【作者】 王美萍；

【导师】 田琦；

【作者基本信息】 太原理工大学 ， 环境工程， 2017， 博士

【摘要】 供热是关系国计民生的重要基础行业和公用事业。随着供热的商品化,供热系统热负荷预测作为供热规划、生产、调度和交易等工作的基础,在供热系统安全和经济运行中起着至关重要的作用。供热负荷预测精度的高低直接影响到供热系统的供热质量、安全性和经济性。随着供热系统节能减排进程的不断推进和智慧供热的需求,使供热负荷预测越来越成为该领域研究的前沿和热点问题,其研究对节能减排、治理雾霾具有重要的意义。由于滞后性、管网的热损失、用户种类复杂程度的差异性等导致了城镇供热系统不同热量结算点热负荷具有不同的规律特点,本文将供热负荷按层级热量结算点来划分,分析各自的影响因素,引入智能算法以及相关组合理论预测技术,以城镇供热系统实测数据为基础,对各层级供热负荷预测的理论与方法进行深入研究,为供热系统运行管理提供较为科学的决策依据。主要研究工作和创新成果如下:(1)对不同层级热量结算点热负荷的特点、影响热负荷的因素及导致热负荷预测误差的相关因素进行了分析;针对历史数据样本的离群数据进行纵向和横向预处理,使其能够更加与热负荷实际运行趋势一致,进而为后期利用这些历史数据样本进行各级热量结算点短期热负荷预测奠定了基础;将相关性分析应用在各层级热量结算点热负荷预测模型输入维数的选择上,使得输入变量与各层级预测热负荷相关性更强,为提高预测结果的准确性和改善预测性能做好进一步的准备;此外,对进入模型的各参数进行归一化处理,避免进入模型的各参数因数值差异大而导致预测性能下降。(2)基于结构风险最小化原则,提出粒子群(pso)优化支持向量机(svm)模型热源热负荷预测方法。该方法对解决系统大热惯性、大时滞性导致热源热负荷随室外温度变化的非线性问题有较好的效果。建立了遗传(ga)优化支持向量机(svm)、标准支持向量机(svm)及粒子群(pso)优化支持向量机(svm)热源热负荷三种预测模型,通过相关性分析并确定预测模型输入变量的维数,证明了粒子群(pso)优化支持向量机(svm)模型在预测精度和泛化能力方面均优于其他两种预测模型。(3)针对一般热交换站用户类型较单一、样本容量大的问题,提出基于adaboost组合多个弱预测器构建出一个强预测器的热交换站热负荷预测方法。弱预测器采用处理大样本、容错能力强的bp神经网络模型,其网络阈值和权值的优化选用经过筛选出的粒子群算法(pso)。利用adaboost理论对9个粒子群(pso)优化bp神经网络预测进行组合构建出一个强预测模型。针对热交换站热负荷及其相关参数历史数据样本进行相关性分析,筛选出与热负荷最相关的影响因素作为预测模型的输入变量维数,最后通过与粒子群(pso)优化bp神经网络方法和未经优化的传统bp神经网络方法进行实验比较,证明本文提出的预测模型有效提高了热交换站热负荷的预测精度和泛化能力。(4)针对建筑热负荷样本数量少及热计量引起的用户调节规律不确定问题,提出两种组合预测方法。将解决小样本非线性问题的粒子群(PSO)优化支持向量机(SVM)模型和容错能力强的粒子群(PSO)优化BP神经网络模型作为组合方法中单一预测模型,基于信息熵理论提取单一预测模型中的有用信息,将提取的有用信息进行融合产生出更强预测能力的组合方法;在基于Adaboost组合粒子群(PSO)优化BP神经网络模型思想上,激发了将处理小样本和非线性问题的支持向量机(SVM)模型作为弱预测器,结合Adaboost理论构建出由8个弱预测器组成的建筑热负荷强预测模型。对建筑热负荷及其相关参数进行相关性分析,找出适合各自预测模型的输入变量,通过实例验证,以上两种组合预测方法均较单一预测模型有较高的预测精度,其中基于信息熵权组合方法更胜一筹,能更好地对住宅建筑热负荷进行预测。更多还原

