In recent years, the global electric vehicle (EV car) industry has experienced exponential growth, driven by technological advancements, policy incentives, and shifting consumer preferences. However, this rapid expansion has exposed the EV cars supply chain to multifaceted risks, including geopolitical tensions, supply disruptions, and technological vulnerabilities. These risks can propagate through complex networks, potentially destabilizing the entire supply chain. In this study, we aim to identify key risks in the EV car supply chain, analyze their propagation dynamics using a coupled physical-information layer network model, and develop an early warning system to enhance risk resilience. By integrating topic modeling, complex network theory, and dynamic simulations, we provide a comprehensive framework for risk management in the EV car industry.
The proliferation of EV cars has reshaped automotive supply chains, introducing dependencies on critical components like batteries and semiconductors. For instance, the global market share of power batteries for EV cars exceeds 60%, highlighting the concentration of resources. Yet, this reliance creates vulnerabilities, as disruptions in raw material supply or geopolitical conflicts can cascade through the network. We begin by reviewing existing literature on EV car supply chains, complex network applications, risk identification methods, and early warning mechanisms. Previous studies have employed models like SEIR and LDA to analyze risk propagation, but few have integrated physical and information layers to capture the full scope of dynamics in EV cars networks.
To address this gap, we design a methodology that combines Latent Dirichlet Allocation (LDA) for risk theme extraction with a UAU-SIR (Unaware-Aware-Undecided/Susceptible-Infected-Recovered)双层网络 model for simulating risk propagation. The LDA model processes multi-source text data, including industry reports and news articles, to identify prevalent risk themes. Subsequently, we construct a双层网络 where the physical layer represents actual supply chain connections among EV car manufacturers and suppliers, and the information layer models the dissemination of risk-related knowledge. The coupling between these layers allows us to simulate how risks spread and interact, providing insights into their propagation rates and impacts.
For risk identification, we collect and preprocess data from sources such as annual reports and news platforms, resulting in over 10,000 text entries. Using jieba for tokenization and custom dictionaries, we refine the data and apply the LDA model to extract key risk themes. The optimal number of themes is determined through coherence scoring, leading to the identification of six core risk categories for EV cars: technical safety, geopolitical issues, supply interruptions, policy compliance, production operations, and infrastructure. Each theme is associated with representative keywords, as summarized in Table 1.
| Risk Theme | Representative Keywords |
|---|---|
| Technical Safety Risks | artificial intelligence, algorithms, cybersecurity, core technology, autonomous driving |
| Geopolitical Risks | Russia-Ukraine conflict, tariffs, international trade, supply chain security, maritime transport |
| Supply Interruption Risks | power batteries, chip shortages, natural disasters, raw materials, lithium prices |
| Policy Compliance Risks | purchase tax exemptions, replacement policies, market regulation, dual-credit systems, incentive policies |
| Production Operation Risks | supply chain, industry chain, technological innovation, equipment failures, informatization |
| Infrastructure Risks | charging infrastructure, 5G, logistics centers, power grid, charging and swapping services |
These risk themes are interconnected; for example, technical safety risks involving AI and autonomous driving in EV cars may exacerbate infrastructure risks related to charging networks, while geopolitical events can trigger supply interruptions. To model their propagation, we analyze the supply chain network of a representative EV car manufacturer, referred to as Company A. The network comprises 95 nodes and 115 edges, with an average path length of 3.278 and a density of 0.013, indicating a loosely connected structure where risks can spread rapidly. The average clustering coefficient is low (0.003), suggesting limited local cooperation, which may amplify vulnerability in EV cars supply chains.
We define the UAU-SIR model parameters based on the LDA theme weights. In the physical layer, nodes can be in susceptible (S), infected (I), or recovered (R) states, representing unaffected, impacted, and resilient entities in the EV car supply chain. The information layer uses unaware (U), aware but not acting (A1), and aware and acting (A2) states to capture risk perception and response. The infection rate ($\beta$) and recovery rate ($\gamma$) are derived from theme weights, as shown in Equation (1) and Table 2:
$$ \beta = 0.05 + \left( \frac{\text{theme weight}}{\text{max weight}} \right) \times 0.10 $$
$$ \gamma = 0.05 – \left( \frac{\text{theme weight}}{\text{max weight}} \right) \times 0.03 $$
| Risk Theme | Weight | $\beta$ | $\gamma$ |
|---|---|---|---|
| Technical Safety | 0.157 | 0.15 | 0.05 |
| Geopolitical | 0.032 | 0.10 | 0.03 |
| Supply Interruption | 0.054 | 0.12 | 0.04 |
| Policy Compliance | 0.048 | 0.08 | 0.02 |
| Production Operation | 0.016 | 0.11 | 0.04 |
| Infrastructure | 0.114 | 0.09 | 0.03 |
| Composite Risk | — | 0.11 | 0.035 |
Simulations reveal distinct propagation characteristics for each risk type in EV cars networks. Technical safety, supply interruption, and production operation risks exhibit rapid spread and high infection rates but quick recovery, akin to “acute illnesses.” For instance, technical safety risks in EV cars, such as cybersecurity breaches, can infect up to 90.99% of nodes within 2 days, with a recovery rate of 0.05. In contrast, geopolitical risks propagate slowly but have prolonged impacts, infecting 95.21% of nodes and requiring over 56 days for recovery, resembling “chronic conditions.” Policy compliance and infrastructure risks show slower propagation but extended recovery periods due to systemic dependencies, such as regulatory adjustments or charging station availability for EV cars.
