Material- based approaches for efficient forecasting and mitigation of air pollution using advanced neural network models

ABSTRACT

Air pollution is a critical environmental problem driven by urbanization and industrialization. Time-series forecasting using previous methods is difficult because models must account for seasonal changes, day-to-day changes, and emergencies that can rapidly affect air quality. Therefore, existing approaches struggle to predict these fluctuations. This research addresses this issue by proposing a material-focused method of air quality prediction using machine learning techniques. The proposed model incorporates feature selection using MS-ANFIS-FS and classification using Unet-RNN (Unet Optimized Recurrent Neural Network). The model focuses on analyzing pollutant interactions with material surfaces, improving prediction accuracy by considering the role of materials in pollutant dispersion and absorption. The Successive Feature Defect Scaling Rate (SFDSR) and Auto-Regressive Integrated Moving Average (ARIMA) methods detect variance dependencies in air quality data. These methods enable the model to identify material traits influencing pollution levels, yielding more accurate results for pollutants like PM2.5 and NO2. The findings demonstrate the critical importance of material properties in environmental management and show how material-based interventions can effectively reduce air pollution. This model has the potential to facilitate real-time pollution monitoring and support the development of sustainable air quality management strategies.

Keywords:
Air pollution forecasting; Materials science; Machine learning; Feature selection; Unet-RNN; Environmental management

1. INTRODUCTION

In recent years, air pollution has emerged as a significant issue leading to millions of deaths, prompting researchers to focus on improving air quality predictions. Subsequently, they utilize the AQI to evaluate the air quality level. Estimating the index value can be enhanced by assessing pollutants similar to PM10, PM2.5, NO2, SO2, CO, O3, NH3, and PB. This method can accurately forecast awareness of the air pollution level in a specific area. Accurate air pollution forecasting provides a reliable basis for determining individual social measures. Accurate predictions of air pollutant concentrations can forecast fluctuations in pollutant levels, enabling authorities to take appropriate corrective action and individuals to plan their activities based on predicted pollutant levels [¹, ²].

Air Quality Indicators Different countries have different air quality indicators, and the current model focuses only on PM2.5 pollutants. Because research shows that PM2.5 causes more problems for humans than other pollutants and causes other pollutants, statistical analysis of PM2.5 pollution prediction can be done using past weather data [³, ⁴]. The data includes real-time PM2.5 and PM10 particulate matter concentrations and six traditional air pollutants, such as gaseous SO2, NO2, CO, and O3. Various pollutants in distinct geographical areas showcase diverse temporal evolution patterns. Spatially, when regions A and B are near each other, they might encounter comparable air pollution conditions [⁵].

Air quality has worsened in recent years, leading to significant health issues. The expansion of modern society has caused a decline in air quality due to urbanization and population growth. As more vehicles and businesses emerge, air pollutants become more hazardous. Furthermore, air pollution negatively impacts an individual’s mental health and emotions. Being exposed to air pollution raises systemic inflammation and oxidative stress, resulting in anxiety [⁶].

Air pollution prediction is crucial for timely warning and management of air pollution, particularly in emergencies where high levels of pollutants are rapidly released into the air, leading to substantial damages. The changing and unpredictable nature of pollution clouds, coupled with limited knowledge of pollution origins and emission levels, pose challenges to accurately modelling this occurrence. Additionally, the immediacy and swift action required have rendered conventional monitoring methods (such as traditional air quality monitoring stations and distributed models) intricate and frequently ineffective [⁷]. The novelty of this research is outlined as follows:

• The proposed preprocessing process is adopted to remove null and missing values from the collective AQI dataset.

• The proposed method used SFDSR and ARIMA to identify air quality defects, patterns, and trends, considering previous problems with AQI patterns.

• To consider the relative features’ importance analysis problem, this work uses the MS-ANFIS method to find the essential features of AQI and reduce the time complexity.

