An Evaluation of the Influence of Meteorological Factors and a Pollutant Emission Inventory on PM 2.5 Prediction in the Beijing–Tianjin–Hebei Region Based on a Deep Learning Method

: In this study, a Long Short-Term Memory (LSTM) network approach is employed to evaluate the prediction performance of PM 2.5 in the Beijing–Tianjin–Hebei region (BTH). The proposed method is evaluated using the hourly air quality datasets from the China National Environmental Monitoring Center, European Center for Medium-range Weather Forecasts ERA5 (ECMWF-ERA5), and Multi-resolution Emission Inventory for China (MEIC) for the years 2016 and 2017. The predicted PM 2.5 concentrations demonstrate a strong correlation with the observed values (R 2 = 0.871–0.940) in the air quality dataset. Furthermore, the model exhibited the best performance in situations of heavy pollution (PM 2.5 > 150 µ g/m 3 ) and during the winter season, with respective R 2 values of 0.689 and 0.915. In addition, the influence of ECMWF-ERA5’s hourly meteorological factors was assessed, and the results revealed regional heterogeneity on a large scale. Further evaluation was conducted by analyzing the chemical components of the MEIC inventory on the prediction performance. We concluded that the same temporal profile may not be suitable for addressing emission inventories in a large area with a deep learning method.


Introduction
In recent years, rapid industrialization, urbanization, and economic development have significantly increased anthropogenic air pollutant emissions [1,2] in China.The unfavorable weather conditions [3], coupled with these factors, have led to a severe air pollution problem in the country.In all air pollution events occurring in China over the past decade, the PM 2.5 concentration consistently held the highest ranking within the Individual Air Quality Index (IAQI) among the six categories of pollutants [4].This elevation in PM 2.5 levels has the potential to exert adverse impacts on atmospheric visibility [5], climate change [6,7], and human health [8][9][10][11][12].In order to mitigate the detrimental impact, the Chinese government has made extensive efforts to decrease pollutant emissions [13][14][15], with the specific goal of reducing the annual PM 2.5 concentrations nationwide.Particularly, significant measures have been implemented in the Beijing-Tianjin-Hebei region (BTH), Yangtze River Delta region (YRD), and Pearl River Delta region (PRD), leading to noteworthy accomplishments.Nonetheless, intermittent occurrences of haze events still take place in certain instances [16,17].Hence, the accurate and timely forecasting of PM 2.5 concentrations is imperative for regions that are impacted by PM 2.5 pollution.
There are two primary approaches to air quality forecasting: knowledge-driven models and data-driven models.Knowledge-driven models rely on complex equations that encompass various atmospheric dynamics.These models, such as CMAQ and WRF-Chem [18,19], require high-performance computers but offer limited accuracy.On the other hand, data-driven models have gained significant attention in recent years, particularly with the advent of the third wave of artificial intelligence (AI).Many researchers have successfully employed statistical models, including tree-based models and machine learning algorithms, to predict air quality and achieve satisfactory results.In a recent study, a group of researchers from Taiwan utilized Continuous Emission Monitoring Systems (CEMSs) to enhance prediction accuracy [20].Their findings demonstrated that incorporating additional features did indeed improve the prediction performance.However, it is important to note that the experiment did not account for the impact of heavily polluted regions.Therefore, in complex air pollutant areas, it is crucial to consider more relevant data to ensure comprehensive analysis and prediction.
The advancement of new information technology, including the Internet of Things (IoT) and AI [21][22][23], has enabled the acquisition and utilization of large-scale data for prompt air quality trend prediction.Meteorological and emission factors are pivotal in the discussion of air pollution causation [24][25][26].For this study, the European Centre for Medium-range Weather Forecasts ERA5 (ECMWF-ERA5) was selected due to its widespread accessibility and rich repository of meteorological factors, which have demonstrated effective performance in the context of the Chinese Mainland environment [27,28].In the realm of emission pollutants, a considerable number of researchers have employed the Multi-resolution Emission Inventory for China (MEIC) to obtain high-resolution emission [29] maps for regional air quality predictions.During the data extraction and conversion process, meteorological factors were extracted, and the Sparse Matrix Operator Kernel Emissions Model (SMOKE) [30] was utilized to allocate the raw MEIC data to finely grained temporal and spatial resolutions.Numerous researchers have attested to the robustness of AI in predicting future air quality [31][32][33][34], particularly through the Long Short-Term Memory network (LSTM) deep learning model, which demonstrates proficient performance [35][36][37] in air quality forecasting.To further leverage the potential of the information contained within these accessible data sources, experiments were designed to assess the applicability of the air quality dataset, ECMWF-ERA5 factors, and MEIC emission inventory in the hourly prediction of PM 2.5 concentrations using the LSTM model.
This study developed a PM 2.5 prediction system by extracting air monitoring datasets, ECMWF-ERA5 datasets, and MEIC emission inventory datasets for the target region.By utilizing the LSTM deep learning model and extracted features, the proposed method successfully predicted next-hour PM 2.5 concentrations at 13 individual air monitoring stations in the BTH region for 2016 and 2017.
To the best of our knowledge, this research represents an attempt to develop a PM 2.5 prediction system that evaluates air monitoring datasets, meteorological fields, and local emission inventories in the BTH region.The study involved the following steps: First, the ECMWF-ERA5 and MEIC datasets were tested using the Mann-Kendall test to select the appropriate variables.Secondly, a single LSTM method was introduced to evaluate its superiority in air quality prediction.Thirdly, based on the LSTM method, the performance of the two aforementioned datasets were tested, and it was found that the meteorological dataset exhibited spatial heterogeneity, while the pollutant emission data displayed temporal disparity.

