Evaluating Model-Agnostic Meta-Learning (MAML) for Adaptive Energy Forecasting in Residential Buildings

Document Type

Conference Proceeding

Publication Date

1-1-2026

Abstract

Accurate forecasting of residential energy consumption is crucial for optimizing energy management, reducing operational costs, and maintaining grid stability. Model-Agnostic Meta-Learning (MAML) has shown promise in enabling rapid adaptation to new prediction tasks, yet its potential for residential energy forecasting remains underexplored. This study investigates the impact of consumption-pattern heterogeneity on forecasting performance and evaluates MAML’s ability to generalize to unseen users. Our approach consists of three key steps: (1) clustering the dataset into multiple groups based on weekly usage patterns to better capture group-level variations in consumption behaviors; (2) training predictive models using both a standard multilayer perceptron (MLP) and MAML with and without the pre-clustered dataset; and (3) assessing model performance to determine how pre-clustering influences predictive accuracy. Experimental results show that MAML-based models reduce prediction error by approximately 3% compared to MLP, achieving improved generalization across diverse datasets and more accurate predictions for users absent from the training set. Moreover, incorporating task clustering prior to MAML training mitigates the impact of training-task distribution, offering a potential pathway to more balanced performance across different user groups.

Publication Source (Journal or Book title)

ASHRAE Transactions

First Page

1178

Last Page

1186

This document is currently not available here.

Share

COinS