• Predicted values from the air quality data’s patterns, trends, and defect rate components are combined to produce the final predictive output. The Unet-RNN approach then obtains the output to extract and fine-tune features accurately.

2. LITERATURE SURVEY

In the current scenario, an Attention-Based Air Quality Prediction (AAQP) approach can be proposed to protect people against air pollution. Furthermore, the AAQP method can predict air quality indices using historical and meteorological data as seq2seq [⁸].

Nevertheless, automated air quality monitoring systems can be designed to increase awareness. Furthermore, the COVID-19 lockdown restrictions are causing severe damage to pollutants in the air [⁹].

After that, Air Quality Monitoring System (AQMS) architectures can be implemented based on the Internet of Things (IoT) technologies candidate analyzed. Furthermore, they provide an outline of classification, typical error sources, and calibration techniques for air pollution sensors, offering significantly lower costs [¹⁰].

Additionally, air quality can be assessed by examining crucial meteorological and pollution-related factors by presenting a Recurrent Air Quality Predictor (RAQP) technique for predicting air pollutant levels. Nonetheless, due to non-linearity and irregularity, the correlation weakens with more extended time intervals [¹¹].

Therefore, the AQI model can be implemented to predict particulate matter (PM 2.5) and other significant air quality variables such as NO, NO2, and O3. Then, Machine Learning (ML) techniques are proposed to provide a valuable method to predict air quality indicators when air quality is critical [¹²].

The impact of weather on AQP can be enhanced through Deep Learning (DL). Thus, hourly concentrations of PM2.5, PM10, and SO2 can be estimated using raw data from an air pollution dataset, and air quality can be forecasted using a Long-Short-Term Memory (LSTM) model [¹³].

The novel offered to classify Air Quality Levels (AQL) using image processing-based image analysis and computer vision techniques. Furthermore, the sky images can be precisely analyzed by employing the AQL technique to extract useful features from them [¹⁴].

Likewise, AQMS can be assessed in natural settings through IoT development and real-time analysis of ambient data measuring carbon monoxide and particulate issues [¹⁵].

Furthermore, they prioritize identifying pollutants by defining the AQI of an area and attempting to analyze trend patterns in pollutants. Nevertheless, there is a pressing need for effective strategies to forecast the adverse effects of air pollution [¹⁶].

To reduce the prediction errors of ML algorithms, prediction results can be improved based on the Support Vector Regression (SVR) model utilizing inputs of multidimensional AQI and weather conditions of multiple cities [¹⁷].

Then, they describe their application using the Air Quality Monitoring and Analysis Network (AQMAN) approach to monitor various air quality parameters, enabling the cost-effective prediction of regional stability [¹⁸]. Health systems can assess the comparable rate of increase in PM2.5 over the past 1-7 days related to air quality policy enforcement and economic changes [¹⁹].

They can accurately predict PM2.5 levels using big data algorithms like Convolutional Neural Networks-LSTM (CNN-LSTM) techniques that operate with different data metrics. Nevertheless, managing air pollution in urban areas presents additional challenges [²⁰].

An Improved Particle Swarm Optimization (IPSO) technique is also suggested to assess PM2.5 attention and use the Echo State Network (ESN) to address Beijing’s air pollution issue. The optimized IPSO algorithm is employed to enhance the search capabilities [²¹]. Accurate assessment of the variations in PM10-75%, PM10-49%, and NO2-96% attention disparities before and after the lockdown can help assess changes in certain air pollutant levels [²²].

Table 1 shows that air pollution can be effectively predicted using feature selection. The accuracy of the feature selection method can be predicted by analyzing its performance evaluation results in the literature. To assess the short- and long-term effects of Attention Air Pollutants (AAP) on respiratory disease morbidity, mortality, and premature mortality, exposure to AAP is conducted. However, the health effects of exposure to AAP are limited [³⁴].

Table 1
Effective forecasting air pollution based on feature selection.

