基于混合模型的非侵入式负荷监测数据的生成OA北大核心

Non-intrusive load monitoring(NILM)is a technique that does not require accessing the internal system of each electrical device to monitor user's equipments,but only to monitor them at the entrance of user's bus.During the investigation of NILM techniques,it is often necessary to collect extensive user load data to confirm the applicability of proposed methods.This requirement inevitably entails a significant burden of data collection and organization.In order to overcome this challenge,we proposed a hybrid approach that combines the principle of frequency invariant transformation for periodic signals(FIT-PS)with time series generative adversarial networks(TimeGAN),denoted as FIT-PS-TimeGAN.Using a Worldwide Household and Industry Transient Energy Dataset(WHITED),we focused on five appliances:air conditioner,microwave oven,hoover,refrigerator and kettle.FIT-PS was employed to segment and stitch the load data aiming to construct training and testing sets for TimeGAN under different states.The validation results on effectiveness of testing sets demonstrated high consistency between the generated waveforms and the real data.Subsequently,FIT-PS was applied to intercept and splice the training data to generate complete single-load waveforms and multi-load waveforms that are able to meet the testing requirements.These generated waveforms were compared with real data in the same state and the comparative results showed that a favorable agreement between the generated and real data.In addition,compared with the other two generation models(autoregressive model and GAN model),FIT-PS-TimeGAN outperforms better in terms of data generation performance.In summary,the FIT-PS-TimeGAN hybrid model is capable of effectively generating waveforms and scenarios that comply with the operational principles of standard appliances.