Research On LS-SVM-based Q-Learning Algorithm For Large-scale Energy Prediction Problem In Iron And Steel Industries

Iron and steel industry is among the most energy-intensive ones in China. With the increasingly pressing of domestic energy situation and deteriorative of environment, the government is attaching more and more importance to energy consumption of iron and steel enterprise. Under such specific background, if the iron and steel enterprise could improve their energy efficiency and innovate technology on the basis of energy-saving, not only could energy costs be improved, but also the competitiveness of enterprises can be improved. In the thesis, the energy media consumption of iron and steel industry is used as the research background, and the hybrid method of Q-learning and Least Squares Support Vector Machine is designed to solve the problem of large-scale energy prediction.(1) During the course of energy media prediction in iron and steel enterprises, using large-scale historical data as the training data can effectively improve the prediction accuracy. But large-scaled prediction problem will always lead to long prediction time, low prediction accuracy and instable prediction model. This thesis uses day as the time unit and analyses historical training data of a year to predict energy consumption future one day or several days into the future.(2) To deal with the problem of long running time and low prediction accuracy, we introduce the Least Squares Support Vector Machine in Q-Learning System. When the algorithm trains the large-scale historical data, the LS-SVM can prevent the Q learning from visiting the all state and speed up the system learning process, thus realize efficient learning in a bigger knowledge space with limited learning experience and memory. Experimental comparison with BP-based prediction algorithm shows that the hybrid algorithm can better follow the actual value under the same error, which proves the validity of this algorithm.(3) To further improve the performance of algorithm in solving the prediction problems with the large-scale historical training data, this thesis introduces online learning technology of minimum time window into the hybrid algorithm to improve the speed and the prediction ability for different data size, also the Doolittle decomposition method is adopted to replace the normal Guass decomposition-method for solving the matrix of the LS-SVM to improve the speed of the algorithm. In the end the simulated annealing idea is introduced into the Q learning algorithm to balance the contradiction between the exploration and utilization and to improve the accuracy of the algorithm. Comparison of the improved algorithm with the original algorithm, the running results show that the speed and the accuracy are improved significantly.(4) Based on the algorithm performance and the practical requirement, the iron and steel enterprise energy forecasting system is developed. The system provides the predict function for different energy media in different operations. The function of operation management energy media management and the management of historical data are designed. Also, rich graphs are presented. This system serves as a convenient tool for the decision makers.