AUTHOR	YEAR	FEATURE SELECTION	PERFORMANCE EVALUATION	DRAWBACK	ACCURACY
WU et al. [23]	2023	Random Forest-Maximum Relevance Minimum Redundancy (RF-mRMR)	Mean Absolute Percentage Error (MAPE),	However, traditional single-block training takes longer to learn and has lower prediction accuracy.	92.02%
UL-SAUFIE et al. [24]	2022	Wrapper Feature Selection	Absolute Error (AE) and Relative Error (RE)	Feature selection has lower accuracy and is more complex in terms of performance.	79%
YAN et al. [25]	2021	LSTM	Root-Mean-Square Error (RMSE)	It is challenging to capture air quality spatial and temporal distribution characteristics sufficiently.	93.07%
GUO et al. [26]	2020	Gated Recurrent Unit (GRU)	Mean Absolute Percentage Error (MAPE)	Most prediction methods focus on model selection.	70.1%
KAYA et al. [27]	2020	CNN	RMSE	Air pollution is an increasing problem, leading to illness and death	81%
GUPTA et al. [28]	2023	CatBoost Regression (CR) Algorithm	Mean Absolute Error (MAE), RMSE	Measuring the direct impact of air pollution on human health is crucial.	90.3%
SUD [29]	2023	Recursive Feature Elimination (RFE)	Accuracy	Air quality testing is complicated because it costs more to implement and analyse.	62.08%
NATARAJAN et al. [30]	2024	Grey Wolf Optimization (GWO)	R-Square, Mean Square Error (MSE),	Air pollution has significant economic consequences and raises health costs.	91.49%
WANG et al. [31]	2022	Multi-objective dragonfly optimization algorithm (MODOA)	Sensitivity, MAPE and MSE	Analysing pollution variability and uncertainty in forecasting methods effectively is challenging.	56.84%
DAO et al. [32]	2022	Random Forest Regression (RFR)	MAE, RMSE	Measuring pollutant levels is critical to air quality	92.8%
PAULPANDI and CHINNASAMY [33]	2023	Multi-objective Staked Feature Selection Approach (MoSFS)	Precision rate, Classification accuracy	Air pollution is a significant environmental problem due to various natural causes.	94.8%

Utilizing Support Vector Machine (SVM) classification methods for addressing feature selection and minimizing the number of features can yield precise binary detection analysis. Nonetheless, the system is costly and encounters issues with data collection, storage, and processing [³⁵].

Common and effective strategies include incorporating feature selection and semi-supervised learning in different layers of the DL network of Deep Air Learning (DAL) models. Moreover, utilizing the pertinent information from unlabeled spatiotemporal data can enhance interpolation and prediction performance [³⁶].

Future PM2.5 values can be predicted using Logistic Regression (LR) techniques on past PM2.5 measurements to create automated modelling for predicting pollution data [³⁷].

These issues can be managed by creating an advanced Large-Scale Optimization (LSO) algorithm for forecasting hourly air pollutant levels using the prior day’s weather data, establishing a continuous hourly prediction model for improved consistency, and suggesting modifications [³⁸].

The system utilizes a novel feature selection model and decomposition techniques to conduct intricate initial sequence analysis, considering model entropy and data preprocessing. The optimization prediction module proposed addresses the issue of air quality monitoring [³⁹]. PM attention can be forecasted using LSTM techniques, with its precision enhancing in contrast to model outcomes relying on root mean square error [⁴⁰].

The article [⁴¹] examined Low-Density Polyethylene (LDPE) plastic bags that were exposed to air for three years. This long-term exposure is crucial for understanding how environmental factors affect degradation over time. The study [⁴²] developed multiple ANN models to predict chloride diffusivity in concrete with varying percentages of PRCA. The methodology includes detailed mix proportions, material properties, and experimental procedures, which are crucial for developing and validating the model.

2.1. Problem statement

• Air quality can change significantly over time and with geographic location. This difference complicates the forecasting process and makes it challenging to develop a universal model, as the model must consider other patterns and trends in different situations.

