FastML-GA: FPGA-accelerated machine learning for real-time energy HVAC optimization in buildings

Mshragi, Mohammed ORCID: https://orcid.org/0009-0003-6511-4594 and Petri, Ioan ORCID: https://orcid.org/0000-0002-1625-8247 2026. FastML-GA: FPGA-accelerated machine learning for real-time energy HVAC optimization in buildings. Neural Computing and Applications 38 (2) , 25. 10.1007/s00521-025-11737-x

PDF - Published Version Download (6MB)

Abstract

Fast machine learning (FastML) has strong potential to enhance energy optimization and operational efficiency in heating, ventilation, and air conditioning (HVAC) systems within building management systems (BMS). Traditional HVAC control approaches frequently depend on static schedules and computationally intensive, CPU-based optimization techniques, which often lack the responsiveness and scalability required for real-time embedded applications. To address these limitations, we propose a fast machine learning framework that integrates a random forest surrogate model implemented as a hardware accelerator on the programmable logic (PL) with a lightweight and adaptive genetic algorithm (GA) executed on the processing system (PS), thereby forming a hybrid PS–PL deployment. This combination of fast machine learning and evolutionary algorithms optimization delivers substantial computational efficiency, achieving over 1.67 million predictions per second on a PYNQ-Z1 FPGA and significantly outperforming recent FPGA-based approaches. By using a case study, we demonstrate how FastML can employ a GA multi-objective fitness function to dynamically optimize hourly airflow rates and supply air temperatures in response to occupancy and seasonal environmental patterns, thereby reducing electricity and thermal energy consumption while maintaining occupant comfort within standard predicted mean vote (PMV) thresholds. Empirical evaluation conducted over 72 days across four distinct seasons reveals consistent electricity savings exceeding 50%, alongside thermal energy reductions of up to 150 kWh per day during heating periods. A comprehensive three-dimensional Pareto front analysis further substantiates the system’s capability to effectively balance energy efficiency and occupant comfort. These results highlight the practicality, scalability, and substantial promise of FPGA-based multi-objective optimization as a robust, real-time solution for intelligent and sustainable building energy management at the edge.

Item Type:	Article
Date Type:	Published Online
Status:	Published
Schools:	Schools > Engineering
Additional Information:	License information from Publisher: LICENSE 1: URL: http://creativecommons.org/licenses/by/4.0/, Type: open-access
Publisher:	Springer
ISSN:	0941-0643
Date of First Compliant Deposit:	2 February 2026
Date of Acceptance:	8 November 2025
Last Modified:	02 Feb 2026 10:00
URI:	https://orca.cardiff.ac.uk/id/eprint/184295

Actions (repository staff only)

Edit Item