KazNU Scientists Develop Multi-Agent AI System for Building Energy Management

Buildings worldwide consume up to 60% of electricity, with this figure reaching 40% in Europe. Rapid urbanization and growing urban populations are intensifying the need to reduce building energy consumption. At the same time, human comfort remains a top priority, requiring both goals to be addressed simultaneously.

Most modern control systems account only for standard temperature, humidity, and CO2 thresholds, without considering a person's actual state of comfort. In addition, many buildings experience simultaneous peak energy demand, placing excessive strain on the power grid and shortening equipment lifespan. Practical implementations of multi-agent reinforcement learning systems remain rare worldwide, making the development of new solutions particularly relevant.

To create an intelligent system for managing building energy consumption and ensuring human comfort using multi-agent reinforcement learning, featuring a new asynchronous energy consumption algorithm and a thermography-based system for determining people's states.

The project will develop a model for determining human comfort levels by combining facial expression recognition and facial temperature (thermography) analysis, regardless of the relative position of the camera and face. A multi-agent control system with reinforcement learning will be created for the first time to enable asynchronous consumption of building energy resources. Artificial intelligence will also be developed to manage a comfort system integrated with photovoltaic systems, and the resulting system will be applied in practice.

The project comprises ten stages: developing an energy monitoring and control system to build a thermodynamic model; developing a thermodynamic environment model for agent training; creating a human state recognition system based on surveillance and infrared cameras; analyzing correlations between input and output data to identify key system parameters; studying the effectiveness of decision-making methods; training the multi-agent system in a virtual environment; testing system flexibility by scaling to more rooms; developing a photovoltaic system and collecting energy generation data; training a neural network to forecast energy generation; and testing the multi-agent system in practice.

The building's thermodynamic model is constructed using COMSOL Multiphysics and EnergyPlus software. Human state recognition employs YOLO (for face detection) and MTCNN (for facial region segmentation) neural networks, along with a Siamese Network architecture. The multi-agent system operates on an Actor-Critic architecture within the framework of Markov game theory. Hyperparameter optimization uses a genetic algorithm, particle swarm optimization, and simulated annealing, while RNN, LSTM, GRU, 1D-CNN, MLP, GNN, and transformer architectures are explored for energy generation forecasting.

The project's results will be published in journals ranked Q1-Q2 in Web of Science or with a CiteScore percentile of at least 65 in Scopus, with one article in a multidisciplinary category addressing tasks from real-sector enterprises in Kazakhstan's economy. The research will contribute to the development of Smart Grid and Microgrid concepts, accelerating the country's transition to green energy.

The developed system will increase competition in the energy market and lower electricity prices for consumers, creating new jobs for servicing decentralized building energy systems. Reduced peak loads will extend the lifespan of energy systems and lower CO2 emissions. The social benefit lies in creating comfortable working conditions for personnel and early detection of workplace burnout, improving personnel management efficiency.

At least 2 articles in journals ranked Q1-Q2 in Web of Science or with a CiteScore percentile of at least 65 in Scopus (one in a multidisciplinary category); at least 2 articles in journals recommended by KOKSON; training of at least 1 PhD candidate; design documentation prepared according to ESKD standards.

Spanning 36 months (2025-2027), the project will generate new knowledge at the intersection of artificial intelligence, thermodynamics, photovoltaics, and psychology. Its main outcome will be an intelligent system for managing energy consumption and comfort in buildings, with high commercialization potential for both individual components and the complete system.