Federated Learning-based Topology-aware Traffic Prediction with Directional Alignment in Industrial IoT

In this paper, we propose a novel Federated Learning framework, named Federated Topology-aware Learning with Directional Alignment Aggregation (F2L-DAA), to address the instability issues in federated traffic prediction for Industrial Internet of Things (IIoT) networks arising from hierarchical data challenges. Specifically, at the local level, we design a Topology-aware Long Short-Term Memory (TO-LSTM) model featuring a graph-harmonic imputation module, which leverages device adjacency via Laplacian regularization to ensure spatial consistency. Subsequently, a robust tensor decomposition step utilizing Wasserstein-guided thresholding and Grassmannian retraction preserves transient temporal patterns, while a sparsetemporal attention mechanism further mitigates missing-data artifacts. At the global level, we introduce a Directional Alignment Aggregation (DAA) strategy that selectively and weightedly aggregates local model updates according to their directional similarity, quantified through cosine similarity, thereby reducing update variance caused by statistical heterogeneity and ensuring unbiased contributions. Experimental evaluations on IIoT uplink scenarios demonstrate that our proposed F2L-DAA framework achieves faster convergence and significantly outperforms base-line methods, reducing Mean Absolute Error (MAE) by 30.0%, Root Mean Square Error (RMSE) by 23.7%, and improving the coefficient of determination by 28.3%.