Institution profile

China Meteorological Administration

Academic institutionasia · cn
Official website
Research library10linked papers
Opportunities0open roles
Selected work

Representative Papers

M$^2$Weather: A Benchmark for Joint Multi-Station and Multi-Variable Weather Forecasting

Sep 30, 2026

This study addresses the limitation of existing meteorological forecasting research, which models station-level spatial dependencies and variable-level physical couplings in isolation without a unified benchmark to evaluate their joint contributions. To bridge this gap, we construct a multi-scale, high-quality meteorological benchmark and introduce the first standardized evaluation framework that jointly assesses multi-station and multi-variable modeling. Furthermore, we design lightweight, plug-and-play adapters that efficiently recover missing spatial and inter-variable correlations without requiring model retraining. Extensive experiments across 16 representative models demonstrate that joint modeling significantly reduces prediction errors, confirming that complementary information across stations and variables constitutes a critical resource for enhancing forecasting accuracy.

0 citationsRead paper

STCFormer: Adaptive Spatio-Temporal Modeling with Dynamic Cluster Transformer for Station-based Weather Forecasting

Sep 30, 2026

This study addresses the challenges of modeling complex spatial dependencies, the neglect of local dynamics by fixed groupings, and the lack of global context in site-specific meteorological forecasting. To this end, we propose an adaptive spatiotemporal Transformer. The method introduces dynamic station grouping within time slices, integrating cluster-guided attention with regional state summaries to achieve synergistic local-global modeling. Furthermore, InfoLoss is incorporated to optimize interaction efficiency, and a theoretical robustness analysis is provided via Lipschitz bounds. Extensive evaluations across 48 comparative settings—spanning eight tasks on three real-world datasets—demonstrate that our approach ranks among the top two in 47 settings and consistently achieves the lowest 24-hour MSE, thereby validating its effectiveness and superiority.

0 citationsRead paper

Beyond the Beam: Constructive Repair and Candidate Completion for Generative Recommendation

Sep 27, 2026

This study addresses the retrieval failure in generative recommendation caused by catalog expansion, which frequently displaces valid identifiers outside the beam during decoding. Building upon T5 and LC-Rec architectures, this work proposes an integer-stream minimal-replacement repair mechanism that unifies generative likelihood with collaborative evidence for candidate completion and global Top-K certification. Furthermore, it introduces output invariance certificates, exact feasibility interval theory, and a prefix-preserving bound stopping criterion to provide rigorous theoretical guarantees for decoding. Evaluated on Amazon datasets, the proposed method yields substantial improvements of 15.5–46.3% in Recall@10 and 15.2–44.4% in NDCG@10, achieving reliable generative recommendation with both provable theoretical soundness and significant empirical performance gains.

0 citationsRead paper
Recent publications

Latest Papers

M$^2$Weather: A Benchmark for Joint Multi-Station and Multi-Variable Weather Forecasting

Sep 30, 2026

This study addresses the limitation of existing meteorological forecasting research, which models station-level spatial dependencies and variable-level physical couplings in isolation without a unified benchmark to evaluate their joint contributions. To bridge this gap, we construct a multi-scale, high-quality meteorological benchmark and introduce the first standardized evaluation framework that jointly assesses multi-station and multi-variable modeling. Furthermore, we design lightweight, plug-and-play adapters that efficiently recover missing spatial and inter-variable correlations without requiring model retraining. Extensive experiments across 16 representative models demonstrate that joint modeling significantly reduces prediction errors, confirming that complementary information across stations and variables constitutes a critical resource for enhancing forecasting accuracy.

0 citationsRead paper

STCFormer: Adaptive Spatio-Temporal Modeling with Dynamic Cluster Transformer for Station-based Weather Forecasting

Sep 30, 2026

This study addresses the challenges of modeling complex spatial dependencies, the neglect of local dynamics by fixed groupings, and the lack of global context in site-specific meteorological forecasting. To this end, we propose an adaptive spatiotemporal Transformer. The method introduces dynamic station grouping within time slices, integrating cluster-guided attention with regional state summaries to achieve synergistic local-global modeling. Furthermore, InfoLoss is incorporated to optimize interaction efficiency, and a theoretical robustness analysis is provided via Lipschitz bounds. Extensive evaluations across 48 comparative settings—spanning eight tasks on three real-world datasets—demonstrate that our approach ranks among the top two in 47 settings and consistently achieves the lowest 24-hour MSE, thereby validating its effectiveness and superiority.

0 citationsRead paper

Beyond the Beam: Constructive Repair and Candidate Completion for Generative Recommendation

Sep 27, 2026

This study addresses the retrieval failure in generative recommendation caused by catalog expansion, which frequently displaces valid identifiers outside the beam during decoding. Building upon T5 and LC-Rec architectures, this work proposes an integer-stream minimal-replacement repair mechanism that unifies generative likelihood with collaborative evidence for candidate completion and global Top-K certification. Furthermore, it introduces output invariance certificates, exact feasibility interval theory, and a prefix-preserving bound stopping criterion to provide rigorous theoretical guarantees for decoding. Evaluated on Amazon datasets, the proposed method yields substantial improvements of 15.5–46.3% in Recall@10 and 15.2–44.4% in NDCG@10, achieving reliable generative recommendation with both provable theoretical soundness and significant empirical performance gains.

0 citationsRead paper