Enhanced short-term prediction of urban PM_2.5 concentrations by improved hybrid deep learning

Abstract

Purpose

The aim of this study was to investigate the impact of improved deep learning model on the predictive performance of PM_2.5 concentration.

Methods

We developed a new model combining one-dimensional convolutional neural network and bidirectional long short-term memory neural network to predict PM_2.5 concentrations at hourly intervals. The air pollution observation data from 2020 to 2022 collected at several national air quality monitoring stations in Shenyang (Liaoning province, China) were employed to train our model. The performance of the proposed model was boosted by connecting the layer of network calculated results with the PM_2.5 sequence data. Furthermore, data of most relevant air quality monitoring stations and PM_2.5 feature factors of the target station were screened. The spatial correlation of major air pollutant and the interaction between PM_2.5 and other pollutant factors were therefore considered to improve the accuracy of the model.

Results

The root mean square error, mean absolute error, mean absolute percentage error of the new method were reduced by 49%, 51%, 44% and the R² was improved by 4.6% respectively compared with the control group for the next hour prediction. The proposed improvement method can reduce the prediction error of the model in the next 6 h.

Conclusions

In this study, the proposed model improvement method can significantly reduce the error of the model in predicting PM_2.5 concentration. The proposed method can improve the model in the next 6 h prediction accuracy. This study provides a new perspective for establishing high-precision models for PM_2.5 prediction.

Supplementary Information

The online version contains supplementary material available at 10.1007/s40201-025-00954-0.

Introduction

In recent years, with the rapid development of global industry, air quality has generally declined and air pollution incidents have occurred more frequently. The phenomenon of haze, in particular, has had serious impacts on people’s daily lives. This results from particulate matter in ambient air, specifically with an aerodynamic equivalent diameter of < 2.5 microns (PM_2.5). PM_2.5 has the dual characteristics of small particle size and large surface area, making it a carrier for the spread of various toxic substances in the air that induce respiratory and lung diseases, seriously endangering people’s health. Shenyang, an important industrial base in China, has a strong foundation in equipment manufacturing and a complete industrial chain. Above all, the logistics industry has developed rapidly in recent years. However, with the development of industry, Shenyang air quality pollution was serious and has exceeded the “Ambient Air Quality Standard” (GB3095-2012) in China. Short-term forecasting refers to the prediction of pollutants in the next six hours [1]. Weather forecasts do not solve the pollution problem at the source, but it can protect people from pollution events through advance warning. Predicting concentration of PM_2.5 in air and to issue early warnings by relevant departments is sensible way to protect people’s lives and property from PM_2.5.

The hybrid neural network, which combines convolutional neural network (CNN) with a long short-term memory neural network (LSTM), is currently the main model structure for air quality prediction. This hybrid model has excellent prediction performance, due to its comprehensive use of the features extraction capability of convolutional neural network and the characteristics of recursive neural network time series prediction [2, 3].

To improve model prediction accuracy ability of air pollution, countless researchers have carried out in-depth research on such a hybrid neural network (CNN-LSTM) [3–5]. The structure of deep neural network is the basis and key of any prediction. Therefore, the model performance of its predictions can be improved by improving the network structure. The common practice was to connect the CNN, LSTM and deep neural network, which were better than the single CNN mixed LSTM to predict PM_2.5 [6]. Bidirectional long short-term memory neural network (BILSTM) can use time series data more efficiently than long short-term memory neural network. Therefore, combining CNN and BILSTM as a model for PM_2.5 prediction leads to a better performance than CNN-LSTM [3]. PM_2.5 will influence by other atmospheric pollutants during the process of transmission. Therefore, PM_2.5 predictions need to consider the effect of multiple pollutants [7, 8]. Meanwhile the screening of air pollutants as feature factors was very important for the prediction of PM_2.5 [9]. In addition, besides using conventional air pollutants as main model factors, meteorological indices can also be used to predict the PM_2.5 concentration such as AQI [10]. Air pollution has spatial attributes, and when dealing with local pollution consideration should be given to the pollution in adjacent areas [11]. Therefore, spatial factors were used to create the CNN-BILSTM network model for better PM_2.5 predictions [12, 13]. Zhang et al. [14] and Li et al. [15] improved the prediction performance of such models by convolutional long and short-term memory neural hybrid network combining spatial factors and the effects of multiple pollutants to prediction.

