The increasing complexity of contemporary control systems, driven by advancements in artificial intelligence, automation, and the widespread use of intelligent systems, demands new control and optimization strategies. Traditional approaches, primarily focusing on individual components or the system as a whole, are increasingly inadequate for control the dynamic interactions and adaptability requirements of these complex systems. This work presents a real-time decentralized clustering method for dynamical multi-agent systems to reduce data dimensionality for adaptive control approaches. Unlike traditional methods that rely on static connections and centralized processing, this method is designed to accommodate dynamic network topologies and varying system states. The method employs compressive sensing and an accelerated consensus approach to acquire aggregated compressed data. A pre-trained neural network then processes this data to forecast cluster centroids. © © 2025 The Authors.