Federated spatial-temporal traffic forecasting with VMD-enhanced graph attention and LSTM

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Accurate spatiotemporal demand forecasting is crucial but challenged by data heterogeneity, non-stationarity, and privacy concerns in distributed systems.
• Traditional federated learning struggles with global model alignment to diverse local data distributions, hindering performance.

Purpose of the Study:

• To propose a novel federated learning framework, VMD-structured LSTM-DSTGCRN with GAT and Client-Side Validation (CSV), for enhanced spatiotemporal demand forecasting.
• To address challenges of data heterogeneity, non-stationarity, and privacy in distributed forecasting environments.

Main Methods:

• Variational Mode Decomposition (VMD) locally decomposes demand signals into Intrinsic Mode Functions (IMFs) to reduce interference.
• An LSTM-MultiHead Attention-AGCRN backbone with Graph Attention Networks (GATs) captures temporal and spatial dependencies.
• A Client-Side Validation (CSV) mechanism selectively integrates global parameters, balancing local optimization and global learning.

Main Results:

• The proposed framework significantly improves prediction accuracy, convergence speed, and robustness compared to baseline federated graph learning models.
• Centralized models showed a 28% reduction in Mean Absolute Error (MAE).
• Federated learning setup achieved a 40.6% decrease in MAE and a 20.1% reduction in Root Mean Square Error (RMSE).

Conclusions:

• The VMD-structured LSTM-DSTGCRN with GAT and CSV offers an effective, privacy-preserving solution for non-stationary, heterogeneous spatiotemporal forecasting.
• The framework demonstrates superior performance in both centralized and federated learning settings.
• This approach enhances the reliability of demand forecasting in complex distributed environments.