Optimized neural network for daily-scale ozone prediction based on transfer learning

Transformer-based ozone multivariate prediction considering interpretable and priori knowledge: A case study of Beijing, China

Ozone pollution is the focus of current environmental governance in China and high-quality prediction of ozone concentration is the prerequisite to effective policymaking. The studied ozone pollution time series exhibits distinct seasonality and secular trends and is associated with various factors. This study developed an interpretable hybrid model by combining STL decomposition and the Transformer (STL-Transformer) with the prior information of ozone time series and global multi-source information as prediction basis. The STL decomposition decomposes ozone time series into trend, seasonal, and remainder components. Then, the three components, along with other air quality and meteorological data, are integrated into the input sequence of the Transformer. The experiment results show that the STL-Transformer outperforms the other five state-of-the-art models, including the standard Transformer. Specially, the univariate forecasting for ozone relies on mimicking the patterns and trends that have occurred in the past. In contrast, multivariate forecasting can effectively capture complex relationships and dependencies involving multiple variables. The method successfully grasps the prior and global multi-source information and simultaneously improves the interpretability of ozone prediction with high precision. This study provides new insights for air pollution forecasting and has reliable theoretical value and practical significance for environmental governance.

Air pollutant prediction model based on transfer learning two-stage attention mechanism

Atmospheric pollution significantly impacts the regional economy and human health, and its prediction has been increasingly emphasized. The performance of traditional prediction methods is limited due to the lack of historical data support in new atmospheric monitoring sites. Therefore, this paper proposes a two-stage attention mechanism model based on transfer learning (TL-AdaBiGRU). First, the first stage of the model utilizes a temporal distribution characterization algorithm to segment the air pollutant sequences into periods. It introduces a temporal attention mechanism to assign self-learning weights to the period segments in order to filter out essential period features. Then, in the second stage of the model, a multi-head external attention mechanism is introduced to mine the network's hidden layer key features. Finally, the adequate knowledge learned by the model at the source domain site is migrated to the new site to improve the prediction capability of the new site. The results show that (1) the model is modeled from the data distribution perspective, and the critical information within the sequence of periodic segments is mined in depth. (2) The model employs a unique two-stage attention mechanism to capture complex nonlinear relationships in air pollutant data. (3) Compared with the existing models, the mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) of the model decreased by 14%, 13%, and 4%, respectively, and the prediction accuracy was greatly improved.

Importance of secondary decomposition in the accurate prediction of daily-scale ozone pollution by machine learning

Machine learning (ML) models have been proven as a reliable tool in predicting ambient pollution concentrations at various places in the world. However, their performance in predicting the maximum daily 8-h averaged ozone (MDA8 O3), the metric often used for O3 pollution assessment and management, is relatively poorer. This is largely resulted from more irregular data fluctuations of the MDA8 O3 levels governed collectively by the synoptic condition, local photochemistry, and long-range transport. In order to improve the prediction accuracy of MDA8 O3, this study developed a secondary decomposition ML model framework which coupled the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) as the primary decomposition, the variational mode decomposition (VMD) as secondary decomposition, and the gate recurrent unit (GRU) ML model. By applying this secondary decomposition model framework on MDA8 O3 prediction for the first time, we showed that the prediction accuracy of MDA8 O3 is largely improved from R2 of 0.46 and RMSE of 30.4 μg/m3 for GRU without decomposition to R2 of 0.91 and RMSE of 12.6 μg/m3 over the Pearl River Delta of China. We also found that the prediction accuracy rate of O3 pollution non-attainments, an essential indicator for initiating contingency O3 pollution control, improved greatly from 14.9 % for GRU without decomposition to 72.5 %. The performance of O3 pollution non-attainment prediction is relatively higher in southwestern PRD, which is mainly due to greater number and severity of O3 non-attainments in southwestern cities located downwind of the emission hotspot area at central PRD. This study underscored the importance of secondary decomposition in accurately predicting daily-scale O3 concentration and non-attainments over the PRD, which can be extended to other photochemically active region worldwide to improve their O3 prediction accuracy and assist in O3 contingency control.

Development of a recurrent spatiotemporal deep-learning method coupled with data fusion for correction of hourly ozone forecasts

Ambient ozone (O3) predictions can be very challenging mainly due to the highly nonlinear photochemistry among its precursors, and meteorological conditions and regional transport can further complicate the O3 formation processes. The emission-based chemical transport models (CTM) are broadly used to predict O3 formation, but they may deviate from observations due to input uncertainties such as emissions and meteorological data, in addition to the treatment of O3 nonlinear chemistry. In this study, an innovative recurrent spatiotemporal deep-learning (RSDL) method with model-monitor coupled convolutional recurrent neural networks (ConvRNN) has been developed to improve O3 predictions of CTM. The RSDL method was first used to build the ConvRNN within a 24-h scale to characterize the spatiotemporal relationships between the monitored O3 data and CTM simulations, and then incorporated the recurrent pattern to achieve 72-h multi-site forecasts based on a pilot study over the Pearl River Delta (PRD) region of China. The results showed that the RSDL method predicted O3 with high accuracy over this case study, with an increase of 27.54% in the correlation coefficient (R) average for all sites as well as an increase in R of 0.14-0.21 for all cities compared to CTM. Moreover, the regional distribution of CTM was further improved by the RSDL predictions with the data fusion technique, which greatly reduced the underpredictions of O3 concentrations, particularly in high O3-level areas (concentrations >160 μg/m3), with a 33.55% reduction in the mean absolute error (MAE).

