基于LSTM和序列到点模型的非侵入式负荷分解技术

Non-Intrusive Load Decomposition Techniques Based on LSTM and Sequence-to-Point Models

随着深度学习在非侵入式负荷监测技术(Non-intrusive Load Monitoring, NILM)的应用,对于负荷识别与功率分解能力有所提升,但是对于多状态用电负荷依旧存在负荷分解准确度低、算法泛化性能低、分解耗时等问题。为此,提出了一种基于LSTM和序列到点的负荷分解模型,滑动总功率数据序列来映射目标设备在功率数据窗口中点的功耗。首先,采用基于滑动窗口的事件探测算法提取激活负荷样本作为序列到点模型的输入数据;利用卷积神经网络自动提取输入负荷总功率的局部负荷特征,引入长短期记忆网络挖掘序列中点前后相关度最高的信息完成负荷辨识。为了验证模型的有效性,将所提出的模型应用于实际家庭能源数据集UK-DALE,并与目前领先的模型进行了比较,综合性能提升了28.8%,结果表明所提出的深度学习模型可以有效地提高负荷监测能力。

With the application of deep learning in non-intrusive load monitoring(NILM),the ability of load identification and power decomposition has been improved, but there are still problems such as low accuracy of load decomposition, low generalization performance of the algorithm and the decomposition being time-consuming for multi-state electricity loads. A load decomposition model based on LSTM and sequence-to-point is proposed to map the power consumption of the target device at the midpoint of the power data window by sliding the total power data sequence. Firstly, a sliding window-based event detection algorithm is used to extract the activated load samples as the input data for the sequence-to-point model. Convolutional Neural Networks is used to automatically extract the local load characteristics of the total power of the input load, and a long and short-term memory network is introduced to mine the information with the highest correlation before and after the points in the sequence to complete the load identification. To verify the effectiveness of the model, the proposed model is applied to a real household energy dataset, UK-DALE,and compared with the current leading models, and the comprehensive performance is improved by 28.8%. The results show that the proposed deep learning model can effectively improve the load monitoring capability.