Introduction

General background of study

Climate change has been a significant issue recently due to its impact on weather conditions and land temperatures, which individuals or natural phenomena can cause. The main causes of air pollution include coal and gas combustion for power generation, transportation, industrial and residential developments1. Global warming is caused by greenhouse gases (GHG), while climate change is caused by global warming. Emerging nations such as India are facing numerous issues related to air pollution and its detrimental environmental and public health consequences2. India is experiencing substantial concerns regarding air quality degradation, such as the massive population growth, industrial companies, the use of fossil fuels for power generation, poor agricultural methods, and motor vehicle emissions3,4. Some particulate matter and gaseous pollutants are produced directly from the source and cause air pollution such as in the size of 2.5 and 10 microns (PM2.5 and PM10), carbon dioxide (CO2), Sulphur dioxide (SO2), Nitrogen oxide (NO2), Ammonia (NH3), benzene, volatile organic compounds (VOCs), carbon monoxide (CO), and ozone (O3)5. The initial air pollutants are used in the general equation to calculate the air quality index (AQI), depending on the geographical region. Besides, these pollutants are the cause of the following issues such as air pollution, depletion of ozone layer, global warming, increased average land temperature, climate change, and acid rain6. The air pollutants concentration changes based on the main atmospheric factors such as precipitation (snowfall, rain, sleet or ice pellets, drizzle, hail, freezing rain, frost, and rime), wind speed (WS), wind direction (WD), relative humidity (RH), solar radiation (SR), and air temperature (T)7. In this context, AI approaches such as machine learning and deep learning can be developed using various datasets collected from different monitoring sites and stations to create strong connections between inputs and outputs.

Literature review

As per the open literature8,9, complex AI models were developed by various researchers to predict the air pollutant levels10,11. For instance, the air quality index of four stations such as New Delhi, Bangalore, Kolkata, and Hyderabad, was predicted by using three different regression algorithms such as support vector regression (SVR), random forest (RF), and CatBoost (CR)12. The RF model showed the lowest root mean square error (RMSE) in Bangalore, Kolkata, and Hyderabad, with a value of 0.5674, 0.1403, and 0.3826, respectively. Meanwhile, the CR model obtained the lowest RMSE value (0.2792) in New Delhi. Additionally, the CR was the superior model in terms of accuracy in New Delhi (R2 = 79.8622%) and Bangalore (R2 = 68.6860%), respectively. Two different approaches such as deep learning and Holt-Winters statistical model were compared to predict the PM2.5 and PM10 concentrations13. The Holt-Winters statistical model exhibited lower RMSE and MSE values than the deep learning model. Also, the PM2.5 concentration was forecasted using LSTM, which was integrated to the particle swarm optimization (PSO) method14. They collected their datasets from 15 monitoring stations and achieved a strong accuracy ranging from R2 = 0.86 to R2 = 0.99. Also, the training and testing sets showed an average percentage error over 15 locations of about 6.6% and 6.9%, respectively. In Chennai, the AQI values were classified using a novel expert system integrated from support vector regression (SVR) and LSTM algorithms15. Their novel expert system showed a superior performance with R2 = 0.97 and RMSE = 10.9.

A case study in Delhi, a dataset was collected from January 2018 to October 2021 to predict the PM2.5 concentration using deep learning model that combined neural networks, fuzzy inference systems (ANFIS), and wavelet transforms16. Their novel model showed an outstanding accuracy, such as: 0.95 < R2 < 0.99 for 1-day short term prediction, 0.85 < R2 < 0.94 for 2-day short term prediction, and 0.81 < R2 < 0.93 for 3-day short term prediction. A novel hybrid model from GRU and LSTM was developed to predict PM2.5 concentration in Delhi17. Also, the dataset was validated using five different standalone models, such as LSTM, linear regression (LR), GRU, K-Nearest Neighbour (KNN), and support vector machine (SVM). The LSTM-GRU was superior to the individual models with an MAE-value of 36.11 and R2-value of 0.84. Moreover, the AQI values were predicted in Chennai using historical data from meteorological locations from 2017 to 20226. The authors developed four different tree-based models such as XGBoost, RF, Bagging Regressor, and LGBM. XGBoost model showed the following prediction metrics: R2 = 0.9935, MAE = 0.02, MSE = 0.001, and RMSE = 0.04. Twelve pollutants and ten meteorological parameters from July 2017 to September 2022 were collected over Visakhapatnam, Andhra Pradesh, India18. This dataset was used to estimate the AQI value using five models such as LightGBM, RF, CatBoost, Adaboost, and XGBoost. The CatBoost model showed the following metrics: R2 = 0.9998, MAE = 0.60, MSE = 0.58, and RMSE = 0.76. Meanwhile, the Adaboost model presented the following metrics with an R2 = 0.9753. Encoder-Decoder (ED) layers were connected to GRU deep learning model for predicting 1-hour, 8-hour, and 24-hour of PM2.5 concentrations in New Delhi, India, and the dataset was collected from 2008 to 201019. The hybrid method showed superior performance over the standalone models such as (RF, XGBoost, ANNs, and LSTM). Standalone versus stacking models were used to estimate 1-hr and 24-hr PM2.520. XGBoost showed higher accuracy than RF and LightGBM with R2 = 0.73. Meanwhile, stacked model (XGBoost as a meta-regressor) improved the accuracy of standalone XGBoost with R2 = 0.77. The eastern region exhibited the best 1-hr prediction with R2 = 0.80 and substantial reduction in Mean Bias (MB = − 0.03 µg m− 3), followed by the northern region with R2 = 0.63 and MB = − 0.10 µg m− 3. In Chandigarh, eight AI models such as RF, KNN, LR, LASSO regression, Decision Tree (DT), SVR, XGBoost, and Deep Neural Network (DNN) with 5-layers were used to predict 24-hr air pollution and outpatient visits for Acute Respiratory Infections (ARI)21. On ARI patients, the RF model performed best, with R2 = 0.606, 0.608 without lag, and 24-hr lag, respectively. Also, on total patients, R2 = 0.872, 0.871 without lag, and 24-hr lag, respectively.

China and India, covering 35% of the global population, face widespread urban air pollution, and real-time and remote sensing assessments remain insufficient. Air quality issues in these rapidly growing economies are increasingly being addressed by the application of machine and deep learning in recent studies. Six models i.e., MLR, SVR, RF, ANN, XGBoost, and LSTM, were used to predict LST for Hyderabad city, India using five-year (2018–2022) data on air pollution and meteorological parameters (from ambient air quality monitoring stations) and MODIS LST data22. Considerable influence of PM2.5 and CO (during summer) and SO2 (during winter) on LST was observed which demonstrated high sensitivity of these parameters on LST. ANN method demonstrated better accuracy with lower error metrics, comprising of RMSE, MAPE, and MSE, compared to the other approaches with ranking in the order ANN > RF > SVR > XGBoost > LSTM > MLR. Based on hourly observations from  2018 in India, integrating temporal and regional features into the LightGBM model led to a notable enhancement in its performance, achieving a 21% reduction in RMSE for PM2.5 estimation and a 19% reduction for PM1023. The ML model predicted an annual nationwide concentration of 68.3 µg/m3 for PM2.5, which was consistent with high satellite aerosol optical depth (AOD) values. A real-time assessment of hazardous atmospheric pollutants across cities in China (Shanghai, Nanjing, Jinan, Zhengzhou and Beijing) and India (Kolkata, Asansol, Patna, Kanpur and Delhi) was conducted using ground observations, Sentinel-5P and NASA satellite data from 2012 to 202324. GMAO’s SO2, NO2 and CO predictions showed high accuracy with near-perfect PC values and low NRMSE proving model reliability. A Unified Spectro-Spatial Graph Neural Network (USS-GNN) designed for forecasting O3-NO2 concentrations for New Delhi, utilized hourly observations for the years 2021 and 202225. The proposed model achieved R2 values of 0.650 and 0.618, RMSE of 13.950 and 16.120 µg/m3, MAE of 10.730 and 12.930 µg/m3 for O3 and NO2 , respectively. Different artificial intelligence models were proposed to simulate climate parameters (1 January 1951–31 December 2022) of Jinan city in China, include ANN, RNN, LSTM, CNN, and CNN-LSTM26. The hybrid CNN-LSTM model significantly reduced the forecasting error compared to the models for the one-month time step ahead. The RMSE values of the ANN, RNN, LSTM, CNN, and CNN-LSTM models for monthly average atmospheric temperature in the forecasting stage were 2.0669, 1.4416, 1.3482, 0.8015 and 0.6292 °C, respectively.

Research objectives and novelty

This study addresses the critical need for an innovative approach to air quality forecasting, focusing on predicting concentrations of eight key air pollutants that significantly impact human health, environmental integrity, and everyday life. The primary objectives are twofold: (1) to develop accurate forecasting models for Carbon Monoxide (CO), Nitrogen Monoxide (NO), Nitrogen Dioxide (NO₂), Ozone (O₃), Sulphur Dioxide (SO₂), Fine Particulate Matter (PM₂.₅), Coarse Particulate Matter (PM₁₀), and Ammonia (NH₃); and (2) to rigorously assess and compare the predictive performance of five advanced standalone artificial intelligence (AI) models tailored for univariate pollutant forecasting.

The novelty of this research lies in its comparative analysis across five AI models—Long Short-Term Memory (LSTM), Bidirectional Long-Short Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost)—in forecasting hourly pollutant concentrations in a high-pollution urban setting. Using extensive data from monitoring stations in Delhi, collected between 25/11/2020 and 24/01/2023, the study performs in-depth statistical analysis to explore patterns and dynamics of each pollutant. Through the use of scatter plots, Taylor diagrams, and forecasting performance matrices, this study provides comprehensive insights into the distinct strengths and limitations of each model.

In highlighting model efficiency and comparing forecast accuracy metrics, this study provides valuable guidance for advancing air quality management strategies. The findings provide practical implications for urban planning and environmental policy, demonstrating that accurate AI-driven forecasting models can be instrumental in environmental mitigation efforts and health risk assessments in rapidly urban areas.

Data collection and description

As outlined above, the datasets of eight air pollutants were collected from the capital city of Delhi, India, spanning from 25/11/2020 to 24/01/2023 including the following elements: Carbon Monoxide (CO), Nitrogen Monoxide (NO), Nitrogen Dioxide (NO2), Ozone (O3), Sulphur Dioxide (SO2), Fine Particles Matter (PM2.5), Coarse Particles Matter (PM10), and Ammonia (NH3). The dataset contains features recorded at an hourly interval, with 18,776 samples. It appears that the datasets are complete, and no missing values were observed. Table 1 shows summary statistics for the eight air pollutants, using various measurements such as Mean, Standard Deviation, Variance, Skewness, Kurtosis, Coefficient of Variation, Median, Interquartile Range (Q3 - Q1), Range (Maximum - Minimum), Median Absolute Deviation, and Robust Coefficient of Variation. Besides, Fig. 1 shows the numerical distributions of eight air pollutants through boxplot formats. Boxplots highlighted extreme outliers that significantly changed the common relationship of the dataset. In this regard, the Winsorizing technique was implemented to manage outliers by replacing extreme values with values closer to the upper or lower limits, resulting in a more robust and reliable analysis of the air quality data.

Table 1 The statistical properties of air pollutants dataset.
Full size table
Fig. 1
figure 1

Boxplots distributions of eight air pollutants.

Full size image

Applied artificial intelligence models

This section outlines the five models employed for predicting air pollutant concentrations. The development and configuration processes of the models are then reviewed. Finally, the section concludes with a summary of the forecasting metrics used to evaluate the accuracy of the model predictions.

Multilayer perceptron (MLP)

An MLP is a fundamental component of ANNs27. It can model complex systems in engineering and data sciences. An MLP is characterized by the number of layers of interconnected neurons and the number of neurons per layer. Every layer transforms its input data and feeds it to the next layer in a purely feedforward fashion, contributing to the network’s decision-making process28. Figure 2a illustrates an MLP with an input layer, a single hidden layer, and an output layer. Every neuron implements a weighted sum with a bias term fed to an activation function, e.g., tanh. Figure 2b shows the structure of a neuron29.

Fig. 2
figure 2

(a) An example architecture of an MLP consisting of an input layer, a single hidden layer, and an output layer of neurons (b) An example architecture of a neuron with two inputs and tanh activation function.

Full size image

The architecture of an MLP, including the number of layers and neurons in every layer, significantly affects its performance. Additionally, the use of hyperparameters, such as the learning algorithm, is crucial in training MLPs to minimize their prediction errors effectively. The objective of the learning algorithm is to solve for the weights and biases that minimize an objective function30. MLPs are capable of capturing non-linear data, making them suitable for various applications, including air pollutant prediction. One crucial property of MLPs is that the signals propagate from the inputs to the outputs in one direction, i.e., MLPs do not contain feedback loops. This property of MLPs limits them from capturing long-lasting relationships of time series. This is precisely why the prediction error rates are higher than the other models.

Long short-term memory (LSTM)

To understand the improvements offered by Long Short-Term Memory (LSTM) networks, the limitations of recurrent neural networks (RNNs) are first examined, as these are addressed by the LSTM architecture. RNNs contain both feedforward and feedback loops that facilitate capturing complicated temporal dynamics of sequential time series data27. The main limitation of RNNs is that they fail to capture long-term dependency in the time series, and this is due to the vanishing and exploding gradients27.

Schmidhuber and Hochreiter31 proposed the LSTM model in 1997 to solve the issue of vanishing and exploding gradients in RNNs. One of the proposed ideas is the cell state, which is intended to enhance the hidden state and to propagate the effect of some data through longer sequences. Also, LSTM introduces three elements that are intended to filter the data flowing through the network in such a way to prevent the vanishing and exploding gradients, and thus maintain that series of long-term dependence. These elements are known as input gate, forget gate, and output gate, as shown in Fig. 3. In particular, the forget gate selects what information to remember or forget. The value of the forget gate is a value between 0 and 1, where 0 means to forget the information and 1 means to remember, and a value in between means partially forget/remember.

Fig. 3
figure 3

Detailed LSTM Cell Architecture. Arrows indicate the flow of information, including the feedback loops for the cell state (\(\:{C}_{t}\)) and hidden state (\(\:{h}_{t}\)). Note that the summation nodes include bias terms.

Full size image

Bidirectional long-short term memory (Bi-LSTM)

As proposed by Schuster and Paliwal in32, Bi-LSTM operates two instances of LSTM, one for processing the sequence in the forward direction and the other for processing the sequence in the reverse direction, enhancing the model to learn from the sequence in both directions. Bi-LSTM then combines the responses of both directions to generate a unified output sequence that can be used to predict data more accurately33.

Gated recurrent unit (GRU)

GRU improves LSTM by reducing the number of gates from three to two34. One gate is the update gate, and the other is the reset gate. The primary objective of these gates is to capture the long-term dependency between the data elements of a time series. The main objective of the update gate is to determine how many dependencies to remember, whereas for the reset gate, the main objective is to determine how many dependencies to forget. GRU utilizes a sequence of computations to generate the new hidden state, given the input and the current hidden state. This sequence of computations, i.e., Eqs. (1)–(4), shown in the following table, is then repeated in the form of a loop, to improve the accuracy of GRU to predict new data and to effectively capture the long-lasting relationship between the elements of the time series.

The update gate decides how much to remember

$$\:{u}^{\left(t\right)}={\upsigma\:}\left({W}_{uh}{h}^{\left(t-1\right)}+{W}_{ux}{x}^{\left(t\right)}+{b}_{u}\right)$$
(1)

The reset gate decides how much to forget

$$\:{r}^{\left(t\right)}={\upsigma\:}\left({W}_{rh}{h}^{\left(t-1\right)}+{W}_{rx}{x}^{\left(t\right)}+{b}_{r}\right)$$
(2)

The candidate hidden state computation according to reset gate feedback

$$\:\stackrel{\sim}{{h}^{\left(t\right)}}=\text{tanh}\left({W}_{hc}\left({r}^{\left(t\right)}\odot\:{h}^{\left(t-1\right)}\right)+{W}_{hx}{x}^{\left(t\right)}+{b}_{h}\right)$$
(3)

The new hidden state computation, according to feedback from the update gate and candidate hidden state

$$\:{h}^{\left(t\right)}=\left(1-{u}^{\left(t\right)}\right)\odot\:{h}^{\left(t-1\right)}+{u}^{\left(t\right)}\odot\:\stackrel{\sim}{{h}^{\left(t\right)}}$$
(4)

Extreme gradient boosting (XGBoost)

All the previous models are proposed to optimize the performance of the gradient boosting machine (GBM)35, XGBoost integrates a number of weak learners to form an efficient predictive model as illustrated in Fig. 436. XGBoost generates k decision trees and validates them using a subset of the training data such that the model generalizes well to new data. Each decision tree creates a residual, i.e., the error in prediction of the current decision tree, and this residual becomes the target variable for prediction in the next decision tree. XGBoost has been proven to be effective in predictions requiring high levels of accuracy, e.g., see28,37,38,39,40,41,42,43.

Fig. 4
figure 4

XGBoost architecture consists of k decision trees that iteratively improves the prediction accuracy.

Full size image

Models development and configuration

Figure 5 shows the development and configuration process. The implementation platform was Google Collaboratory, utilizing Python and powered by GCC 9.4.0. Many libraries were utilized to complete this project, including NumPy, Pandas, Scikit-learn, and the APIs for TensorFlow Keras. The input dataset contains the hours of concentrations of eight air pollutants in Delhi between November 25th, 2020, and January 24th, 2023. Prior to any actions being executed on the input dataset, it was standardized using scikit-learn and then split into 80–20% training and testing sets.

The training set uses a filter to substitute outliers by values that are closer to the upper or lower bound of the dataset when these outliers are excluded. After that, the filtered data is processed to each of the five models to complete the learning process. The test data is utilized to evaluate the accuracy of the trained models by generating a set of six performance metrics, such as RMSE, R2, MAE, MSE, standard deviation and WI .

Prior to the training, hyperparameter optimization and tuning were conducted to maximize the performance of each model. This tuning process was conducted using scikit-learn algorithms specifically developed for such purposes. The GridSearchCV algorithm was employed for neural network-based models, while for the XGBoost model, the RandomizedSearchCV algorithm was employed. The optimized hyperparameters for all models are presented in Table 2.

Fig. 5
figure 5

Models development and configuration.

Full size image
Table 2 Hyperparameters, along with the best values, were tested for LSTM, Bi-LSTM, GRU, ANN-MLP, and XGBoost models.
Full size table

Forecasting metrics

There are several statistical performance metrics were adopted for the forecasting accuracy evaluation for each developed predictive model. The mathematical expression of each metric are as follows:

Coefficient of Determination (R2)

$$\:{R}^{2}\left(y,\widehat{y}\right)=1-\frac{{\sum\:}_{i=1}^{N}{\left({y}_{i},\widehat{y}\right)}^{2}}{{\sum\:}_{i=1}^{N}{\left({y}_{i},\stackrel{-}{y}\right)}^{2}}$$
(5)

Root Mean Square Error (RMSE)

$$\:RMSE\left(y,\widehat{y}\right)=\sqrt{\frac{{\sum\:}_{i=0}^{N-1}{\left({y}_{i},\widehat{y}\right)}^{2}}{N}}$$
(6)

Mean Absolute Error (MAE)

$$\:MAE\left(y,\widehat{y}\right)=\frac{{\sum\:}_{i=0}^{N-1}\left|{y}_{i},{\widehat{y}}_{i}\right|}{N}$$
(7)

Mean Squared Error (MSE)

$$\:MSE\left(y,\widehat{y}\right)=\frac{{\sum\:}_{i=0}^{N-1}{\left({y}_{i},{\widehat{y}}_{i}\right)}^{2}}{N}$$
(8)

Standard Deviation (σ)

$$\:{\upsigma\:}=\frac{\sqrt{{\sum\:}_{i=1}^{N}{({y}_{i}-\mu\:)}^{2}}}{N}$$
(9)

Willmott Index (WI)

$$\:WI\left(y,\widehat{y}\right)=1-\frac{{\sum\:}_{i=0}^{N-1}{\left({\widehat{y}}_{i},{y}_{i}\right)}^{2}}{{\sum\:}_{i=0}^{N-1}{(\left|{\widehat{y}}_{i}-mean\left(y\right)\right|+\left|{y}_{i}-mean\left(y\right)\right|)}^{2}}$$
(10)

where \(\:\widehat{y}\) is the predicted air pollutant, \(\:y\) is the actual air pollutant, \(\:mean\left(y\right)\) is the mean of actual target output, \(\:{\upsigma\:}\) is the Standard Deviation, \(\:\mu\:\) is the air pollutant mean, \(\:N\) is the number of observations.

Results and discussion

This section presents the results and analysis of various ML models (LSTM, Bi-LSTM, GRU, MLP, and XGBoost) applied to forecast concentrations of eight airborne particles: Carbon Monoxide (CO), Nitrogen Monoxide (NO), Nitrogen Dioxide (NO₂), Ozone (O₃), Sulphur Dioxide (SO₂), Fine Particulate Matter (PM₂.₅), Coarse Particulate Matter (PM₁₀), and Ammonia (NH₃). Each model was evaluated both graphically and statistically. To assess the predictive performance, a Taylor diagram was used to compare the models based on three key statistical metrics—correlation, RMSE, and standard deviation—in relation to the benchmarked observational dataset. Additionally, scatter plots were generated to visualize the deviations between actual and forecasted values along the 1:1 line.

Carbon monoxide (CO)

Carbon monoxide was forecasted due to its highly poisonous concern for human health44. Figure 6 presents a scatter plot and Taylor diagram along with the forecasting metrics. According to the scatter plot, the CO concentration was varying between (0-23000). In general, all models showed a good variation around the 45° line. However, MLP and XGBoost revealed some observations from the identical line. According to Fig. 6a, the obtained determination coefficient based on the regression scatter formula was (LSTM = 0.97, Bi-LSTM = 0.97, GRU = 0.97, MLP = 0.94 and XGBoost = 0.96). Moreover, the Taylor diagram revealed that the observational dataset of the CO was located at the standard deviation of 1 as per Fig. 6b. Overall, Bi-LSTM surpassed others with the highest R2 (0.979), lowest RMSE (438.898), and lowest MAE (244.543), indicating precise and reliable predictions (Fig. 6c). GRU also performed well, with an R² of 0.977 and RMSE of 456.729. LSTM followed closely with an R² of 0.971. XGBoost and MLP had a relatively lower accuracy, with MLP having the lowest R2 (0.941) and the highest RMSE (740.927). Therefore, Bi-LSTM demonstrated the best performance, making it the most effective model for accurate air quality forecasting in this research.

Fig. 6
figure 6

Actual and forecasted values of CO using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Nitrogen monoxide (NO)

NO is an extremely reactive gas that is initiated during high-temperature fuel burning. It is emitted by automobiles and non-road vehicles (e.g., boats and construction equipment). Breathing NO with a high concentration can cause respiratory diseases such as asthma, which could lead to respiratory infections44. According to the scatter plots visualization (Fig. 7a), the proposed models differed from the identical best-fit line. However, LSTM, GRU, and MLP demonstrated some observation scatters from a similar line. Numerically, the attained determination coefficient based on the regression scatter formula was (LSTM = 0.94, Bi-LSTM = 0.96, GRU = 0.95, MLP = 0.91 and XGBoost = 0.94). In addition, based on the Taylor diagram presentation (Fig. 7b), the observational dataset of the NO was located at the standard deviation of 1. In this comparison of air quality forecasting models (Fig. 7c), Bi-LSTM emerged as the top performer with the highest R² (0.961), lowest RMSE (13.922), and MAE (7.854), indicating accurate and consistent predictions. GRU and LSTM followed, with R² values of 0.956 and 0.948, respectively, and slightly higher RMSEs and MAEs. XGBoost performed well, but had a slightly lower accuracy (R2 = 0.943). The lowest performance was achieved by MLP, with an R2 of 0.913 and the highest RMSE (20.852). Overall, Bi-LSTM proved the most reliable model for precise air quality forecasting in this study.

Fig. 7
figure 7

Actual and forecasted values of NO using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Nitrogen dioxide (NO2)

Increases in mortality and hospital admissions for respiratory diseases are also associated with air nitrogen dioxide concentrations44. The lungs’ ability to combat microorganisms can be compromised by nitrogen dioxide, leaving the air more vulnerable to illnesses. Additionally, it may cause asthma to be worse. However, LSTM, GRU, and MLP revealed some observation scatters from the identical line. According to (Fig. 8a), the attained determination coefficient based on the regression scatter formula was (LSTM = 0.95, Bi-LSTM = 0.95, GRU = 0.94, MLP = 0.94, and XGBoost = 0.94). The Taylor diagram in Fig. 8b indicated that the observational dataset of the NO2 was located at the standard deviation of 0.90 to 1. In this analysis of air quality forecasting models (Fig. 8c), Bi-LSTM achieved the highest R2 (0.956) and the lowest RMSE (11.055) and MSE (122.215), suggesting its superior accuracy. Also, LSTM performed well, with an R² of 0.953 and a similar RMSE (11.452), making it a strong contender. GRU followed closely, though with slightly lower accuracy (R² = 0.946) and higher error metrics. XGBoost and MLP had a lower accuracy, with MLP having the lowest R2 (0.937) and the highest RMSE (13.275). Overall, Bi-LSTM provided the most precise and consistent air quality predictions in this comparison.

Fig. 8
figure 8

Actual and forecasted values of NO2 using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Ozone (O3)

Many health problems, such as congestion, coughing, throat irritation, and chest pain, can be caused by breathing in ground-level ozone44. Bronchitis, asthma, and emphysema can all become worse due to it. Additionally, ozone can irritate the lining of the lungs and impair lung function. Long-term lung tissue scarring could result from repeated exposure. Increased vulnerability to diseases, pests, and other stressors such as severe weather results from elevated ozone levels, which also reduce the yields of commercial forests and crops and the growth and survival of tree seedlings. As per Fig. 9a, XGBoost and MLP revealed some observation scatters from the identical line. Numerically, the attained determination coefficient based on the regression scatter formula was (LSTM = 0.95, Bi-LSTM = 0.96, GRU = 0.96, MLP = 0.95, and XGBoost = 0.96). The observational dataset of the O3 was located at the standard deviation of 1 according to Taylor diagram in Fig. 9b. In this air quality forecasting comparison (Fig. 9c), both GRU and XGBoost excelled with the highest R² values (0.962). GRU achieved a lower RMSE (12.050) and MSE (145.191), while XGBoost had the lowest MAE (6.409), indicating excellent accuracy and minimal error. Bi-LSTM and LSTM also performed well with R² values of 0.959 and 0.958, respectively, but with slightly higher RMSE and MAE values. MLP, while accurate (R² = 0.957), showed the highest RMSE (12.869). In this part, GRU and XGBoost proved the most effective models for accurate and consistent air quality predictions.

Fig. 9
figure 9

Actual and forecasted values of O3 using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Sulphur dioxide (SO2)

The combustion of sulphur-containing fuels releases sulphur dioxide (SO₂), impacting ecosystems and human health. SO₂ harms plants, streams, and forests, and in humans, it worsens respiratory conditions like asthma and bronchitis, especially during exercise44. Studies also link SO₂ exposure to higher cardiovascular disease risks. However, XGBoost and MLP revealed some observation scatters from the identical line. As per Fig. 10a, the attained determination coefficient based on the regression scatter formula was (LSTM = 0.95, Bi-LSTM = 0.95, GRU = 0.95, MLP = 0.92, and XGBoost = 0.91). Based on the Taylor diagram in Fig. 10b, the observational dataset of the SO2 was located at the standard deviation of 1 except MLP (SD = 0.86). In this comparison of air quality forecasting models (Fig. 10c), Bi-LSTM achieved the highest performance with the highest R2 (0.955), the lowest RMSE (11.110), and MAE (6.076), indicating superior accuracy and low error. LSTM also performed well with an R² of 0.950 and slightly higher RMSE (11.648). GRU followed closely, with similar metrics (R² = 0.949). The accuracy of MLP and XGBoost was lower, with MLP achieving an R2 of 0.928 and XGBoost the lowest R2 (0.916) and highest RMSE (15.214). In conclusion, Bi-LSTM was the most effective model for precise air quality predictions in this study.

Fig. 10
figure 10

Actual and forecasted values of SO2 using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Fine particles matter (PM2.5)

Many scientific investigations have demonstrated that particulate matter decreases visibility and negatively impacts materials, ecosystems, and climate. PM, particularly PM2.5, alters how light is absorbed and scattered in the atmosphere, which can affect visibility45,46. Long-term exposure to fine particles may also increase the risk of heart disease and be linked to a higher incidence of chronic bronchitis, deteriorated lung function, and lung cancer, according to studies. However, XGBoost and Bi-LSTM revealed some observation scatters from the identical line. According to Fig. 11a, the attained determination coefficient based on the regression scatter formula was (LSTM = 0.97, Bi-LSTM = 0.97, GRU = 0.96, MLP = 0.93, and XGBoost = 0.96). The Taylor diagram illustration (Fig. 11b) demonstrated that, the observational dataset of PM2.5 was located at the standard deviation of 1.05. In this analysis of air quality forecasting models (Fig. 11c), LSTM demonstrated the best performance with the highest R² (0.973), the lowest RMSE (37.549), and MAE (22.183), indicating strong accuracy and precision. Bi-LSTM closely followed with an R² of 0.971 but had a slightly higher RMSE (38.519). GRU also performed well, with an R² of 0.969 and comparable metrics. XGBoost achieved a satisfactory R2 of 0.961 but with higher RMSE (45.189) and MAE (25.134). MLP had the lowest overall performance, with an R² of 0.936 and significantly higher RMSE (57.665) and MAE (37.627). In this study, LSTM emerged as the most effective model for accurate air quality predictions.

Fig. 11
figure 11

Actual and forecasted values of PM2.5 using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Coarse particles matter (PM10)

Particle size contributes to the health and ecological impacts of particulate matter (PM). PM10 particles (10 micrometer or smaller) can reach the lungs upon inhalation, resulting in serious health risks to heart and lung health. Additionally, PM deposition affects ecosystems by impairing water quality and altering plant growth, especially due to the metal and organic compounds within PM47. However, XGBoost revealed some observation scatters from the identical line. As per Fig. 12a, the attained determination coefficient based on the regression scatter formula was (LSTM = 0.97, Bi-LSTM = 0.97, GRU = 0.97, MLP = 0.94, and XGBoost = 0.96). Also, Taylor diagram showed in Fig. 12b that, the observational dataset of PM10 was located at the standard deviation of 1.05. In this comparison of air quality forecasting models (Fig. 12c), Bi-LSTM achieved the highest performance with the highest R2 (0.975), lowest RMSE (43.126), and MAE (24.414), indicating excellent accuracy. LSTM and GRU were closely followed, both with R2 values of 0.974, but slightly higher RMSE and MAE values. XGBoost had a moderate performance, with an R2 of 0.962 but a higher RMSE (53.085). MLP showed the lowest accuracy with an R² of 0.949 and the highest RMSE (61.782). Overall, Bi-LSTM proved to be the most accurate model for air quality prediction.

Fig. 12
figure 12

Actual and forecasted values of PM10 using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Ammonia (NH3)

Ammonia is a major contributor to nitrogen pollution. The effects of nitrogen buildup on plant species diversity and composition within impacted environments are an essential component of ammonia pollution’s impacts on biodiversity. Ammonia can cause irritation to the skin or eyes or irritation to them when it comes into contact with it. If ammonia is inhaled, it can cause coughing, wheezing, and shortness of breath, as it can cause irritation to the respiratory system. Furthermore, ammonia inhalation may cause irritation to the throat and nose. However, XGBoost revealed some observation scatters from the identical line. According to Fig. 13a, the attained determination coefficient based on the regression scatter formula was (LSTM = 0.96, Bi-LSTM = 0.97, GRU = 0.96, MLP = 0.93, and XGBoost = 0.93). As per Fig. 13b, Taylor diagram showed that the observational dataset of NH3 was located at the standard deviation of 1.05. In this assessment of air quality forecasting models (Fig. 13c), Bi-LSTM demonstrated the highest performance with an R2 of 0.971, the lowest RMSE (5.283), and the lowest MAE (2.743), indicating superior accuracy and minimal error. LSTM and GRU were followed closely, with R2 values of 0.963 and 0.965, respectively, but slightly higher RMSE and MAE scores. The accuracy of MLP and XGBoost was lower, with MLP having an R2 of 0.939 and the highest RMSE (7.645), while XGBoost had the lowest R2 (0.933) and a high RMSE (8.036). In conclusion, Bi-LSTM was the most effective model for precise air quality predictions in this study.

Fig. 13
figure 13

Actual and forecasted values of NH3 using different AI-models; (a) Scatter plots, (b) Taylor diagram, (c) Forecasting metrics.

Full size image

Exposure to air pollution is a global public health hazard, with a considerable body of evidence linking short-term and long-term exposures to a range of health outcomes, including all-cause and cause-specific mortality, respiratory and cardiovascular conditions, neurodevelopmental deficiencies, and adverse pregnancy and birth out-comes44. The current research was fueled by the robustness of deep learning models that predictability performance of recent research development on machine learning establishment, deep learning revealed superior performance to the other ML models. The eight air quality parameters were chosen due to their seriousness as environmental and atmospheric indices. In this regard, the study’s unique methods for predicting air pollution provide valuable results in various circumstances. Models that are considered to be suitable and beneficial for daily prediction include LSTM, Bi-LSTM, GRU, MALP and XGBoost. In order to discuss the reliability and validity of the current intelligence models, Table 3 summarizes the latest air quality studies using multiple machine and deep learning models under different meteorological conditions.

Table 3 Summary of previous studies on air pollutant prediction using intelligent learning approaches across India.
Full size table

This study’s significance lies in its model variety, particularly Bi-LSTM, which achieved higher R2 values than other methods. Comparatively, previous studies achieved comparable R2 values but were limited to fewer pollutants or simpler seasonal analyses. In addition, studies in Hyderabad (2022) and Kolkata (2019–2022) showed strong results with models such as ANN and XGBoost, but focused on fewer pollutants and seasonal variations. In contrast, the Delhi study utilized Bi-LSTM, LSTM, and GRU extensively across pollutants and extended the model’s applicability across multiple seasons, enhancing predictive power for dynamic air quality conditions in a complex urban area. The high performance and adaptability of the Bi-LSTM model across pollutants highlight its potential as a robust choice for multi-pollutant air quality forecasting, enabling policymakers and environmental agencies to make data-driven decisions and mitigate pollution’s impact on public health.

Conclusion, limitations and future directions

In this study, different air quality parameters were proposed, including CO, NO, NO2, O3, SO2, PM2.5, PM10, and NH3. Datasets were collected for monitoring stations located near Delhi, India, for the duration of (25-11-2020/24-01-2023) with an hourly rate. For this purpose, various AI models were introduced, including Long Short-Term Memory (LSTM), Bidirectional Long-Short Term Memory (Bi-LSTM), Gated Recurrent Unit (GRU), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). The following findings can be drawn from the current study:

  1. i.

    In CO forecasting, the AI models showed the following accuracy: Bi-LSTM = 0.979, GRU = 0.977, LSTM = 0.971, XGBoost = 0.963, and MLP = 0.941, respectively.

  2. ii.

    In NO forecasting, the AI models showed the following accuracy: Bi-LSTM = 0.961, GRU = 0.956, LSTM = 0.948, XGBoost = 0.943, and MLP = 0.913, respectively.

  3. iii.

    In NO2 forecasting, the AI models showed the following accuracy: Bi-LSTM = 0.956, GRU = 0.953, LSTM = 0.946, XGBoost = 0.940 and MLP = 0.937, respectively.

  4. iv.

    In O3 forecasting, the AI models showed the following accuracy: GRU = 0.9624, XGBoost = 0.9619, Bi-LSTM = 0.9588, LSTM = 0.9579 and MLP = 0.9571, respectively.

  5. v.

    In SO2 forecasting, the AI models showed the following accuracy: Bi-LSTM = 0.955, LSTM = 0.950, GRU = 0.949, MLP = 0.928 and XGBoost = 0.916, respectively.

  6. vi.

    In PM2.5 forecasting, the AI models showed the following accuracy: LSTM = 0.973, Bi-LSTM = 0.971, GRU = 0.969, XGBoost = 0.961 and MLP = 0.936, respectively.

  7. vii.

    In PM10 forecasting, the AI models showed the following accuracy: Bi-LSTM = 0.9751, LSTM = 0.9744, GRU = 0.9737, XGBoost = 0.9622 and MLP = 0.9488, respectively.

  8. viii.

    In NH3 forecasting, the AI models showed the following accuracy: Bi-LSTM = 0.971, GRU = 0.965, LSTM = 0.963, MLP = 0.939 and XGBoost = 0.933, respectively.

This study highlights the potential of AI-based models in air quality forecasting, essential for proactive urban pollution management. Accurate predictions support real-time alert systems, preventing public health as cities expand. While high model accuracy was achieved, limitations remain, such as reliance on pollutant data alone, lacking real-time adaptability to sudden changes in pollution. Future research should incorporate meteorological and socioeconomic data, enabling more robust, responsive models. Real-time monitoring and long-term trend analyses could further enhance air quality management, enabling sustainable policy decisions. Interdisciplinary collaboration will be crucial in advancing these models for efficient, data-driven air quality solutions.