Published post: 1 Feb 2026
Weather forecasting plays an essential role in modern society. Accurate predictions of wind speed are especially important for applications such as renewable energy management, aviation, and climate monitoring.
Traditional machine learning approaches often treat weather data simply as time series, ignoring the spatial relationships between measurement stations. However, wind patterns naturally propagate across geographical regions.
In this work, we explore how Graph Convolutional Networks (GCNs) can model these spatial dependencies to improve wind speed prediction.
Weather stations are distributed across geographical locations. Instead of treating them as independent data sources, we represent them as nodes in a graph.
In this graph:
each node corresponds to a weather station
edges represent relationships between stations
weather variables serve as node features
This graph representation allows the model to learn how atmospheric conditions propagate between nearby regions.
Rather than defining these relationships manually, the model learns the strength of connections between stations directly from data through a learnable adjacency matrix.
Wind speed prediction is inherently a spatio-temporal problem.
Two types of relationships must be modeled:
Spatial relationships
interactions between neighboring weather stations
Temporal relationships
how weather variables evolve over time
To capture both aspects, the proposed model represents the data as a spatio-temporal graph, where nodes are connected not only across space but also across time steps.
The model uses graph spatial convolution to aggregate information from neighboring stations.
Instead of relying on fixed geographic relationships, the network learns the adjacency matrix that determines how strongly different stations influence each other.
During training:
connections between stations are automatically discovered
spatial dependencies between cities are learned from historical weather data
This allows the model to capture realistic atmospheric interactions across regions.
To capture the evolution of weather over time, the model also performs temporal convolution.
Temporal convolution allows the network to learn patterns such as:
gradual wind changes
storm development
delayed interactions between atmospheric variables
By combining spatial and temporal operations, the model learns how wind speed evolves across both space and time.
The architecture combines spatial and temporal processing using spatio-temporal blocks (ST-blocks).
Each block contains:
graph spatial convolution
temporal convolution
batch normalization
ReLU activation
residual connections
Stacking several of these blocks enables the network to learn increasingly complex patterns in the weather data.
The model is evaluated using real-world datasets collected from weather stations in Denmark and the Netherlands.
The task is to predict wind speed several hours into the future for selected cities.
Prediction horizons include:
6–24 hours ahead for Denmark
2–10 hours ahead for the Netherlands
The proposed model — called WeatherGCNet — consistently outperforms several baseline deep learning approaches on these datasets.
One advantage of the model is that the learned adjacency matrix reveals how weather stations influence each other.
Visualizations of the learned graph show meaningful spatial patterns between cities, highlighting how wind dynamics propagate across regions.
This provides both improved prediction accuracy and interpretability.
Weather prediction is fundamentally a spatio-temporal problem.
By modeling weather stations as nodes in a graph and learning spatial relationships directly from data, Graph Convolutional Networks provide a powerful framework for forecasting weather variables such as wind speed.
This work demonstrates how combining graph neural networks with temporal modeling can significantly improve wind speed prediction across geographically distributed weather stations.
Stańczyk, T., & Mehrkanoon, S. Deep Graph Convolutional Networks for Wind Speed Prediction, European Symposium on Artificial Neural Networks (ESANN), 2021.