A HYBRID LEARNING APPROACH COMBINING GRAPH AND TRANSFORMER MODELS FOR COMMUNICATION EFFICIENT DISTRIBUTED CONTROL AND ESTIMATION IN NETWORKED SYSTEMS

The operation of the Networked Control Systems underpins modern-automated industries including cyber-physical systems, large-scale infrastructure and process applications. However, classical methods of Distributed Model Predictive Control and Event-Triggered Distributed Estimation suffers from inefficiencies such as communication overhead, poor state estimation, slow adaptation to system variations, and short predicting horizons. The methods currently in use fail to capture the spatial and temporal dependencies within the networked subsystems, resulting in suboptimal control decisions and excessive computational. To address this, the proposed Integrated Model for DMPC and ETDE includes the following five enhanced methodologies: (1) Graph Neural Network-Based Predictive Control; (2) Attention-driven Event-Triggered Estimation; (3) Transformer-Based Predictive Observer; (4) Meta-Learning-Based Adaptive Control; and (5) Variational Autoencoder-Based Communication-Efficient Control (VAE-CEC). GNN-PC allows for an efficient approach to modeling interdependencies between subsystems, thus strengthening decentralized control decisions. AET-E employs attention mechanisms to focus on updates from relevant subsystems, therefore preventing unnecessary transmissions. TPO utilizes transformers for accurate predictions of long-range states to adopt resilience against data losses. MLAC guarantees the required robustness under non-stationary conditions by promoting quick adaptation to changing environments through meta-learning. VAE-CEC realizes effective communication by compressing high-dimensional state information at a low cost, which does not affect control performance. The integrated model reported communication overhead saving of 50%, better control adaptive capacity by 40%, improvement in accuracy of state prediction by 35%, and control error reduction by 30%. This proves that the proposed work significantly enhances the efficiency and reliability of DMPC and ETDE methods, thereby making real-time distributed control technology more scalable, adaptive, and resource efficient.