Reduced-Order Multi-Physics Modeling and Active Power Derating for Preventing Voltage Collapse Under Constant-Power Loads

Yuwei Zhou, Changsong Wang, Zihao Wan

Accepted to IEEE I&CPS Asia 2026. The paper asks why a lithium-ion cell can collapse under constant-power loads before its stored charge is fully used. We derive a Padé-approximated single-particle model and a closed-form Voltage Margin Governor that derates power before the system reaches an unsafe region. Against a DFN reference at -10°C, the reduced model achieves 5.911 mV voltage RMSE and 5.842 mV end-of-discharge MAE, with an average online update time of 85.60 microseconds per step. My contribution centered on the mathematical derivation, theorem proof, and manuscript writing. 录用于 IEEE I&CPS Asia 2026。论文关注一个具体问题：在恒功率负载下，锂电池为什么会在电量尚未完全耗尽前发生电压崩溃。我们推导 Padé 近似单粒子模型，并设计闭式电压裕度调速器，使系统在进入危险区域前主动降额。在 -10°C DFN 对照实验中，降阶模型电压 RMSE 为 5.911 mV，放电末端 MAE 为 5.842 mV，在线更新时间平均 85.60 微秒/步。本人主要负责数学推导、定理证明和论文写作。

Adaptive Simplicial-Complex Epidemic Modeling with HAPA Behavioral Memory行为记忆与高阶传播耦合的适应性单纯复形流行病模型

Zihao Wan, Longxing Qi

This preprint studies how behavioral memory and higher-order social interactions change epidemic dynamics. I start from an agent-based model, reduce it to a six-dimensional mean-field ODE, and analyze thresholds, stability, hysteresis, and oscillatory transitions. The bifurcation analysis identifies a Hopf-to-homoclinic oscillation window. In a Shanghai 2022 COVID-19 case study, the HAPA-SEIR model reaches test R² = 0.9515 and reduces RMSE by 56.02% compared with the SEIR baseline. The project taught me to separate curve fitting from mechanism-level explanation. 这篇预印本研究行为记忆和高阶社会交互如何改变疫情传播动力学。我从智能体模型出发，将系统约化为六维均场 ODE，并分析阈值、稳定性、磁滞和振荡转迁。分岔分析揭示了由 Hopf 分岔到同宿轨道组织的振荡窗口。在上海 2022 年疫情案例中，HAPA-SEIR 测试段 R² = 0.9515，相比 SEIR 基线 RMSE 降低 56.02%。这段工作也让我更早意识到，拟合曲线和解释机制是两件需要同时认真对待的事。

State-Adaptive Machine Learning for Systemic Financial Risk Measurement状态自适应机器学习的系统性风险测度

Zihao Wan, et al.

This working paper studies systemic risk measurement under changing market regimes. We use a Student-t HMM to infer macro-financial states, train regime-conditional LightGBM experts for rolling VaR/MES prediction, combine them through soft MoE gating, and use TreeSHAP to interpret state-dependent risk drivers. The part that stayed with me is calibration: a model should not speak as if the market were unchanged when the data indicate a different regime. 这篇工作论文研究市场区制变化下的系统性风险测度。模型使用 Student-t HMM 识别宏观金融状态，在不同状态下训练 LightGBM 专家模型进行滚动 VaR/MES 预测，并通过 MoE 软门控聚合，同时用 TreeSHAP 解释状态依赖的风险驱动因素。这个项目让我印象最深的是校准问题：当数据已经显示市场进入不同区制时，模型不应仍以同一种口吻作判断。

• Introduced HAPA behavioral factors into epidemic modeling to describe how perception, memory, and action can reshape transmission.将 HAPA 行为因素引入传染病建模，用来刻画感知、记忆和行动如何反过来改变传播过程。
• Built a multi-layer simplicial-complex simulation, then reduced it to a six-dimensional ODE system for threshold, stability, fold/Hopf bifurcation, hysteresis, and oscillation analysis.构建多层单纯复形网络仿真，并进一步约化为六维 ODE 系统，用于分析阈值、稳定性、fold/Hopf 分岔、磁滞和振荡。
• Led proposal writing, model design, algorithm implementation, numerical simulation, result analysis, and manuscript preparation; the Shanghai case test reached R² = 0.9515.负责项目申报、模型设计、算法实现、数值仿真、结果分析与论文准备；上海案例测试段 R² 达到 0.9515。
• What stayed with me:这段经历留下的影响： It made me more careful about the gap between matching data and understanding the process that generated it.它让我更谨慎地看待“拟合数据”和“理解数据生成过程”之间的差别。

• Proposed a spatiotemporal manifold approximation to make a high-dimensional, non-smooth optimization problem tractable.提出时空流形近似方法，将高维、非光滑优化问题转化为更可处理的形式。
• Designed a hierarchical strategy for single-target, multi-UAV, and multi-target settings, and completed modeling, coding, validation, and paper writing.设计单目标、多无人机和多目标场景下的分层策略，并完成建模、代码实现、验证和论文写作。
• What I learned:这段经历让我学到： A good algorithm often starts with seeing the structure of the problem clearly, before choosing the tool.好的算法往往不是先选工具，而是先看清问题本身的结构。

• Used HMM to describe market regimes and trained LightGBM expert models for rolling VaR and MES prediction under different states.使用 HMM 描述市场区制，并在不同状态下训练 LightGBM 专家模型进行 VaR 与 MES 滚动预测。
• Worked on data cleaning, feature engineering, model construction, code implementation, rolling evaluation, and TreeSHAP interpretation.负责数据清洗、特征工程、模型构建、代码实现、滚动评估与 TreeSHAP 解释分析。
• What interested me most:我最感兴趣的地方： The useful question was not only how to predict risk, but when the model should stop trusting a pattern learned from another regime.我最感兴趣的不只是如何预测风险，而是模型何时应当停止信任来自另一个区制的旧模式。

• Worked on literature review, technical route design, and proposal writing for SOH diagnosis, lifetime prediction, and adaptation across operating conditions.参与 SOH 诊断、寿命预测和多工况适应方向的文献调研、技术路线设计和项目技术方案撰写。
• Designed a PI-DTCNN route combining segment-level CNN features, transfer learning, SOH regression, degradation-rate modeling, and physical consistency constraints. The phase-one targets are MAE/MAPE/RMSE ≤ 2% and R² ≥ 98%.设计 PI-DTCNN 技术路线，结合片段级 CNN 特征、迁移学习、SOH 回归、退化速率建模与物理一致性约束；一期验收目标为 MAE/MAPE/RMSE ≤ 2%、R² ≥ 98%。
• What this changed for me:这段经历改变了我对模型的理解： Accuracy is only one part of usefulness; a model also has to respect constraints and make clear what its estimate rests on.准确率只是有用性的一部分；模型还需要尊重约束，并说明估计结果建立在什么依据之上。