• Determining the exact characteristics that affect air quality is complex. Also, selecting the best features requires expertise and is time-consuming.

• Balancing model complexity and ensuring model generalizability to unseen data is always a challenge.

• Determining the relative importance of different factors affecting air quality can be difficult. Some traits may have subtle effects that may be difficult to quantify, complicating the selection process.

• The training data may not reflect sudden changes, which may lead to performance degradation when such events occur.

3. PROPOSED SYSTEM

This section elaborates on the efficiency of the proposed Unet-ORNN algorithm. It illustrates the process of air pollution forecasting through preprocessing, defect analysis, variation dependencies, choosing the best features, and finally, classification. The detailed architecture diagram is represented in Figure 1.

Figure 1
Architecture diagram for forecasting air pollution.

In the proposed method, the first step is data preprocessing to normalize the collected air quality dataset. Then, we apply the SFDSR method to analyze the defect rate of AQI of the processed dataset. Next, the ARIMA method is used to identify the air variation dependencies based on defect rate data. After that, the MSANFISFS technique will be used to select the best features for air indexing. Finally, the proposed Unet-ORNN method was applied to classify the AQI based on the best features.

3.1. Data preprocessing

In this section, a dataset is gathered for feature analysis and preprocessed using the input records. Additionally, the average value of the data preprocessing can be analyzed to ensure that all data values are within the record. It is essential to ensure that all record values associated with attribute properties fall within the specified range of each record. Additionally, these marginal values are checked and used in data processing to fill, remove and clean noisy data. To reduce the total number of variables or dimensions in an air quality index dataset, data preprocessing collects and cleans all the data in the dataset.

(1) $$\text{Mean (}{M}^n)={\displaystyle {\sum }_{Q=1}^{W=1}\frac{O_b}{T_R}}$$

As Equation 1 describes, the mean value of the data set can be analyzed and calculated. Let’s assume O_b– Observed values in the dataset, T_R– presents total records in the dataset, W is a row, and Q is a column, S_D– standard deviation,

(2) $$S_D=\sqrt{{\displaystyle \sum _{Q=1}^{w=1}\frac{{(O^b-m^n)}^2}{T_R}}}$$

In this section, data is collected and preprocessed for feature analysis using input records. Additionally, data preprocessing is performed to ensure that all data values are accurately recorded.

As shown in Equation 2, compute the mean and the standard deviation from the observed values.

Algorithm step 1

Input: Air quality dataset Id Output: Preprocessed dataset ND Begin function Import AQI dataset Id $$\forall W=0donthen:$$ $$\forall Q=0donthen:$$ Check for noise and null values in the dataset. ←(NL) $$N_L={\displaystyle \sum \frac{I_d-M^n}{S^d}}$$ For elimination of the records ← ri For standard data set record ← ND $$N_D={\displaystyle \sum \frac{N_L-minimum(f_v)}{maximum(f_v)-minimum(f_v)}}$$ Update the dataset’s records End for End for Return ←ND End function

In algorithm step 1, the first step is to read the collected dataset. Then, each row and column checks the noise and values. If null values are present, they are eliminated from the dataset. The final step is to obtain the normalized records. The proposed algorithm removes null values and noisy records from air quality datasets. Let’s assume fv – feature values and n- noisy values in the dataset.

3.2. Successive Feature Defect Scaling Rate (SFDSR)

In this section, our proposed method is applied to the SFDSR technique from the processed dataset. This technique is used to find the defect and non-defect features of AQI. The proposed SFDSR method to identify the air pollution for each day in the dataset.

(3) $$O_I=I_{upper}-\left((b_{pupper}-r_c)\ast \frac{I_{upper}}{b_{pupper}-b_{plower}}\right)$$

The above equation is used to identify the oxygen level of air based on the upper and lower index of oxygen in the dataset. Let us assume that bpupper and blower are the breaking points of the upper and lower indexes, respectively.

