Spark-based Parallel OS-ELM Algorithm Application for Short-term Load Forecasting for Massive User Data

Abstract

The data type and quantity of user load data show an exponential growth, so that the traditional load forecasting methods can hardly meet the load forecasting requirements of massive users. Aiming at this problem, a parallel OS-ELM short-term load forecasting model based on Spark is proposed in this article. By analyzing the characteristics of the Spark framework and the MapReduce framework, the Spark big data processing framework is determined as the basic framework for processing massive user load data, and a parallel K-means load clustering model based on Spark is designed. The on-line sequential learning machine OS-ELM makes the hidden layer data of computing each incremental training dataset mutually independent, therefore, a Spark-based parallel OS-ELM (SBPOS-ELM) algorithm is put forward. The proposed model is applied under the smart electricity big data environment and the training samples are selected using the incremental training dataset to make a short-term prediction of the millions of users’ smart meter electricity load, which verifies the feasibility and effectiveness of the proposed model. At last, comparing with other commonly used short-term load forecasting algorithms, the experimental results show that SBPOS-ELM algorithm has higher accuracy and operation efficiency.