| Improving energy efficiency and reducing environmental pollution is one of the most challenging issues facing society today,and it is the core issue that drives countries around the world to promote the construction of smart grids.Non intrusive load monitoring technology can extract energy consumption information of each equipment,guide users to use electricity,save resources,and maintain power system supply without knowing only the total power consumption data.The current non-invasive load monitoring methods for multi-state equipment and overlapping equipment during startup and shutdown cycles are not ideal,and cannot meet practical application requirements.In addition,deep learning methods also have problems in load decomposition,such as the disappearance of gradients,and large decomposition errors for low-power appliances.To this end,this article in-depth analyzes the characteristics of user total energy consumption data,conducts non-invasive load decomposition research based on deep learning methods,and verifies the effectiveness of the algorithm on public datasets REDD and UK-DALE.The main contributions include the following two points:（1）A load decomposition algorithm based on equipment operation status is proposed.To solve the problem of difficulty in decomposing single user multi-state devices and overlapping startup and shutdown cycles,thesis divides the load decomposition task into two stages.The first stage uses load identification methods to identify the events of the target electrical appliance,and the second stage uses data filling and power correction methods to achieve load decomposition based on the event identification results.The experimental results show that the method proposed in this thesis can achieve good results on common household appliance decomposition tasks.Compared to the comparison model,the overall decomposition average absolute error decreases by 44%to 3.331;The average F1 value increased by 24%to 0.897.（2）A load decomposition algorithm based on multi scale feature fusion is proposed.In order to solve the problem of poor generalization of single user load decomposition method and difficulty in obtaining high-precision labels for different users,this thesis adopts an encoder decoder structure.The encoder module completes the task of extracting different scale features and timing features,and uses jump connections to fuse the hierarchical features of the encoder with high-level features of the decoder.Through learning the change patterns of similar appliances for multiple different users,The target electrical power consumption value on the bus load data can be accurately decomposed in an event tag free environment.Experimental results show that compared to existing research methods,the proposed method reduces the average absolute error of decomposition of the target electrical appliance on unknown house data by 36%to 6.734,verifying the effectiveness of the proposed method. |
