关键词: 空气质量, 卷积神经网络, 长短期记忆神经网络, 深度学习, 时空预测

Abstract:

Air pollution has become a pressing global environmental issue with severe implications for human health, economic growth, and societal wellbeing. Long-term exposure to key pollutants, such as particulate matter (PM), ozone (O₃), sulfur dioxide (SO₂), and carbon monoxide (CO), is strongly linked to respiratory and cardiovascular diseases. According to the World Health Organization (WHO), 90% of the world’s population resides in areas where air quality fails to meet health standards, contributing to approximately seven million premature deaths annually. Urban areas, characterized by dense populations and concentrated industrial activities, are particularly vulnerable to high levels of pollution. Accurately predicting air quality trends is essential for mitigating health risks, informing policy decisions, and effectively managing pollution sources. Traditional forecasting methods, such as numerical weather prediction models (WRF-Chem) and statistical approaches (ARIMA), have limitations owing to the complexity of atmospheric processes and inherent model uncertainties. These methods often struggle to capture the nonlinear dynamics of pollution dispersion and interactions among various environmental factors. In contrast, recent advances in machine and deep learning have offered promising solutions by leveraging large, diverse datasets. These approaches integrate information from multiple sources, uncover hidden patterns, and generate reliable predictions. The aim of this study was to predict trends in air pollutant concentrations within the Poyang Lake Ecological and Economic Zone (POLEZ) using deep learning models, specifically Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM).This study is based on daily monitoring data of major air pollutants collected from 77 ambient air quality automatic monitoring stations (36 national control points and 41 provincial control points) within the POLEZ, covering the period from January 1, 2020, to April 6, 2024. Data from January 1, 2020, to December 31, 2023, were used as the training set, while the data from January 1, 2024, to April 6, 2024, served as the testing set. Initially, the spatial and temporal evolution of six major air pollutants (CO, NO₂, SO₂, ozone, PM_2.5, and PM₁₀) from 2020 to 2023 were analyzed using spatial visualization techniques. This analysis provides a comprehensive overview of pollutant trends and regional variations, helping identify key patterns and anomalies across the POLEZ. Building on this spatiotemporal analysis, a CNN-LSTM model was developed to predict pollutant concentrations in 2024. The CNN component of the model extracts time-series features from historical air quality data, whereas the LSTM component captures long-term dependencies, thereby enabling more accurate predictions of future trends. The performance of the CNN-LSTM model was compared with that of a traditional LSTM model to evaluate the added benefit of combining CNN and LSTM for air quality prediction. The key findings from this study are as follows: 1) Seasonal Patterns of Pollutants: Analysis of data from 2020 to 2023 reveals that concentrations of nitrogen dioxide (NO₂), respirable particulate matter (PM_2.5), inhalable particulate matter (PM₁₀), and carbon monoxide (CO) follow a “U”-shaped seasonal pattern, with lower levels in the summer and higher levels in the winter. This pattern reflects the influence of seasonal weather conditions on air quality, particularly in winter when temperature inversions and weak vertical mixing hinder the dispersion of pollutants. In contrast, ozone (O₃) concentrations peaked in summer and fall, driven by enhanced photochemical reactions due to increased sunlight and warmer temperatures. Monthly trends in sulfur dioxide (SO₂) concentrations show less variability, with a gradual decline over time, indicating the effectiveness of recent pollution control efforts in the region. 2) Spatial and Temporal Distribution of Pollutants: The spatial distribution of pollutants across the POLEZ revealed that Nanchang, as a regional economic hub, experienced the highest pollutant concentrations, primarily due to intensive industrial and transportation activities. Pollutants such as carbon monoxide (CO), nitrogen dioxide (NO₂), and PM_2.5, exhibit a west-to-east gradient, with higher pollution levels in western cities, including Nanchang and Jiujiang. Sulfur dioxide (SO₂) concentrations remained relatively low across the region, particularly around Poyang Lake. However, PM_2.5 and PM₁₀ concentrations increased in 2023, likely driven by economic recovery following the pandemic. Jiujiang, in particular, is the most polluted city for ozone (O₃), with concentrations showing a northwest-southeast gradient. 3) Model Performance: The CNN-LSTM model demonstrated superior prediction accuracy compared to the LSTM model. Specifically, the Mean Absolute Error (MAE) for CO decreased from 0.22 to 0.17, representing a 22.73% improvement; for SO₂, the MAE decreased from 1.67 to 1.28, a 23.35% improvement; for NO₂, the MAE decreased from 10.96 to 7.15, showing a 34.76% improvement; for O₃, the MAE decreased from 17.21 to 12.39, an improvement of 28.01%; for PM_2.5, the MAE decreased from 21.09 to 13.35, reflecting a 36.70% accuracy improvement; and for PM₁₀, the MAE decreased from 28.42 to 19.51, representing a 31.35% improvement. Additionally, the CNN-LSTM model showed significant improvements in both the Root Mean Squared Error (RMSE) and R² values, indicating that the model provided more reliable and accurate predictions of air pollution trends. The Pearson correlation coefficients between the predicted and observed values for most monitoring stations exceeded 0.8, further validating the high accuracy and reliability of the CNN-LSTM model for air quality forecasting. 4) Projections for 2024: The 2024 forecasts indicate an upward trend in the concentrations of all six major pollutants. Carbon monoxide levels are expected to rise slightly from 2023 but will remain below the levels observed in 2020. Nitrogen dioxide concentrations are projected to remain stable, underscoring the need for continued control measures to maintain air quality. Sulfur dioxide concentrations are anticipated to increase significantly, warranting increased attention and stricter control measures. Similarly, the concentrations of PM_2.5, PM₁₀, and O₃ are projected to rise, with notable increases in both the PM₁₀ and ozone levels. These findings emphasize the need for more effective and coordinated control measures, particularly for PM₁₀ and ozone, to prevent further deterioration of air quality. The results of this study underscore the effectiveness of the CNN-LSTM model in predicting the spatial and temporal trends of air pollutants and provide valuable insights for future air quality management in the Poyang Lake Ecological and Economic Zone. These findings emphasize the necessity for sustained and coordinated air pollution control efforts, particularly in high-risk areas, such as Nanchang and Jiujiang. By incorporating deep learning models into air quality prediction systems, policymakers can make informed decisions that will ultimately contribute to improving public health and environmental quality.

Key words: air quality, convolutional neural network (CNN), long short-term memory (LSTM) network, deep learning, spatiotemporal prediction