Various measures to improve the prediction performance of the model have been made, including better consideration of the spatial effects of air pollutants, the influence between PM_2.5 and multiple pollutants, and improvements the model structure [11, 16, 17]. However, enhancing the prediction accuracy of the model remains a research challenge. Therefore, in this study, the combination of feature and the neural network layer output results was used to improve the model prediction performance. Here a CNN-BILSTM hybrid neural network containing the combination of PM_2.5 feature and the neural network layer output results was used to predict upcoming concentration of PM_2.5 in the next few hours. For this purposed data was collected from 8 national monitoring stations in Shenyang (Liaoning Province, China). The proposed network consists of three-layer one-dimensional (1D) CNN, two-layer BILSTM, a full connected layer, and a structure of the combination of PM_2.5 feature and the neural network layer output results. The innovative contributions provided by the new methodology are as follows: (i) a novel hybrid deep neural network structure with combination of feature and the neural network layer output results was developed, (ii) the combined feature and the neural network layer output results was used to improve the prediction accuracy of concentration of PM_2.5, and (iii) the proposed hybrid model structure can save the number of features used by model training.

Materials and methods

Study area and data source

Shenyang, the capital of Liaoning province in the south of Northeast China, is located between 41°48 ‘11.75 “- 123°25’ 31.18 “. It is an important industrial base in northeast China, where the problem of air pollution is prominent.

Hourly data of PM_2.5 and other pollutant (PM₁₀, NO₂, SO₂, CO, O₃, AQI) concentrations between 1th January 2020 to 31th December 2022 were collected from 8 national air quality monitoring stations for the model training of proposed method, i.e., Hunnandonglu (HNDL), Xinxiujie (XXJ), Donglinglu (DLL), Lingdongjie (LDJ), Wenhualu (WHL), Xiaoheyan (XHY), Taiyuanjie (TYJ), Jingshenjie (JSJ). The data sourced from the National Urban Air Quality Real Time Release Platform of China Environmental Monitoring Station. The concentrations of PM_2.5 of each monitoring station in Shenyang for 2020–2022 were showed in Table A1. The locations of the national air quality monitoring stations were illustrated in Fig. 1.

Fig. 1.

The locations of the national air quality monitoring stations in Shenyang

Proposed model

The proposed model in this study was based on a combination of 1D convolutional network, bidirectional long short-term memory neural network (BILSTM) and full connect network. The general structure of the proposed model was showed in Fig. 2.The sample data was input into 1D convolutional neural network and the result were output. Then the result was connected with PM_2.5 data of the target station, after that it was input into the bidirectional long short-term memory neural network for calculation to obtain the prediction result. Meanwhile, the calculation results of each layer of the bidirectional long short-term memory neural network were connected with the PM_2.5 data of the target station and then enter the next layer for calculation. Finally, the PM_2.5 data connected to the target station with the prediction result was input into the full connect network to output the predicted value. This hybrid model was designed to use CNN to extract data features, BILSTM to implement prediction using feature data, and full connect network for predicted value output. The PM_2.5 data of the target station was connected with neural network layer result many times during the calculation process to improve the accuracy of the model prediction.

Fig. 2.

The framework of the prediction model

Data preprocessing

The quality of data significantly affects the performance of the model. Employing incomplete data directly in predictive modeling could compromise the accuracy of predictions [18]. Therefore, it is necessary to preprocess the data, which encompassing tasks such as quality control of data, missing value interpolation and normalization.

The hourly air quality data used in this study was sourced from the National Urban Air Quality Real Time Release Platform of China Environmental Monitoring Station. At the time of data publication, the data quality control operations such as removing outliers have been carried out, and it meeting the minimum requirements for the validity of pollutant concentration data in China’s “Environmental Air Quality Standards” (GB3095-2012).

Several widely recognized interpolation methods are available for addressing missing data [7, 19, 20]. In this study, the K-nearest neighbor interpolation method was used to interpolate the data. The value near the vacancy value will be considered for interpolation, while the data missing for a long time and unable to be interpolated by the algorithm will be replaced by pollutant data at the same time the next day. Meanwhile, stations with a substantial amount of missing data will be directly excluded from the analysis to ensure the integrity of the dataset.

The objective of normalization is to standardize the scale of each variable. In this study, maximum and minimum normalization was used, which the maximum and minimum values in the data column were standardized, and the standardized values are between [0, 1] [21].

In deep learning, the data set used for model training often needed to be divided into training sets and test sets. Typically, 70% of the dataset is designated as the training set, with the remaining 30% serving as the test set [21]. However, the manner in which the original dataset is partitioned can significantly impact the evaluation metrics computed on the validation set. The K-fold cross-validation, wherein the data is randomly divided into K equal parts, with K-1 parts chosen randomly as the training data and the remaining part utilized as the validation data, is unsuitable for training time series models [1]. Therefore, in this study, the time series split cross validation was used as validation mothed. The first 80% of the data set was divided into the training set for time series split cross validation and the last 20% was divided into the test set.

Method for constructing model samples

It is assumed that there were m features and each feature was arranged into 1 column according to the time series t and the m column features were arranged together to form a data set. Then the sliding window method was used to construct the sample. The t_i−t display used the features in the t_i-t time period as the training features, the value at t_i+1 as the label and ti-t was called the window size. Supposing the data set was of length t including t₁, t₂… t_i …t_n. Where X_i was the pollutant PM_2.5 that we want to predict. If the window was t_i-t₁, then for sample 1 it had features between t₁, t₂… t_i as its sample feature, X_{i (ti+1)} as its label. For sample 2, it had t₂, t₃… t_i+1 as its features, and Xi _(t1+2) as its label. Sample 3 and Sample 4 were presented in a similar manner (Fig. A1).

Stations and features selection

The concentration of PM_2.5 in the vicinity, along with other predictive factors, plays a substantial role in influencing the PM_2.5 concentration at the anticipated location [22]. In the context of station analysis, a prevalent approach involves assessing the Pearson correlation coefficient of PM_2.5 levels across various monitoring stations [15, 23, 24]. This enables the selection of stations that are pertinent to the prediction task. Subsequently, the air quality data from these highly correlated stations are incorporated into the model alongside the data from the target stations, following methodologies outlined in prior studies [25, 26].

In this study, the selection process involved calculating the Pearson correlation coefficient values between each feature factor (PM₁₀, NO₂, SO₂, CO, O₃, AQI) and PM_2.5 concentration. This analysis aimed to discern the correlation between PM_2.5 concentration and individual features, facilitating the identification of the most optimal combination of feature factors for model training.

Model validation

Root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R²) indices were used here to evaluate the performance of the proposed model. RMSE, MAE, MAPE and R² are given as:

1

2

3

4

Where y_i represents the observed air pollutant concentration, represents the observed average air pollutant concentration, represents the predicted air pollutant concentration, and n represents the total number of samples.

Experimental design

Among the eight collected monitoring stations, WHL had the strongest correlation with neighboring stations in all the collected monitoring stations. Therefore, WHL was selected as the target monitoring station of the study. TYJ, XHY and XXJ were selected as auxiliary stations according to the Pearson correlation coefficient of PM_2.5 concentration at each monitoring stations in Shenyang from 2020 to 2022 (Fig. A2).

Two treatment groups were set up in this experiment, and each group was set up with two treatments. To ensure uniformity in comparisons, identical training, testing, and evaluation datasets were employed across both groups. The test methods were configured and structured as follows:

1. Control group

The control neural network contained three convolutional layers with 64, 64, and 32 filters with causal padding, successively, the size of the convolution filter was 3. All was followed by a MaxPooling1D layer with a pool size of 3. The last convolutional layer was linked to BILSTM, which contained two layers with 200 and 100 units per layer, and then it was linked to tree full connected layers with 100, 50, 50 neurons. Furthermore, the single neuron was used to output results.

• (2)Experimental group

The experimental group neural network structure was same as the control group. However, in the experimental network, the final layer of the convolutional neural network and the BILSTM results from each layer were interconnected with the PM_2.5 data of the target station. This integration produced a new feature, which was subsequently fed into the subsequent layer for further computation.

Training dataset was constructed by data from four stations (WHL, TYJ, XHY and XXJ), incorporating 7 feature factors for each site and totaling 28 feature factors. These were input into the control network, designated as CK-28, and the experimental network, labeled as Model-28. Simultaneously, the data from three stations (TYJ, XHY, and XXJ) contributed to the construction of another training dataset, encompassing 7 feature factors for each site and amounting to 21 feature factors. This subset was input into the control network as CK-21 and into the experimental network as Model-21.

Each iteration was trained for 100 epochs, the training set loss and test set loss of each epoch was recorded and a model was saved for each training epoch. After 100 training sessions, the model with the smallest test set loss cycle was retained as the result, and the above training was repeated for 4 times. The model with the smallest test set loss cycle in the 4 repetitions was selected as the best model for the processing. The development environment and hyperparameter for experiments were showed in Table A2 and Table A3.

Results and discussion

Hybrid model prediction and time lags selection

Compared with control models, the root mean square error (RMSE) of each treatment groups in single-step prediction of proposed model were significantly decreased (Fig. 3a). It revealed that the proposed structure significantly improved the prediction accuracy of the model. There was no significantly difference of RMSE for Model-21 and Model-28 (Fig. 3a), which showed that used the data of three stations to train the model was same as used the data of four stations. Our results showed that the new model structure can save the number of features input to the model.

Fig. 3.

The comparison of minimum root mean square error (RMSE) in each treatment group (a) and RMSE changes of each treatment group under different time lags(b)

In addition, The RMSE of Model-21 decreased first and then increased with the increase of time lag, and the time lag of 24 h had the lowest RMSE (Fig. 3b). This observation suggests that a shorter data length may provide insufficient effective information to the model, whereas an excessively long data input introduces a surplus of irrelevant information, diminishing the effectiveness of the input data [27].

The multi-step prediction RMSE of Molde-28 and Ck-28 showed that the proposed method improved the model performance not only in the next hour, but also in the next 12 h (Table 1). With the increase of prediction step size, the improvement of model performance decreases. The model performance was steadily improved within the 6 h, and the improvement rate fluctuated between 6 and 12 h.

Table 1.

The multi-step root mean square error of Model-28 and Ck-28 in 12 h time lag

Prediction steps	Ck-28 (ug/m³)	Model-28 (ug/m³)	Improvement (%)
1 h	5.74	2.94	48.7
2 h	8.18	7.30	10.8
3 h	11.28	11.21	0.6
4 h	13.60	13.58	0.2
5 h	16.20	14.74	9.1
6 h	17.81	16.17	9.2
7 h	18.22	19.55	−7.3
8 h	18.95	18.98	−0.1
9 h	19.44	19.65	−1.1
10 h	20.42	19.45	4.7
11 h	19.49	19.60	−0.6
12 h	20.04	19.99	0.3

Improvement denotes the improvement of Model-28 over Ck-28

Effect of multi-step and strong correlation factors on prediction

The strong correlation was defined by a correlation coefficient greater than 0.7 [28]. According to the Pearson correlation between each feature factor and PM_2.5 in WHL (Fig. A3). The correlation between PM₁₀ and AQI and PM_2.5 was extremely significant (P < 0.001). Therefore, PM_2.5, PM₁₀ and AQI were selected for model training, the three stations with the strongest correlation had total 9 feature factors. The 9 factors were input into the proposed model with time lag 24 h named Model-9.

The RMSE trends for both Model-9 and Model-21 exhibited a consistent upward trajectory, with Model-9 consistently demonstrating smaller errors compared to Model-21, as illustrated in Fig. 4a. It can be seen that the RMSE of the model can be reduced by screening the features of the input model. Meanwhile, the MAPE of Model-9 was also lower than that of Model-21 (Fig. 4b). The mean absolute error and coefficient of determination of Model-9 and Model-21 were showed in Fig. A4. Not all feature factors were helpful to the prediction, and the no or low correlation factors will lead to the deterioration of the model’s prediction performance [29]. The error of Model-9 was lower than that of Model-21, because the feature factors used by Model-9 were strongly related to PM_2.5. Some studies have showed that feature screening can effectively reduce model prediction errors [17].

Fig. 4.

The root mean square error (RMSE) (a) and mean absolute percentage error (MAPE) (b) of model-9 and model-21

To conduct a comprehensive evaluation of the proposed model in this study, the performance index of some researchers’ PM_2.5 prediction models based on deep learning theory in the next hour was collected. The comparison will be made from RMSE, MAE, MAPE and R² and it followed as Table 2. It was showed that most of the indexes of the proposed model were better than those of the references listed in the table.

Table 2.

The evaluate index of proposed method compared with the reference methods

Methods	RMSE (μg/m3)	MAE (μg/m3)	MAPE (%)	R2 (%)	Author/Year
LSTM	-	5.78	-	89.00	Gokul et al./ 2023 [30]
CNN-LSTM	19.09	-	-	92.90	Yang et al./ 2021 [31]
CNN-LSTM	7.75	11.77	-	-	Suriya et al./ 2023 [32]
Lag-FLSTM	3.48	1.85	25.63	-	Ma et al./ 2020 [33]
IDW-BILSTM	8.24	4.80	9.01	-	Ma et al./ 2019 [34]
EMD-SE-BILSTM	2.77	1.88	-	98.70	Jin et al./ 2020 [35]
MI-RI-BILSTM	12.16	8.51	-	-	Narkhede and Hiwale/ 2023 [36]
1D-CNN-BILSTM	3.88	2.52	-	94.00	Zhu and Xie/ 2023 [13]
this paper	2.54	1.72	11.72	98.81

Abbreviation: RMSE: Root mean square error. MAE: Mean absolute error. MAPE: Mean absolute percentage error. R2 : Coefficient of determination

Effect of different time lags on multi-step prediction

In this set of tests, time lags of 4 h to 72 h were used to predict the concentration of PM_2.5 in the next 1 h to 24 h by Model-9. In the 1–24 h predicted step size, the minimum RMSE did not all appeared in the optimal time lag 24 h (Table 3). This contrasts with previous studies utilizing hybrid deep learning models for PM_2.5 concentration prediction, where the time lag yielding the minimum RMSE for predicting the next hour also served as the optimal time lag for subsequent prediction steps [37]. The concentration of PM_2.5 varies significantly in different seasons, leading to instability in the concentration data [31]. The unstable data lead to the different results under the same training process [24]. Thus, the anticipated minimum RMSE at time lag 24 appears in different time lags due to the inherent instability in the data.

Table 3.

The minimum root mean square error of different time lag and prediction steps

	1 h	2 h	3 h	4 h	5 h	6 h	12 h	24 h
time lag 4	2.73	7.61	10.58	13.25	15.64	17.44	21.64	22.10
time lag 8	2.96	7.88	10.65	13.56	15.14	17.62	20.76	21.77
time lag 12	2.98	7.65	10.81	13.87	15.60	17.13	21.63	22.07
time lag 16	3.42	7.39	10.89	13.58	15.19	16.71	19.98	21.77
time lag 20	2.79	7.30	11.01	13.34	14.96	16.51	19.87	22.11
time lag 24	2.83	7.23	10.39	13.18	15.28	16.20	20.31	21.73
time lag 28	2.55	7.40	10.56	13.53	15.14	16.89	20.30	22.00
time lag 32	2.83	7.71	10.84	13.69	15.05	16.20	21.09	22.31
time lag 36	2.86	8.04	11.15	13.58	15.29	16.45	20.67	21.67
time lag 48	2.62	7.57	11.16	13.58	15.17	16.77	20.43	22.39
time lag 72	3.43	7.24	11.93	13.86	15.36	16.64	21.03	24.38

The bold values represent the minimum root mean square error in this step

The forecasting performance of Model-9 decreased as the forecast hours increased (Fig. A5). The 24 h time lag was optimal for most predicted step sizes. The average RMSE for the Model-9 forecast for the next 1–12 h was 14.93 ug/m³. Some studies believed that the predicted average RMSE of PM_2.5 concentration in the future 1–12 h was less than 20 ug/m³, which can be regarded as good prediction performance [25]. Therefore, the Model-9 with 24 h time lag showed excellent performance in short-term prediction.

The model presented in this study has unparalleled performance in short-term forecasting. However, the accuracy of long-term prediction of concentration of PM_2.5 also needs to be further improved. In addition, the trained model has not been validated with data in other regions. Future research will prioritize achieving accurate long-term predictions for PM_2.5 concentration and validating the model’s performance with data from diverse geographical regions.

Model generalization

In order to evaluate the PM_2.5 concentration prediction performance of Model-9 at other air quality monitoring stations. Data of the same time period and size were selected from 8 air quality monitoring stations and input into Model-9 for PM_2.5 concentration prediction. The predicted values of each air quality monitoring station were compared with the true values, and the predictive performance of Model-9 was evaluated using four indicators including RMSE, MAE, MAPE and R²(Table 4).

Table 4.

Performance of Model-9 for single-step prediction on different stations

stations	RMSE (µg/m3)		MAE (µg/m3)		MAPE		R 2
WHL	2.46	-	1.66	-	0.10	-	0.99	-
DLL	2.44	(−1.05%)	1.53	(−8.70%)	0.07	(−41.81%)	0.98	(−0.54%)
HNDL	3.33	(26.00%)	1.65	(−0.28%)	0.07	(−38.26%)	0.98	(−1.03%)
JSJ	2.80	(11.97%)	1.59	(−4.06%)	0.08	(−28.35%)	0.99	(−0.39%)
LDJ	2.39	(−2.99%)	1.42	(−16.69%)	0.10	(−1.34%)	0.99	(−0.09%)
TYJ	2.42	(−1.68%)	1.45	(−14.64%)	0.06	(−61.27%)	0.99	(−0.07%)
XHY	2.41	(−2.34%)	1.43	(−16.01%)	0.10	(−2.56%)	0.99	(−0.150%)
XXJ	2.54	(2.94%)	1.56	(−6.10%)	0.08	(−22.06%)	0.99	(−0.40%)

The values in parentheses represent the reduction rate of error between the data of this station and WHL data. Negative numbers indicate a decrease in prediction error, while positive numbers indicate an increase in prediction error. The calculation formula is: Reduction rate = ((other station data – WHL station data)/WHL station data) * 100%

Among numerous stations, when Model-9 was applied to the three air quality stations of HNDL, JSJ, and XXJ, their root mean square errors (RMSE) increased by 26.00%, 11.97%, and 2.94% respectively compared to the RMSE of the WHL air quality monitoring station. However, the RMSE of the model did not increased when applied to other stations. In addition, the three indicators of mean absolute error (MAE), mean percentage error of determination (MAPE), and coefficient of determination (R²) did not increase but decreased. The comparison results of these indicators indicate that Model-9 is suitable for PM_2.5 concentration prediction across all regions of Shenyang. This finding validates the robust generalization ability of our proposed model within the study area.

Conclusions

Our hybrid deep learning network based on 1D-CNN-BILSTM and structure of the combination of feature and the neural network layer output results was proposed to predict the concentration of PM_2.5 in Shenyang in the short term. The proposed model obtained more accurate prediction performance through the combination of feature and the neural network layer output results. Furthermore, the proposed structure saved 21.4% of input data to achieve the same predictive performance. In addition, screening the stations with higher PM_2.5 correlation coefficient and strong correlation feature factors also improved 4.5% of prediction accuracy. The proposed structure improved the prediction accuracy of the model in the next 6 hours.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data Availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declaration

CRediT authorship contribution statement

Yin Zhou: Conceptualization, Methodology, Software, Writing - Original Draft. Yaping Lyu: Writing - Review & Editing. Xiuli Dang: Supervision, Project administration, Resources, Writing - Review & Editing. Roland Bol: Writing - Review & Editing. Peng Zhang: Writing - Review & Editing. Na Yu: Formal analysis. Yuling Zhang: Visualization.

Competing Interests

The authors have no relevant financial or non-financial interests to disclose.