Estimation of the atmospheric boundary layer height by means of machine learning techniques using ground-level meteorological data

Aerosols affect air quality, weather and climate through many mechanisms and are dangerous to human health. They are mostly concentrated within the atmospheric boundary layer (ABL) its height is affected by the radiation emitted by the surface, causing turbulence and evolving along the day, influencing the vertical mixing of the air pollutants generated near the surface and therefore, their ground-level concentration from local sources. Lidars have demonstrated their capabilities to study the aerosol vertical distribution and their spatio-temporal evolution can provide very complete information on the ABL dynamics. In this work, machine learning techniques are employed to predict the ABL height. The meteorological variables measured at ground-level are used as features of the algorithm and the ABL height estimated by the STRATfinder algorithm using ceilometer profiles, a small lidar instrument with enhanced characteristics for unassisted continuous operation, are considered the truth in the supervised regression algorithm. The machine learning models allow considering combination of features in the regression algorithm and also allow characterizing the importance of each of the predictors to determine the final result. This property is used to study different boundary layer regimes. The ABL is difficult to study in certain parts of the day due to transitions between atmospheric regimes. In order to improve the performance of the model, each day was divided in four parts (nighttime, morning, daytime and evening). The Madrid ceilometer profile database has been studied for the year 2020, splitting the training datasets for the machine learning algorithm into season and part of the day, and the importance of predictors analyzed. Major influence of tem-perature and relative humidity is found in most of the situations, but also wind velocity in certain circumstances and pressure. The influence of radiation is small, contrary to expected. The main advantage of the proposed method is that MLHs and ABLHs can be retrieved directly from widely available ground-level meteorological data. Future work will focus on more relevant predictors, as latent heat or turbulence.

Automated summary

In this study, machine learning techniques were used to predict the height of the atmospheric boundary layer (ABL). Ground-level meteorological data and lidar instrument readings were used as features for the algorithm. The study found that temperature, relative humidity, and wind velocity are important predictors for ABL height, while radiation has a smaller influence.