Research Domain and Period Chosen
This research focuses on the Beijing-Tianjin-Hebei (BTH) region (113.45°E~119.85°E and 36.03°N~42.62°N), encompassing 13 cities: Beijing (BJ), Tianjin (TJ), Shijiazhuang (SJZ), Tangshan (TS), Qinhuangdao (QHD), Handan (HD), Baoding (BD), Zhangjiakou (ZJK), Chengde (CD), Langfang (LF), Cangzhou (CZ), Hengshui (HS), and Xingtai (XT).This region, which spans diverse terrains from the seaside to mountainous areas, constitutes the largest economic hub in northern China, characterized by a high emission density (excluding CD and ZJK).The unique topographic features, variable meteorological fields, and substantial emissions make it prone to elevated PM2.5 concentrations.Within these cities, the following stations were selected as research targets: Mansu Saigon (MS), Forward Road  The study period spanned from 1 January 2016 to 31 December 2017, totaling 17,544 hours, during which comprehensive measures were implemented.Figure 2 depicts the monthly PM2.5 concentrations at the selected stations.Specifically, for HV and DZ, the PM2.5 concentrations remained below 75 µg/m 3 throughout the two-year period.In 2017, the monthly PM2.5 concentrations at QMS also stayed below the 75 µg/m 3 threshold.Conversely, The study period spanned from 1 January 2016 to 31 December 2017, totaling 17,544 h, during which comprehensive measures were implemented.Figure 2 depicts the monthly PM 2.5 concentrations at the selected stations.Specifically, for HV and DZ, the PM 2.5 concentrations remained below 75 µg/m 3 throughout the two-year period.In 2017, the monthly PM 2.5 concentrations at QMS also stayed below the 75 µg/m 3 threshold.Conversely, for SH, STP, RBC, BMS, MMS, TRS, and NCIA, the monthly PM 2.5 concentrations exhibited a significant increase from August to December 2017, eventually surpassing 150 µg/m 3 .The trend observed in the two-year line plot indicates the influence of meteorological characteristics on air quality.Furthermore, notable trough values were observed in May and August 2016, and in April, June, and August 2017, across all stations, indicating the implementation of stringent and timely control measures by the Chinese government in 2017.A comparative analysis revealed that after February 2017, the PM 2.5 concentrations consistently remained below 150 µg/m 3 , including in December of the same year, indicating the substantial positive impact of the measures.In terms of data training, the selection of 2016 for the training process and 2017 for testing was deemed reasonable.
for SH, STP, RBC, BMS, MMS, TRS, and NCIA, the monthly PM2.5 concentrations exhibited a significant increase from August to December 2017, eventually surpassing 150 µg/m 3 .The trend observed in the two-year line plot indicates the influence of meteorological characteristics on air quality.Furthermore, notable trough values were observed in May and August 2016, and in April, June, and August 2017, across all stations, indicating the implementation of stringent and timely control measures by the Chinese government in 2017.A comparative analysis revealed that after February 2017, the PM2.5 concentrations consistently remained below 150 µg/m 3 , including in December of the same year, indicating the substantial positive impact of the measures.In terms of data training, the selection of 2016 for the training process and 2017 for testing was deemed reasonable.

