Page 85 - 《高原气象》2026年第2期

2 期 黄嘉雯等：基于机器学习的青海湖水位变化模拟研究 385

Simulation of Qinghai Lake Water Level Fluctuations

Using Machine Learning

HUANG Jiawen， LONG Yinping， MA Qimin， XU Weixin， BIAN Yuxia，
CHEN Hui， TAN Xiwen， LI Suowu
（Chengdu University of Information Technology， Chengdu 610225， Sichuan， China）

Abstract： A comprehensive analysis was conducted to examine the processes and impacts of water level varia‐
tions in Qinghai Lake under changing climatic conditions. The study utilized monthly mean water level data from
1959 to 2017， sourced from the Qinghai Lake Basin， in conjunction with meteorological and climate variables
derived from the ERA5 reanalysis dataset developed by the European Centre for Medium-Range Weather Fore‐
casts （ECMWF）. Several large-scale atmospheric circulation indices were also incorporated to investigate their
influence on lake dynamics. This integrated dataset enabled a systematic assessment of the dominant climatic
drivers and facilitated the development of predictive models to simulate future water level changes. To identify
the most relevant influencing factors， the Random Forest （RF） algorithm was employed to perform feature selec‐
tion and importance ranking. This process allowed for an evaluation of the relationship between feature relevance
and model performance. Subsequently， a comparative analysis was undertaken using five machine learning mod‐
els： RF， Support Vector Machine （SVM）， Multi-layer Perceptron （MLP）， Long Short-Term Memory （LSTM）
networks， and Multiple Linear Regression （MLR）. The models were trained and validated to simulate monthly
water level fluctuations and to assess the influence of model complexity and temporal learning ability on predic‐
tive accuracy. The analysis revealed that key drivers of Qinghai Lake water levels include the North Atlantic Os‐
cillation （NAO）， Atlantic Multidecadal Oscillation （AMO）， Niño 3. 4 index， relative humidity at 400 hPa， 450
hPa， and 100 hPa （RH400， RH450， RH100）， precipitation， temperature at 1000 hPa （T1000）， vertical wind
velocity at 1000 hPa （W1000）， and longwave radiation （LW）. Among the models tested， the LSTM network ex‐
hibited superior performance due to its ability to capture complex nonlinear and sequential dependencies in the
data. When the ten most significant features were selected， the LSTM model achieved a Pearson correlation coef‐
ficient （R） of 0. 95， Nash-Sutcliffe Efficiency （NSE） of 0. 96， Normalized Root Mean Square Error （NRMSE）
of 0. 14， and Kling-Gupta Efficiency （KGE） of 0. 87. The MLP model demonstrated the second-best perfor‐
mance， while RF and SVM yielded comparable but slightly lower results. MLR performed the worst， reflecting
its limitations in modeling nonlinear and temporal relationships. Projections based on the LSTM model indicate
that the water level of Qinghai Lake is likely to rise by approximately 2. 55 m between 2017 and 2030. This antic‐
ipated increase reflects the continuing influence of climate change and underscores the importance of adaptive
water resource management strategies in plateau lake regions. The findings offer a reliable methodological frame‐
work for modeling and forecasting hydrological changes in alpine lake systems under future climate scenarios.
Key words： atmospheric circulation； water level variation； Qinghai Lake Basin； machine learning； correlation
analysis