Atmospheric visibility prediction by using the DBN deep learning model and principal component analysis.

Area of Science:

• Environmental Science
• Computer Science
• Atmospheric Science

Background:

• Accurate atmospheric visibility prediction is crucial for urban air quality management and transportation safety.
• Existing methods may lack the precision required for both short-term and long-term forecasting.

Purpose of the Study:

• To develop and validate a novel deep learning model for predicting atmospheric visibility.
• To assess the model's performance against established neural network architectures.

Main Methods:

• Utilized principal component analysis (PCA) to identify key meteorological and environmental factors influencing visibility.
• Developed a deep belief network (DBN) with optimized parameters (double hidden layer, 70/50 nodes, Z-score normalization, tanh activation, Adam optimizer).
• Trained and tested the DBN model using data from 2016-2019, including visibility meter, particle spectrometer, and meteorological station data.

Main Results:

• The DBN model achieved an average prediction accuracy of 0.84 and a coefficient of determination of 0.96.
• DBN performance significantly surpassed that of back propagation (BP) neural networks and convolutional neural networks (CNN).
• The model demonstrated good short-term (within 3 days) prediction accuracy of 0.79 and practical applicability across various weather conditions.

Conclusions:

• The proposed deep learning model, integrating PCA and DBN, offers an effective and feasible solution for atmospheric visibility prediction.
• The model shows significant potential for applications in transportation, navigation, meteorology, and environmental research.
• This approach provides a robust technical framework for predicting atmospheric parameters vital for public safety and environmental monitoring.