Research On Non-intrusive Load Monitoring Method Based On Deep Learning

Non-intrusive load monitoring(NILM)is an important part of the smart grid.The purpose of it is to monitor the energy consumption and the status of appliance from the total power consumption and its signal changes at the power entrance in the home.This method not only helps user to improve their habits of electricity usage,but also assists the power grid to dispatch power resources more efficiently.It can further promote the development of smart cities,and gradually helps human beings to achieve the major goals of energy conservation and sustainable development.The core part of NILM include load disaggregation and load identification.The load disaggregation method can directly determine the load component on the power aggregation signal by disaggregating it.But the drawback of this method is that the procession of disaggregating needs to establish multiple models cause this method unsuitable for real-time detecting.The load identification method is to identify the type of unknown appliance by calculation the difference in power signal when the appliance put into use,which makes it has the advantage of fast real-time detection and high accuracy.However,the shortage of this method is that cannot directly detect the status of appliance from the aggregation signal.The main research work of this paper is based on load disaggregation and load identification,and both them are combined with deep learning algorithm to explore and implement.But due to the limitation of the load data,the UK-DALE and REDD data sets are used in the work of load disaggregation,and the PLAID and WHITED data sets are used in the work of load identification.The main work of this paper is as follows:The work of load disaggregation based on deep learning include:(1)Selecting some commonly used appliance such as kettle,fridge and microwave as experimental object and take the power activation curve as the feature of the load,then building three neural network models to do load disaggregation experiments on the selected appliances.The first model is a fully-connected auto-encoding neural network(AENN)that can effectively remove signal noise;the second one is a convolutional neural network(CNN)that can capture signal feature through a convolution kernel;the third one is a long-short term memory neural network(LSTM),which can use it hidden layers to retain the past time information.The result of the experiment found that the AENN and CNN in five indicators of precision,recall,F1 score,mean absolute error and relative error of total energy were significantly better than the LSTM.(2)Designing a convolutional auto-encoding neural network that combines the ideas of auto-encoder and the convolution kernel that effectively captures the feature of power signal.This model first uses a multi-channel convolution layer to obtain the features from the power signal sequence,then designs the auto-encoder convolution layer to compress the data to filter the signal noise,and then uses the form of deconvolution to reconstruct the target load power activation curve.Through the comparison of experiments,it is found that the convolutional auto-encoding neural network designed in this paper improves the F1 score about 3% and reduces the mean absolute error 3 Watt of the selected load.It proves that the designed model can have more advantages in load disaggregation.The work of load identification based on deep leaning include:(1)Selecting the binary voltage and current(VI)trajectory with strong robustness and less calculation as the research object of appliance identification,and improving the CNN model for appliance identification,and in the experimental comparison,it is found that the improved model is superior to the current recognition results of the load recognition algorithm based on the binary VI trajectory(2)Designing a dynamic VI trajectory as the signature of appliance.This signature presents the trajectory generation process through multiple frame images and each frame shows the shape and spatial position of the VI trajectory at different times,which adds the time information to the static VI trajectory and overcoming the shortage of that.(3)Proposing a normalize trajectory mapping algorithm,which compared with other existing trajectory binary mapping algorithms,it is found that the normalized mapping is more suitable for the dynamic voltage-current trajectory.(4)Building a 3D-CNN model that can capture the spatiotemporal information of the dynamic voltage-current trajectory proposed in this paper,and the experiment on the PLAID and WHITED dataset show that the loss curve by this method can quickly converge,and the best F1-macro score have reached 95% and 95.4%,respectively.Compare with the method of the static voltage-current trajectory combined with convolutional neural network,the F1 score of the experiment results is improve 2.5% and 1.5% in PALID and WHITED datasets,and the experiment results are also comparable to the highest recognition results on the PLAID and WHITED data sets through color-coded load trajectories and transfer learning methods,which reaches the F1 scores are 95.4% and 98.66%,respectively.Moreover,the amount of parameters of the algorithm model designed in this paper is significantly less than the existing load identification algorithm,which verifies that the method proposed in this paper has certain advanced and superiority,which provides the feasibility of deploying deep learning model capabilities on edge devices with low computing and storage capabilities.