Large flow compressed air load forecasting based on Least Squares Support Vector Machine within the Bayesian evidence framework

Abstract

Energy-saving of compressed air system was very important for the sustainable development of enterprises, which could be achieved though fast and accurate load forecasting. In this paper, according to the distribution rules and characteristics of 24 hours compressed air supply, the 24h compressed air flow demand model was firstly built with least square support vector machine (LSSVM). In order to avoid the long time consumption for determining the model parameters in the traditional cross validation method, Bayesian evidence framework was selected to train the parameters, and then identified and optimized them. Meanwhile, Nyström low- rank approximation decomposition algorithm was used to accelerate kernel matrix decomposition process. Though the experimental verification with real industrial data, the modeling time of LSSVM within Bayesian evidence framework is reduced to 1/20 compared with traditional cross-validation method; in the contrast with Practical Swarm Optimization (PSO), the modeling time is reduced to 80%, and the prediction accuracy can increase 14.3%, proving this method quite suitable for fast and accurate forecasting for large flow compressed air load.