多模态理解，
从研究到实现。

The rapid proliferation of electric vehicles (EVs) necessitates accurate charging load forecasting to support urban energy infrastructure planning and grid stability. However, existing approaches inadequately capture the complex interplay between charging demand and heterogeneous contextual factors, particularly dynamic meteorological conditions and spatial heterogeneity across stations, limiting their predictive accuracy and robustness under evolving urban environments. To address these challenges, we propose ChargeGPT, a multimodal large language model framework that systematically integrates charging loads, meteorological variables, and station-level geographic attributes for spatiotemporal forecasting. The framework employs STL-based temporal decomposition to disentangle multi-scale patterns, a meteorology-aware Cross-Attention Mixture-of-Experts (CA-MoE) module to adaptively fuse weather factors via interpretable attention mechanisms, and a GPT-based backbone with LoRA fine-tuning for contextual reasoning over unified multimodal tokens. Experiments on real-world data from Shenzhen demonstrate that ChargeGPT reduces forecasting error by 7% compared to the best baseline under full-sample conditions and maintains superior accuracy with limited historical observations, validating its generalization capability for newly constructed stations. The interpretable CA-MoE module further reveals dynamic weather impacts across spatiotemporal contexts, providing actionable insights for proactive capacity allocation and dynamic pricing strategies, thereby advancing sustainable urban mobility systems.

Xiao Luo, Wenyu Yin et al.
摘要来源 · CC BY 4.0

With the continuous development of smart cities and human-oriented transportation, the streets, as a key component of urban public space, have a significant impact on the behavior and urban vitality of residents. Traditional street perception methods based on semantic segmentation face challenges such as limited semantics, insufficient understanding of relationships, and poor task adaptability in open-world environments. Therefore, enhancing the accuracy and fine-grained perception capabilities of street environment perception is particularly important. This study proposes the introduction of Vision-Language Models to overcome the limitations of traditional frameworks. Using the Qwen2.5-VL-72B-Instruct model combined with the MIT Place Pulse 2.0 Dataset, experiments were carried out to verify the superiority of Vision-Language Models in six dimensions of street perception: safe, lively, clean, wealthy, depressing and beautiful. The research results indicate that Vision-Language Models not only enables fine-grained street space perception but also flexibly adapts to various tasks through prompt-based approaches, further enhancing perception accuracy and personalization. Future work can optimize Vision-Language Models performance through low-parameter fine-tuning (e.g., LoRA), improve perception speed, reduce computational consumption, and thus promote its application in urban space perception.

Wenyu Yin, Chonghao Zhang, Hao Yu, Xiao Luo, Jinyi Huang, Jialei Zhang, Shuwen Qi
摘要来源 · CC BY-SA 4.0

该研究构建了融合图像、兴趣点语义和地理信息的多模态检索增强生成框架，以统一嵌入表达多源上下文，并通过结合空间邻近关系与多模态特征的匹配机制回应个性化查询。实验与消融分析评估了各模态在检索准确性、语义表达和推荐效果中的作用。

Chonghao Zhang, Hao Yu, Xiao Luo, Wenyu Yin, Jinyi Huang, Xuanyu Liu, Zhe Liu
查看原文 Abstract ↗

Efficient taxi repositioning is essential to address the spatiotemporal mismatch between urban travel demand and vehicle supply. Traditional approaches often rely on heuristic rules or reinforcement learning, which suffer from limited generalization and poor interpretability. In this paper, we propose an LLM-driven framework for intelligent taxi repositioning that integrates three key components: (1) LLM-based decision reasoning, (2) spatiotemporal clustering for demand pattern abstraction, and (3) experience-driven responsibility assignment for adaptive policy optimization. A city-scale simulator is developed to evaluate the effectiveness of our approach. Experimental results show that our method achieves the highest response rate of 70.59%, the lowest average pickup time of 153.3 seconds, and the highest vehicle utilization rate of 48.94%, outperforming four baselines including clustering and optimization-based methods. Our simulator and code are publicly available at: https://github.com/clover-Saber/TrafficSimulator.

Hao Yu, Chonghao Zhang, Xiao Luo, Wenyu Yin, Zhe Liu, Jinyi Huang, Xuanyu Liu
摘要来源 · CC BY-SA 4.0