The dynamic propagation is modeled using differential equations. In the physical layer, the change in susceptible nodes ($S$) over time ($t$) is given by:
$$ \frac{dS}{dt} = -\beta S I $$
where $I$ represents infected nodes. The infected nodes change as:
$$ \frac{dI}{dt} = \beta S I – \gamma I $$
and recovered nodes evolve as:
$$ \frac{dR}{dt} = \gamma I $$
In the information layer, the transition from unaware ($U$) to aware ($A1$) is governed by the risk perception rate ($\lambda_U$), and from $A1$ to acting ($A2$) by the action conversion rate ($\lambda_A$). The coupling between layers is captured by cross-layer influence parameters, such as the risk perception delay ($\tau_{IA}$) and action protection efficiency ($\eta_{AS}$). For example, a supply disruption in the physical layer (e.g., lithium price spike for EV cars batteries) can trigger awareness in the information layer, leading to preventive actions that mitigate further spread.
Based on these dynamics, we develop a multi-level early warning指标体系 for EV cars supply chains. The indicators are derived from UAU-SIR network properties, including risk perception rates, infection pressure, propagation acceleration, and recovery coefficients. For instance, the risk perception rate $\lambda_U$ measures how quickly nodes become aware of risks, calculated as:
$$ \lambda_U = \frac{A1_t – A1_{t-1}}{U_{t-1}} $$
Similarly, the infection pressure $P_i(t)$ in the physical layer assesses the likelihood of node infection:
$$ P_i(t) = \frac{\text{number of infected neighbors}}{\text{total neighbors}} $$
We set warning thresholds using an XGBoost model, which processes feature vectors from network dynamics and text data. The model outputs risk probabilities, categorized into three levels: Level 1 (high risk, probability above upper threshold $Y$), Level 2 (medium risk, between mean $M$ and $Y$), and Level 3 (low risk, below $M$). This approach allows for real-time monitoring and proactive interventions in EV car supply chains.
To validate the warning system, we apply it to a case study involving Company A during the 2021-2022 lithium price surge. As demand for EV cars batteries soared, lithium prices increased sharply, posing supply interruption risks. Our model processes related news data and network simulations, generating risk probabilities. The results show that 22% of periods triggered medium-risk warnings (Level 2), indicating initial supply chain stress, while 18% reached high-risk levels (Level 1), corresponding to actual disruption events. For example, on day 5 of the simulation, the infection pressure peaked at 0.9302, and propagation acceleration reached 0.8, signaling an imminent crisis. The early warning system issued Level 1 alerts 72 hours before significant disruptions, enabling preemptive measures such as inventory adjustments or supplier diversification for EV cars production.
In conclusion, our study demonstrates the effectiveness of integrating LDA theme modeling with UAU-SIR双层网络 simulations for EV car supply chain risk management. The identified risk themes—technical safety, geopolitical, supply interruption, policy compliance, production operation, and infrastructure—highlight the interconnected nature of vulnerabilities in EV cars networks. Simulations reveal that risks with high thematic weights, such as technical safety and supply interruptions, propagate rapidly but are manageable with timely recovery efforts, whereas geopolitical and infrastructure risks cause prolonged disruptions. The early warning system, with its multi-level indicators, provides a robust tool for detecting and mitigating risks in EV car supply chains, enhancing resilience against dynamic threats.
We recommend that EV car manufacturers and stakeholders adopt a proactive approach by monitoring key risk indicators, diversifying suppliers, investing in technology resilience, and fostering collaboration across the supply chain. Future research could explore real-time data integration and machine learning enhancements to further refine risk predictions for the evolving EV cars industry. By addressing these challenges, the sustainable growth of EV cars can be supported, ensuring stability in the face of global uncertainties.