Scholarlink

<p>Weather forecasting and spatio-temporal data mining share a fundamental objective: discovering the temporal dynamics and spatial interactions within complex observational data. Despite recent advances, two primary challenges remain to be addressed. (1) Data distribution imbalances: due to heterogeneous observational infrastructures across geographic regions, some areas suffer from sparse or incomplete data. (2) Lack of interpretability and generalization: many existing models are overly specialized and lack causal reasoning mechanisms, which limits their adaptability to unseen conditions. In this study, we propose a novel plug-and-play framework termed Spatio-Temporal Predictive Coding (STPC), which incorporates causal-inspired discovery and contrastive pretraining to simultaneously enhance both model structure and input data. STPC can be seamlessly integrated into a variety of downstream forecasting backbones to improve performance across weather-related scenarios. It comprises two stages: causal-inspired discovery and model-data co-update. In the first stage, we construct a self-supervised learning-based pretraining module using Vision Transformers to identify causal associations in spatio-temporal data. In the second stage, we freeze the parameters of the causal-inspired discovery module and use a generative model (VQ-VAE) to augment the environmental components of the data. The combined original and augmented data are then fed into downstream models to improve predictive robustness and generalization. We conduct extensive experiments on four real-world spatio-temporal benchmarks, including WeatherBench and SEVIR, and the results consistently demonstrate the effectiveness and versatility of our method across multiple weather forecasting tasks. Our code is available at <a href="https://github.com/anshen666/STPC">https://github.com/anshen666/STPC</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Spatio-temporal predictive coding for weather forecasting and Governance: a plug-and-play module for causal discovery and data augmentation

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • An Zhang,
  • Lilan Peng,
  • Sheng Chen,
  • Pengcheng Wu

摘要

Weather forecasting and spatio-temporal data mining share a fundamental objective: discovering the temporal dynamics and spatial interactions within complex observational data. Despite recent advances, two primary challenges remain to be addressed. (1) Data distribution imbalances: due to heterogeneous observational infrastructures across geographic regions, some areas suffer from sparse or incomplete data. (2) Lack of interpretability and generalization: many existing models are overly specialized and lack causal reasoning mechanisms, which limits their adaptability to unseen conditions. In this study, we propose a novel plug-and-play framework termed Spatio-Temporal Predictive Coding (STPC), which incorporates causal-inspired discovery and contrastive pretraining to simultaneously enhance both model structure and input data. STPC can be seamlessly integrated into a variety of downstream forecasting backbones to improve performance across weather-related scenarios. It comprises two stages: causal-inspired discovery and model-data co-update. In the first stage, we construct a self-supervised learning-based pretraining module using Vision Transformers to identify causal associations in spatio-temporal data. In the second stage, we freeze the parameters of the causal-inspired discovery module and use a generative model (VQ-VAE) to augment the environmental components of the data. The combined original and augmented data are then fed into downstream models to improve predictive robustness and generalization. We conduct extensive experiments on four real-world spatio-temporal benchmarks, including WeatherBench and SEVIR, and the results consistently demonstrate the effectiveness and versatility of our method across multiple weather forecasting tasks. Our code is available at https://github.com/anshen666/STPC.