Short-term Load Forecasting Based On Combination Model With SVR

Short-term load forecasting is an important part in electric power system operation and control.Modern society becomes more and more unable to separate from electric energy.An accurate report about short-term load forecasting can guide the department of power grid to coordinate the balance between power production and power load consumption,which ensures that residents and enterprises could use electricity normally and safely.With the vigorous development of electric power industry,the accuracy of short-term load forecasting is required to improve step by step.Compared with the past,the law of power load change is becoming more and more complicated.How to use some advanced technologies for improving the accuracy of short-term load forecasting is currently a hot research in power-related industries.In this paper,a short-term load forecasting method based on combination model with Support Vector Regression(SVR)was proposed,which focused on solving two aspects that were paid little attention to in previous methods:(1)Feature selection.A feature selection method with classification was proposed in this paper.Firstly,the feature variables in the training samples were divided into two categories according to their respective characteristics.Secondly,the Spearman Correlation Coefficient and the Max-Relevance & Min-Redundancy(m RMR)algorithm were used for intra-class selection respectively.Thirdly,the optimal dimension of the comprehensive feature set was determined according to the Bayesian Information Criterion(BIC).This method is able to reduce the redundancy between feature variables,determine the appropriate number of feature variables.and ensure the prediction performance of the model without making the model too complex.(2)Model combination.In this paper,the Recursive-Support Vector Regression(RSVR)load forecasting model and the Gated Recurrent Unit Networks(GRU)load forecasting model were constructed as two single forecasting models.Among them,the RSVR model is an improved model of the SVR model.The historical load data of the previous moment was introduced to alleviate the problem of weak sensitivity of the SVR model to the time series data.The GRU model is a popular recurrent neural network model.In this paper,the General Regression Neural Network(GRNN)combination model which combined the similar day algorithm was used to combine the two single forecasting models together to complement each other’s advantages.The similar day algorithm could reduce the number of training samples needed to construct the GRNN combination model and is able to improve the data quality.The GRNN model not only has the self-learning ability shared by other neural networks,but also has a clear theoretical basis,which is suitable for model combination.This method makes the combination mode of the two single forecasting models closer to the optimal.The power load data of Singapore from April to December 2019 were used in this experiment,in which the Mean Absolute Percentage Error(MAPE)and the Root Mean Squared Error(RMSE)were applied as the evaluation standards.The experiment results show that the forecasting method proposed in this paper has higher prediction accuracy than other single forecasting models and other combination models.In addition,a visual system of power load forecasting was designed and implemented in this paper.The system is developed based on Java Script,adopts browser/server architecture and uses Mongo DB database to store data.In this system,the load forecasting results are obtained by the proposed forecasting method from this paper.This visual system can visually display historical sample data,abnormal data,similar-day sample data,load forecasting results and errors,which is convenient for all kinds of users to carry out the works of analysis and research according to their needs.