Machine learning model to accurately estimate the planetary boundary layer height of Beijing urban area with ERA5 data

Abstract

The planetary boundary layer height (PBLH) is one of the most important parameters in the environmental, weather, and climatic research. Therefore, it is of great significance and application value to accurately estimate the PBLH by using the available conventional reanalysis meteorological datasets. This study established a hybrid machine learning (HML) method by combining three different ensemble algorithms with parallel training to estimate the urban PBLH of Beijing with ERA5 reanalysis data. The Mean Absolute Percentage Error (MAPE) between the ERA5-BLH and the observed BLH is as high as 68%, whereas it decreased to 10% for the PBLH estimated by the HML model. The radiation-related thermal factors of surface temperature and dew point temperature in ERA5, play a critical role in summer in regulating the PBLH, while the dynamic factors of wind speed and pressure dominate in the other seasons. The MAPE for all the seasons decreases by 0.3–1.6% after introducing the measured temperature and humidity profiles. Shapley Additive Explanations (SHAP) method shows that higher RH contributes less for PBLH estimated in spring, while 2 m temperature is positively correlated with HML performance in summer. Finally, an optimal HML model of PBLH is synthesized as the contribution of meteorological elements. The MAPE drops to 5.2%–15.1% throughout the year. The Mean Absolute Error (MAE) is <50 m in autumn and winter, and the maximum MAE is up to 80 m in the summer afternoon, when the convection is intensely developed. Therefore, the HML is capable of accurately estimating the urban PBLH in high resolution, which provides great significances and references for the investigations regarding the atmospheric environment capacity, as well as for advancing weather forecasting.