Research On Distributed SVR In Power Load Forecasting

The load forecasting of electric power system not only plays a very important role in people's production and life,but also in the efficient and economical operation of power network.With the popularity of smart grid,most of the power grid data generated by the city every day in millions of,and the dimensions of the data is also expanding.Now under the environment of traditional single load forecasting method has been stretched,distributed computing platform is to solve the single computing resources and storage resources bottleneck.The traditional single SVR algorithm is suitable for small data,and it is not suitable for arge data sets because of long training time.The parallel SVR algorithm has become a hot research topic in recent years.This paper combines the most popular distributed computing platform Hadoop and the parallel SVR algorithm for short-term load forecasting.The main work has the following 4points:(1)Building a Hadoop cluster,using the MapReduce framework and Libsvm algorithm package to achieve a 2 layers iterative structure of parallel SVR algorithm.(2)The related data attributes that affect the load forecasting are compared with the experimental data,extracts the feature of the sample data and constructes the structure of the training data.(3)Using UD-SVR method to optimize the parameters,and further improve the accuracy of the distributed SVR algorithm.(4)Three experiments are designed to contrast the training time and the prediction accuracy between single SVR and distributed SVR.The parameter optimization effect of UD-SVR is verified and the results are analyzed.Experimental results show that: With the increase of the amount of data and computation time of the single SVR is exponential growth,and the computation time of the distributed SVR has increased not obvious.By comparing the error,the accuracy of distributed SVR prediction is not significantly decreased compared with that of the single machine SVR prediction;With the increase of the Hadoop cluster parallel Map task,the speedup of the distributed SVRcomputing is also increasing.