MULTI-AGENT ARCHITECTURE OF SOFT SENSORS BASED ON DEEP LEARNING FOR PROACTIVE ENERGY EFFICIENCY MANAGEMENT OF SUBMERGED ARC FURNACES

Abstract

The relevance of this study is determined by the high energy intensity of ferroalloy production and the limited ability to directly monitor internal bath parameters in submerged arc furnaces (SAF), which significantly complicates effective process control under conditions of uncertainty and noisy industrial data. Existing automated control systems are predominantly based on reactive strategies that focus on correcting already occurred deviations, leading to increased energy consumption and higher operational risks. Therefore, there is a growing need to implement proactive approaches capable of predicting system states and generating control actions in advance.

The aim of this study is to develop a concept of an intelligent control system based on a multi-agent architecture using soft sensors to improve energy efficiency and operational safety. The research employs deep learning methods, including recurrent neural networks (LSTM/GRU), physics-informed neural networks (PINN), as well as multi-agent coordination and decision-making approaches. The proposed system architecture includes specialized agents focused on energy balance, thermal state monitoring, and technological process parameters, whose interaction is coordinated by an orchestrator agent responsible for synthesizing optimal control actions.

The results demonstrate the feasibility of transitioning to proactive control with a prediction horizon of 1–3 hours, enabling reduction of specific energy consumption and mitigation of operational risks. In addition, the proposed approach provides improved robustness to noisy industrial data and enhances the interpretability of model outputs through the integration of explainable artificial intelligence techniques. It is established that the combination of a multi-agent approach with hybrid neural network models improves prediction accuracy, consistency of control decisions, and overall stability and reliability of the technological process in energy-intensive industrial environments.