Leveraging Graph Neural Networks to Forecast Electricity Consumption

TL;DR

The paper addresses short-term electricity demand forecasting in decentralized grids with high renewable penetration. It proposes a graph-based approach using Graph Neural Networks to model spatial dependencies, along with data-driven graph inference and explainability via GNNExplainer, extending Generalized Additive Models. Through synthetic and real-France regional data, it demonstrates that learned graphs and GNNs can outperform GAM baselines, especially when inter-regional links are present, and that expert aggregation adds robustness. The work offers practical guidance for incorporating relational structure into load forecasting and highlights directions for temporal modeling and richer data sources.

Abstract

Accurate electricity demand forecasting is essential for several reasons, especially as the integration of renewable energy sources and the transition to a decentralized network paradigm introduce greater complexity and uncertainty. The proposed methodology leverages graph-based representations to effectively capture the spatial distribution and relational intricacies inherent in this decentralized network structure. This research work offers a novel approach that extends beyond the conventional Generalized Additive Model framework by considering models like Graph Convolutional Networks or Graph SAGE. These graph-based models enable the incorporation of various levels of interconnectedness and information sharing among nodes, where each node corresponds to the combined load (i.e. consumption) of a subset of consumers (e.g. the regions of a country). More specifically, we introduce a range of methods for inferring graphs tailored to consumption forecasting, along with a framework for evaluating the developed models in terms of both performance and explainability. We conduct experiments on electricity forecasting, in both a synthetic and a real framework considering the French mainland regions, and the performance and merits of our approach are discussed.

Figures (8)

• Figure 1: Example of a message passing layer in a GNN. $V_n$, $E_n$ and $U_n$ respectively refer to node, edge, and global level at stage $n$. $\phi$ are update functions and $\rho$ are propagation functions.
• Figure 2: Graph corresponding to $\boldsymbol{W}_{\lambda}$ with $\lambda = 0.71$ and kernel bandwidth $\sigma=478.3$.
• Figure 3: Inferring a graph from data using a dimension reduction algorithm and a statistical transformation.
• Figure 4: Generated temperature and load using pairwise influence between the regions ($\mathbf{\Sigma} = \boldsymbol{\rho}(\mathbf W_\lambda)$).
• Figure 5: Error variation by model on the synthetic test set.