HOT Dataset: HVAC Operations Transfer

Overview

The HOT (HVAC Operations Transfer) dataset is the first large-scale open-source dataset purpose-built for transfer learning research in building control systems. Buildings account for approximately 10-15% of global energy consumption through HVAC systems, making intelligent control optimization critical for energy efficiency and climate change mitigation. Beyond technical advances, the field needs standardized evaluation protocols analogous to machine learning's model cards - comprehensive building data cards that systematically document building characteristics, performance baselines, and transfer learning suitability. Such standardisation would accelerate deployment decisions and enable practitioners to confidently assess transfer feasibility without extensive pilot studies. HOT establishes the foundation for this transformation.

Key Statistics

• 159,744 unique building-weather combinations
• 15,808 building models with controllable zone-level setpoints
• 19 ASHRAE climate zones across 76 global locations
• 16 commercial building types (office, retail, school, hospital, etc.)
• 12 systematic occupancy patterns
• 3 thermal performance scenarios
• 15-minute timestep resolution for control applications

Dataset Description

HOT addresses the critical infrastructure gap in building control transfer learning by providing systematic variations across four key context dimensions:

🏢 Building Geometry (16 Types)

• Office: Small (511 m²), Medium (4,983 m²), Large (46,323 m²)
• Retail: Standalone (2,294 m²), Strip Mall (2,294 m²)
• Educational: Primary School (6,874 m²), Secondary School (19,592 m²)
• Healthcare: Hospital (22,443 m²), Outpatient (3,805 m²)
• Hospitality: Small Hotel (3,726 m²), Large Hotel (11,349 m²)
• Residential: Midrise Apartment (2,825 m²), Highrise Apartment (7,063 m²)
• Food Service: Sit-down Restaurant (511 m²), Fast Food (232 m²)
• Industrial: Warehouse (4,835 m²)

🌡️ Thermal Performance (3 Scenarios)

• Default: Baseline thermal properties (R_mult = 1.0)
• High Performance: Enhanced insulation (R_mult = 2.0, U_mult = 0.5)
• Low Performance: Minimal insulation (R_mult = 0.5, U_mult = 2.0)

🌍 Climate Conditions (76 Locations)

• Complete ASHRAE coverage: All 19 climate zones (0A through 8)
• Global diversity: From tropical (Ho Chi Minh) to subarctic (Fairbanks)
• Weather data types: TMY (Typical Meteorological Year) + Real historical (2014-2024)
• Temperature range: -44.4°C to 47.0°C
• HDD range: 0 to 7,673 heating degree days
• CDD range: 6 to 4,301 cooling degree days

👥 Occupancy Patterns (12 Schedules)

• Standard: Traditional office hours (8 AM - 5 PM weekdays)
• Low/High Occupancy: 50%/150% intensity variations
• Shift Operations: Early (6 AM-3 PM), Late (2 PM-11 PM)
• Sector-Specific: Retail (10 AM-9 PM), School (7 AM-4 PM + evening)
• Healthcare: Hospital 24/7 with shift patterns
• Modern Work: Flexible hybrid with staggered hours
• Specialized: Gym (morning/evening peaks), Warehouse logistics
• Continuous: 24/7 operations

Dataset Structure

HOT/
├── data/
│   ├── base/                          # Raw building models by geometry type
│   │   ├── ApartmentHighRise_STD2013/
│   │   ├── ApartmentMidRise_STD2013/
│   │   ├── Hospital_STD2013/
│   │   ├── OfficeSmall_STD2013/
│   │   └── ...                        # 16 building geometry folders
│   ├── processed/
│   │   └── base/                      # Processed EPJSONs ready for control
│   │       ├── ApartmentHighRise_STD2013.epJSON
│   │       ├── Hospital_STD2013.epJSON
│   │       └── ...                    # All processed buildings
│   ├── variations/                    # Building variations
│   │   ├── occupancy/                 # Occupancy schedule variations
│   │   │   ├── standard/
│   │   │   ├── low_occupancy/
│   │   │   ├── hospital/
│   │   │   └── ...                    # 12 occupancy patterns
│   │   ├── thermal/                   # Thermal performance variations
│   │   │   ├── default/
│   │   │   ├── high_performance/
│   │   │   └── low_performance/
│   │   └── combined/                  # Multi-variable combinations
│   │       ├── occupancy_24_7_thermal_default/
│   │       ├── occupancy_hospital_thermal_high_performance/
│   │       └── ...                    # All combinations
│   ├── weather/                       # Weather data files (.epw)
│   │   ├── base/                      # Base TMY weather files (19 locations)
│   │   ├── expanded/                  # Extended TMY files (57 additional locations)
│   │   ├── real_base/                 # Historical weather (2014-2024)
│   │   └── tables/                    # Weather metadata tables
│   └── tables/                        # Dataset metadata and combinations
│       ├── buildings.csv              # Building characteristics
│       └── building_weather_combinations.csv  # All 159,744 pairings