Ground Air Quality Concentration
Hourly ground-level data for six primary air pollutants (PM 2.5 , PM 10 , SO 2 , CO, NO 2 , and O 3 ) were obtained from the China National Environmental Monitoring Center for the duration of 1 January 2016 to 31 December 2017.Throughout this period, the hourly air quality monitoring data from all the stations were stored in a MySQL 5.5 database.Subsequently, data from 13 stations in the BTH region were extracted from this database.

ECMWF-ERA5 Meteorological Factors
The meteorological data utilized in this study were obtained from the ECMWF-ERA5 hourly data on single levels and ERA5-Land hourly data from the ECMWF reanalysis, encompassing the entire years of 2016 and 2017.These datasets were acquired at a spatial resolution of 0.25 degree × 0.25 degree.Specifically, the surface pressure (SPRE), relative humidity (RH), 2 m temperature (TMP), as well as the u and v components of wind were selected as input factors for the model.To ensure spatial alignment with the corresponding air quality data, the nearest distance algorithm was employed to calculate the corresponding positions.
Throughout the training process, it was observed that the u and v components of wind exhibited adverse effects on predictions.As a result, in the present study, these components were converted into wind speed (WS) and wind direction (WD) using the following functions: where u and v represent horizontal and vertical wind speeds, respectively

MEIC Emission Dataset
To accurately procure the emission data for the BTH region, the MEIC version 1.3, encompassing CO, NOx, NO 2 , VOCs, NO 2 , PM 2.5 , PM-coarse, BC, and OC, with a spatial resolution of 0.25 × 0.25 throughout 2016 and 2017, was obtained from its official website.In order to attain a representative distribution of emission characteristics, the SMOKE model was employed to process the data, with spatial and temporal resolutions adjusted to 36 km × 36 km and 1 h, respectively.Moreover, for the species allocation of VOCs, the CB05 chemical mechanism was selected [38,39].Finally, the nearest distance algorithm was utilized to determine the corresponding positions of the air quality stations.

Methods and Technical Roadmap 2.3.1. The Structure of the Experimental Model
The LSTM neural network, a derivative of the recurrent neural network (RNN), effectively addresses the issues of gradient explosion and vanishing gradients.In this study, our primary focus was not on the algorithm itself.Instead, the research aimed to develop a system that is capable of utilizing air quality data, meteorological factors, and pollutant emission inventories to forecast future air quality and advance the management of emission pollutants using innovative strategies.
Using a single-layer LSTM model, we assessed the performance of the PM 2.5 concentration prediction.The chosen optimizer and loss function were "Adam" [40] and "MAE", respectively.Throughout the experiments, the inclusion of additional related features led to an expansion of the design input factors, as outlined in detail in Table 1.In this study, our target for prediction was the PM 2.5 concentration measured at 13 selected monitoring stations.The prediction tests consisted of three datasets: (1) TEST1, which included hourly air quality monitoring data; (2) TEST2, comprising the coupling of TEST1 with ECMWF-ERA5's meteorological factors; and (3) TEST3, which encompassed the combination of TEST2 with the chemical components of the MEIC emission dataset.To mitigate the inherent randomness in the system, we conducted the tests 100 times and then calculated the average results.Also, to determine the appropriate input variables, we employed the Mann-Kendall test.Detailed information can be seen in Figure 3.

