Time-series imaging for improving the accuracy of short-term load forecasting

Accurate short-term load forecasting (STLF) is essential for the efficient operation and planning of power systems. However, various factors increase the complexity of load data, making it more difficult to extract features and make accurate forecasts. Therefore, based on a thorough consideration of load characteristics, this paper spatially encodes load series and leverages feature extraction principles from computer vision to perform STLF. First, unlike the traditional one-dimensional (1D) feature extraction approach, the load series is transformed into two-dimensional (2D) Gramian Angular Field (GAF) and Recurrence Plot (RP) representations. Second, comparative experiments are conducted to evaluate the feature encoding and temporal representation capabilities of the 2D representations. Finally, an improved forecasting framework based on traditional STLF is proposed. Experimental results show that, compared with 1D input, using 2D load data reduces the mean absolute percentage error (MAPE) by up to 23.54 % and 73.9 % in CNN-based and LSTM-based models, respectively. By applying the proposed imaging-based STLF framework to existing published models, the prediction accuracy is further improved. Without parameter tuning, the proposed model still achieves better performance than existing methods on the ISO-NE dataset.