[Paper Review] End-to-end deep metamodeling to calibrate and optimize energy loads

Max Cohen, Maurice Charbit+3 more|arXiv (Cornell University)|Jun 19, 2020

Building Energy and Comfort Optimization28 references4 citations

TL;DR

This paper proposes an end-to-end deep metamodeling framework using a Transformer-based neural network to calibrate and optimize energy loads in large buildings. Trained on EnergyPlus simulations and refined with real sensor data via CMA-ES, the model enables multi-objective optimization of HVAC settings to reduce energy use while maintaining thermal comfort and indoor air quality, achieving significant computational efficiency and energy savings without physical retrofitting.

ABSTRACT

In this paper, we propose a new end-to-end methodology to optimize the energy performance and the comfort, air quality and hygiene of large buildings. A metamodel based on a Transformer network is introduced and trained using a dataset sampled with a simulation program. Then, a few physical parameters and the building management system settings of this metamodel are calibrated using the CMA-ES optimization algorithm and real data obtained from sensors. Finally, the optimal settings to minimize the energy loads while maintaining a target thermal comfort and air quality are obtained using a multi-objective optimization procedure. The numerical experiments illustrate how this metamodel ensures a significant gain in energy efficiency while being computationally much more appealing than models requiring a huge number of physical parameters to be estimated.

Motivation & Objective

• To develop a computationally efficient method for optimizing energy performance in large buildings without costly physical renovations.
• To address the challenge of calibrating high-dimensional physical models using real-time sensor data.
• To integrate thermal comfort, indoor air quality, and energy load minimization into a unified multi-objective optimization framework.
• To reduce the computational burden of traditional physics-based simulation and optimization pipelines.
• To enable real-time, data-driven HVAC control through a surrogate metamodel trained on simulation data and fine-tuned with real building measurements.

Proposed method

• A Transformer-based neural network is trained on a large dataset generated by EnergyPlus simulations to predict indoor temperature and energy consumption across multiple HVAC and weather variables.
• The metamodel is calibrated using real sensor data (indoor temperature and heat consumption) via the CMA-ES optimization algorithm to adjust physical parameters and building management settings.
• A multi-objective optimization procedure, combining CMA-ES with the calibrated metamodel, identifies optimal HVAC settings that minimize energy loads while maintaining target comfort and air quality levels.
• The model is trained using a grid search over hyperparameters including latent dimension, number of heads, layers, and attention size to maximize predictive accuracy.
• The framework uses a hybrid approach: simulation-based data generation for training and real-world data for calibration, enabling robustness to real building variability.
• Output variables include indoor temperature, cooling, heating, and equipment consumption, which are aggregated and compared to real measurements during calibration.

Experimental results

Research questions

• RQ1 Can a deep Transformer-based metamodel accurately predict building energy loads and indoor temperatures using simulation-generated training data?
• RQ2 How effectively can real sensor data be used to calibrate a physics-informed metamodel to reflect actual building behavior?
• RQ3 To what extent can the calibrated metamodel reduce energy consumption while maintaining thermal comfort and indoor air quality in large buildings?
• RQ4 How does the proposed end-to-end approach compare in computational efficiency to traditional physics-based simulation and optimization pipelines?
• RQ5 What is the impact of hyperparameter tuning on the performance of the Transformer-based metamodel in building energy prediction tasks?

Key findings

• The Transformer-based metamodel achieves high predictive accuracy for indoor temperature and energy loads across diverse HVAC and weather conditions, as validated by distributional analysis in Figure 5.
• The calibration step using real sensor data significantly improves model fidelity, reducing discrepancies between simulated and actual indoor temperatures and heat consumption.
• The multi-objective optimization procedure successfully identifies HVAC settings that reduce total energy loads while maintaining target thermal comfort and air quality levels.
• The framework demonstrates substantial computational advantages over traditional physics-based simulation, enabling faster optimization without requiring extensive physical parameter estimation.
• The model’s performance is robust across different building configurations, as evidenced by consistent results across multiple days and weather conditions in the dataset.
• Hyperparameter tuning via grid search identified optimal settings (e.g., d_emb=64, h=8, N=4, Δ=12), which contributed to the model’s predictive stability and accuracy.