基于LS-SVM的Q学习算法求解钢铁企业大规模能源预测问题的研究

发布时间：2018-01-03 17:11

本文关键词：基于LS-SVM的Q学习算法求解钢铁企业大规模能源预测问题的研究　出处：《东北大学》2014年硕士论文　论文类型：学位论文

【摘要】：钢铁企业是中国能源消耗大户,随着国内能源形势的不断紧迫和自然环境的逐步恶化,政府对钢铁企业能源消耗问题越来越重视,在这样一个特殊背景下,钢铁企业提高自身能源利用率,开发节能减排技术,不仅仅节约能源成本,而且提高企业综合竞争力。本文以典型钢铁企业能源介质的消耗预测为背景,设计Q学习与最小二乘支持向量机混合方法来对大规模能源预测问题进行求解,以达到节省能源,提高效率的目的。主要内容如下：(1)钢铁企业对其能源介质消耗进行预测时,大规模的历史数据作为训练数据可以有效提高预测结果精度,但大规模的预测问题,会存在预测时间长、预测精度差、预测模型不稳定等问题。本文以天为时间单位,利用钢铁企业一年的历史训练数据进行分析测试,对未来一天或者多天的能耗情况进行预测。(2)本文提出了基于最小二乘支持向量机的Q学习方法来解决钢铁企业大规模能源预测时存在的预测时间长、预测精度差等问题。大规模历史数据进行训练时,最小二乘支持向量机可以避免Q学习系统遍历所有状态,加快系统学习速度,使其在有限的学习经验和有限的记忆中实现对一个更大范围空间知识的有效学习和表达。混合算法与基于PSO的神经网络预测算法作比较实验显示,在误差相差不大情况下,本文混合算法的预测值可以更好的跟随实绩值,为此证明了本文算法的有效性。(3)为了进一步提高混合算法的性能,将最小时间窗的在线学习技术引入训练学习集中增加系统的训练规模,将Doolittle分解法代替普通高斯分解法求解最小二乘支持向量机的线性方程加快算法的预测速度；将模拟退火算法引入Q学习算法中提高算法精度。数值试验显示,改进后的算法相对与原算法速度和精度显著提高。(4)根据算法预测性能以及用户需求,本文开发了钢铁企业工序能源预测系统。系统主要设计了对不同工序中所含介质的预测功能,提供了工序管理、能源介质管理、历史实绩数据管理等基础性管理,并且提供了丰富的图表功能,方便决策者管理。
[Abstract]:Iron and steel enterprises are the major users of energy consumption in China. With the constant urgency of domestic energy situation and the gradual deterioration of the natural environment, the government pays more and more attention to the energy consumption of iron and steel enterprises, under such a special background. Iron and steel enterprises to improve their own energy utilization, the development of energy-saving emission reduction technology, not only to save energy costs, but also to improve the comprehensive competitiveness of enterprises. A hybrid method of Q learning and least square support vector machine is designed to solve the large-scale energy prediction problem in order to save energy. The purpose of improving efficiency. The main contents are as follows: 1) when iron and steel enterprises predict their energy consumption, large-scale historical data as training data can effectively improve the accuracy of prediction results. However, large scale prediction problems, such as long prediction time, poor prediction accuracy, unstable prediction model and so on. This paper takes days as the time unit, using the historical training data of iron and steel enterprises to analyze and test. This paper proposes a Q-learning method based on least squares support vector machine (LS-SVM) to solve the problem of large scale energy prediction in iron and steel enterprises. The least square support vector machine can avoid the Q learning system traversing all states and accelerate the system learning speed. The hybrid algorithm is compared with the neural network prediction algorithm based on PSO to realize the effective learning and expression of a larger range of spatial knowledge in the limited learning experience and limited memory. In the case of little error difference, the prediction value of the hybrid algorithm can better follow the performance value, which proves the effectiveness of the algorithm. 3) in order to further improve the performance of the hybrid algorithm. The minimum time window online learning technology is introduced into the training learning focus to increase the training scale of the system. The Doolittle decomposition method is used instead of the ordinary Gao Si decomposition method to solve the linear equations of the least squares support vector machine (LS-SVM). The simulated annealing algorithm is introduced into the Q-learning algorithm to improve the accuracy of the algorithm. Numerical experiments show that the improved algorithm improves the speed and accuracy of the algorithm significantly compared with the original algorithm. In this paper, a process energy prediction system for iron and steel enterprises is developed. The system mainly designs the prediction function of medium contained in different processes, and provides process management and energy medium management. Historical performance data management and other basic management, and provides a wealth of chart functions for decision-maker management.
【学位授予单位】：东北大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：F426.31;TP18

【相似文献】