Research And Application Of Transfer Learning In Non-Intrusive Load Monitoring

Non-Intrusive Load Monitoring(NILM)as a method to obtain fine-grained power consumption information can improve the service level of advanced grid applications such as demand response,load forecasting,and user behavior analysis.Current NILM studies on transfer learning usually pre-train the model and then fine-tune the model to fit the target domain,which requires new training data and consumes computational resources,affecting the usefulness of NILM.To this end,this paper uses a multi-task approach to construct a one-to-many load disaggregation model,and improves the model generalization performance by improving the feature representation based on device on/off and operating states.After training the model using the dataset REFIT,it is transferred directly without fine-tuning the model,and the effectiveness of the method is verified in the datasets UKDALE and REDD.The main research work includes the following three points:(1)A cross-domain load disaggregation algorithm with collaborative multi-device state information is designed.Based on the encoder parameter sharing form of multi-task learning,one-to-many load disaggregation model is constructed and the performance of the model is compared when four different network structures are used as encoders.It breaks through the limitation of considering only single target device in the past load disaggregation algorithm,and improves the situation of over-fitting the original data domain in the single-task model which leads to misidentification in the target domain.Experimental results show that the average MAE values of the proposed algorithm for medium and high-power device disaggregation decrease by 20.3%,23.2%,and 29.1%on the three houses of the UKDALE dataset,and by 5.4%,25.9%,and 14.8%on the three houses of the REDD dataset,respectively,when compared with the single-task model.(2)A representation learning method based on equipment on/off and operation states is designed.Firstly,a multi-label load identification task is introduced to improve the situation that low-power device refrigerators cannot be effectively identified under the fluctuation of high-power devices;secondly,a load feature similarity learning task is introduced,which enables the encoder to ignore the surface factors and focus on learning the intrinsic information.The experimental results show that,by optimizing the parameters of the encoder,the fitting effect is better than the one-to-many model proposed in this paper for devices in low-power operation,and the average MAE values decrease by 23.3%,14.4%and 11.8%respectively for the three houses in the UKDALE dataset.(3)A non-intrusive load monitoring simulation system is designed and implemented to realize monitoring the status information and power consumption of the target devices.