Abstract:Weather forecasting requires not only accuracy but also the ability to perform probabilistic prediction. However, deterministic weather forecasting methods do not support probabilistic predictions, and conversely, probabilistic models tend to be less accurate. To address these challenges, in this paper, we introduce the \textbf{\textit{D}}eterministic \textbf{\textit{G}}uidance \textbf{\textit{D}}iffusion \textbf{\textit{M}}odel (DGDM) for probabilistic weather forecasting, integrating benefits of both deterministic and probabilistic approaches. During the forward process, both the deterministic and probabilistic models are trained end-to-end. In the reverse process, weather forecasting leverages the predicted result from the deterministic model, using as an intermediate starting point for the probabilistic model. By fusing deterministic models with probabilistic models in this manner, DGDM is capable of providing accurate forecasts while also offering probabilistic predictions. To evaluate DGDM, we assess it on the global weather forecasting dataset (WeatherBench) and the common video frame prediction benchmark (Moving MNIST). We also introduce and evaluate the Pacific Northwest Windstorm (PNW)-Typhoon weather satellite dataset to verify the effectiveness of DGDM in high-resolution regional forecasting. As a result of our experiments, DGDM achieves state-of-the-art results not only in global forecasting but also in regional forecasting. The code is available at: \url{<a class="link-external link-https" href="https://github.com/DongGeun-Yoon/DGDM" rel="external noopener nofollow">this https URL</a>}.

What problem does this paper attempt to address?

The paper aims to address two main issues in weather forecasting: accuracy and the ability to make probabilistic predictions. Traditionally, deterministic weather forecasting methods can provide relatively precise forecast results but cannot make probabilistic predictions, which means they struggle to capture the uncertainty and randomness in atmospheric processes. On the other hand, probabilistic models can offer a variety of possible future scenarios to reflect uncertainty but often fall short in terms of accuracy compared to deterministic models. To resolve this contradiction, the paper proposes a new framework called the "Deterministic Guidance Diffusion Model" (DGDM), which combines the advantages of both deterministic and probabilistic methods. Specifically, DGDM consists of two branches: the Deterministic Branch and the Probabilistic Branch. The Deterministic Branch is responsible for improving forecast accuracy, while the Probabilistic Branch accounts for the inherent uncertainty in weather forecasting, ensuring the prediction of a range of possible future scenarios. During training, these two branches are jointly trained end-to-end. In the inference phase, the forecast results from the Deterministic Branch are used as an intermediate starting point for the reverse process of the Probabilistic Branch. This method adjusts the diversity of probabilistic forecasts, thereby controlling the range of possible future weather scenarios. This approach not only improves forecast accuracy but also maintains the ability to estimate the diversity of possible future scenarios. To validate the effectiveness of DGDM, the authors conducted experimental evaluations on multiple datasets, including the global weather forecasting dataset (WeatherBench), the video frame prediction benchmark dataset (Moving MNIST), and introduced a new high-resolution regional forecasting dataset (Pacific Northwest Windstorm (PNW)-Typhoon weather satellite dataset). The experimental results show that DGDM not only achieves state-of-the-art performance in global low-resolution weather forecasting but also excels in high-resolution regional weather forecasting and performs well in general video frame prediction tasks such as Moving MNIST.

Deterministic Guidance Diffusion Model for Probabilistic Weather Forecasting

Weather Prediction with Diffusion Guided by Realistic Forecast Processes

Generative emulation of weather forecast ensembles with diffusion models

CogDPM: Diffusion Probabilistic Models via Cognitive Predictive Coding

A Multi-model Combination Approach for Probabilistic Wind Power Forecasting.

SEEDS: Emulation of Weather Forecast Ensembles with Diffusion Models

CoDiCast: Conditional Diffusion Model for Weather Prediction with Uncertainty Quantification

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

Probabilistic and deterministic wind speed forecasting based on non-parametric approaches and wind characteristics information

Nonparametric Stochastic Differential Equations for Ultra-Short-Term Probabilistic Forecasting of Wind Power Generation

Diff-MGR: Dynamic Causal Graph Attention and Pattern Reproduction Guided Diffusion Model for Multivariate Time Series Probabilistic Forecasting

Deterministic and Probabilistic Wind Power Forecasting Using a Variational Bayesian-based Adaptive Robust Multi-Kernel Regression Model

A novel wind power forecast diffusion model based on prior knowledge

DYffusion: A Dynamics-informed Diffusion Model for Spatiotemporal Forecasting

Predict, Refine, Synthesize: Self-Guiding Diffusion Models for Probabilistic Time Series Forecasting

Adaptive Weighted Combination Approach for Wind Power Forecast Based on Deep Deterministic Policy Gradient Method

On the generation of probabilistic forecasts from deterministic models

Reliable precipitation nowcasting using probabilistic diffusion models

Residual Corrective Diffusion Modeling for Km-scale Atmospheric Downscaling

Diffusion Models for High-Resolution Solar Forecasts

PreDiff: Precipitation Nowcasting with Latent Diffusion Models