Generative Modeling of High-resolution Global Precipitation Forecasts James Duncan

2025-05-06 0 0 3.41MB 9 页 10玖币

侵权投诉

Generative Modeling of High-resolution Global

Precipitation Forecasts

James Duncan

UC Berkeley

Berkeley, CA 94720

jpduncan@berkeley.edu

Shashank Subramanian

Lawrence Berkeley National Laboratory

Berkeley, CA 94720

shashanksubramanian@lbl.gov

Peter Harrington

Lawrence Berkeley National Laboratory

Berkeley, CA 94720

pharrington@lbl.gov

Abstract

Forecasting global precipitation patterns and, in particular, extreme precipitation

events is of critical importance to preparing for and adapting to climate change.

Making accurate high-resolution precipitation forecasts using traditional physical

models remains a major challenge in operational weather forecasting as they incur

substantial computational costs and struggle to achieve sufﬁcient forecast skill.

Recently, deep-learning-based models have shown great promise in closing the gap

with numerical weather prediction (NWP) models in terms of precipitation forecast

skill, opening up exciting new avenues for precipitation modeling. However, it is

challenging for these deep learning models to fully resolve the ﬁne-scale structures

of precipitation phenomena and adequately characterize the extremes of the long-

tailed precipitation distribution. In this work, we present several improvements

to the architecture and training process of a current state-of-the art deep learning

precipitation model (FourCastNet) using a novel generative adversarial network

(GAN) to better capture ﬁne scales and extremes. Our improvements achieve

superior performance in capturing the extreme percentiles of global precipitation,

while comparable to state-of-the-art NWP models in terms of forecast skill at 1–2

day lead times. Together, these improvements set a new state-of-the-art in global

precipitation forecasting.

1 Introduction

Precipitation is a fundamental climate phenomenon with major impact on crucial infrastructure such

as food, water, energy, transportation, and health systems [

]. On timescales relevant to weather

prediction, making accurate forecasts of precipitation and extreme events is critical to the planning

and management of these systems, along with disaster preparedness for extreme precipitation events,

which are greatly ampliﬁed by climate change [

]. Unfortunately, extreme precipitation remains one

of the most challenging atmospheric phenomena to forecast accurately, due to high spatiotemporal

variability and the myriad of complex multi-scale, multi-phase processes that govern its behavior

[2, 20, 22].

State-of-the-art NWP models, such as the Integrated Forecast System

(IFS), produce operational

forecasts by combining physics-based PDE solvers with assimilated observations from a variety

1https://www.ecmwf.int/en/forecasts/documentation-and-support

Preprint. Under review.

arXiv:2210.12504v1 [cs.LG] 22 Oct 2022

of sources. The complexity of moisture physics requires many parameterizations in these models

for processes like turbulent mixing, convection, subgrid clouds and microphysics [

], and such

parameterizations can lead to large biases in NWP precipitation forecasts [

]. As a result, global

precipitation forecasts generally achieve inadequate forecast skill [

] and, hence, there has been an

increasing interest in fully data-driven solutions, primarily using deep learning, in recent years.

Data-driven models can be orders of magnitude faster with the potential to learn complex parame-

terizations between input and output function spaces directly from data, reducing model bias. With

such models, major advances have been made in the area of precipitation “nowcasting”, where

forecasts are made over limited spatial regions with lead times on the order of minutes to hours. Deep

learning models trained directly on radar and/or satellite observations now outperform traditional

methods for nowcasting [

]. However, until recently, there has been limited progress for

models predicting precipitation at larger spatiotemporal scales (e.g., over the full globe up to days in

advance), mainly due to computational limitations on resolution [

]. FourCastNet [

] is the ﬁrst

deep-learning-based global weather model running at

∼

30km scale, which outperforms IFS in terms

of precipitation forecast skill up to

∼

2 day lead times and is the current state-of-the-art. However,

despite using a dedicated network just for precipitation due to its unique challenges, FourCastNet

predictions still lack ﬁne-scale details (see Figure 1) and thus underestimate the extreme percentiles

of the precipitation distribution.

In this work, we aim to overcome some of these limitations using generative models to advance the

state-of-the-art in deep learning-based precipitation forecasts. In particular, our contributions are as

follows: (i) we apply a state-of-the-art generative adversarial network [

] that integrates multi-scale

semantic structure and style information, allowing us to synthesize physically realistic ﬁne-scale

precipitation features; (ii) we show that capturing ﬁne-scale phenomena leads to improved predictions

of extreme precipitation while also preserving forecast skill, attaining comparable skill at 1–2 lead

day times with respect to IFS.

2 Methods

2.1 Dataset

We replicate the data preparation pipeline of the original FourCastNet precipitation model [

relying on the European Center for Medium-Range Weather Forecasting (ECMWF) global reanalysis

dataset ERA5 [

], which combines archived atmospheric observations with physical model outputs.

Following FourCastNet’s time steps of length 6 hours, we model the 6hr accumulated total precipita-

tion

T P

(this also makes for easier comparisons against IFS, which archives

T P

forecasts in 6 hourly

accumulations as well). A sample snapshot of

T P

, in log-normalized units for easy visualization of

ﬁne-scale details, is shown in the bottom-right inset of Figure 1. We use the years 1979-2015 and

2016-2017 as training and validation sets, respectively, and set aside 2018 as a test set. We refer the

reader to the original FourCastNet paper [14] for further details on the dataset.

2.2 Model

Given the success of adversarial learning for high-resolution precipitation models in localized regions

[

], we explore the utility of conditional generative adversarial networks (cGANs) for

modeling

T P

over the entire globe using prognostic atmospheric variables as conditional input

for operational diagnosis. In particular, we adopt the TSIT architecture [

] for our task, due to its

success in varied image tasks and ability to ﬂexibly condition intermediate features at a variety of

scales in the generator. TSIT’s generator

employs symmetric downsampling and upsampling

paths, fusing features from the downsampling path via feature-wise adaptive denormalization, and

uses multi-scale patch-based discriminators [

] for adversarial training. We take a constant latitude

embedding channel and the 20 atmospheric variables output by FourCastNet (corresponding to various

different prognostic variables such as wind velocities, total column water vapor and temperature at

different pressure levels) as input to the network, as shown in Figure 2. Randomness is injected at

intermediate scales via elementwise additive noise, and we can thus generate “zero-shot” ensembles

for probabilistic forecasting given a single input, which we explore in Appendix A. We refer the reader

to [

] for additional details on the TSIT architecture, and list hyperparameters in Appendix A, together

with the particulars of our setup. In this work, we focus on three model versions for demonstrating

the impact of the adversarial training: (i) FourCastNet: the precipitation baseline model from [

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

GenerativeModelingofHigh-resolutionGlobalPrecipitationForecastsJamesDuncanUCBerkeleyBerkeley,CA94720jpduncan@berkeley.eduShashankSubramanianLawrenceBerkeleyNationalLaboratoryBerkeley,CA94720shashanksubramanian@lbl.govPeterHarringtonLawrenceBerkeleyNationalLaboratoryBerkeley,CA94720pharrington@lbl.go...

展开>> 收起<<

Generative Modeling of High-resolution Global Precipitation Forecasts James Duncan.pdf

共9页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Generative Modeling of High-resolution Global Precipitation Forecasts James Duncan

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: