cs.LGApr 11, 2026

A Diffusion-Contrastive Graph Neural Network with Virtual Nodes for Wind Nowcasting in Unobserved Regions

Authors: Jie Shi, Siamak Mehrkanoon

Organizations: Department of Information and Computing Sciences, Utrecht University, Utrecht, The Netherlands

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Accurate weather nowcasting remains one of the central challenges in atmospheric science, with critical implications for climate resilience, energy security, and disaster preparedness. Since it is not feasible to deploy observation stations everywhere, some regions lack dense observational networks, resulting in unreliable short-term wind predictions across those unobserved areas. Here we present a deep graph self-supervised framework that extends nowcasting capability into such unobserved regions without requiring new sensors. Our approach introduces "virtual nodes" into a diffusion and contrastive-based graph neural network, enabling the model to learn wind condition (i.e., speed, direction and gusts) in places with no direct measurements. Using high-temporal resolution weather station data across the Netherlands, we demonstrate that this approach reduces nowcast mean absolute error (MAE) of wind speed, gusts, and direction in unobserved regions by more than 30% - 46% compared with interpolation and regression methods. By enabling localized nowcasts where no measurements exist, this method opens new pathways for renewable energy integration, agricultural planning, and early-warning systems in data-sparse regions.

Figures & tables

Figure 1: Overview of meteorological station locations and the spatiotemporal coverage of training and testing datasets across the Netherlands.
VariableVariable NameUnits
ddWind direction (10-min mean)degree
ffWind speed at 10 m (10-min mean)m s -1
gffWind gust at 10 m (10-min max)m s -1
taAir temperature (1-min mean)°C
rhRelative humidity (1-min mean)%
ppAir pressure at sea level (1-min mean)hPa
Table 1: List of meteorological variables used in this study
Figure 2: Overview of the model architecture and input graph generation process. Step A: Graph construction with real (blue) and virtual (red) nodes linked by diffusion-weighted edges. Step B: Model architecture combining supervised wind prediction on real nodes and self-supervised contrastive learning on virtual nodes.
ModelDiffusionContrastiveMoCoSelf-supervised
ContraVirt (Augmented MoCo)✓✓✓✓
ContraVirt (Multi-step MoCo)✓✓✓✓
Augmented✓✓X✓
Multi-step✓✓X✓
w/o Contrastive✓XX✓
w/o Contrastive & DiffusionXXXX
Table 2: Ablation and baseline model comparison. Each column denotes whether the corresponding component is included in the model.
Figure 3: Graph diffusion and virtual node dependence. (a) Diffusion process transforming the spatial graph structure and variable distribution across the Netherlands. (b) Influence distribution showing how virtual nodes aggregate information from multiple nearby real stations.
ModelMAERMSE
DirectionSpeedGustDirectionSpeedGust
ContraVirt (Augmented MoCo)29.981.481.7545.161.852.25
ContraVirt (Multi-step MoCo)29.491.521.7944.521.912.29
Augmented31.061.621.8945.942.022.41
Multi-step31.451.681.9746.452.082.50
w/o Contrastive33.411.541.8848.371.922.41
Table 3: Forecasting performance (MAE and RMSE) across models.Models are evaluated on wind direction (°), wind speed, and wind gust (m/s).
Figure 4: Model performance across metrics, lead times, and test stations. (a) MAE/RMSE across models and targets. (b) MAE in different lead time ( t+1t{+}1 – t+6t{+}6 ). (c) Spatial distribution of MAE across the eight testing stations (10 variants). Our methods achieve lower short-lead errors and degrade more slowly, especially Augmented MoCo.
Figure 5: Seasonal MAE results. (a) Lead–time MAE for wind direction, wind speed, and wind gust. (b) Station-wise MAE patterns and seasonal variations across the Netherlands.
Figure 6: Case studies of short-horizon wind forecasts. (a) Wind-direction roses at horizon t+1t{+}1 for four representative test stations (6283, 6242, 6340, 6391), comparing ground truth (black) with six models during a storm-prone winter month in the Netherlands. (b,c) Time–series comparisons of wind speed and wind gust at the same stations, showing close alignment between predictions and observations.
Figure 7: Embedding distance dynamics and projections. (a,b) Positive and negative pair-distance evolution for augmented and multi-step MoCo. (c,d) t-SNE and UMAP projections showing progressively tighter contrastive embeddings.
More in this topic

Weather Forecasting

Weather forecasting predicts future atmospheric conditions, from local storms to global patterns, using physical models, observations, and machine learning.

See all 176 papers

Explore similar work

CardsList

  1. Zhinv: Real-time hub-height wind field reconstruction using only local sparse observations

    Jul 28, 2026Zongwei Zhang, Chin Chun Ooi, Lianlei Lin +8Sparse Recovery

    School of Electronic and Information Engineering, Harbin Institute of Technology, Harbin, 150001, Heilongjiang, China · Centre for Frontier AI Research, Agency for Science, Technology and Research (A*STAR), Singapore 138632, Singapore · Technological Innovation Center of Littoral Test, China +2

  2. GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance

    Jan 16, 2026Francisco Giral, Álvaro Manzano, Ignacio Gómez +2Diffusion Model GuidanceData Assimilation

    Applied Mathematics Department, ETSIAE-School of Aeronautics, Universidad Politécnica de Madrid, Spain · Department of Aerospace Engineering, University of Michigan, Ann Arbor, MI 48109, USA.

  3. When Does Advection-Aware Graph Nowcasting Help? A Controlled Study of Distributed Solar Ramp Forecasting with a Self-Supervised Cloud-Motion Estimator

    Sep 8, 2026Phillip JiangSolar Power ForecastingSolar Irradiance Forecasting

    Appsofa LLC