Graph Neural Networks for Electricity Load Forecasting

TL;DR

This paper tackles electricity load forecasting in modern, decentralized power systems by leveraging Graph Neural Networks (GNNs) that capture spatial-temporal dependencies. It introduces a unified framework that combines graph-based representations, diverse GNN architectures (including GCN, GraphSAGE, APPNP, and GAT variants), attention-driven interpretability, and ensemble aggregation to boost accuracy, robustness, and transparency. Through experiments on French regional data, UK residential data, and synthetic benchmarks, the study shows graph-aware models generally outperform baselines and foundation models, with bottom-up aggregation and diffusion-based methods offering robustness in heterogeneous conditions. The work highlights the value of integrating structural modeling with interpretable explanations (GNNExplainer, attention analysis, ALE plots) and discusses practical trade-offs between accuracy, complexity, and transparency, while outlining directions for dynamic graphs and probabilistic forecasting to enhance operational reliability.

Abstract

Forecasting electricity demand is increasingly challenging as energy systems become more decentralized and intertwined with renewable sources. Graph Neural Networks (GNNs) have recently emerged as a powerful paradigm to model spatial dependencies in load data while accommodating complex non-stationarities. This paper introduces a comprehensive framework that integrates graph-based forecasting with attention mechanisms and ensemble aggregation strategies to enhance both predictive accuracy and interpretability. Several GNN architectures -- including Graph Convolutional Networks, GraphSAGE, APPNP, and Graph Attention Networks -- are systematically evaluated on synthetic, regional (France), and fine-grained (UK) datasets. Empirical results demonstrate that graph-aware models consistently outperform conventional baselines such as Feed Forward Neural Networks and foundation models like TiREX. Furthermore, attention layers provide valuable insights into evolving spatial interactions driven by meteorological and seasonal dynamics. Ensemble aggregation, particularly through bottom-up expert combination, further improves robustness under heterogeneous data conditions. Overall, the study highlights the complementarity between structural modeling, interpretability, and robustness, and discusses the trade-offs between accuracy, model complexity, and transparency in graph-based electricity load forecasting.

Figures (7)

• Figure 1: Example of a message passing layer in a GNN. $V_\ell$, $E_\ell$ and $U_\ell$ respectively refer to node, edge, and global level at layer $\ell$. $\phi$ and $\rho$ are respectively update and propagation functions.
• Figure 2: Temporal splits for training, validation, and testing on the French and UK datasets.
• Figure 3: Weights associated with the experts on the synthetic datasets in the single-step regression setup. $\mathbf{\Sigma} = \boldsymbol{\rho}(\mathbf W_\lambda)$ (left), $\mathbf{\Sigma} = \boldsymbol{I}$ (right). GAM is the main expert, followed by mutliple SAGE-gl3sr and SAGE-dtw.
• Figure 4: Explanation graphs for the GAT convolutional layer on the French dataset.
• Figure 5: Spline (dashed line) and predicted (scatter plot) effects. The distribution of the generated temperatures is represented in gray.