【作者】 徐萌；

【导师】 王雷； 王新立；

【作者基本信息】 山东大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 在“碳达峰、碳中和”目标背景下,综合能源系统(Integrated Energy Systems,IES)能够将各种能源紧密联系起来,进一步提升能源利用效率,是能源转型过程中的有效手段。负荷预测作为需求侧能量管理的重要环节,对IES规划和运行调度至关重要。现有研究方法难以精确捕捉每种负荷固有的不确定性,也未能充分提取不同负荷之间复杂的耦合关系,导致预测性能受到影响。本文旨在解决综合能源系统中多元负荷不确定性组分的量化难题,通过深入分析负荷数据,并构建基于深度学习时序特征挖掘的概率预测模型,用以量化预测所具有的不确定性,进而达到对IES负荷的概率预测。本文主要研究工作及创新点如下:(1)针对电负荷时序数据中潜在的时间偏移特征提取的问题,提出基于周期转移注意力网络的电负荷概率预测模型(Q-PSANet)。该模型利用GRU学习负荷的短期依赖特征,并在周期动态特征提取模块中对历史天中同一时刻的相邻时段进行动态赋权,捕捉关键时序信息,以处理潜在的时间偏移特征;其次,将分位数回归引入到模型中,通过设置不同的分位点,能够构建出在不同置信水平下的预测区间,从而有效地捕捉数据和模型的不确定性。实验结果表明,Q-PSANet模型可以充分提取周期动态相似特征,相较于当前主流预测模型,能够有效提升概率预测的准确性和可靠性。(2)针对多元负荷时序数据中的复杂耦合特征提取的问题,提出基于双自注意力网络的多元负荷概率预测模型(Q-DSANet)。该模型采用TCN和GRU分别捕获多元负荷的全局长期和局部短期特征,并通过自注意力机制以提高对关键特征的识别,从而学习电、冷和热负荷序列之间的复杂耦合关系;同时使用自回归模块,来处理神经网络针对输入尺度不够敏感的问题。最后,利用基于多任务学习的分位数回归框架,实现多元负荷联合概率预测。实验结果验证该模型在确定性预测准确度、概率预测精度以及性能等方面均优于当前主流预测模型,可以有效量化综合能源系统中的多元不确定性成分。(3)基于上述模型,设计并构建IES多元负荷预测系统。基于Spring Boot、Vue.js等技术,设计数据库和系统整体架构,并完成图形化操作软件的前后端搭建工作。实现的功能包括登录认证、系统权限管理、电、冷和热负荷数据管理和预测等。管理人员只需选择相应服务,系统即可对历史数据进行多元负荷概率预测,并生成可视化的预测结果,有助于能源部门制定科学的用能管理计划。更多还原

【Abstract】 In the context of the goal of "carbon peaking and carbon neutrality",the integrated energy systems closely link various energy sources,thereby enhancing energy utilization efficiency and serving as a pivotal mechanism in the energy transition process.As an indispensable part of demand-side energy management,load forecasting is crucial to comprehensive energy system planning and operation scheduling.Current research methodologies encounter challenges in precisely capturing the inherent uncertainty associated with each load,and they also fail to fully extract the complex coupling relationships between different loads,causing the prediction performance to be affected.This thesis aims to solve the uncertainty problem in multi-energy load forecasting of integrated energy systems.By in-depth analysis of load data and building a probabilistic prediction model based on deep learning time series feature mining,this research can quantify the uncertainty of predictions and achieve load probabilistic forecasting in integrated energy systems.The principal contributions and novel aspects of this thesis are outlined as follows:(1)Aiming at the problem of potential time offset feature extraction in electrical load time series data,an electrical load probabilistic prediction model based on periodic shifting attention network(Q-PSANet)is proposed.The model employs the GRU to learn the short-term dependency characteristics of electrical load.Additionally,it dynamically weights adjacent periods in the same time slot across historical days within the periodic dynamic feature extraction module,which captures key temporal information and addresses potential time offset characteristics effectively;secondly,quantile regression is introduced into the model,and by setting different quantile points,prediction intervals at different confidence levels can be constructed,thereby effectively capturing data and model uncertainty.Experimental results show that the Q-PSANet model can fully extract periodic dynamic similarity features,and can effectively improve the accuracy and reliability of probabilistic predictions compared with current mainstream prediction models.(2)Aiming at the problem of complex coupling feature extraction in multi-energy load time series data,a multi-energy load probabilistic forecasting model(Q-DSANet)based on a dual self-attention network is proposed.The model uses TCN(Temporal Convolutional Networks)and GRU to capture the global long-term and local short-term characteristics of multi-energy loads respectively,and introduces a self-attention mechanism to enhance the focus on important features,thereby learning the complex coupling relationship among electrical,cooling and heating load sequences;Simultaneously,it applies an autoregressive model in parallel,to solve the problem of insensitivity of neural networks to input scale.Finally,quantile regression based on multi-task learning framework is used to achieve joint probabilistic prediction of multi-energy loads.Experimental results verify that the model is superior to current mainstream prediction models in terms of deterministic prediction accuracy,probabilistic prediction performance and reliability,which can effectively quantify multivariate uncertainty components in integrated energy systems.(3)A multi-energy load forecasting system for the integrated energy system has been designed and developed based on the above model.This system is developed using Spring Boot,Vue.js,and various other technologies to create comprehensive graphical operating software.It covers front-end and back-end development,including database design and system architecture.Functions implemented include login authentication,system authority management,electrical,cooling and heating load data management and prediction,etc.Managers only need to select the corresponding service,and the system can conduct multi-energy probabilistic load forecasting on historical data and generate visual prediction results,which will help the energy department formulate scientific energy management plans.更多还原