Short-Term Load Forecasting Based on Graph Convolution and Dendritic Deep Learning

Abstract

Short-term load forecasting (STLF) is a significant task to the planning, operation and control of future power systems. The increasing number of devices connected to the system has led to more complex characteristics and forms of load, which has brought considerable difficulties to the relevant methods in achieving higher load prediction accuracy and reliability. In this regard, this study proposes a deep learning model that combines graph convolutional network (GCN), gated recurrent unit (GRU), and dendritic neural model (DNM) to forecast electric load more accurately. Firstly, the sample load data is constructed into graph data with individual time steps as nodes. A GCN is used to extract the hidden features while allowing a full communication between the time steps feature data. A GRU is then used to capture the time-dependent relationship of the data. Finally, a dendritic layer instead of a fully connected layer is used as the output to integrate data features in depth. Experiments are conducted to verify the validity of the proposed model and compared it with several effective deep learning models, including CNN_LSTM, Transformer and Kolmogorov-Arnold Networks (KAN). The results show a significant improvement in prediction compared to the baseline models, with mean absolute percentage error($MAPE$) of 1.62% and 3.98%, coefficient of determination($R^{2}$) of 0.983 and 0.928 respectively on two load datasets at different levels of aggregation, nationally and regionally.