Unveiling the Correlation Between Lyapunov Coefficients and Deep Learning Performance Using Ceilometer Data

The planetary boundary layer (PBL) is a crucial parameter to investigate for characterizing the atmosphere, particularly concerning aerosol concentrations. Understanding the PBL allows us to estimate air quality, provide weather forecasts, and establish correlations with astronomical seeing conditions and atmospheric turbulence intensity. The PBL can be defined in many ways, but its importance remains constant as it is the atmospheric layer where most socioeconomic activities occur. In this letter, we present a method to predict the stochastic PBL height (SPBLH) using ceilometer data and a deep learning approach based on a fully connected neural network (NN). We found a correlation between the Lyapunov coefficient calculated for each SPBLH time series and the loss function, which is influenced by various factors such as atmospheric parameters, pollution, aerosols, and more. The performance of a typical NN used to predict a time series is significantly affected by the degree of chaos, quantified by the largest Lyapunov exponents (LLEs). Our results show a decrease in accuracy as a function of increasing LLE. Moreover, an increased number of virtual neurons in the NN can be detrimental to SPBLH prediction for the complex dynamics of the PBL due to atmospheric conditions and unforeseen events.