Research Of STL-SVR Short-term Electric Energy Consumption Forecasting Algorithm Based On Spark

With the continuous advancement of energy conservation and emission reduction,the demand for accurate forecasting.of electric energy consumption has become increasingly urgent.At the same time,the continuous development of Internet of Things technology has enabled the scale of electric energy consumption data collection to expand continuously,and forecasting based on massive data.The stand-alone environment is bound to encounter bottlenecks in computing resources.How to quickly process these massive data for forecasting while meeting the accuracy requirements of electric energy consumption forecasting has become a hot research direction in recent years.Based on this,this paper establishes a STL-SVR electric energy consumption forecasting model,and implements this model to quickly process large-scale data through the Spark distributed processing platform.The main innovations of this paper are as follows:(1)Use STL time series decomposition algorithm combined with SVR to establish STL-SVR power consumption prediction model.For the problem that the single model can't achieve accurate electric energy consumption forecasting,the electric energy consumption data is decomposed into trend item,periodic item and residual item by introducing STL time series decomposition algorithm,and the support vector regression algorithm and other methods is adopted according to the characteristics of each item,the overall model is integrated to achieve accurate forecasting of electric energy consumption data.(2)Use simulated annealing algorithm to optimize super-parameter selection of SVR.Aiming at the problem of slow optimization of parameters of SVR algorithm using grid search,the parameter optimization of SVR is realized by introducing simulated annealing algorithm,which is suitable for large-scale combinatorial optimization problem,which greatly accelerates the parameter optimization efficiency of SVR algorithm.(3)Parallelization of STL-SVR Power Consumption Prediction Model.Aiming at the problem of insufficient computational resources and slow operation of the algorithm in the single-machine environment under massive data,the Spark distributed processing platform was introduced.Parallelization of data preprocessing,feature engineering processing and SVR algorithm was implemented on Spark,and Spark-based STL-SVR power consumption prediction model.While ensuring the accuracy of prediction,the training time of the model under massive data is reduced.In this paper,based on the proposed STL-SVR prediction model,three experiments are carried out,which are the comparison experiments of regression forest,SVR and STLSVR models;the comparison experiment of SVR parameters optimization using simulated annealing algorithm and grid search;the stand-alone environment and Spark Comparative experiment of the STL-SVR model under the environment.The experimental results show that the Spark-SVR prediction model based on Spark is more competitive than the traditional algorithm in the single-machine environment,and has a significant improvement in prediction accuracy and parameter optimization efficiency.It can also maintain high training efficiency under massive data.