Research on Low Visibility Forecast Model of Sea Fog in Beibu Gulf Based on Attention Mechanism-Embedded LSTM Deep Learning

To address the complexities associated with forecasting low-probability, low-visibility fog events and the underlying nonlinear interdependencies among various influencing variables, we present an attention mechanism-embedded long short-term memory (ATT-LSTM) deep learning model for sea fog visibility hazard prediction. This architecture seamlessly incorporates ATT into the conventional LSTM neural network framework. This integration enables the model to adaptively assign weights to the input features, thereby distinguishing between salient and non-salient variables. This targeted allocation enhances the contribution of considerable factors within the LSTM forecasting algorithm, optimizes input data, and assigns varying levels of attention to each variable. Consequently, the model substantially mitigates prediction errors in multivariate scenarios. An empirical analysis employing an independent dataset encompassing 303 foggy days over a biennial period confirmed the superior performance of the proposed ATT-LSTM model. Comparative evaluations with LSTM, logistic classification regression, and support vector machine classification regression models revealed that the ATT-LSTM model achieved a recall rate of 37%, a precision rate of 48%, an accuracy rate of 91%, and a threat score (TS) of 0.26. Among the assessed methodologies, the ATT-LSTM model outperformed the others in terms of recall, accuracy, and TS metrics. These findings confirm that the ATT-LSTM model offers a potent and innovative deep learning approach for enhancing the accuracy of low-visibility sea fog hazard predictions.