(4) $$A{P}_I=I_{upper}+\left((r_c-b_{pupper})\ast \frac{I_{upper}}{I_{upper}-b_{plower}}\right)$$

The above equation is used to find the air pollutant index defect API.

(5) $$A_w={\displaystyle \sum \frac{F(r_1,r_q)-F{\left(\frac{r_1,r_q}{n}\right)}^2}{\left(\frac{r_1,r_q}{n}\right)}}$$

The above equation is used to find the strength of the relationship among r₁, r_q, and n, which is the number of attributes.

3.3. Auto Regressive Integrated Moving AVERAGE (ARIMA)

In this section, the ARIMA technique is used to estimate based on the defect ratio to analyze the dependence of changes. The ARIMA is a statistical estimation model that uses time series data to predict future air index based on the defect rate. It is a type of regression analysis that aims to forecast the predicted direction of a random walk by analyzing the discrepancies in values within a series rather than utilizing the actual data values. The lags of the serial difference are named auto-regressive (A_r), and the L_s of the prediction data are named the moving average (M_a). The moving average of data is as follows below: equation.

(6) $$T_s=\left\{s_{d1},s_{d2},\dots ,s_{dt},\dots \right\}$$

Where t is the index set (t =1,2,3…), t is the actual number, and s_d is a scaling defect rate.

(7) $$I_t=f(x)=\left\{\begin{array}{l}i,iftheithmonthfor{o}_{b}airindexattime{m}_e,\\ 0,otherwise\end{array}\right\}$$

The above expression is used to analyze the daily air quality value prediction. Here, we assume I_t is the observed air index level at time m_e, for i=1,2…. p (p is the possible outcomes), and o_b is an observed value.

(8) $$A_{r\left(me\right)}\left(c_s\right)={\varnothing }_0+{\varnothing }_1{P}_{me-1}+{\varnothing }_2{P}_{me-2}+\dots .+{\varnothing }_p{P}_{me-p}+e_{r\left(me\right)}$$

The above expression identifies the current m_e CS air index value in the dataset based on auto-regressive. Where Pm_e-1–previous day’s AQI value p-auto regressive part, A_r, ∅-coefficient of constant level of series and e_r–error rate.

(9) $$M_{a\left(me\right)}\left(c_1\right)=\mu +e_{r(}{}_{me)}-{\theta }_1{e}_{r(}{}_{me)-1}-{\theta }_2{e}_{r(}{}_{me)-2}-{\theta }_q{e}_{r(}{}_{me)-q}$$

The above expression is analysis the current air index (C_I) values of moving average at time m_e in the dataset. Here, we assume μ is the coefficient of the constant level of the series, and q is the part of the moving average M_a at time.

Autoregressive moving average expression is followed by,

(10) $$\left(1-{\displaystyle \sum _{n=1}^p\varnothing L{s}^n}\right)T_s=\left(1+{\displaystyle \sum _{n=1}^q\theta L{s}^n}\right)e_r$$

Let us assume p and q are the autoregressive and moving average, respectively, n is the number of iterations, Ls is the Lag, ∅ is the autoregressive parameter, θ is the moving average parameter model and e_r error rate.

(11) $$\left(1-{\displaystyle \sum _{n=1}^p\varnothing L{s}^n}\right)T_s=\left(1+{\displaystyle \sum _{n=1}^q\theta L{s}^n}\right)e_r$$

The above expression identifies the integration of I_n air index moving variation based on labs and integrates the autoregressive and moving averages.

Rewrite the above equation as:

(12) $$\left(1-{\displaystyle \sum _{n=1}^p\varnothing L{s}^n}\right){(1-L{s}^n)}^{I_n}{T}_s=\left(1+{\displaystyle \sum _{n=1}^{p-q}\theta L{s}^n}\right)e_r$$

This section proficiently identifies the AQI variation dependencies using the ARIMA technique. The ARIMA is generally described as (p, q, I_n), p is the autoregressive process, I_n is the degree of differencing, and q is the moving average.

