Proportional Identification of Residential Air-Conditioning Loads Based on Graph Theory

Fine-grained demand-side management relies on a sound understanding of appliance-level power usage patterns, especially the air conditioning (AC) loads which can even account for 80% of the total power consumption. Considering the customers' concern of privacy leakage and the cost-saving needs for power supply companies, we apply a graph convolutional network (GCN)-based approach to infer the proportion of AC loads' power consumption for an entire residential load dataset only through a small ratio of labeled data. After the time index alignment and the data repair, the visibility graph ( VG) algorithm is adopted to convert time-series power consumption data to a non-Euclidean graph. Then a GCN model utilizing the Chebyshev convolutional network is trained, where the inputs are the adjacent matrix of the constructed VG along with the node feature matrix that is composed of the total power consumption and the local temperature data. Accordingly, the proportional interval of AC loads can be inferred from the node-level output of the GCN model. Finally, the proposed approach is tested with the real-world electricity data from Pecan Street Dataport and the temperature data from NOAA. Different settings of intermittent AC proportion labels are employed to simulate the limited availability of fine-grained AC consumption data on the demand side, while all the numerical results validate the effectiveness of the proposed approach.