【作者】 王琎；

【导师】 宫铭举；

【作者基本信息】 天津理工大学 ， 信息与通信工程， 2020， 硕士

【摘要】 区域供热系统保障了公共建筑、住宅建筑等各类建筑群的冬日采暖需求,是我国北方城镇采暖供热的主要组成部分。但其系统智能化程度普遍较低,容易受到外界因素干扰,导致其供热不均,能源浪费现象严重。对热负荷供应量的准确规划与控制是保证供热质量和节能减排的重要前提。为此,本文以供热系统短期热负荷预测为研究对象,具体内容包括:首先,热负荷容易受到不同因素的影响,利用Pearson与LASSO特征选择方法,对不同特征变量与热负荷之间的关系进行分析,并以此为依据构建特征集。针对区域供热系统非线性、时滞性、热惯性与时变性的特点,利用离散小波变换(Discrete Wavelet Transform,DWT)对非线性信号多尺度分辨的优势,结合集成学习算法中的极端随机树(Extreme Random Tree,ETR)、梯度提升树(Gradient Boosting Decision Tree,GBDT),对未来一小时的热负荷建立混合预测模型,并将人工神经网络与支持向量回归作为对照实验。结果表明,混合模型比单一模型预测精度更高,LASSO方法所选择特征集下的DWT-ETR模型可以更加准确地预测热负荷。此外,将历史热负荷作为模型输入可以提高预测精度。其次,由于热负荷序列具有时间序列的特征,所以利用长短时记忆网络(Long and Short-Term Memory,LSTM)处理时间序列上的优势,对未来1小时和24小时的热负荷建立预测模型。同时利用经验模态分解(Empirical Mode Decomposition,EMD)对热负荷序列进行分解,由分解后的本征模态函数和残差作为模型输入。实验结果表明,EMD-LSTM模型在两种提前预测时间下均具有良好的预测精度。然后,根据换热站实测数据,提出一种基于供回水温度的系统延迟时间估计方法。对不同特征与二次回水温度之间的相关性进行分析,并构建不同特征集来分析不同特征对二次回水温度预测结果的影响。针对供热系统时滞性对换热站温度控制的影响,将系统延迟时间作为预测时间步长。同时针对GBDT算法易积累训练误差的缺点,引入Light GBM(Light Gradient Boosting Machine,LGBM)算法,结合DWT对二次回水温度进行预测。实验结果表明,在高相关系数组成的特征集下的DWT-LGBM模型可以更准确地预测二次回水温度,有助于提高管理人员对系统温度控制效率。最后,对本文的工作与主要创新点进行了总结,同时对本课题研究以后的发展前景给出了自己的见解。更多还原

【Abstract】 The district heating system ensures the winter heating demand of public buildings,residential buildings and other buildings.It is the main way of heating in northern cities and towns of China.However,its system intelligence is generally low,and it is susceptible to external interference,resulting in uneven heating and serious energy waste.Accurate planning and control of energy supply is the important prerequisite to ensure the quality of heating and energy conservation and emission reduction.Therefore,this paper takes the heat load prediction of heating system as the research object,including:First of all,the heat load is easy to be affected by different factors.The relationship between different feature variables and heat load is analyzed based on Pearson and lasso feature selection method,and the feature sets are constructed based on the analysis result.According to the characteristics of non-linear,time-delay,thermal inertia and time-varying of district heating system,using the advantage of discrete wavelet transform(DWT)in multi-scale resolution of non-linear signals,and the extreme random tree(ETR)and gradient boost decision tree(GBDT)in the ensemble learning algorithm,hybrid prediction models for the heat load in the next hour are established.Meanwhile,artificial neural network and support vector regression was used as the control experiment.The experimental results show that the hybrid model has higher prediction accuracy than the single model.The DWT-ETR model under the selected feature set of LASSO method can predict the heat load more accurately.In addition,the prediction accuracy can be improved by using the historical heat load as the input parameter of the model.Secondly,because of the characteristics of time series on the heat load series,using the advantages of long and short-term memory network(LSTM)to deal with time series,prediction models of heat load in the next hour and 24 hours are established.simultaneously,empirical mode decomposition(EMD)is used to decompose the heat load series,and the EMD component and residual after decomposition are used as the model input.The experimental results show that the EMD-LSTM model has good prediction accuracy under both pre-prediction times.Then,according to the measured data of heat exchange station,a method of system delay time estimation based on supply and return temperature is proposed.The correlation between different features and secondary return temperature is analyzed,and different feature sets are constructed to analyze the influence of different features on the secondary return temperature prediction results.In view of the influence of time delay of heating system on the temperature control of heat exchange station,the system delay time is taken as the prediction aheat time.Meanwhile,aiming at the shortcomings of GBDT algorithm which is easy to accumulate training errors,Light GBM(LGBM)algorithm is introduced to predict the secondary return temperature combined with DWT.The experimental results show that the DWT-LGBM model can predict the secondary return temperature more accurately under the feature set of high correlation coefficient,so as to improve the temperature control efficiency of the system.Finally,the work and main innovations of this paper are summarized.Simultaneously,several views on the future development of this research are prospected.更多还原