Modern IoT environments are increasingly evolving into system-of-systems, where independently managed subsystems interconnect and operate over shared IoT devices and infrastructures. As these heterogeneous systems evolve autonomously, the potential for IoT conflicts rises, particularly when they issue overlapping or competing control requests. This growing complexity underscores the need for a robust, dynamic, and real-time conflict management framework that can adapt to changing contexts and system behaviors. Traditional resolution strategies, such as fixed priorities or first-come-first-served, often fail to consider contextual factors and the effective impact of decisions, resulting in degraded non-functional properties. To address these challenges, we propose an impact-aware, attention-based conflict management framework. Our approach resolves conflicts by jointly considering request importance and their predicted system-level consequences. By leveraging real-time contextual data and historical conflict patterns, the model dynamically selects resolution actions that minimize negative impacts. We demonstrate the effectiveness of this framework through extensive evaluations in a smart transportation scenario, using energy consumption and \(CO_2\) emissions as key non-functional metrics.