Wangzhi Yu 余望之
Ph.D. Student · College of Computer Science and Technology
Tongji University, Shanghai, China · Advised by Prof. Dawei Cheng
About
I am a Ph.D. student in the College of Computer Science and Technology at Tongji University, advised by Prof. Dawei Cheng. My research focuses on time series forecasting and spatio-temporal data mining, with applications to electricity load forecasting and traffic flow prediction.
My recent work studies peak-aware electricity load forecasting and the use of large language models for time series analysis.
Research Interests
Time Series Forecasting Spatio-temporal Data Mining Electricity Load Forecasting Graph Neural Networks LLMs for Time Series
Publications
Underline marks me; ★ marks advisor. Ranks follow the CCF recommendation list.
-
2026Large Language Models for Time Series Analysis: Methodologies, Applications, and Emerging Challenges
TechRxiv preprint
Abstract
Large language models (LLMs), pre-trained on massive corpora, have driven remarkable breakthroughs in natural language processing and computer vision. Building upon these successes, recent studies have examined the potential of pre-trained LLMs for time series analysis (TSA), focusing on downstream tasks such as time series forecasting, classification, anomaly detection and imputation across domains including finance, traffic and energy. In this survey, we systematically categorize the roles of LLMs for TSA into three principal types and further abstract their unified workflows: (1) fine-tune-based inference engines, where LLMs serve as core models for TSA tasks, with workflows involving tokenization, prompt engineering, semantic alignment, and fine-tuning; (2) TSA-methods-based enhancers, where LLMs serve as auxiliary modules within TSA methods; (3) hybrid collaborators, representing emerging approaches that integrate LLMs with lightweight models and multi-agent systems to enable advanced collaborative solutions. Then, we systematically review how LLMs empower downstream tasks across diverse domains. Specifically, we organize these applications according to the underlying mechanisms through which LLMs augment domain-specific tasks. This approach provides clear insights into the multifaceted roles of LLMs in advancing TSA across industries. Finally, we analyze the principal challenges faced by LLMs for TSA, including limitations such as cross-domain generalization, privacy preservation, and multi-modal data alignment. By establishing a rigorous foundation, we aim to guide future research efforts and foster methodological innovation within the field.
-
2026PeakFocus: Bridging Peak Localization and Intensity Regression via a Unified Multi-Scale Framework for Electricity Load Forecasting
arXiv preprint arXiv:2605.21550
Abstract
Electricity load peak forecasting (ELPF), simultaneously predicting peak timing and intensity, is a prerequisite for effective grid scheduling and risk management. However, existing methods face three limitations. First, they adopt a two-stage predict-then-locate paradigm, which severs the link between temporal localization and intensity regression. Second, they still struggle with the multi-scale representation conflict, leading to peak misjudgment and timing misalignment. Third, the lack of explicit peak timing context during intensity regression causes intensity smoothing because predictions are dominated by global smoothing trends. To address these limitations, we propose PeakFocus, a unified framework for ELPF. (i) A Unified Peak-Aware Pipeline (UPAP) utilizes a triple hybrid loss to jointly supervise temporal localization and intensity regression, alongside a tolerance-based evaluation protocol. (ii) A Multi-Scale Mixing Peak Locator (MSM-PL) exploits coarse-grained features to mitigate peak misjudgment caused by local fluctuations, and injects them into fine-grained features via a cascade mechanism to resolve timing misalignment. (iii) A Location-Aware Decoder (LAD) injects peak timing context into the intensity regression process, providing explicit guidance to counteract intensity smoothing and improve peak intensity estimation. Extensive experiments on the public Electricity (ELC) dataset and our industrial-scale World Large-scale Electricity Load (WLEL) dataset show that PeakFocus outperforms baselines in both timing precision and intensity estimation.
-
2024Interactive dynamic diffusion graph convolutional network for traffic flow prediction
Information Sciences, vol. 677, 120938
Abstract
Capturing the temporal and spatial features, and the spatiotemporal correlations in traffic networks is the essential task for accurate traffic flow prediction. Although most existing models have explored the temporal features deeply, the periodicity and dynamics of diffusion signals in spatial features are neglected. Besides, most previous methods cannot capture the heterogeneous spatiotemporal correlations of traffic sequences. In order to tackle the above two issues, a novel deep learning-based model called interactive dynamic diffusion graph convolutional network is proposed to realize accurate traffic flow prediction. Firstly, a new periodic and dynamic diffusion graph convolutional network is proposed to extract the periodicity and dynamics of diffusion signals in spatial features. Secondly, a new spatiotemporal interactive dynamic graph generator is proposed to generate a spatiotemporal dynamic graph through a spatiotemporal interactive operation, and it can comprehensively capture the heterogeneous spatiotemporal correlations of traffic sequences. Finally, the proposed model attains an accurate traffic flow prediction through extensive experiments on two real-world traffic flow datasets, which confirm the superiority of the proposed model compared with twelve baseline models.
-
2023GSTC-Unet: A U-shaped multi-scaled spatiotemporal graph convolutional network with channel self-attention mechanism for traffic flow forecasting
Expert Systems with Applications, vol. 232, 120724
Abstract
Accurate forecasting of traffic flows remains a significant challenge owing to its complex spatiotemporal dependencies. Although existing methods capture some spatiotemporal dependencies and stack them in a channel dimension, long-range sequences, implicit patterns, hidden fine-grained features and multi-scaled spatiotemporal dependencies are often ignored, which makes it difficult to represent the spatiotemporal dependencies comprehensively. To overcome these limitations, a novel deep learning model named Graph Spatiotemporal Channel Unet (U-shaped network) is proposed to achieve accurate and reliable traffic flow forecasting. First, a new temporal encoder-decoder module with causal convolution and transposed convolution is proposed, which can efficiently alleviate the gradient explosion/vanishing in capturing long-range sequences during encoding and decoding. Secondly, a new channel self-attention mechanism is proposed, which can efficiently capture the implicit patterns and hidden fine-grained features between channels and enhance the representation ability of the spatiotemporal dependencies. Thirdly, a new U-shaped multi-scaled spatiotemporal graph convolutional network is proposed to effectively capture the multi-scaled spatiotemporal dependencies. Experiments on two real-world datasets show that the proposed model outperforms baseline models and achieves accurate traffic flow forecasting.
Contact
- Email wangzhiy@tongji.edu.cn
- GitHub @Yuwangzhi
- Scholar Wangzhi Yu