Assessing uncertainty and heterogeneity in machine learning-based spatiotemporal ozone prediction in Beijing-Tianjin- Hebei region in China

Accurate prediction of spatiotemporal ozone concentration is of great significance to effectively establish advanced early warning systems and regulate air pollution control. However, the comprehensive assessment of uncertainty and heterogeneity in spatiotemporal ozone prediction remains unknown. Here, we systematically analyze the hourly and daily spatiotemporal predictive performances using convolutional long short term memory (ConvLSTM) and deep convolutional generative adversarial network (DCGAN) models over the Beijing-Tianjin-Hebei region in China from 2013 to 2018. In extensive scenarios, our results show that the machine learning-based (ML-based) models achieve better spatiotemporal ozone concentration prediction performance with multiple meteorological conditions. A further comparison to the air pollution model-Nested Air Quality Prediction Modelling System (NAQPMS) and monitoring observations, the ConvLSTM model demonstrates the practical feasibility of identifying high ozone concentration distribution and capturing spatiotemporal ozone variation patterns at a high spatial resolution (here 15 km × 15 km).

Ambient particle characteristics by single particle aerosol mass spectrometry at a coastal site in Hong Kong: a case study affected by the sea-land breeze

The sea-land breeze (SLB) circulation plays a vital role in the transport of atmospheric pollutants in coastal cities. In this study, a single particle aerosol mass spectrometer (SPAMS) and combined bulk aerosol instruments were deployed to investigate the ambient particle characteristic at a suburban coastal site in Hong Kong from February 22 to March 10, 2013. Significant SLB circulations were captured from March 6-10, 2013, during the campaign. During the SLB periods, air quality worsened, with PM_2.5 concentrations reaching a peak of 55.6 μg m^-3 and an average value of 42.8 ± 4.5 μg m^-3. A total of 235,894 particles were measured during the SLB stage. Eight major sources were identified by investigating the mixing states of the total particles, including the coal-burning related particles (48.1%), biomass burning particles (6.7%), vehicle emission-related particles (16.4%), sea salt (9.2%), ship emission particles (2.7%), dust/steeling industries (3.7%), waste incineration (6.3%), and road dust (3.9%). It was noteworthy that the PM_2.5 concentrations and particle numbers increased sharply during the transition of land wind to the sea breeze. Meanwhile, the continental sourced pollutants recirculated back to land resulting in a cumulative increase in pollutants. Both individual and bulk measurements support the above results, with high contributions from coal burning, biomass burning, bulk K⁺, and NO₃ ^-, which were probably from the regional transportation from the nearby area. In contrast, the ship and vehicle emissions increased during the SLB period, with a high sulfate concentration partially originating from the ship emission. In this study, field evidence of continental-source pollutants backflow to land with the evolution of sea breeze was observed and helped our current understanding of the effect of SLB on air quality in the coastal city.

PM2.5 forecasting for an urban area based on deep learning and decomposition method

Rapid growth in industrialization and urbanization have resulted in high concentration of air pollutants in the environment and thus causing severe air pollution. Excessive emission of particulate matter to ambient air has negatively impacted the health and well-being of human society. Therefore, accurate forecasting of air pollutant concentration is crucial to mitigate the associated health risk. This study aims to predict the hourly PM2.5 concentration for an urban area in Malaysia using a hybrid deep learning model. Ensemble empirical mode decomposition (EEMD) was employed to decompose the original sequence data of particulate matter into several subseries. Long short-term memory (LSTM) was used to individually forecast the decomposed subseries considering the influence of air pollutant parameters for 1-h ahead forecasting. Then, the outputs of each forecast were aggregated to obtain the final forecasting of PM2.5 concentration. This study utilized two air quality datasets from two monitoring stations to validate the performance of proposed hybrid EEMD-LSTM model based on various data distributions. The spatial and temporal correlation for the proposed dataset were analysed to determine the significant input parameters for the forecasting model. The LSTM architecture consists of two LSTM layers and the data decomposition method is added in the data pre-processing stage to improve the forecasting accuracy. Finally, a comparison analysis was conducted to compare the performance of the proposed model with other deep learning models. The results illustrated that EEMD-LSTM yielded the highest accuracy results among other deep learning models, and the hybrid forecasting model was proved to have superior performance as compared to individual models.