Research On Key Technologies Of Non-Intrusive Power Load Identification

Intelligent power load detection is an important part of smart grid management and monitoring at present.Accurate detection of power load is the premise of detailed load management and grid power supply quality monitoring,which is conducive to improving power management of user terminals and improving power operation efficiency.However,the traditional electricity detection method requires monitoring modules to monitor a single electrical appliance separately,and the efficiency conversion rate is not high.Non-intrusive Load Monitoring(NILM)recognizes the total load of electricity consumption into status information of each electrical equipment through intelligent perception.This thesis analyzes the overall framework of the NILM system,conducts in-depth research on event detection,feature extraction,and load identification,and proposes a secondary detection method based on difference averaging and an improved temporal convolutional network.Firstly,aiming at the problem of low accuracy of the input recognition algorithm caused by using only a single load feature as the identification feature,this thesis proposes to combine the transition state envelope data after the load switching event with the steady-state waveform data as the load identification feature.The four input characteristics of power parameters,harmonic parameters,steady-state current waveform data,and the combination of current waveform transition state envelope and steady-state current waveform data are studied.Through the verification of public data sets and measured data sets,accuracy,Evaluation indicators such as precision and recall rate evaluate the results.Secondly,to solve the problem of low detection accuracy of the sliding window based bilateral Cumulative Sum(CUSUM)algorithm for small current equipment switching events,this thesis proposes a quadratic detection method based on difference average.This method improves the bilateral CUSUM algorithm based on sliding window,and performs secondary detection on all detected valid change point segments.Set the state threshold for switching events of low-current equipment,and take a period value before and after the time recording point to calculate the difference average.The performance of the event detection algorithm is evaluated by evaluation indicators such as accuracy,precision and recall.And the improved CUSUM algorithm improves the accuracy by 6.8% and 2.5%,the precision by 2.8% and 1.7%,and the recall rate by 5.5% and 3.1%,respectively,on the public data set and the measured data set.Thirdly,in view of the access of a large number of distributed new load devices on the power consumption side and the diversified development of devices,the simple load identification algorithm is difficult to meet the multi-device identification and the identification accuracy is not high.This thesis proposes a time-based convolution Improved algorithms for Temporal Convolutional Networks(TCN).This algorithm not only ensures that the output contains long-term and effective feature information,but also effectively avoids the problem of gradient explosion caused by deep networks.The algorithm inputs the sequence signal into the one-dimensional convolutional neural network(One-dimensional Convolutional Neural Networks,1DCNN)to initially learn the features,then passes through the TCN and finally inputs the network reinforcement learning features of the attention mechanism.The multi-index results show that the improved algorithm in this thesis improves the F1 score index by 3.5% and 8.2% compared with the unimproved algorithm,improves the accuracy of load identification,and can reach the application standard in actual household life.