Multichannel Image Encoding for Short Time-Series Feature Representation Applied to System Peak Demand Forecasting Model

Over the years, researchers explored several approaches to improve system peak demand forecasting subjected to short-time series data. In this paper, we introduced a novel framework to address this problem. The framework involves applying Seasonal-Trend Decomposition by LOESS (STL), creating a multichannel image using a modified Gramian Angular Field, extracting features from the multichannel image using Autoencoder, and performing forecasting using Long Short-Term Memory (LSTM). The STL-GAF-LSTM model with image encoding shows the best performance with an MAE ≈ 2.70, MSE ≈ 7.30, RMSE ≤ 10
⁻⁰³
, and R
²
≈ 0.999987 across all dataset. Comparatively, other models like STL-LSTM without image encoding, Linear Regression, and different ARIMA models show higher errors and lower accuracy scores. The comparative analysis highlights the adaptability and strength of the model in different geographical contexts and time series characteristics. Experimental results show that it is consistent across different datasets. Multichannel image encoding enables the forecasting model to generalize the temporal relations and is insensitive to residuals. Despite the different temporal characteristics of the data, the approach performs better than benchmarking models.