External Data-Enhanced Meta-Representation for Adaptive Probabilistic Load Forecasting

TL;DR

This paper introduces a novel meta-representation framework using hypernetworks and Mixture-of-Experts to adaptively incorporate external data into load forecasting models, significantly improving accuracy and robustness.

Contribution

It proposes a new paradigm where external data acts as meta-knowledge to dynamically adapt the forecasting model via hypernetworks and MoE, outperforming existing methods.

Findings

Enhanced forecasting accuracy across diverse datasets.

Improved robustness and efficiency with limited external data.

Outperforms state-of-the-art load forecasting models.

Abstract

Accurate residential load forecasting is critical for power system reliability with rising renewable integration and demand-side flexibility. However, most statistical and machine learning models treat external factors, such as weather, calendar effects, and pricing, as extra input, ignoring their heterogeneity, and thus limiting the extraction of useful external information. We propose a paradigm shift: external data should serve as meta-knowledge to dynamically adapt the forecasting model itself. Based on this idea, we design a meta-representation framework using hypernetworks that modulate selected parameters of a base Deep Learning (DL) model in response to external conditions. This provides both expressivity and adaptability. We further integrate a Mixture-of-Experts (MoE) mechanism to enhance efficiency through selective expert activation, while improving robustness by filtering redundant external inputs. The resulting model, dubbed as a Meta Mixture of Experts for External data (M2oE2), achieves substantial improvements in accuracy and robustness with limited additional overhead, outperforming existing state-of-the-art methods in diverse load datasets. The dataset and source code are publicly available at https://github.com/haorandd/M2oE2\_load\_forecast.git.