Abstract: Insufficient consideration of the spatial heterogeneity of wind speed in existing methods for fog visibility prediction leads to low accuracy and stability. To address this issue, this paper presents a long short-term memory neural network with an attention mechanism (LSTM-AM) model for fog visibility prediction that considers the spatial heterogeneity of wind speed. This model quantifies the variation characteristics of wind speed at different spatial locations using a semi-variogram, integrating spatial distribution of neighboring points and differences in wind speed. It employs wind direction angles and variation values to weigh and effectively extract the features of wind speed spatial heterogeneity. In addition, the attention mechanism enhances the LSTM model by improving its focus on key information, enabling it to effectively capture and reflect the impact of critical meteorological factors on fog visibility. This enhances the model’s ability to consider important temporal information and improves accuracy under wind speed spatial heterogeneity. The results indicate that the proposed model improves R² by 10%–20%, reduces root mean square error by 25%–40%, and decreases mean absolute error by 26.3%–39.1%, demonstrating high accuracy and stability in fog visibility prediction.