Self-adaptive graph learning with multiscale decomposition for long-horizon load forecasting
Year: 2026
Type: article
Abstract: Accurate short-term load forecasting has become a critical task for maintaining stability and an efficient economy in modern power systems. Traditionally, deep learning methods used to forecast short-term load often fails to capture the complex interdependencies present among various factors, degrading in performance when forecasting over extended time windows. To address these challenges, our study proposes a Graph Learning-based Multiscale Framework (TR-GNN) . The framework uses a self-adaptive Graph Learning Module to construct a graph structure autonomously from tabular data. The framework builds on an established graph-learning backbone for load forecasting using a self-adaptive Graph Learning Module combined with node-independent dilated temporal convolutions to extract a temporal representation for each variable in the data, alongside a residual graph convolutional network that propagates the representations over the learned graph. To prevent error accumulation in long-horizon forecasting, we introduce a Multiscale Forecasting head that decomposes the prediction into a linear baseline and a deep, nonlinear residual. Experimental results across three real-world benchmark datasets (ISO-NE, AT, and SH) illustrate the proposed framework’s applicability across different load forecasting settings. TR-GNN achieves remarkable performance through a Mean Squared Error (MSE) of 0.1682 on the ISO-NE dataset, 0.0907 on the AT, and 0.1623 on the SH dataset, reducing MSE by approximately 22%, 23%, and 12% on the respective datasets. Ablation studies show that the framework can effectively learn complex temporal dynamics when trained on raw input sequences without any manual feature engineering.
Institution: North South University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235607186
