I am an Algorithm Development Expert at SUPCON. I received my Ph.D. from the College of Computer Science and Technology at Zhejiang University in December 2025, advised by Prof. Yang Yang and Yin Zhang.
My research focuses on time series modeling and intelligent decision-making for the process industries. I aim to build models that learn transferable dynamics from multivariate time series and support forecasting, simulation, and industrial decisions. My current interests include:
- Time series foundation models: learning temporal patterns and cross-variable dependencies at scale, with a focus on multivariate forecasting, conditioning on future-known covariates, and generalization across operating conditions and industrial processes.
- Dynamics and causal modeling: studying how models learn process responses to control inputs, how causal structure and process knowledge can inform modeling, and how to evaluate models for simulation and control beyond forecasting accuracy.
- Industrial AI agents: combining time series models, LLMs, and engineering tools to support predictive maintenance, anomaly diagnosis, and operational decision-making.
My earlier work spans spatiotemporal forecasting in chemical sensor networks and traffic networks, foundation models for power systems, and cross-subject seizure detection. My work has appeared at NeurIPS and KDD 
I welcome collaborations on time series foundation models, dynamical systems, and industrial AI. Please feel free to contact me at shihao.tu@zju.edu.cn.
我是涂世豪，现任中控算法开发专家。2025年12月，我在浙江大学计算机科学与技术学院获得博士学位，师从杨洋教授、张引教授。
我的研究聚焦于面向流程工业的时间序列建模与智能决策，希望从多变量时序数据中学习可迁移的动力学规律，支撑预测、仿真与工业决策。目前主要关注：
- 时间序列基础模型：通过大规模预训练学习时间模式与变量间依赖，重点研究多变量预测、未来已知协变量条件下的建模，以及跨工况、跨装置的泛化能力。
- 动力学与因果建模：研究模型如何学习工业过程对控制输入的响应，如何利用因果结构与工艺知识辅助建模，以及如何超越预测误差，从仿真与控制效果评价模型能力。
- 工业智能体（Agent）：结合时序模型、大语言模型与工程工具，探索预测性维护、异常诊断与运行决策辅助。
此前，我的研究涉及化学传感器网络与交通网络的时空预测、电力系统基础模型，以及跨受试者癫痫检测，相关工作发表于 NeurIPS、KDD 等国际会议
欢迎就时序基础模型、动力系统建模与工业 AI 开展交流合作：shihao.tu@zju.edu.cn。
🔥 News🔥 新闻
- 2026.09: 🎉🎉 Our paper “Can Time-Series Foundation Models Serve as Simulators for Process Control?” has been accepted by the FMTS Workshop at NeurIPS 2026. 🎉🎉 论文《Can Time-Series Foundation Models Serve as Simulators for Process Control?》被 NeurIPS 2026 FMTS Workshop 录用。
- 2026.05: 🎉🎉 Our paper
STUNethas been accepted by KDD 2026. 🎉🎉 论文STUNet被 KDD 2026 录用。 - 2026.01: Joined SUPCON as an Algorithm Development Expert. 入职中控，担任算法开发专家。
- 2025.12: Received my Ph.D. from Zhejiang University. 于浙江大学获得博士学位。
- 2025.05: 🎉🎉 Our paper
ASTNethas been accepted by KDD 2025. 🎉🎉 论文ASTNet被 KDD 2025 录用。 - 2024.09: 🎉🎉 Our papers
PowerPMandDMNethave been accepted by NeurIPS 2024. 🎉🎉 论文PowerPM与DMNet被 NeurIPS 2024 录用。
📝 Publications📝 论文
Can Time-Series Foundation Models Serve as Simulators for Process Control?
Shihao Tu, Wenyue Ding, Peixuan Ding, Zhaoxiang Wu, Huiyang Wang, Xu Zhang, Yuxin Chen, Yicheng Lu, Xiang Shao, Yangjie Zheng
This study evaluates whether time-series foundation models can serve as simulators for process control, examining physical response fidelity, candidate evaluation, and closed-loop performance. On the tested noiseless linear processes, lower forecasting error does not always translate into better control, highlighting the need to jointly evaluate prediction, decision-making, and execution.本研究评估时序基础模型能否作为过程控制的仿真器，从物理响应保真度、候选控制方案评估和闭环控制效果三个层面进行检验。在所测试的无噪声线性过程中，较低的预测误差并不总能带来更好的控制效果，表明面向控制的模型评估需要同时考察预测、决策与执行。
Unified Spatio-Temporal Tokens are Bases for Generalizable Traffic Forecasting
Yujun Chen*, Shihao Tu*, Wenyue Ding, Yicheng Lu, Qingkai Ren, Yangjie Zheng, Yang Yang (*: equal contribution)
STUNet (Spatio-Temporal Unified Network) explicitly encodes spatial structure into unified tokens and integrates them with temporal representations for generalizable traffic forecasting. It tokenizes the adjacency matrix into spatial patches and uses query-aggregate attention to capture upstream–downstream dependencies across traffic networks.STUNet 面向交通流预测中的跨路网泛化问题，把空间结构显式编码为统一表征，再与时间信息结合。模型将邻接矩阵划分成空间 token，并用查询–聚合注意力刻画上下游依赖，从而在不同交通网络上保持较好的预测能力。
ASTNet: Asynchronous Spatio-Temporal Network for Large-Scale Chemical Sensor Forecasting
Shihao Tu, Yang Yang, Wenyue Ding, Yicheng Lu, Qingkai Ren, Yupeng Zhang, Yin Zhang
ASTNet proposes a novel approach for real-time spatiotemporal forecasting in chemical sensor networks, addressing computational latency and complex spatial dependencies. It features asynchronous modeling to reduce latency and dynamic graph fusion to enhance robustness.ASTNet 面向大规模化学传感器网络的实时时空预测：用异步建模降低计算时延，并用动态图融合更好地刻画复杂的空间依赖。
PowerPM: Foundation Model for Power Systems
Shihao Tu*, Yupeng Zhang*, Jing Zhang, Zhendong Fu, Yin Zhang, Yang Yang (*: equal contribution)
PowerPM is an advanced model for electricity time series (ETS) analysis, designed to address the complexities of hierarchical and temporal data for power systems applications, focusing on electricity usage forecasting, grid stability and consumer behaviour analysis.PowerPM 是面向电力系统电量时间序列的基础模型，同时建模层次结构与时间依赖，可用于负荷预测、电网稳定评估和用户行为分析。
DMNet: Self-comparison Driven Model for Subject-independent Seizure Detection
Shihao Tu, Linfeng Cao, Daoze Zhang,
Junru Chen, Lvbin Ma, Yin Zhang, Yang Yang
Difference Matrix-based Neural Network (DMNet) addresses the domain shift in iEEG signals across different subjects by leveraging a self-comparison mechanism for subject-independent automatic seizure detection.DMNet 针对不同患者颅内脑电信号的分布差异，借助自比较机制缓解域偏移，从而实现跨受试者的癫痫发作自动检测。
💼 Work💼 工作
- 2026.01.06 - Present, Algorithm Development Expert, SUPCON, Hangzhou, China.2026.01.06 – 至今，算法开发专家，中控，杭州。