Non-Invasive Load Monitoring Algorithm Based On Variable-Length Time-Series Waveform Positioning

Improving energy efficiency and achieving carbon neutralization is a long-term sustainable development goal in China.With the in-depth development of smart grid construction,smart grid applications such as ultra-short-term load forecasting and advanced demand response combined with fine-grained energy consumption data have not only improved energy efficiency,but also brought better economic and environmental benefits.The demand for large-scale measurement and collection of fine-grained electricity consumption data is becoming increasingly urgent.Non-intrusive load monitoring is a technology that decomposes appliance-level energy consumption data and usage information from aggregated load data.Compared with the intrusive method that uses the submeter to assist measurement,it has fewer requirements for hardware facilities and lower implementation cost.It is a more feasible data acquisition method for large-scale application based on fine-grained electricity data.In recent years,non-intrusive load monitoring has gradually become a research hotspot in the field of smart grids.Adapting to the low-frequency data collection environment of the existing smart grid is the basis and prerequisite for the large-scale application of non-invasive load monitoring methods.Based on a large number of manual and accurate annotations on the two most commonly used public data sets in the field,REDD and UKDALE,this article clusters and analyzes the specific waveform data of different types of electrical appliances,and analyzes the results and limitations of existing research.The key tasks of the non-intrusive load monitoring method in low frequency scenarios are given,and the two algorithms of supervised and semi-supervised methods are studied.The contributions are as follows:(1)A non-invasive load monitoring algorithm based on variable-length time-series waveform positioning is proposed.By learning the power waveform change mode,the electrical switch actions on bus load data can be detected under relatively low-frequency data collection,and load identification can be completed with event waveform.The algorithm is a two-stage one-dimensional time series data target detection algorithm based on end-to-end in-depth learning.The experimental results on public datasets show that the algorithm can achieve good results in the detection and identification of common household appliances.(2)Aiming at the problem that it is difficult to obtain a large number of high-precision artificial labels in actual use scenarios,this paper improves the training strategy of deep learning networks based on the aforementioned deep learning methods,and proposes event detection and classification algorithm based on transfer learning and semi-supervised clustering methods.The proposed semi-supervised algorithm can get similar event detection and load identification results to the supervised method in a small number of samples and labels,which greatly reduces the requirement of the training sample size and widens the application range of the model in the actual scene.