兰州大学机构库 >大气科学学院
机器学习提供下边界场改进气候预测的研究
Alternative TitleThe Application of Machine Learning in the Lower Boundary Conditions for Improving Climate Prediction
魏森涛
Subtype硕士
Thesis Advisor王澄海
2023-05-30
Degree Grantor兰州大学
Place of Conferral兰州
Degree Name理学硕士
Degree Discipline大气科学
Keyword机器学习 machine learning 海表面温度 sea surface temperature 土壤湿度 soil moisture 短期气候预测 short-term climate prediction 降水 precipitation
Abstract

由于气候系统的复杂性,气候预测往往是非常困难的。气候模式作为气候预测的重要手段,如何获取模式准确的初边界场是一个基本且具有挑战性的科学问题。大气模式的初边界条件中,下垫面的海表面温度、积雪、海冰、土壤湿度等信息是必不可少的,也是气候变化重要的外部强迫。其中,海表面温度既是大气的下边界条件,其变化直接影响海洋和大气的相互作用,又是在不同时间尺度上对区域和全球气候产生显著影响的外部强迫。因此,获得精确的海表面温度以在数值模式中生成准确的下边界条件对提高短期气候预测效果起着至关重要的作用。研究表明,土壤湿度在气候系统中的作用仅次于海表面温度,在远离海洋的陆地上,它的作用甚至超过海表面温度,是气候预测中的重要因子。因此,本文利用机器学习中的卷积神经网络算法,对季节尺度上的海表面温度、土壤湿度进行预测。在此基础上,将机器学习预测得到的海表面温度尝试作为下边界条件更新到WRF模式中,对2022年夏季短期气候进行预测试验和检验。得到以下初步结论:

(1)就全球海洋平均而言,海温和海盐异常信号可持续6个月或更长时间; 海表温度和深层海洋之间存在着滞后相关关系。海温在 5.02 m、15.08 m、25.16 m、35.28 m、45.45 m和76.55 m深度层上具有显著的年际变化差异,各层海温的平均标准差分别为1.57 K、1.50 K、1.40 K、1.28 K、1.18 K、0.94 K,随着深度加深,海温变化逐渐变小。表层海温的滞后自相关系数最高,滞后一个月的自相关系数达到0.7,随着深度加深,滞后相关系数逐渐减小,表层海温与76 m处海温滞后一个月的相关系数低于0.4,但二者的相关系数滞后12个月仍能通过99%显著性检验。次表层海温异常的持续性在不同月份具有较大差异,海盐异常的持续性不随月份而改变,而海温异常的持续性在春季有较明显的减弱,存在“春季障碍”。

(2)利用海洋不同深度的温度和盐度的相关关系及其持续性,可以通过机器学习,获得表层海温未来的变化信号。卷积神经网络可以有效地提取深层海温和海盐的时空分布特征,以预测全球范围较长时期的海表面温度。以前六个月的深层海温和海盐作为预测因子,预测的未来六个月海表面温度平均偏差约为0~0.8 K,预测偏差在海岸线附近较大,而在远离海岸线的洋面小于0.5 K。同时,利用深层海温和海盐信号构建的卷积神经网络模型能再现海表面温度的主要异常变化特征,如IPO、AMO、IOD和EI Nino/La Nina事件等,这为海温预测提供了新的参考途径。

(3)利用不同深度土壤温度和湿度的相关及其持续性,通过机器学习,获得了土壤湿度未来的变化信号。卷积神经网络可以有效提取土壤温、湿度的时空特征,可以用于预测全球范围较长时间的土壤湿度。用前六个月的0~7 cm、7~28 cm、28~100 cm、100~289 cm共四层土壤温、湿度作为因子预测未来六个月的土壤湿度,其中浅层土壤湿度的偏差小于0.05 ,深层土壤湿度偏差在0.02内。利用土壤温湿度信号建立的卷积神经网络模型可用于预测不同干湿地区的土壤湿度,并能再现其主要异常变化特征,对干旱和湿润区的土壤湿度预测的平均偏差在0.02 内,对湿润区域的预测效果略优于干旱区域的预测效果。

(4)利用机器学习获得的海表未来时刻的温度作为气候模式未来时刻的下边界场,可有效提高模式的预测能力。WRF模式对于全球短期气候的预测能力仍有待提高,模拟的降水空间分布与观测有较大差异,尤其对赤道附近降水模拟的偏差明显。WRF模拟的气温存在明显的偏差,平均偏差最大可达到4.5 K。将机器学习预测的海表面温度作为外强迫输入到WRF模式中,能有效提高短期气候预测的效果,降水量的相对偏差减小到50%以下,气温的绝对偏差基减小到3.5 K范围内。