Mann-Kendall Correlation for Input Variables
In order to obtain the appropriate input variables, input feature selection is a key procedure.In this research, the Mann-Kendall test was chosen to evaluate the relationship among air quality, ECMWF-ERA5's meteorological factors, and the chemical components of the MEIC emission dataset and then determine the number of input variables.From the datasets mentioned above, we can see that there were a total of fifty-four input features in To assess the influence of different datasets and the impact of neuron cells on the prediction accuracy, we designed six experiments utilizing a single-layer LSTM model with an unchanged optimizer and loss function.Experiments 1-3 evaluated the impact of adding different datasets on the PM 2.5 prediction performance.The first experiment focused on the performance of the air quality dataset alone.The second and third experiments added ECMWF-ERA5's meteorological factors and the chemical components of the MEIC emission dataset coupled with ECMWF-ERA5 meteorological factors, respectively.
To assess the simulated results derived from the aforementioned experiments, we utilized the determination coefficient (R 2 ), mean absolute error (MAE), and root mean square error (RMSE) as statistical criteria.The specific formulations for evaluating these parameters are listed below.
where y i and ŷi represent the values of the monitoring data and predicted data at i time, respectively, y denotes the mean value of the monitoring data, and N illustrates the number of samples.

Mann-Kendall Correlation for Input Variables
In order to obtain the appropriate input variables, input feature selection is a key procedure.In this research, the Mann-Kendall test was chosen to evaluate the relationship among air quality, ECMWF-ERA5's meteorological factors, and the chemical components of the MEIC emission dataset and then determine the number of input variables.From the datasets mentioned above, we can see that there were a total of fifty-four input features in Table 1, namely, six, five, and forty-three, corresponding to the three datasets, respectively.During the process, the relationship between the t-th PM 2.5 and (t − 1)-th variables are tested by the Mann-Kendall correlation coefficient, where ** and * represent 0.01 and 0.05 levels of significance, respectively.We set a Kendall correlation coefficient absolute value > 0.08 as the threshold value to choose the appropriate features (Table 1).
The air quality dataset shows that PM 2.5 , PM 10 , CO, SO 2 , and NO 2 have a positive influence on the performance of the PM 2.5 prediction in the next time step.The highest influence is observed for PM 2.5 , with correlation coefficients ranging from 0.756 to 0.865, followed by PM 10 , CO, SO 2 , and NO 2 .Conversely, O 3 has a negative influence, with correlation coefficients ranging from 0.056 to 0.136.The influence of the dataset is relatively stable.Regarding the ECMWF-ERA5 meteorological factors, RH and WS have a consistent influence.RH has a positive impact, with correlation coefficients ranging from 0.134 to 0.317, while WS has a negative impact, with coefficients ranging from 0.129 to 0.238.The influence of the remaining factors (SPRE, TMP, and WD) is unstable.WD, for example, did not pass the Kendall 0.01 test on PB and had positive functions for the HV, DZ, and SH factors.RH is conducive to the formation of PM 2.5 , while WS is favorable for PM 2.5 diffusion.Therefore, RH and WS perform well in predicting air quality.The Mann-Kendall test results for chemical species show that sixteen species listed in Table 2 passed the test at the 13 stations.NO 2 passed at ten stations, and HONO passed at five stations, indicating that these chemical components are useful for predicting PM 2.5 concentrations.