3.4. Multi-scaling anfis feature selection approach

After identifying the AQI variation dependencies, this section applies the MS-ANFIS-FS approach to select the finest attributes of AQI. The proposed ANFIS technique has 5 rules for choosing mutual features for efficient classification results. In the first case, the ANFIS approach should create rules according to the attribute selection, stability and completeness of the fuzzy rule base. This is when there is a fuzzy rule between the input and the output to get the finest features for AQI. Once the optimal feature set is determined, these refined features can be used to develop an ANN model. This model can also capture more complex patterns and relationships.

Figure 2 illustrates the MS-ANFIS approach architecture model diagram. This method efficiently integrates the finest features of air indexing. ANFIS is well suited for processing air indexing datasets due to its high-performance ability. The symbols of A and B are the ARIMA input values of the r-Rule.

Figure 2
Diagram for MS-ANFIS approach.

(13) $$R_1=\beta x_i(A)\text{,}i=1,2\text{,}{R}_{\text{1}}=\beta y_{i-2}(B)i=3,4$$

The above expression identifies the first rule, R1. Let us assume β is the Gaussian membership function, and x_i and y_i are the membership values of β.

Where β,

(14) $$\beta (A)=ex{p}^{-{\left(\frac{A-{\rho }_i}{{\sigma }_i}\right)}^2}$$

Here, ρ_i and σ_i present a premise parameter set.

(15) $$R_3={\overline{W}}_l=\frac{W_i}{{\displaystyle \sum W}}$$

For estimating the third rule R3, the above expression can be used to analyze the ${\overline{W}}_i$. Let us assume W is the weight.

(16) $$R_3={\overline{W}}_l=\frac{W_i}{{\displaystyle \sum W}}$$

For estimating the third rule R3, the above expression can be used to analyze the ${\overline{W}}_i$. Let us assume W is the weight.

(17) $$R_4={\overline{W}}_{l}F={\overline{W}}_l(l_{i}A+k_{i}B+n_i)$$

The above expression finds the R4 based on the weight functions. Where, l_i and k_i is the consequent parameters of node i, and n_i represent network as input.

(18) $$R_5={\displaystyle \sum {\overline{W}}_l{F}_i}$$

The above expression estimate the rule 5 (R₅). In algorithm step 2, the first step is to read the input of AQI variation from the previous stage. The second step is to process the fuzzy rules to identify the complex AQI patterns and relationships of AQI’s finest features. The next step is performed to select essential features based on the relationship of AQI patterns in the dataset.

Algorithm 2

Input: ARMIA variation (A, B) and fuzzy rules Output: AQI Finest feature (F2) Begin function Import ARMIA variation (A, B) $$\forall Idon\text{,}then:$$ If A is βxi and B is βyi, then F1 = l1A + k1B + n1 If A is βxn and B is βyn then Fn = lnA + knB + nn End for each $$F^2=1-{\displaystyle \sum _{i=1}^n\frac{{(P_{ti}-o_{bi})}^2}{{(|P_{ti}-{\overline{o}}_b|+|o_b-{\overline{o}}_b|)}^2}}$$ Obtain the finest features←F2 Stop function

This phase algorithm produces the best result of the finest features F² selection for AQI classification results from the ARIMA dataset. Let us assume that P_ti is the predicting value, o_b is the observed value of the ARIMA dataset, and ob is the mean value of observed values. By focusing on the most influential inputs, the proposed ANFIS method understands the relationship between air quality indicators and their determinants, thereby improving forecast accuracy.

3.5. Unet-Optimized Recurrent Neural Network (UNET-RNN)

Afterwards, for the finest features analysis, we apply the Unet-RNN algorithm for AQI categorization. Unet Optimized Recurrent Neural Network algorithm work based on DFNC. The training process of the RNN method is a collection of AQI information based on the LSTM technique. This method has input, recurrent hidden and output neurons to classify the AQI. LSTM provides to avoid long processes by replacing the RNNs hidden neuron. Each layer is related to other neurons with particular importance. Therefore, this proposed method’s input layer obtains the values from the MS-ANFIS-FS technique. Then, this layer is fed into the hidden layer to extract important weight values for air quality using LSTM. Based on the hidden layer weight values, the high weight values for air indexing are predicted using the Unet-RNN algorithm’s output layer.

(19) $$\sigma =\frac{ex{p}^{F^2}}{{\displaystyle {\sum }_{i=1}^{cs}ex{p}^{F{i}^2}}}$$

The above expression is used to estimate the activation function process to predict air pollution indexing. Let us assume exp^F2 denotes the exponential function of input values, exp^Fi2 denotes the exponential function of output values, and cs expresses the number of features in the dataset.

Algorithm step3

Input: AQI Finest feature (F2) Output: Classification of AQI Ca Start function Import AQI Finest feature (F2) into the input layer Read features $F^2\leftarrow F_1^2,F_2^2\mathrm{...}{F}_n^2$ $$\forall idon\text{,}then:$$ Evaluating the Hidden neuron of the LSTM process Estimate input gate process LInput $$L_{Input}=\sigma \left(W_i{F}_{me}^2+W_i{h}_{me-1}+B_s\right)$$ Estimate forget gate process Lforget $$L_{forget}=\sigma \left(W{1}_{forget}{F}_{me}^2+W{2}_{forget}{h}_{me-1}+B_s\right)$$ Estimate output gate Lout process LCell $$L_{out}=L_{forget(me)}\otimes C_{me-1}+L_{input}\tanh \otimes \left(W{1}_{cell}{F}_{me}^2+W{2}_{cell}{h}_{me-1}+B_s\right)$$ Computing the output layer process Ca $$C_a=\sigma \text{(}{W}_i\tanh \otimes \text{(}{L}_{out}\text{))}$$ End for Obtain Classification of AQI← Ca Stop function

In algorithm step 3, the first step is to read the selected feature set $F^2\leftarrow F_1^2,F_2^2\mathrm{...}{F}_n^2$, then for each record, evaluate the hidden neuron process with the LSTM process. The LSTM estimates the input (L_Input), forget (L_forget) and output (L_out) gate for the hidden layer. The next step is the output layer process for AQI prediction. The proposed algorithm steps efficiently classifies the AQI from the finest features. Let us assume W is a weight factors for each node i, h_me−1 is a previous hidden state at time me, B_s is bias, ⊗ is the element multiplication. The LSTM technique have three gates there are: input, forget and output.

4. RESULT AND DISCUSSION

The seasonal air quality measurements are assessed against a real-time composite dataset monitored by the Central Pollution Control Board (CPCB). The algorithm under consideration is implemented within the Python framework utilizing the DL library and indexed to the code quality assessment thresholds set by the government of India.

Performance results can be assessed using a variety of functions, such as precision, recall, rate, F-Measure, error rate, time complexity, and loss rate, depending on processed data sets and environmental factors. The variance measures that are used to test asymptotic results for performance evaluation compute these functions. Table 2 presents the simulation parameters for sir pollution forecasting.

Table 2
Environment setup and its consideration.

LIMITS	VARIABLE
Tool	Jupiter Notebook
Dataset Name	AQI Dataset
Language	Python Framework
Features measured	Random attributes < = 20
Quality By class	Year by Index class.

