FedInv: A Semi-Asynchronous Federated Learning Framework with Dynamic Aggregation for Privacy-Preserving Industrial Energy Forecasting

Accurate forecasting of industrial energy consumption is essential for optimizing energy management and reducing operational costs. However, traditional centralized and synchronous federated learning (SFL) methods face significant limitations in heterogeneous environments characterized by nonIID data distributions, device variability, and communication bottlenecks, ultimately affecting prediction accuracy and training efficiency. To address these challenges, we introduce FedInv, a semi-asynchronous federated learning framework specifically designed for industrial energy forecasting. FedInv incorporates a dynamic gradient aggregation mechanism that adaptively weights client updates based on gradient staleness and directional consistency, thereby mitigating the effects of stragglers and accelerating convergence. Moreover, an adaptive buffer adjustment strategy is proposed to optimize buffer size dynamically, balancing communication overhead and model stability in largescale heterogeneous systems. Experimental evaluations on realworld industrial datasets demonstrate that FedInv consistently outperforms existing FL frameworks, achieving higher accuracy and robustness in scenarios with pronounced seasonal variation and long-term dependencies.