Prediction Performance of the Next-Hour PM 2.5 Concentration
To assess the prediction accuracy of the proposed model in forecasting the PM 2.5 concentration using air quality data, hourly air quality forecasting was conducted.Throughout the training period using the year of 2016, there were slight changes in the standard deviations of R 2 and RMSE.The mean values of R 2 and RMSE were 0.907 ± 0.0029 and 17.54 ± 0.273, respectively.For instance, the maximum standard deviations of R 2 and RMSE for TRS were 0.009 and 0.725.Additionally, we observed a range of values for R 2 between 0.871 and 0.940, and for RMSE between 12.309 and 23.660.This information is presented in Figure 4, where the predicted results from the 13 stations demonstrate a strong consistency.Consequently, we have confidence that both underfitting and overfitting did not occur throughout the modeling run (Figure 5).Therefore, the model is deemed capable of effectively predicting PM 2.5 concentrations within the BTH region.(a) The results based on the four seasons and different pollution levels were analyzed to evaluate the forecasting performance of the model.As shown in Table 3a, good performances were achieved by the model.The average R 2 values appeared in the sequence of summer (0.755), spring (0.861), autumn (0.888), and winter (0.915), with the best performance presented by winter.Additionally, it is noted that we can capture the high pollution episodes in winter (Table 3b).Specifically, the PM2.5 levels are divided into low (0~75 µg/m 3 ), middle (75-150 µg/m 3 ), and high (>150 µg/m 3 ), and the model performed well for the majority of the stations under the high PM 2.5 pollution level.This is because the PM 2.5 concentration for these stations throughout 2016 and 2017 remained at a high concentration, featuring a high pollution frequency.Correspondingly, the average R 2 , MAE, and RMSE for the low PM2.5 level were 0.464, 7.417, and 12.102, respectively.These values were found to be lower than the corresponding metrics for the high PM 2.5 concentration level, which were 0.689, 24.512, and 40.406, respectively.

Evaluation Performance of Air Monitoring, Meteorological Factors, and Emission Inventory on PM 2.5 Prediction
By employing the same approach as TEST1 did, an assessment was conducted on the predictive capabilities of TEST2 and TEST3.The R 2 , MAE, and RMSE values obtained from the TEST1 simulation were 0.907, 10.349, and 17.550, respectively, which were correspond-ingly 0.833, 16.577, and 23.297 for the TEST2 simulation and 0.755, 21.584, and 28.554 for TEST3 (seen in Figure 6), respectively.Notably, significant improvements were not achieved by introducing meteorological factors and the MEIC emission dataset.The reason for this phenomenon is attributed to discrepancies in the temporal emission characteristics of the emission inventory.The employment of uniform temporal allocation coefficients results in an inadequate representation of the temporal emission characteristics across different regions, consequently constraining the predictive accuracy [41].Regional heterogeneity in meteorology a large scale can also be responsible for this phenomenon, resulting in improved predictive performance for some meteorological sites, while others experience less improvement [42].Additionally, the terrain of the observation sites, located in coastal, plain, and mountainous areas, indicates that the differences in terrain have a certain impact on or interference with the predictive performance of atmospheric pollution [43].Thus, in future regional air quality forecasting processes, it is necessary to fully consider the effects of terrain variations, heterogeneity in meteorological conditions, and differences in emission inventory emission characteristics.Based on the above discussion, the R 2 , MAE, and RMSE values obtained from TEST3 ranged from 0.411 to 0.918, 7.329 to 46.787, and 12.827 to 56.960, respectively.It is obvious that the prediction performance of PM 2.5 concentrations was superior to knowledge-driven models that also utilized meteorological models, such as the Weather Research and Forecasting model (WRF), to forecast air quality by using the air pollutants [44].When using air monitoring quality datasets and ECMWF-ERA5 meteorological factors to forecast PM 2.5 concentrations, we achieved a timely hourly prediction.We noted an improvement in prediction performance for the three stations (NCIA, HV, and PB), possibly due to the alignment of the temporal profiles of chemical components with the true situation.Thus, employing the same temporal profile may not be suitable for addressing the emission inventory in a large area.This implies that different regional temporal profiles in the emission inventory could enhance the performance of knowledge-driven models in air quality prediction.
improvement [42].Additionally, the terrain of the observation sites, located in coastal, plain, and mountainous areas, indicates that the differences in terrain have a certain impact on or interference with the predictive performance of atmospheric pollution [43].Thus, in future regional air quality forecasting processes, it is necessary to fully consider the effects of terrain variations, heterogeneity in meteorological conditions, and differences in emission inventory emission characteristics.

