BLCM: a BP-LGBM-based atmospheric visibility forecasting model

Abstract

The atmospheric visibility is not only related to environmental quality and public health, but also has a significantly impact on industries such as navigation and aviation. The conventional Numerical weather prediction (NWP) method is run by a supercomputer with high computational cost. On the other hand, due to the inhomogeneity of the visibility distribution, most of machine learning models always analyze and predict visibility on a seasonal basis. To address these issues, we propose a visibility prediction model called BP-LGBM Combination Method (BLCM), which combines the Back Propagation (BP) neural network and the Light Gradient Boosting Machine (LGBM) classifier. By leveraging the advantages of regression and classification algorithms, this model achieves high accuracy predictions of visibility values while significantly reducing computation costs. Meanwhile, in order to resolve the seasonal issue, the data decision filtering process was proposed. It can output different categories of visibility prediction in any season, which expands the applicability of visibility forecasting to any period throughout the year. We also designed a visual analysis system for domain scientists to interactively explore the prediction results and their causes. Finally, the effectiveness of the proposed method has been demonstrated through several ablation experiments, contrast experiments and case studies.

Graphical abstract