Reinforcement Learning-Guided Federated Multi-Task Learning for Adaptive Energy Management in IIoT-Enabled Industrial Microgrids

Industrial microgrids are evolving into highly heterogeneous and dynamic environments, increasingly enabled by edge intelligence and distributed control. Traditional Federated Learning (FL) approaches remain shackled by static aggregation policies, single-task learning goals, centralized control, scalability, and robustness issues in non-IID data regimes with time-varying industrial energy demands. To overcome these challenges, the paper proposes a Reinforcement Learning-guided Federated Multi-Task Learning (FMT-RL) framework that integrates multitask representation learning and adaptive aggregation. An RL agent learned using Proximal Policy Optimization (PPO), dynamically chooses FL clients to participate and allocates aggregation weights based on real-time key performance indicators such as demand forecasting accuracy, cost efficiency, battery utilization, and CO2 emission reduction. Experimental evaluations on standard industrial microgrid datasets demonstrate that the proposed FMT-RL framework significantly outperforms popular baselines such as FedAvg, FedProx, and Single-Task Learning, achieving 0.098 RMSE with 2.7% MAPE in demand forecasting, 14.9% reduction in electricity cost, 87.7% battery efficiency, and 18.2% reduction in emissions. These findings indicate that FMT-RL as a scalable, decentralized, and adaptive approach for resilient and sustainable energy optimization in IIoT-enabled industrial microgrids, and has significant potential for deployment in next-generation industrial process control systems.