Гібридний фреймворк для адаптивного неперервного контролю на базі нейронних операторів Фур’є

• Datum: 2026
• Hauptverfasser: Петровський, Олександр, Лясковська, Соломія
• Format: Artikel
• Sprache: Englisch
• Veröffentlicht: Kamianets-Podilskyi National Ivan Ohiienko University 2026
• Online Zugang: https://mcm-tech.kpnu.edu.ua/article/view/354699
• Назва журналу: Mathematical and computer modelling. Series: Technical sciences

Hybrid adaptive control methods are of high scientific interest and industrial urgency due to their ability to address the weaknesses of both model-driven and data-driven controllers, as the former are reliable and predictable, but rigid and often suboptimal, while the latter provide eventual optimality guarantees given enough time and exploration, but are effectively unapplicable to systems with a high cost of operation and failure. Thus, this work aims to develop and evaluate a hybrid framework for building reinforcement-learning-based adaptive controllers that leverages the unique properties of Fourier Neural Operators (FNOs) to achieve higher accuracy and reliability in control systems. To achieve this, we propose an architecture of a TD3-based agent that uses pretrained FNO as a world model. The framework includes training the FNO network on historical data, pretraining the agent on the obtained environment surrogate, adjusting the world model in real time based on new observations, and dynamically balancing between Dreamer-like planning and Q-network estimates depending on how well the world model can predict the system's response to the agent's actions. The effectiveness of the method was evaluated on the simulation of a baker’s yeast fermenter. Experimental results show that the proposed method significantly outperforms popular algorithms such as SAC, TD3, and TD3+CQL, as well as a pure FNO-based controller: the agent successfully reaches the target biomass concentration without risky exploration of the environment and demonstrates the ability to overcome model-reality mismatches, which proves its effectiveness and great potential. The practical value of this work lies in developing a method that enables the creation of reliable adaptive controllers for complex nonlinear processes with a high cost of failure, which do not require analytical models and can continually adjust themselves to real conditions, given the availability of historical data.