Research On Non-invasive Load Monitoring Technology Based On Deep Learning

Non-invasive load monitoring(NILM)realizes the monitoring of indoor load types,operation status,energy consumption by analyzing the data at the power entrance of users.It is a key technology for the smart grid to realize the overall perception of state and the bidirectional flow of information.It has a wide range of application scenarios and important social value.In recent years,Deep Learning(DL)in the field of artificial intelligence has attracted widespread attention from all walks of life due to its excellent performance.To this end,this thesis combines the NILM and DL to carry out algorithmic research on the identification and decomposition of non-intrusive load monitoring based on DL and to develop a prototype system of NILM,and the main work is as follows.1)To overcome the limitations of a single steady-state feature and the difficulties of fusing different types of features,a non-invasive load identification algorithm based on deep transfer learning was proposed.This algorithm uses convolution and BP advanced feature extraction network to realize the abstract expression and fusion of V-I trajectory and power features,and uses BP classification neural network to identify the fusion features.The proposed algorithm has improved the defect that the V-I trajectory feature cannot reflect the load power and improves the load identification ability of the V-I trajectory feature.Experiments have been designed to verify the effectiveness of the proposed algorithm by using public datasets and real test scenarios.2)Aiming at the problem of the high cost of data annotation in the field of non-invasive load identification and the existence of a large number of low-value samples in the actual data,a load identification sample selection algorithm based on deep active learning is proposed.The deep active learning technology is used to select a small number of high-value samples worthy of labeling from a large number of unlabeled samples.The proposed algorithm can achieve high accuracy with a small number of samples,ensuring the accuracy of load identification and significantly reducing the cost of data labeling.The effectiveness of the algorithm is verified using a mixed dataset and the applicability of the pool-based and stream-based sampling approaches to NILM is analyzed in comparison.3)Aiming at the problem that traditional algorithms cannot consider the temporal correlation of the power sequence and the DL algorithms based on Long Short-Term Memory(LSTM)are very slow,a non-invasive energy decomposition algorithm based on the sequence-to-sequence(Seq2seq)model is proposed.In the Seq2 seq model based on bidirectional GRU,the local attention mechanism and beam search algorithm are introduced to optimize the coding and decoding process of the model,and the decomposition accuracy and computation speed are improved.Experiments have been designed to verify the effectiveness of the proposed algorithm by using public datasets and real test scenarios.4)The prototype system of NILM with strong expansibility is developed.The modular design is adopted in every part of the system,which is convenient for expansion and reuse.The visual interface is designed following closely the trend of technology and modern design concept.The developed prototype system of NILM has the advantages of cross-platform,cross-terminal,intuitive interface and strong expansibility.