【作者】 张玉森；

【导师】 吴皓；

【作者基本信息】 山东大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 非侵入式负荷监测可以通过单只智能观测总表监测下级电路中负荷的运行状态和功率消耗,相比于侵入式的负荷监测,具有成本低、非侵入性和易于维护等优势。近年来随着人工智能技术的迅速崛起,为非侵入式负荷监测中存在的复杂问题提供了有效的解决方案,从而加快了负荷监测的研究进程,促进了该项技术的应用与推广。在基于人工智能技术的非侵入式负荷监测研究中,存在以下难点问题:①构建具有显著特征和区分性的负荷指纹。②对于基于深度学习的非侵入式负荷辨识,模型具有较强的特征建模能力,在拥有良好识别效果的情况下保持模型的轻量性和高效性。③总功率序列中负荷的波动功率与单个负荷分解功率的幅值存在显著偏差。④模型的实际应用环境复杂多变,其数据分布与模型训练数据的分布存在未知差异。针对上述难点问题,基于深度学习方法对非侵入式负荷监测问题展开研究。主要的研究内容包括以下四个方面。(1)针对目前手工构建的负荷指纹表示和挖掘负荷潜在特征的能力不足问题,提出一种基于图像的可学习负荷指纹构建方法。首先基于时序建模方法提取负荷电流特征,并将提取的特征序列映射到二维空间,然后采用图像识别方法对二维图像特征进行分类得到负荷类别。为了证明可学习负荷指纹方法的有效性,采用可学习的递归图、可学习的格拉姆矩阵以及生成图三种特征映射方式,在公开的非侵入式负荷辨识数据集上进行实验验证。实验结果表明,负荷指纹的特征表达能力明显提高,并能取得更准确的负荷辨识结果。通过多个有效输入的对比实验证明了可学习的负荷指纹对输入具有鲁棒性。在建立图像分类网络的基础上,进行了不同时序建模网络的实验,结果表明基于残差学习的时间卷积神经网络更适用于该负荷指纹的构建。(2)为了提高负荷辨识模型的特征提取能力,提出一种对长距离特征之间的依赖关系进行建模的方法。首先建立基于残差学习时间卷积神经网络提取负荷电流的局部特征,然后引入改进的非局部注意力模块对局部特征间的关联关系进行建模,挖掘负荷潜在的特征信息,最后在公开的数据集上对提出的方法进行验证。实验结果表明改进的非局部注意力模块有效提高了模型的负荷辨识能力,与目前基于深度神经网络的负荷辨识模型相比,构建的基于残差学习的时间卷积神经网络具有轻量性、高效性和准确性。(3)为了解决负荷开启状态时的预测幅值与实际聚合功率的波动幅值存在的偏差问题,根据序列到点模型对负荷状态边界识别的可靠性,提出一种基于开关状态分类的负荷分解方法。该方法首先将负荷功率预测的回归模型转换为负荷状态的分类模型,利用分类模型得到负荷开关状态序列,为了减少模型在推理阶段对计算资源的需求,采用分段预测的方式降低模型推理的次数。然后基于开启时间和关闭宽限时间阈值法剔除干扰状态和填充间断区间。最后通过预测的负荷开关状态序列从总功率序列中提取单个电器的功率消耗。(4)针对用于训练的源域数据与目标域数据分布存在的差异性问题,提出一种基于数据分布参数校正的负荷分解方法。该方法通过在现有负荷分解网络的基础上增加一个分布参数预测模块,用于获取目标域与源域数据分布的相关系数,利用该系数对模型的预测结果进行校正。该方法不需要增加大规模的模型参数,利用较少的计算资源提升原有模型的负荷分解准确性。本文对非侵入式负荷监测中负荷辨识和负荷分解的深度学习方法进行了研究,具体探究了负荷指纹的构建方法,轻量精准的负荷辨识模型和高效准确的负荷分解框架。与基于手工构建的负荷指纹相比,提出的可学习负荷指纹有效提高了负荷辨识的精度;通过引入改进的非局部注意力模块,增强了深度神经网络的特征提取能力;与基于回归模型的负荷分解框架相比,提出的基于开关状态的分解框架有效提高了负荷分解的精度。与基于序列到点的负荷分解模型相比,提出的分段预测方法在保持分解准确性的同时大幅降低了模型的推理次数,对非侵入式负荷分解的实际应用具有重要意义。更多还原

【Abstract】 Non-intrusive load monitoring can monitor the operating status and power consumption of the load in the lower circuit through a single intelligent observation meter.Compared with intrusive load monitoring,it has the advantages of low cost,non-intrusive and easy maintenance.In recent years,with the rapid development of artificial intelligence technology,it provides an effective solution to the complex problems in non-intrusive load monitoring.At the same time,the research process of load monitoring is accelerated by artificial intelligence,which promotes the application and promotion of this technology.In the research of non-intrusive load monitoring based on artificial intelligence technology,there are the following difficult problems:①Construct load fingerprints with discriminative and distinctive features.②For non-intrusive load identification based on deep learning,the model has strong feature modeling ability and superior identification effect while maintaining light weight and minimum computational complexity.③There is a significant deviation between the fluctuating power of loads in the total power sequence and the decomposed power amplitude of a single load.④The actual application environment of the model is complex and changeable,and there are unknown distribution differences between the actual data and the training data.Aiming at the above difficult problems,the non-intrusive load monitoring problem is studied based on the deep learning method.The main research contents include the following four aspects.(1)Aiming at the insufficient ability of the load fingerprint representation and mining the latent features of load constructed manually at present,an image-based learnable load fingerprint construction method is proposed.Firstly,the load current features are extracted based on the time series modeling method,and the extracted feature sequence is mapped to the two-dimensional space,and then the two-dimensional image features are classified by the image recognition method to obtain the load category.In order to prove the effectiveness of the learnable load fingerprint method,three feature mapping methods,namely,Learnable Recursive Graph,Learnable Gramma Matrix,and Generative Graph,are used to conduct experiments on the public non-intrusive load identification dataset.The experimental results show that the feature expression ability of the load fingerprint is obviously improved,and more accurate load identification results can be obtained.The robustness of learnable load fingerprints to inputs is demonstrated through comparative experiments with multiple valid inputs.Based on the establishment of the image classification network,different time series modeling network experiments are carried out,and the results show that the temporal convolutional neural network based on residual learning is more suitable for the construction of the load fingerprint.(2)In order to improve the feature extraction ability of the load identification model,a method for modeling the dependencies between long-range features is proposed.Firstly,a temporal convolutional neural network based on residual learning is established to extract the local features of the load current,then an improved non-local attention module is introduced to model the correlation between local features,and the potential feature information of the load is mined.The proposed method is verified on the set.The experimental results show that the improved non-local attention module effectively improves the load identification ability of the model.Compared with the current load identification model based on deep neural network,the temporal convolutional neural network based on residual learning is lightweight,efficient and accurate.(3)In order to solve the problem of deviation between the predicted amplitude and the actual aggregated power fluctuation amplitude when the load is on,a load decomposition method based on switch state classification is proposed according to the reliability of the sequence-to-point model for load state boundary identification.This method first converts the regression model of load power prediction into a classification model of load state,and uses the classification model to obtain the load switch state sequence.In order to reduce the demand for computing resources in the inference stage of the model,a segmented prediction method is used to reduce inference times.Then based on the on-time and off-grace time thresholds,disturbing states are eliminated and discontinuous intervals are filled.Finally,the power consumption of individual appliances is extracted from the total power sequence through the predicted load switching state sequence.(4)Aiming at the difference between the distribution of target domain data and source domain data used for training,a load decomposition method based on data distribution parameter correction is proposed.In this method,a distribution parameter prediction module is added to the existing load disaggregation network,and the correlation coefficient between the target domain and source domain data distribution predicted by the module is used to correct the prediction results of the model.This method does not need to increase large-scale model parameters,and uses less computing resources to improve the load disaggregation accuracy of the original model.This paper studies the deep learning method of load identification and load decomposition in non-intrusive load monitoring,specifically explores the construction method of load fingerprints,a lightweight and accurate load identification model,an efficient and accurate load disaggregation framework.Compared with the load fingerprint based on manual construction,the proposed learnable load fingerprint effectively improves the accuracy of load identification;By introducing an improved non-local attention module,the feature extraction ability of the deep neural network is enhanced;Compared with the load decomposition framework based on regression model,the proposed disaggregation framework based on switch state effectively improves the accuracy of load disaggregation.Compared with the sequence-to-point based load disaggregation model,the proposed segmental forecasting method greatly reduces the number of inferences of the model while maintaining the accuracy of the disaggregation,which is of great significance for the practical application of non-intrusive load decomposition.更多还原