the input variable features by TCN layer and captures the time series relationship of power load data by GRU model.Then the root mean square error value of the TCN-GRU model is used as the fitness function for iterative optimization search,and the optimal solution corresponding to GWO at the end of the iteration is the best parameter of the model.Based on the same data,a particle swarm optimization TCN-GRU model(PSO-TCN-GRU),TCN-GRU,and CNN-GRU models are established for comparison experiments,and the results show that the proposed model has lower prediction error,thus verifying the effectiveness of the proposed model.