[Paper Review] NeurOpt: Neural network based optimization for building energy management and climate control

Achin Jain, Francesco Smarra+3 more|arXiv (Cornell University)|Jan 21, 2020

Building Energy and Comfort OptimizationEngineering18 citations

TL;DR

This paper proposes NeurOpt, a data-driven model predictive control (MPC) framework using neural networks to learn building energy and temperature dynamics from historical sensor data, enabling energy savings and improved occupant comfort without requiring physics-based models or system modifications. In a real two-story building in Italy, NeurOpt reduced energy use by 5.7% and improved temperature stability compared to a baseline rule-based controller.

ABSTRACT

Model predictive control (MPC) can provide significant energy cost savings in building operations in the form of energy-efficient control with better occupant comfort, lower peak demand charges, and risk-free participation in demand response. However, the engineering effort required to obtain physics-based models of buildings is considered to be the biggest bottleneck in making MPC scalable to real buildings. In this paper, we propose a data-driven control algorithm based on neural networks to reduce this cost of model identification. Our approach does not require building domain expertise or retrofitting of existing heating and cooling systems. We validate our learning and control algorithms on a two-story building with ten independently controlled zones, located in Italy. We learn dynamical models of energy consumption and zone temperatures with high accuracy and demonstrate energy savings and better occupant comfort compared to the default system controller.

Motivation & Objective

• To reduce the high engineering cost of physics-based modeling for building energy management.
• To enable scalable, data-driven model predictive control (MPC) without requiring domain expertise or system retrofitting.
• To improve energy efficiency and occupant comfort in real buildings using neural networks to model both energy consumption and zone temperature dynamics.
• To demonstrate that black-box neural network models can effectively replace traditional physics-based models in MPC for real-world buildings.
• To validate the approach on a real building with ten independently controlled zones, showing measurable energy savings and comfort improvements.

Proposed method

• Neural networks are trained on historical data from building automation systems (BAS) and weather stations to model energy consumption and zone temperature dynamics.
• The learned neural network models are integrated into a nonlinear MPC framework that optimizes control inputs to balance energy use and thermal comfort.
• Two MPC variants are implemented: MPC-min to minimize energy use while respecting comfort bounds, and MPC-tracking to reduce temperature variance around a setpoint.
• The MPC controller acts as a supervisory layer, dynamically adjusting setpoints without modifying the existing heating system.
• Real-time data flow is enabled via a LabVIEW-based local server communicating with a cloud-based Elasticsearch database through Modbus TCP/IP and OPC protocols.
• The system synchronizes setpoints and measurements every 15 seconds between the local server and remote cloud database for real-time monitoring and control.

Experimental results

Research questions

• RQ1 Can neural networks accurately model both energy consumption and zone temperature dynamics in a real building using only historical sensor and weather data?
• RQ2 Can a neural network-based MPC controller achieve significant energy savings compared to a rule-based baseline controller without modifying the existing HVAC system?
• RQ3 How does the performance of the neural network MPC compare to rule-based control in terms of occupant comfort and energy use stability?
• RQ4 Can the controller effectively trade off energy savings and thermal comfort through dynamic setpoint adjustment?
• RQ5 What is the impact of using black-box neural networks instead of physics-based models on MPC scalability and deployment effort?

Key findings

• NeurOpt achieved a 5.7% reduction in energy consumption over the baseline rule-based controller during a three-day experiment in a real two-story building.
• The MPC-min controller maintained zone temperatures closer to the lower comfort bound, reducing energy use while preserving thermal comfort.
• The MPC-tracking controller significantly reduced temperature variance compared to the baseline, demonstrating improved occupant comfort without system modifications.
• The neural network models learned from historical data accurately captured the nonlinear dynamics of energy consumption and zone temperatures.
• The approach eliminated the need for physics-based modeling, reducing engineering effort and enabling faster deployment of MPC in real buildings.
• The system demonstrated real-time operability through a cloud-based data pipeline with 15-second synchronization between control commands and sensor measurements.