Short-Term Prediction of Evaporation Duct Height Using a Bayesian Optimization-Based CNN-LSTM-Attention Model

Evaporation ducts are special electromagnetic wave propagation channels formed by abnormal gradients in atmospheric refractivity caused by water vapor evaporation in the marine near-surface atmospheric layer. It enable over-the-horizon propagation of radar waves and communication signals, making their high-precision prediction crucial. Traditional prediction methods often use reanalysis data with insufficient accuracy and resolution as input, combined with evaporation duct diagnostic models that contain diagnostic errors, to predict evaporation duct height, leading to poor reliability in results. This study proposes a shortterm prediction method based on a Bayesian-optimized CNN-LSTM-Attention framework for evaporation duct height, developed by constructing a deep hybrid neural network architecture. Firstly, the convolutional module extracts spatiotemporal correlation features from multisource meteorological elements. Secondly, an attention mechanism is innovatively introduced to dynamically quantify the influence weights of different feature layers on evaporation duct formation. Then, a Long Short-Term Memory network captures the temporal evolution patterns of meteorological parameters. Finally,Bayesian optimization is used to tune network hyperparameters. This framework proposes a novel approach for short-term evaporation duct height prediction by learning the nonlinear mapping relationship between meteorological elements and evaporation duct height. Validation using measured maritime data demonstrates that the proposed method achieves an evaporation duct height prediction root mean square error of 0.80. Moreover, it significantly reduces reliance on the measurement precision of any single parameter and exhibits enhanced interference resistance and prediction robustness, particularly under non-stationary marine conditions.