DiffFlex: A Diffusion-Based Model for Medium-Term Load Forecasting Over Flexible Intervals

Abstract

Accurate medium-term load forecasting (MTLF) is critical for capacity planning, maintenance scheduling, and intelligent resource allocation in modern power systems. However, existing approaches are typically designed for fixed-horizon, contiguous prediction and often suffer from multistep error accumulation, limited spatiotemporal modeling capacity, and insufficient uncertainty quantification. This article proposes DiffFlex, a diffusion-based probabilistic forecasting framework that reformulates MTLF as a temporal outpainting problem. By decoupling forecasting from temporal contiguity, DiffFlex enables direct parallel prediction over arbitrary, noncontiguous future intervals without cumulative error propagation. The framework integrates fused historical blocks to capture cyclical consumption patterns, models spatiotemporal-interval dependencies across multivariate load series, and employs an interval-aware diffusion module to generate full predictive distributions. Experiments on the Independent System Operator New England, New York Independent System Operator, and European Network of Transmission System Operators for Electricity datasets across 1- to 12-week horizons show that DiffFlex outperforms state-of-the-art baselines in both point and probabilistic accuracy, achieving average mean absolute percentage errors of 8.88%, 9.71%, and 4.86%, respectively. These results indicate that DiffFlex can serve as a practical tool for scenario-based planning and risk-informed decision-making in power system operations.