Key Features

🎮 Reinforcement Learning Ready

• Controllable setpoints: Zone-level heating/cooling temperature control
• Gymnasium interface: Standard RL environment wrapper
• Comprehensive state space: Zone temperatures, outdoor conditions, energy consumption
• Multi-objective rewards: Energy efficiency + thermal comfort + control stability
• EnergyPlus integration: Physics-based building simulation

🔬 Transfer Learning Framework

• Similarity metrics: Quantitative compatibility assessment across 4 dimensions
• Zero-shot evaluation: Direct policy transfer without retraining
• Systematic variations: Single and multi-variable transfer scenarios
• Benchmark protocols: Standardized evaluation methodology

🌐 Global Climate Coverage

• All inhabited regions: Complete ASHRAE climate zone representation
• Real vs. synthetic: TMY baseline + historical weather variability
• Extreme conditions: From subarctic (-44°C) to desert (+47°C)
• Transfer analysis: Climate adaptation and geographic deployment

📊 Research Infrastructure

• Standardized formats: Consistent EnergyPlus epJSON structure
• Processing pipeline: Automated building modification tools
• Validation tools: Building model verification and testing
• Similarity analysis: Transfer feasibility assessment toolkit

Research Applications

🤖 Reinforcement Learning

• Multi-agent control: Coordinate multiple HVAC zones
• Meta-learning: Fast adaptation to new buildings (MAML, Reptile)
• Foundation models: Pre-train on diverse building types
• Safe RL: Constraint-aware control with comfort guarantees

🔄 Transfer Learning

• Domain adaptation: Geographic and climate transfer
• Few-shot learning: Minimal data adaptation for new buildings
• Cross-building generalization: Policy transfer across archetypes
• Similarity-guided selection: Optimal source building identification

📈 Building Analytics

• Energy benchmarking: Performance comparison across climates
• Occupancy analysis: Usage pattern impact on energy consumption
• Envelope optimization: Thermal performance sensitivity analysis
• Climate resilience: Building adaptation to changing conditions

Dataset Statistics

Dimension	Count	Range	Examples
Building Types	16	232-46,323 m²	Office, Hospital, School
Climate Zones	19	-44°C to +47°C	0A (Tropical) to 8 (Subarctic)
Occupancy Schedules	12	53-168 hrs/week	Office, Retail, Hospital, 24/7
Thermal Scenarios	3	0.5-2.0× resistance	High/Default/Low performance
Weather Files	192	TMY + Real (2014-2024)	Geographic + temporal variation

File Formats

Building Models (.epJSON)

• Format: EnergyPlus JSON input files
• Features: Zone-level controllable setpoints, comprehensive meters
• Compatibility: EnergyPlus 24.1+
• Size: ~50-500 KB per building

Weather Files (.epw)

• Format: EnergyPlus Weather format
• Frequency: Hourly meteorological data
• Variables: Temperature, humidity, solar, wind
• Size: ~1-2 MB per location-year

Metadata Tables (.csv)

• Buildings: Physical characteristics, variations, file paths
• Weather: Climate statistics, location data, file paths
• Combinations: Valid building-weather pairings (159,744 total)

Benchmarks and Baselines

Control Algorithms

• Static Baseline: Seasonal ASHRAE setpoint schedules
• PPO: Proximal Policy Optimization with building-specific tuning

Transfer Learning Methods

• Zero-shot: Direct policy application without retraining
• Fine-tuning: Limited adaptation with target building data
• Meta-learning: MAML and Reptile for fast adaptation

Evaluation Metrics

• Transfer Performance Ratio: Transferred vs. target-trained performance
• Energy Efficiency: HVAC consumption reduction vs. baseline
• Comfort Violations: Hours outside desired temperature range
• Training Acceleration: Reduced learning time through transfer

Citation

If you use the HOT dataset in your research, please cite:

@inproceedings{2025hot,
  title={A HOT Dataset: 150,000 Buildings for HVAC Operations Transfer Research},
  author={anonymous},
  booktitle={x},
  year={2025},
  publisher={x}
}

License

This dataset is released under the MIT License. See LICENSE file for details.

Contributing

We welcome contributions to expand and improve the HOT dataset:

• New building types: Additional commercial/residential archetypes
• Climate expansion: More geographic locations and weather data
• Enhanced metadata: Additional building characteristics
• Analysis tools: Transfer learning evaluation scripts
• Bug reports: Issues with building models or processing

Support and Contact

• Discussions: Hugging Face Discussions
• Email: anonymous for now

---

HOT Dataset - Advancing building energy research through comprehensive, standardized, and globally-representative data for intelligent HVAC control systems.