Collaborative City Digital Twin for the COVID-19 Pandemic: A Federated Learning Solution

The novel coronavirus, COVID-19, has caused a crisis that affects all segments of the population. As the knowledge and understanding of COVID-19 evolve, an appropriate response plan for this pandemic is considered one of the most effective methods for controlling the spread of the virus. Recent studies indicate that a city Digital Twin (DT) is beneficial for tackling this health crisis, because it can construct a virtual replica to simulate factors, such as climate conditions, response policies, and people’s trajectories, to help plan efficient and inclusive decisions. However, a city DTsystem relies on long-term and high-quality data collection to make appropriate decisions, limiting its advantages when facing urgent crises, such as the COVID-19 pandemic. Federated Learning (FL), in which all clients can learn a shared model while retaining all training data locally, emerges as a promising solution for accumulating the insights from multiple data sources efficiently. Furthermore, the enhanced privacy protection settings removing the privacy barriers lie in this collaboration. In this work, we propose a framework that fused city DT with FL to achieve a novel collaborative paradigm that allows multiple city DTs to share the local strategy and status quickly. In particular, an FL central server manages the local updates of multiple collaborators (city DTs), providing a global model that is trained in multiple iterations at different city DT systems until the model gains the correlations between various response plans and infection trends. This approach means a collaborative city DT paradigm fused with FL techniques can obtain knowledge and patterns from multiple DTs and eventually establish a "global view" of city crisis management. Meanwhile, it also helps improve each city’s DT by consolidating other DT’s data without violating privacy rules. In this paper, we use the COVID-19 pandemic as the use case of the proposed framework. The experimental results on a real dataset with various response plans validate our proposed solution and demonstrate its superior performance.