Abstract:In recent years, machine learning has established itself as a powerful tool for high-resolution weather forecasting. While most current machine learning models focus on deterministic forecasts, accurately capturing the uncertainty in the chaotic weather system calls for probabilistic modeling. We propose a probabilistic weather forecasting model called Graph-EFM, combining a flexible latent-variable formulation with the successful graph-based forecasting framework. The use of a hierarchical graph construction allows for efficient sampling of spatially coherent forecasts. Requiring only a single forward pass per time step, Graph-EFM allows for fast generation of arbitrarily large ensembles. We experiment with the model on both global and limited area forecasting. Ensemble forecasts from Graph-EFM achieve equivalent or lower errors than comparable deterministic models, with the added benefit of accurately capturing forecast uncertainty.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve two key challenges in high - resolution weather forecasting: 1. **Accurately capturing weather processes at different spatial scales**: Weather systems involve multi - scale phenomena from local to global. How to effectively represent these processes at different scales in the model is an important issue. 2. **Modeling the uncertainty of chaotic weather systems**: Due to the chaotic nature of weather systems, there is a high degree of uncertainty in future weather forecasts. Traditional deterministic models only provide a single prediction result, ignoring various uncertainty factors in the prediction. To solve these problems, the author proposes a probabilistic weather forecasting model based on Hierarchical Graph Neural Networks (HGNN) - Graph - EFM. This model combines a flexible latent variable formulation and a successful graph - based forecasting framework to solve the problems in the following ways: - **Hierarchical graph structure**: By constructing a hierarchical grid - graph structure, the model can effectively capture weather processes at different spatial scales. Each layer of the graph structure processes information at different length scales, thus improving the ability to model multi - scale weather phenomena. - **Probabilistic modeling**: Introduce latent variable \(Z_t\) to capture the uncertainty at each time step, enabling the model to generate multiple possible future weather state samples, forming an ensemble forecast. This not only provides more abundant prediction information but also better reflects the uncertainty in the prediction. Specifically, the Graph - EFM model is implemented through the following steps: 1. **Hierarchical graph construction**: Define a grid - graph \(G_1,\ldots,G_L\) with multiple levels and add connections between adjacent levels. This structure allows information to spread efficiently at different scales and imposes a latent variable distribution at the top level to capture global uncertainty. 2. **Single - step prediction model**: Decompose the weather forecasting problem into a series of single - step prediction problems, and use latent variable \(Z_t\) and conditional variables \(X_{t - 2:t - 1},F_t\) to model the conditional distribution \(p(X_t|X_{t - 2:t - 1},F_t)\) at each time step. 3. **Training objective**: Adopt the idea of Variational Auto - Encoder (VAE), use the variational lower bound (ELBO) as the training objective, and add the Continuous Ranked Probability Score (CRPS) loss term to improve the calibration of the ensemble forecast. Through these methods, Graph - EFM can not only generate high - quality deterministic forecasts but also accurately capture the uncertainty in the forecasts, thus providing users with more valuable weather prediction information.

Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks

Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks

Regional Ocean Forecasting with Hierarchical Graph Neural Networks

Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts

Forecasting Global Weather with Graph Neural Networks

DeepHGNN: Study of Graph Neural Network based Forecasting Methods for Hierarchically Related Multivariate Time Series

HiSTGNN: Hierarchical Spatio-temporal Graph Neural Networks for Weather Forecasting

Hierarchical Joint Graph Learning and Multivariate Time Series Forecasting

Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting

DGFormer: a physics-guided station level weather forecasting model with dynamic spatial-temporal graph neural network

Multi-Information Aggregation and Estrangement HyperGraph Convolutional Networks for Spatiotemporal Weather Forecasting

Multi-modal graph neural networks for localized off-grid weather forecasting

Parallel Multi-Scale Dynamic Graph Neural Network for Multivariate Time Series Forecasting

Ensemble weather forecast post-processing with a flexible probabilistic neural network approach

Learning skillful medium-range global weather forecasting

GraphCast: Learning skillful medium-range global weather forecasting

Hierarchical U-net with re-parameterization technique for spatio-temporal weather forecasting

It is all Connected: A New Graph Formulation for Spatio-Temporal Forecasting

Decomposing weather forecasting into advection and convection with neural networks

Probabilistic hierarchical forecasting with deep Poisson mixtures

Graph Hawkes Neural Network for Forecasting on Temporal Knowledge Graphs