Research On State Identification And Power Estimation Of Air-conditioning Loads Based On Graph Neural Network

With the construction of the new power system dominated by renewable energy,the output uncertainty of the generation side increases.Meanwhile,the growth of load demands accelerates,which imposing heavy pressure on the power balance and challenging the planning and operation of the power system.The flexible and controllable resources on user side can interact with the power grid in a bi-directional way through demand-side response projects.However,the regulation potential of massive and scattered flexible loads is left to be estimated.Under this circumstance,from the standpoint of utility company,the thesis applies graph neural network technology to cope with three types of users who attempt to prevent privacy leakage during the implementation of demand-side response.The state identification and power estimation of airconditioning loads are realized based on the available data in these scenarios.The effectiveness of the proposed algorithms is verified on a real data set with a sampling frequency of 15 minutes per time.Finally,some suggestions are put forward for the utility company based on the analysis of the case study.The main work of the thesis is as follows:The second chapter is aimed at conservative users,from whom only intermittent real power of air-conditioning loads can be obtained on a regular basis.In this situation,an algorithm based on the spectral graph convolutional neural network is proposed to analyze time-series power data.Firstly,the time-series power data is converted into a non-Euclidean graph structure through the visibility graph,and then the Chebyshev graph convolutional neural network is applied to identify the user’s air-conditioning power states in a long term.The results show that the utility company can purchase air-conditioning loads depending on the cost,and the best identification accuracy can be achieved by purchasing load labels for 7 consecutive days each time.The third chapter is oriented to friendly users,from whom the utility company can obtain all the historical air-conditioning power consumption data.Considering the delay of data collection and transmission,an algorithm for processing streaming data based on spatial graph neural network is proposed.Construct the dynamic graph topology through the visibility graph.Then utilize the spatial graph neural network with memory and attention mechanism to complete the pre-training.Afterwards,calculate the identification result of the air-conditioning loads power in real time.Once the delayed labels are obtained,the model will be updated.The results show that the model can accurately identify the air-conditioning power,and has a high tolerance for the delay of the load label.The fourth chapter is targeted at confrontational users,from whom the utility company obtains nothing of the household power data.In this scenario,considering the cluster characteristics of regional users,a semi-supervised transductive algorithm based on spatialtemporal graph neural network is proposed.First,the graph topology is constructed based on the total power consumption similarity of users.And then the semi-supervised loss function is constructed,and the spatial-temporal graph neural network is applied to estimate the airconditioning load power of multiple confrontational users at the same time.The results show that constructing the graph topology with a similar number of friendly users and confrontational users can not only ensure the identification efficiency,but also achieve the ideal accuracy.