Smart Cyber-Physical Systems (sCPS) operate in dynamic and uncertain environments, where anticipation to adverse situations through effective decision-making is crucial and decentralization is often necessary due to scalability, resilience, and effiency reasons. Addressing the limitations related to the lack of foresight of (decentralized) reactive self-adaptation (e.g., slower response, sub-optimal resource usage), this paper introduces a novel method that employs Predictive Coordinate Descent (PCD) to enable decentralized proactive self-adaptation in sCPS. Our study compares our proactive PCD approach with a reactive Deep Q-Network (DQN)-based strategy on Unmanned Aerial Vehicles (UAVs) in wildfire tracking adapta- tion scenarios. Results demonstrate the effectiveness of PCD, which outperforms DQN both under standard operational conditions, as well as in challenging scenarios with limited observability of the environment.