Conclusions
In this research, a single-layer LSTM model was constructed to evaluate the influence of meteorological factors from ECMWF-ERA5 and the emission inventory from MEIC.The results showed that the air quality dataset can achieve the best performance, reaching an R 2 of 0.907 on average in the chosen 13 stations, while for different seasons and levels, winter and a high level are superior to the other seasons and levels on average.The ECMWF-ERA5 and MEIC datasets can equally obtain an approximate prediction performance, which is perhaps due to the chosen meteorological factors of ECMWF-ERA5 and the chemical components of the MEIC dataset, which cannot accurately reflect the spatial and temporal characteristics with the changing PM 2.5 trend, but is still better than knowledge-driven model for long-term prediction.
The contribution of this work was to construct a new system which included both the external and internal causes of air pollution using an LSTM model, such as meteorological factors and the local emission inventory, which were commonly used in knowledge-driven models.Although the prediction performance of the new method did not improve greatly, the results were still superior to those of knowledge-driven models.Moreover, this research provided an idea for optimizing knowledge-driven models by improving the accuracy of the prediction performance.In addition, the proposed method had a certain capacity to predict the regional air quality of each monitoring station and manage the local pollutant emission for the future reduction plan.

Figure 2 . 2 .
Figure 2. Two years of monthly PM2.5 concentrations in the 13 monitoring stations in the BTH region.The blue dotted line (lower one) represents the 75 µg/m 3 concentration level and the red line (upper Figure 2. Two years of monthly PM 2.5 concentrations in the 13 monitoring stations in the BTH region.The blue dotted line (lower one) represents the 75 µg/m 3 concentration level and the red line (upper one) represents the 150 µg/m 3 concentration level.(a) represents the air quality situation in 2016, and (b) represents the air quality situation in 2017.
and  represent the values of the monitoring data and predicted data at i time, respectively,  denotes the mean value of the monitoring data, and N illustrates the number of samples.

Figure 4 .
Figure 4. Comparison of forecasting performance among chosen 13 stations.

Figure 4 .
Figure 4. Comparison of forecasting performance among chosen 13 stations.

Figure 4 .
Figure 4. Comparison of forecasting performance among chosen 13 stations.

Figure 5 .
Figure 5. STD of (a) R 2 and (b) RMSE in chosen 13 stations based on 100 rounds of prediction.Figure 5. STD of (a) R 2 and (b) RMSE in chosen 13 stations based on 100 rounds of prediction.

Figure 5 .
Figure 5. STD of (a) R 2 and (b) RMSE in chosen 13 stations based on 100 rounds of prediction.Figure 5. STD of (a) R 2 and (b) RMSE in chosen 13 stations based on 100 rounds of prediction.

Figure 6 .
Figure 6.An evaluation of the performance of the PM 2.5 concentration with different datasets.The smaller symbols represent the prediction performance of the air quality dataset, and the bigger ones represent the impact of adding the ECMWF-ERA5 meteorological factors dataset, ECMWF-ERA5 and meteorological factors, and MEIC chemical components dataset, respectively.(a) The comparative performance of air quality with and without the ECMWF-ERA5 meteorological factors dataset; (b) the comparative performance of air quality and ECMWF-ERA5 meteorological factors without and with the chemical components of the MEIC dataset.

Table 2 .
Kendall test results of different datasets for PM 2.5 concentration prediction.

Table 3 .
PM 2.5 prediction performance of air quality dataset in different seasons and levels.