buildyn-thermal-dynamics-control

name	buildyn-thermal-dynamics-control
description	Excitation-driven data generation framework for building thermal dynamics modeling and control. BuilDyn enables customizable excitation strategies for control-oriented data generation, supports sampling from representative building distributions, and provides Python interface for ML pipelines. Use when: (1) training data-driven building models, (2) ensuring sufficient state-space exploration, (3) preparing training data for building control applications, (4) generating excited vs stationary operation datasets, (5) developing building-specific foundation models.
license	Complete terms in LICENSE.txt
metadata	{"arxiv_id":"2605.29849","published":"2026-05-28","authors":"Felix Koch, Thomas Krug, Fabian Raisch, Benjamin Schäfer, Benjamin Tischler","tags":["building-thermal","excitation-driven","control-oriented","machine-learning","BuilDyn","foundation-models","systems-control"]}

BuilDyn: Excitation-Driven Data Generation for Building Thermal Dynamics

Framework for generating control-oriented training data through systematic excitation of building thermal systems.

Problem Context

Challenge: Existing building datasets and simulations reflect stationary operation under fixed control policies, leading to:

• Limited exploration of control-driven state space
• Poor generalization to unseen operating conditions
• Reduced robustness in downstream ML tasks (fault detection, energy-efficient control)

Solution: BuilDyn enables excitation-driven data generation to ensure sufficient state-space coverage.

Core Concepts

1. Excitation Strategies

Goal: Explore control-driven system state space systematically.

Types of excitation:

• Random excitation: Stochastic control inputs
• Periodic excitation: Sinusoidal or square-wave signals
• Step excitation: Sudden control changes
• Optimized excitation: Design signals to maximize information gain

Key parameters:

• Excitation amplitude (Δu)
• Excitation frequency (f)
• Excitation duration (T)
• Coverage metrics (state-space volume)

2. Building Sampling

BuilDyn supports sampling from representative building distributions:

# Sample buildings from distribution
building_params = {
    'thermal_mass': sample_thermal_mass(),
    'window_area': sample_window_area(),
    'insulation': sample_insulation(),
    'occupancy_patterns': sample_occupancy(),
    'climate_zone': sample_climate()
}

This enables:

• Training on diverse building characteristics
• Transfer learning across building types
• Foundation model development

3. Python Interface

Integration into ML pipelines:

from buildyn import BuilDynGenerator

# Initialize generator
generator = BuilDynGenerator(
    building_distribution='representative_buildings.json',
    excitation_strategy='optimized',
    duration_hours=168
)

# Generate excited dataset
dataset = generator.generate(
    n_buildings=100,
    n_scenarios_per_building=10
)

# Export for ML training
dataset.save('excited_building_data.h5')

Implementation Workflow

Step 1: Define Excitation Strategy

excitation_config = {
    'type': 'optimized',  # 'random', 'periodic', 'step', 'optimized'
    'control_variables': ['heating_power', 'cooling_power', 'ventilation_rate'],
    'amplitude_range': [0.0, 1.0],  # Normalized control range
    'duration': 168,  # Hours (1 week)
    'sampling_rate': 3600  # Seconds (1 hour intervals)
}

Step 2: Sample Building Parameters

building_distribution = {
    'thermal_mass': {
        'distribution': 'uniform',
        'range': [1e6, 5e7]  # J/K
    },
    'window_area_fraction': {
        'distribution': 'normal',
        'mean': 0.15,
        'std': 0.05
    },
    'insulation_R': {
        'distribution': 'lognormal',
        'mean': 2.5,  # m²·K/W
        'std': 0.5
    }
}

Step 3: Generate Training Data

# Generate excited data
excited_data = generator.generate_excited(
    n_samples=10000,
    excitation_config=excitation_config,
    building_distribution=building_distribution
)

# Generate baseline (non-excited) data for comparison
baseline_data = generator.generate_stationary(
    n_samples=10000,
    control_policy='fixed_setpoint',
    building_distribution=building_distribution
)

Step 4: Train ML Models

# Compare performance
model_excited = train_model(excited_data)
model_baseline = train_model(baseline_data)

# Evaluate on unseen operating conditions
test_scenarios = generate_challenging_scenarios()
performance_excited = evaluate(model_excited, test_scenarios)
performance_baseline = evaluate(model_baseline, test_scenarios)

# Expected: excited model > baseline model on unseen conditions

Key Advantages

1. State-space coverage: Systematic exploration of control inputs
2. Robustness: Better generalization to unseen operating conditions
3. Diversity: Sampling from representative building distributions
4. Integration: Python interface for easy ML pipeline integration
5. Foundation models: Enables building-specific foundation model development

Excitation Design Principles

Maximizing Information Gain

Excitation signals should:

• Cover wide range of control inputs (heating, cooling, ventilation)
• Explore different thermal dynamics regimes
• Avoid redundant trajectories
• Maximize Fisher information or mutual information

Coverage Metrics

Quantify state-space coverage:

def compute_coverage(dataset):
    state_vectors = dataset['states']  # [temperature, humidity, CO2, etc.]
    
    # Volume coverage
    volume = convex_hull_volume(state_vectors)
    
    # Density coverage
    entropy = state_entropy(state_vectors)
    
    return {'volume': volume, 'entropy': entropy}

Use Cases

• Fault detection & diagnosis: Training robust FDD models
• Energy-efficient control: Model predictive control (MPC) training data
• Building energy simulation: Validating simulation models
• Transfer learning: Pre-training foundation models for specific buildings
• Climate adaptation: Training models for diverse weather conditions

Comparison: Excited vs Stationary Data

Criterion	Stationary Operation	Excited Operation
State-space coverage	Limited (fixed setpoints)	Broad (variable inputs)
Generalization	Poor (unseen conditions)	Strong (diverse scenarios)
Robustness	Low (fragile to changes)	High (handles variations)
Information density	Sparse	Dense
Model performance	Suboptimal	Optimal

Practical Considerations

• Excitation duration: Balance exploration vs stability (1-2 weeks typical)
• Sampling rate: Match downstream model requirements (hourly for thermal dynamics)
• Building diversity: Sample from realistic distributions
• Climate variation: Include diverse weather scenarios
• Occupancy patterns: Account for usage variability

Integration with Building Simulation

BuilDyn builds on BuilDa (Building Data simulator):

# BuilDyn extends BuilDa
from builda import BuildingSimulator
from buildyn import ExcitationModule

simulator = BuildingSimulator(building_config)
exciter = ExcitationModule(strategy='optimized')

# Excited simulation
control_sequence = exciter.generate_controls(duration=168)
simulation_results = simulator.run(control_sequence)

Research Directions

1. Foundation models: Building-specific foundation models via large-scale excited data
2. Transfer learning: Cross-building knowledge transfer
3. Adaptive excitation: Online excitation based on model uncertainty
4. Multi-objective excitation: Balance energy cost vs information gain

Activation Keywords

• building thermal dynamics
• excitation-driven data
• control-oriented modeling
• building ML training data
• BuilDyn framework
• building foundation models
• thermal system excitation