Title: RESCAST-100K: A Comprehensive Dataset for Cross-Domain Residential Load and Indoor Temperature Forecasting

Abstract

Precise short-term predictions of indoor temperatures and residential energy consumption are critical components of home energy management systems, grid-level demand response initiatives, and broader community efficiency strategies. While domain adaptation and transfer learning offer promising solutions for enhancing forecasting accuracy in the face of data scarcity and heterogeneity typical of residential environments, advancements have been hindered by the absence of robust, comprehensive residential datasets. Current benchmarks often suffer from limited target coverage and fail to facilitate structured cross-domain evaluations.

To address this gap, we present RESCAST-100K, a large-scale benchmark designed to investigate cross-domain generalization in residential forecasting. This platform features a configuration-driven interface that allows for the instantiation of source and target domains along interpretable dimensions, such as geography, climate zone, wall construction materials, and heating equipment. This setup enables the systematic assessment of zero-shot generalization, domain adaptation, and transfer learning under controlled domain shifts.

The dataset comprises approximately 100,000 EnergyPlus-simulated U.S. homes, derived from ResStock. For each residence, it provides 15-minute time series data for three coupled targets: total load, HVAC load, and indoor temperature. These predictions are supported by weather channels, HVAC setpoints, and more than 40 static building covariates. Furthermore, RESCAST-100K incorporates five real-world residential datasets within a unified schema, thereby facilitating sim-to-real evaluation on identical tasks.

In our benchmarks, we evaluated recurrent, attention-based, and MLP-mixer architectures across various domains, forecasting tasks, and missing-input conditions. The results indicate that cross-attention and MLP-mixer models consistently surpass recurrent networks and classical transformer baselines when subjected to domain shifts. RESCAST-100K aims to assist the machine learning and building analytics communities in advancing cross-domain residential forecasting capabilities at the home, community, and grid levels.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC