Visibility Prediction and Diagnostic Interpretation Based on Comparative Modelling for Sustainable Urban Environmental Management: A Chengdu Study from November 2021 to October 2024

Abstract

Reliable visibility forecasting is important for transportation safety and sustainable urban environmental management, particularly in the Sichuan Basin, where poor visibility remains a persistent concern. This study developed a comparative diagnostic framework for one-day-ahead prediction of continuous daily visibility in Chengdu using meteorological and air-quality observations from November 2021 to October 2024. Grey relational analysis and temporal diagnostics were used to characterize variable associations and temporal dependence. Seven methods—seasonal autoregressive integrated moving average with exogenous variables (SARIMAX), CatBoost, long short-term memory (LSTM), Transformer, convolutional neural network–LSTM (CNN–LSTM), CNN–Transformer, and Transformer–LSTM—were evaluated under four input configurations, yielding 28 model–input combinations. SARIMAX achieved the best overall balance between predictive accuracy and generalization stability. CatBoost was the most accurate data-driven method and obtained the highest coefficient of determination (R2 = 0.644) under meteorological-only inputs. Meteorological-only inputs outperformed pollutant-only and full multivariate inputs for all models. Among the deep-learning models, LSTM performed better under univariate, meteorological-only, and pollutant-only inputs, whereas Transformer benefited more from the full multivariate input. CNN preprocessing improved LSTM mainly under the full multivariate and pollutant-only configurations. The moderate maximum R2 suggests that routine observations did not fully capture aerosol composition, particle number and size distributions, hygroscopic growth, aerosol–water interactions, and fog-related processes. These findings support model and input selection for daily visibility forecasting.