Federated Dynamic Modeling and Learning for Spatiotemporal Data Forecasting

TL;DR

This work tackles the challenge of privacy-preserving spatiotemporal forecasting in distributed transport contexts by integrating an LSTM-based replacement for GRU into the DSTGCRN backbone and introducing a Client-Side Validation (CSV) mechanism within a federated learning (FL) framework. The LSTM-DSTGCRN model combines temporal modeling (LSTM) with attention and dynamic graph embeddings (AGCRN) to capture long-term dependencies and evolving spatial interactions, while CSV filters server-provided updates at the client side to boost robustness and convergence. Empirical results on multimodal transport-demand and OD-matrix datasets show that FL with CSV achieves faster convergence and improved accuracy compared to local training and standard FL baselines, with noticeable gains in MAE and more stable training dynamics. The framework preserves data privacy across heterogeneous, region-specific datasets and demonstrates practical applicability to real-time, region-specific forecasting, with open-source code enabling reproducibility and wider adoption.

Abstract

This paper presents an advanced Federated Learning (FL) framework for forecasting complex spatiotemporal data, improving upon recent state-of-the-art models. In the proposed approach, the original Gated Recurrent Unit (GRU) module within previous Dynamic Spatial--Temporal Graph Convolutional Recurrent Network (DSTGCRN) modeling is first replaced with a Long Short-Term Memory (LSTM) network, enabling the resulting model to more effectively capture long-term dependencies inherent to time series data. The resulting architecture significantly improves the model's capacity to handle complex temporal patterns in diverse forecasting applications. Furthermore, the proposed FL framework integrates a novel Client-Side Validation (CSV) mechanism, introducing a critical validation step at the client level before incorporating aggregated parameters from the central server into local models, ensuring only the most effective updates are retained and improving both the robustness and accuracy of the forecasting model across clients. The efficiency of our approach is demonstrated through extensive experiments on real-world applications, including public datasets for multimodal transport demand forecasting and private datasets for Origin-Destination (OD) matrix forecasting in urban areas. The results demonstrate substantial improvements over conventional methods, highlighting the framework's ability to capture complex spatiotemporal dependencies while preserving data privacy. This work not only provides a scalable and privacy-preserving solution for real-time, region-specific forecasting and management but also underscores the potential of leveraging distributed data sources in a FL context. We provide our algorithms as open-source on GitHub

Figures (9)

• Figure 1: The workflow of FL in multimodal transport demand forecasting. Annotation of the steps: ➊ client feeds pre-processed data into its local model; ➋ client sends trained parameter to server; ➌ server aggregates parameters received from clients; ➍ server sends aggregated parameter to clients; ➎ client feeds pre-processed testing/new data into its model; ➎ client makes prediction and transforms into the required format.
• Figure 2: Flow of the Client-Side Validation mechanism.
• Figure 3: Forecasts given by LSTM-DSTGCRN at some randomly chosen nodes of the transport demand datasets. The dashed lines present the predicted values, the solid lines present the true values.
• Figure 4: Training losses of FL for LSTM-DSTGCRN with FedAvg and FedAvg+CSV.
• Figure 5: Module replacements map resulted by LSTM-DSTGCRN + FedAvg with CSV approach. The colored boxes indicate that at that FL round, the corresponding module was replaced by the aggregate one received from server.