【Abstract】 Heating is the important basic industries and public utility to the national economy and people’s life.With the commercialization of heating,heating load forecasting plays a vital role in the safety and economic operation of heating system,and undertakes the basis work of heating system planning,production,dispatch and transaction.The accuracy of heat load forecasting directly affects the quality,safety and economy of heating system.With the heating system continuous improvement of energy-saving emission reduction and intelligent heating demand,the heat load forecasting has become a hot and leading research topic,its study is of great significance on energy-saving emission reduction and management of haze.Because of the difference in big delay,pipeline heat loss,and complexity degree of user categories,the heat load of different heat settlement site have different characteristics in urban heating system.This paper divides the heat load forecasting by the heat settlement site,and analyzes their influence factors.Based on measured data of urban heating system,it studies the theory and method of different scale level heat load forecasting in depth through a variety of intelligent algorithms and the combination of theoretical prediction techniques,and provides a more scientific basis for decision-making of the operation and management of heating system.The main research works and innovations are as follows:(1)Analyze the characteristics of different scale of heat load,the factors affecting the thermal load,and the factors that cause the heat load forecasting error.Through vertically and horizontally pretreated the historical data samples,so that to be more consistent with the actual running trend of the heat load,and then lay the foundation for the short and midum term heat load forecasting by the scale heat settlement site.The correlation analysis is applied to the selection of the input dimension of the heat load forecasting model of each heat settlement site.By this method,the input variable parameters are more correlated with the predicted heat load and make preparations for enhancing the accuracy of the model.Moreover,the parameters of the model are normalized,so as to avoid the decline of the prediction performance.(2)Based on structural risk minimization,Support Vector Machine model optimized by Particle Swarm Optimization algorithm(PSO-SVM model)was proposed which is to forecast heat load of the heat source.The method has a better effect for solving the nonlinear problem of the thermal load of heat source with the outdoor temperature caused by the system with large thermal inertia and large time delay.The paper established three heat load forecasting model,namely,Genetic algorithm(GA)optimizing support vector machine(SVM),Standard support vector machine(SVM)and PSO optimizing support vector machine(SVM),determined the dimension of the input variables of three models by correlation analysis,and verified that the PSO-SVM model is superior to the other two prediction models in both accuracy and generalization ability.(3)For the problems of simplex user type and of large data samples,the paper proposed a heat load forecasting method for heat exchange station which is based on Adaboost combined with multiple weak predictors to construct a stronger predictor.The weak predictor uses BP neural network model with large sample handling and strong fault tolerance,which select PSO to optimize network threshold and weight.Apply Adaboost theory to construct a strong forecasting model by combining nine weak predictors of PSO-BP neural networks.According to the correlation analysis of the thermal load of the heat exchange station and its influence factors,these influence factors of the heat load are selected as the input variables of forecasting model.Finally,through the experiment to compare the method of PSO-BP neural network to the method of traditional BP neural network to proved that the prediction model proposed in this paper can effectively improve the prediction accuracy and generalization ability of heat load in heat exchanging station.(4)In view of the fact that the number of building heat load samples is small and the regulation of the users caused by heat metering is uncertain,two kinds of combination forecasting methods are proposed.Taking the better prediction performance of PSO-SVM model and PSO-BP neural network model as a single method in combination forecasting model,and basing on information entropy theory to extract useful information from single prediction model,then combining the extracted useful information produces a combination method of stronger predictive capabilities.Based on the thought of Adaboost algorithm,the PSO optimizing support vector machine model which will deal with small sample problem is used as weak predictor.utilizing Adaboost theory,it structures a strong prediction model of building heat load composed of eight weak predictors.Making the correlative analysis of the building heat load and its influence factors and finding the suitable input variables for the respective forecasting models.Through the example verification that the above two kinds of combination forecasting methods have higher prediction accuracy than the single prediction model,and the combination methods based on information entropy even better,it can be better on heat load prediction of the residential building.更多还原