Title: EnergyMamba: An Uncertainty-Aware Graph-Enhanced Selective State Space Model for Energy Consumption Prediction

Abstract:

Accurate forecasting of energy usage is critical for optimizing grid operations, managing demand, and supporting sustainable energy strategies. While sophisticated machine learning techniques have been deployed to enhance prediction outcomes, current approaches suffer from two primary shortcomings. First, they typically treat the problem as a standalone time-series task, neglecting the spatial correlations between different geographic areas. Second, they lack the ability to generate trustworthy forecasts with uncertainty bounds during anomalies, such as severe weather conditions.

To address these gaps, we introduce EnergyMamba, a novel framework designed for robust and precise energy consumption forecasting that accounts for uncertainty. This spatiotemporal learning system integrates two main components: (i) a Graph-Enhanced Selective State Space Model (GE-Mamba), which embeds spatial information derived from the grid’s topology into temporal processes, thereby facilitating joint spatiotemporal analysis; and (ii) an Adaptive Sequential Conformalized Quantile Regression (AS-CQR) module. The latter features local adaptive normalization and an online feedback loop to adjust prediction intervals in real-time, adapting to potential shifts in data distribution.

We tested EnergyMamba using four extensive real-world datasets sourced from Florida, New York, and California. The experimental results demonstrate that EnergyMamba outperforms 15 leading baseline models, delivering approximately a 5% gain in prediction accuracy and a 6% enhancement in uncertainty quantification.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC