🤖 machine learning

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

This paper introduces RESCAST-100K, a large-scale benchmark comprising 100,000 simulated and five real-world residential datasets designed to systematically evaluate and improve cross-domain generalization for short-term energy load and indoor temperature forecasting using advanced machine learning architectures.

Original authors: Jainam Dhruva, Yousaf Raza, A. B. Siddique, Simone Silvestri

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jainam Dhruva, Yousaf Raza, A. B. Siddique, Simone Silvestri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a student how to predict the weather inside a house. The problem is that every house is different: some are in hot Florida, others in cold Minnesota; some have brick walls, others have wood; some use gas furnaces, others use heat pumps.

If you only teach the student using data from one specific house in one specific city, they will likely fail when you ask them to predict the temperature in a totally different house. This is the big problem in energy forecasting: we don't have enough data from every type of house to train a perfect model for each one.

Enter RESCAST-100K.

Think of this paper as introducing a massive, super-smart "training gym" for AI models. Here is what they built, explained simply:

1. The "Virtual City" of 100,000 Homes

Instead of waiting to collect data from real houses (which is slow, expensive, and often private), the authors used a high-tech simulator called EnergyPlus to create a digital twin of 100,000 American homes.

The Variety: They didn't just make 100,000 identical boxes. They made a diverse city. They varied the location (geography), the climate, the wall materials, the heating systems, and the size of the homes.
The Data: For every single virtual home, they recorded three things every 15 minutes for a whole year:
1. Total Energy Load: How much electricity the whole house used.
2. HVAC Load: How much electricity the heating and air conditioning used specifically.
3. Indoor Temperature: How hot or cold it was inside.
The Context: They also gave the AI "clues" like the outdoor weather, the thermostat settings, and the static details of the house (like "it has a brick wall").

2. The "Training Test" (Cross-Domain Challenge)

The real magic of this dataset isn't just the size; it's how they use it to test the AI.

Imagine you train a student on houses in New York (cold, brick walls, gas furnaces). Then, you immediately test them on houses in Arizona (hot, stucco walls, heat pumps). This is called a "Cross-Domain" test.

Most old datasets only let you test the student on houses they've already seen. RESCAST-100K lets you set up specific "challenges":

Geography Shift: Train in the North, test in the South.
Climate Shift: Train in a humid zone, test in a dry zone.
Equipment Shift: Train on houses with furnaces, test on houses with heat pumps.

This helps researchers see if their AI is actually "learning the rules of physics" or just memorizing the specific houses it studied.

3. The "Missing Puzzle Piece" Problem

In the real world, data is often messy. Sometimes a sensor breaks, or a house doesn't have a smart thermostat, so the AI is missing key clues (like the outdoor temperature).

The authors tested their AI models by intentionally hiding some of this data during training. They wanted to see: Can the AI still make a good guess if it's missing the weather report or the thermostat setting?

4. The Results: Who Won the Game?

They tested several different types of AI "brains" (architectures) to see which one could handle these tough, changing conditions.

The Losers: Traditional models (like simple Recurrent Neural Networks) and standard "Transformer" models (the kind that power many modern AI tools) struggled when the house type changed. They got confused when the rules shifted from cold climates to hot ones.
The Winners: Two specific types of models performed the best:
1. MLP-Mixers (TSMixer): These models are great at looking at all the different clues (weather, wall type, etc.) at once and mixing them together to find patterns. They were the best at predicting energy usage.
2. Cross-Attention Models (TimeXer): These models are good at focusing on the specific "story" of the house's temperature history while still paying attention to outside factors. They were the best at predicting the actual indoor temperature.

Key Finding: The models that were good at handling "missing data" and "different house types" were the ones that used these advanced mixing and attention techniques.

5. The "Sim-to-Real" Reality Check

Finally, the authors took the models trained on their 100,000 virtual homes and tested them on five real-world datasets (actual houses with real people living in them).

The Result: As expected, the models weren't perfect on real houses because real people are unpredictable (they open windows, leave lights on, etc., in ways the simulator didn't guess).
The Silver Lining: Even without retraining on the real data, the models trained on the virtual city were surprisingly good at guessing the general shape of the energy usage (probabilistic forecasting). They were better at saying, "It's likely to be between 50 and 60 kWh," than at guessing the exact number.

Summary

The paper introduces RESCAST-100K, a massive, configurable dataset of 100,000 simulated US homes. It acts as a rigorous testing ground to see if AI can learn to predict energy use in any house, even if it has never seen that specific type of house before. They found that advanced AI models (Mixers and Cross-Attention) are much better at adapting to new environments and handling missing data than older models, offering a promising path for smarter energy management in the future.

Technical Summary: RESCAST-100K

Problem Statement

Accurate short-term forecasting of residential energy load and indoor temperature is critical for home energy management, grid balancing, and demand response. However, developing robust models is hindered by data scarcity, fragmentation, and privacy concerns in residential settings. Existing deep learning approaches typically require large amounts of historical data from the target domain, which is rarely available. Furthermore, residential environments exhibit multi-dimensional heterogeneity (geography, climate, building envelope, HVAC systems) that induces distinct distribution shifts, making cross-domain generalization difficult.

Current benchmarks suffer from significant limitations:

Narrow Scope: Many datasets focus on non-residential buildings or lack key targets like HVAC load and indoor temperature.
Missing Covariates: Existing resources often omit exogenous variables (weather, control signals) or static building properties necessary for generalization.
Lack of Cross-Domain Support: Few datasets support systematic evaluation under controlled domain shifts or integrate synthetic and real-world data under a unified schema, making "sim-to-real" evaluation difficult.

Methodology and Dataset Construction

The authors introduce RESCAST-100K, a large-scale benchmark designed to study cross-domain generalization in residential forecasting.

Dataset Composition

Scale: Approximately 100,000 U.S. homes derived from ResStock and simulated using EnergyPlus.
Resolution: 15-minute time series data.
Targets: Three coupled targets per home:
1. Total electric load (kWh)
2. HVAC load (kWh)
3. Indoor temperature (°F)
Features:
- Exogenous Dynamic Inputs: Weather channels (dry/wet bulb temp, humidity, wind, solar radiation) and HVAC setpoints.
- Static Covariates: Over 40 building properties including geography, climate zone, wall construction, foundation type, and HVAC equipment.
Real-World Integration: The benchmark integrates five real-world residential datasets (ECOBEE, HEAPO, IDEAL, NEST, REFIT) under a unified schema to enable sim-to-real evaluation.

Configuration-Driven Domain Abstraction

A core feature of RESCAST-100K is its configuration-driven interface (YAML-based) that allows users to instantiate source and target domains along interpretable axes:

Geography (State, Region)
Climate Zone (ASHRAE/IECC)
Building Envelope (Wall type, Construction)
HVAC Equipment (Heating technology)

This enables systematic evaluation of transfer learning, domain adaptation, and zero-shot generalization under controlled distribution shifts.

Experimental Setup

The authors formulate the problem as probabilistic forecasting with missing covariates.

Task: Predict a trajectory of length $L_f$ (96 timesteps/1 day) given a context of $L_c$ (384 timesteps/4 days) and static attributes.
Robustness: A stochastic masking operator simulates missing dynamic channels (weather, setpoints) and static covariates to test model robustness.
Models Benchmarked: The study evaluates six model families:
- Recurrent: Encoder LSTM (LSTM-E), Encoder-Decoder LSTM (LSTM).
- Transformer-based: Transformer, TimeXer, PatchTST.
- Mixer-based: TSMixer.
- Robust Variants: Models were modified (e.g., LSTM-R, Transformer-R) to handle missing data.
Metrics: Normalized Root Mean Square Error (NRMSE) for point accuracy and Continuous Ranked Probability Score (CRPS) for probabilistic accuracy.

Key Results

1. In-Domain Performance

Under full data availability, Transformer-based and TimeXer models generally achieved the best accuracy. However, when data was missing, TSMixer and TimeXer consistently outperformed classical LSTMs and standard Transformers, demonstrating superior robustness to incomplete inputs.

2. Cross-Domain Zero-Shot Generalization

Evaluated across four domain shift categories (State, Climate Zone, Wall Type, Heating Technology):

Load Forecasting (Total & HVAC): TSMixer-R consistently achieved the lowest NRMSE and CRPS across all domain categories, followed closely by TimeXer-R.
- Observation: MLP-mixer models appear to better learn dependencies on exogenous features by processing them multiple times during mixing, compared to single attention operations.
Temperature Forecasting: TimeXer-R achieved the best accuracy.
- Observation: TimeXer's use of self-attention on the target (maintaining thermal inertia) combined with cross-attention for exogenous influence proved superior for capturing smooth, setpoint-driven thermal dynamics.
Baseline Performance: Recurrent models (LSTM) and vanilla Transformers degraded significantly (often by an order of magnitude) under domain shifts, particularly for temperature forecasting.

3. Sim-to-Real Evaluation

When models trained on synthetic data were tested zero-shot on five real-world datasets:

Point Accuracy: NRMSE degraded substantially due to the sim-to-real gap, as expected without fine-tuning.
Probabilistic Accuracy: CRPS values remained relatively low and consistent. TSMixer-R and TimeXer-R maintained the best performance.
Implication: Probabilistic pretraining on synthetic data transfers more reliably to real homes than point forecast accuracy, suggesting that models capture the general distributional shape of the target signal even without target-domain data.

Contributions and Significance

The paper claims the following primary contributions:

RESCAST-100K Dataset: A large-scale (100k homes), multi-target dataset with a configuration-driven interface for studying cross-domain forecasting along axes like geography, climate, and construction.
Unified Schema: Integration of high-fidelity physics-based simulations with five real-world datasets, enabling rigorous sim-to-real evaluation previously unavailable at this scale.
Benchmarking Insights: Systematic evaluation of recurrent, attention, and mixer architectures reveals that cross-attention and MLP-mixer based models consistently outperform recurrent and classical transformer baselines under domain shifts and missing data conditions.

Significance:
The authors position RESCAST-100K as a necessary resource to advance the machine learning and building analytics communities. By providing a standardized, large-scale benchmark for cross-domain generalization, it addresses the "absence of comprehensive residential datasets" that has limited progress in transfer learning and domain adaptation for residential energy systems. The findings suggest that specific architectural choices (Mixer and Cross-Attention) are critical for handling the heterogeneity inherent in residential energy forecasting.