Other Abstract

Climate prediction is a challenging task due to the complexity of the climate system, and climate models are crucial for accurate predictions. However, obtaining precise initial fields and boundary conditions for these models is a basic yet difficult scientific problem. The initial fields and boundary conditions of atmospheric models require essential information such as sea surface temperature, snow cover, sea ice, and soil moisture on the underlying surface. These external forcings play a significant role in climate change and have a massive impact on regional and global climate on different time scales. Of these factors, sea surface temperature is particularly important as it directly affects the interaction between the ocean and atmosphere, making it a necessary lower boundary condition. Obtaining accurate sea surface temperature is critical for generating precise initial and boundary conditions for numerical models, which in turn improves short-term climate predictions. Soil moisture also plays a vital role in the climate system, ranking second only to sea surface temperature. In some regions far from the ocean, soil moisture's role even exceeds that of sea surface temperature, making it an essential factor in climate prediction. Therefore, this paper uses the convolutional neural network algorithm in machine learning to predict sea surface temperature and soil moisture on a seasonal scale. On this basis, the sea surface temperature and soil temperature predicted by machine learning are updated to the WRF model as the lower boundary conditions and initial conditions, and the prediction test and verification of the short-term summer climate in 2022 are carried out. The preliminary conclusions are as follows:

(1) SST has significant interannual variability at depths of 5.02 m, 15.08 m, 25.16 m, 35.28 m, 45.45 m, and 76.55 m. The mean standard deviation of temperature in each layer is 1.57 K、1.50 K、1.40 K, 1.28 K, 1.18 K, 0.94 K respectively. As the depth increases, the sea temperature changes gradually become smaller. The lagged autocorrelation coefficient of the surface sea temperature is the highest, and the autocorrelation coefficient with a one-month lag reaches 0.7. As the depth deepens, the lagged correlation coefficient gradually decreases. The correlation coefficient between the surface sea temperature and the one-month lag at 76 m is lower than 0.4, but the correlation coefficient between the two can still pass the 99% significance test with a lag of 12 months. The persistence of subsurface sea temperature anomalies varies greatly in different months. The persistence of sea salt anomalies does not change with the month, and the persistence of sea temperature anomalies is significantly weakened in spring, showing a "spring barrier". Overall, the Shanghai temperature and sea-salt anomaly signal can last for 6 months or more in the global ocean.

(2) The convolutional neural network can effectively extract the temporal and spatial distribution characteristics of deep sea temperature and sea salt to predict the global sea surface temperature over a long period of time. The deep sea temperature and sea salt in the past six months are used as predictors, and the average deviation of the predicted sea surface temperature in the next six months is about 0-0.8K. The prediction deviation is larger near the coastline, while it is less than 0.5K in the ocean far away from the coastline. At the same time, the convolutional neural network model constructed using deep sea temperature and sea salt signals can reproduce the main anomalous changes in sea surface temperature, such as IPO, AMO, IOD, and El Niño/La Niña events, providing a new basis for sea temperature prediction.

(3) The convolutional neural network can effectively extract the spatiotemporal characteristics of soil temperature and humidity and can be used to predict the global soil moisture for a long time. Use the soil temperature and humidity of four layers, 0~7 cm, 7~28 cm, 28~100 cm, and 100~289 cm, in the first six months as factors to predict the soil moisture in the next six months. The deviation of shallow soil moisture is less than 0.05 , and the deep soil moisture deviation is within 0.02 . The convolutional neural network model established using soil temperature and humidity signals can be used to predict soil moisture in different dry and wet areas and can reproduce its main abnormal change characteristics. The average deviation of soil moisture prediction in arid and humid areas is within 0.02 , and the prediction effect of the wet area is slightly better than that of the dry area.

(4) The ability of the WRF model to predict global short-term climate still needs improvement, and the simulated spatial distribution of precipitation is quite different from the observation, especially for the simulated precipitation near the equator. The air temperature simulated by WRF has an obvious deviation, and the average deviation can reach up to 4.5 K. The sea surface temperature predicted by machine learningis input into the WRF model as an external forcing, which can effectively improve the short-term climate prediction. The relative deviation of precipitation is reduced to less than 50%, and the absolute deviation of air temperature is reduced to within 3.5 K.

MOST Discipline Catalogue理学 - 大气科学 - 气象学
URL查看原文
Language中文
Other Code262010_220200902081
Document Type学位论文
Identifierhttps://ir.lzu.edu.cn/handle/262010/537248
Collection大气科学学院
Affiliation
兰州大学大气科学学院
Recommended Citation
GB/T 7714
魏森涛. 机器学习提供下边界场改进气候预测的研究[D]. 兰州. 兰州大学,2023.
Files in This Item:
There are no files associated with this item.
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Altmetrics Score
Google Scholar
Similar articles in Google Scholar
[魏森涛]'s Articles
Baidu academic
Similar articles in Baidu academic
[魏森涛]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[魏森涛]'s Articles
Terms of Use
No data!
Social Bookmark/Share
No comment.
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.