基于深度数据挖掘的用电行为关联提取与分类技术OA
Extraction and classification technology of power consumption behavior correlation based on deep data mining
[目的]用电行为的精准识别与监测,对电力系统调度计划的科学制定至关重要.当前,用电行为识别算法主要分为机器学习与深度学习两大类.其中,机器学习算法的识别精度尚存提升空间,且模型泛化能力相对不足.深度学习算法虽具备独特优势,但对数据时序特征的挖掘不够充分,易出现过拟合现象,进而降低识别精度.[方法]针对用电行为分类算法中时序特征挖掘不充分、分类精度有待提升的问题,本文提出了一种基于改进双层双向长短期记忆网络与注意力机制(IDL-BiLSTM-Atten)的用电行为分类模型.该模型的核心实现流程为:首先,采用融合累积和(CUSUM)算法精准识别用电设备的投切事件;其次,整合稳态特征、暂态特征与非电量参数,构建全面的用电设备行为特征集.针对传统神经网络因逐层串行连接架构导致的性能瓶颈,本文对双层双向长短期记忆(IDL-BiLSTM)网络进行改进,引入数据跨层传输策略,允许数据跳过部分或全部中间层,直接传递至深层网络进行处理,从而有效增强模型的稳定性与分类精度.同时,为强化关键信息对输出结果的驱动作用,将注意力机制嵌入双层双向长短期记忆(LSTM)模型,进一步提升模型在复杂任务场景下的计算效率与分类精度.[结果]利用典型用电设备生成了 1 500组训练样本并以此作为输入数据,对IDL-BiLSTM-Atten模型进行了训练.结果表明:增加暂态特征与非电量参数作为输入数据,有利于提升模型分类精度;相较于传统双向长短期记忆(BiLSTM)和长短期记忆(LSTM)模型,IDL-BiLSTM-Atten模型在收敛速度和分类精度方面具有更优性能;在实际应用案例中,IDL-BiLSTM-Atten模型能够实现用电行为的精确分类,验证了其在实际场景中应用的可靠性.[结论]相较于传统模型,本文模型的改进之处主要体现在两个方面:其一,拓展输入特征维度,将暂态特征与非电量参数纳入模型输入数据,有效提升模型的分类精度;其二,优化模型结构设计,通过改进BiLSTM模型并嵌入注意力机制,使得模型在收敛速度与分类精度方面均具备更优的性能表现.
[Objective]Accurate identification and monitoring of power consumption behavior are crucial for the scientific formulation of power system dispatching plans.Currently,algorithms for power consumption behavior identification are mainly divided into two categories:machine learning and deep learning algorithms.The classification precision of machine learning algorithms still has room for improvement,and the generalization ability of the models is relatively insufficient.Although deep learning algorithms have unique advantages,they do not fully exploit the temporal features of data and are prone to overfitting,which thereby reduces classification precision.[Methods]To address the problems of insufficient mining of temporal features and the need to improve classification precision in power consumption behavior classification algorithms,a power consumption behavior classification model based on an improved double-layer bidirectional long short-term memory network and attention mechanism(IDL-BiLSTM-Atten)was proposed.The core implementation process of the model was as follows:first,the fused cumulative sum(CUSUM)algorithm was adopted to accurately identify the switching events of electrical equipment;second,steady-state features,transient-state features,and non-electrical parameters were integrated to construct a comprehensive feature set of electrical equipment behavior.To overcome the performance bottleneck of traditional neural networks due to the layer-by-layer serial connection architecture,an improved double-layer bidirectional long short-term memory(IDL-BiLSTM)network was built by introducing a data cross-layer transmission strategy,which allowed data to skip some or all intermediate layers and be directly transmitted to deep networks for processing,thus effectively enhancing the stability and classification precision of the model.Meanwhile,to strengthen the driving effect of key information on output results,the attention mechanism was embedded into the double-layer bidirectional long short-term memory model,further improving the computational efficiency and classification precision of the model in complex task scenarios.[Results]A total of 1 500 sets of training samples were generated using typical electrical equipment and used as input data to train the IDL-BiLSTM-Atten model.The results show that adding transient-state features and non-electrical parameters as input data is conducive to improving the classification precision of the model.Compared with traditional bidirectional long short-term memory(BiLSTM)and long short-term memory(LSTM)models,the IDL-BiLSTM-Atten model has superior performance in convergence speed and classification precision.In practical application cases,the IDL-BiLSTM-Atten model can achieve accurate classification of power consumption behavior,which verifies its application reliability in practical scenarios.[Conclusions]Compared with traditional models,the proposed model shows improvements mainly in two aspects:first,expanding the dimension of input features by incorporating transient-state features and non-electrical parameters into the input data of the model,which effectively improves the classification precision of the model;second,optimizing the model structure design by improving the BiLSTM model and embedding the attention mechanism,which enables the model to have better performance in both convergence speed and classification precision.
王伟;钱旭盛;俞阳;李明;朱萌
中国科学院大学软件研究所,北京 100190||国网江苏省电力有限公司营销服务中心,江苏南京 210019国网江苏省电力有限公司营销服务中心,江苏南京 210019国网江苏省电力有限公司营销服务中心,江苏南京 210019国网江苏省电力有限公司营销服务中心,江苏南京 210019国网江苏省电力有限公司营销服务中心,江苏南京 210019||东南大学电气工程学院,江苏南京 210096
信息技术与安全科学
长短期记忆网络稳态特征暂态特征非电量参数用电行为分类注意力机制事件检测
long short-term memory networksteady-state featuretransient-state featurenon-electrical parameterpower consumption behavior classificationattention mechanismevent detection
《沈阳工业大学学报》 2026 (5)
49-56,8
国家自然科学基金项目(12003056,11903066).
评论