The MAE is represented by,

(20) $$MAE=\frac{1}{n}{\displaystyle \sum _{i=1}^n|y_i-{\widehat{y}}_i|}$$

The performance rating can be evaluated by considering the loss rate shown in RMSE.

$$RMSE=\frac{1}{n}{\displaystyle \sum _{i=1}^n{(y{}_i-{\widehat{y}}_i)}^2}$$

The Prediction Interval Coverage Probability (PICP) measures the mean error rate y_i and ŷ_i of the corresponding eigenvalues, and the interpolation rate loss uncertainty principles can analyze performance. Moreover, the Mean Prediction Interval Width (MPIW) scale has a high predictive value of the probability of features. The evaluation be carried out,

(21) $$PICP=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{C}_i\mathrm{}with}\mathrm{}\left\{\begin{array}{c}{C}_i=1,if{y}_i\in [L_i,U_i]\\ C_i=0,else\end{array}\right\}MPIW=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n(U_i-L_i)}$$

According to the fuzzy membership function, the difference functions between the upper bound actual value changes L_i and U_i and the lower bound are calculated respectively.

Figure 3 above shows the actual air mass ratios, with an average ratio of 0 to 0.25, in the latest definitions under No2, Co and So2.

Figure 3
Present the monthly variations of three critical air pollutants monitored in Chennai during 2020: (a) NO2 concentration trends displayed using pink bars, showing values ranging from 0-15 μg/m^³; (b) CO levels represented through green bars, indicating concentrations between 0–1.0 ppm; (c) SO2 measurements depicted via orange bars, demonstrating values between 0–14 μg/m^³.

The median mean ratio varies based on the ARIMA ratio of the UNET-RNN classifier. Figure 4 illustrates the continuous characteristic level and trend value of seasonal change according to NO2, CO and SO2 parameters. Let’s examine the fundamental characteristics of carbon monoxide (CO), Sulfur Dioxide (SO2), and Nitrogen Dioxide (NO2). Elevated concentrations of NO2 and SO2 may cause throat and eye irritation.

Figure 4
Illustrate the temporal patterns of the same pollutants for Chennai in 2019: (a) NO2 temporal distribution exhibited through pink bars; (b) CO seasonal variations shown using blue bars; (c) SO2 monthly trends represented by green bars.

Follow-up rates differ from AQI and provide a linear index of year-to-year change, as illustrated in Figure 5. Furthermore, the remaining observations can increase the mean ratio value by creating a mutual feature scale based on seasonal trends and analyzing continuous feature change levels and seasonal change trends.

Figure 5
Presents a green line graph showing temporal fluctuations/variations between 2015-2020: (a) Time series visualization of successive features showing characteristic peaks and troughs; (b) Four-panel decomposition plot displaying:

– Observed values demonstrating raw data patterns

– Trend component indicating long-term variations

– Seasonal component revealing cyclical patterns

– Residual component showing remaining variations after removing trend and seasonality.

The median mean ratio varies based on the ARIMA ratio of the UNET-RNN classifier. Figure 4 illustrates the continuous characteristic level and trend value of seasonal change according to NO2, CO and SO2 parameters. Let’s examine the fundamental characteristics of carbon monoxide (CO), Sulfur Dioxide (SO2), and Nitrogen Dioxide (NO2). Elevated concentrations of NO2 and SO2 may cause throat and eye irritation.

Follow-up rates differ from AQI and provide a linear index of year-to-year change, as illustrated in Figure 5. Furthermore, the remaining observations can increase the mean ratio value by creating a mutual feature scale based on seasonal trends and analyzing continuous feature change levels and seasonal change trends.

As shown in Figure 6, the continuous loss rate function helps estimate the error rate. Based on this, the drop rate criterion is improved by 0.25 compared to the feature threshold. Additionally, ANFIS can attain higher efficiency and lower loss rates to minimize the loss based on the failure rate of the proposed system. It performs exceptionally well in terms of prediction accuracy with minimal loss rates.

Figure 6
Successive loss rate function.

Figure 7 above can improve the accuracy of the classification results through the proposed hair to analyze the overall performance of the classification accuracy. Moreover, the proposed UNET-RNN technique achieves a specific ratio by mean sensitivity and source protection.

Figure 7
Classification accuracy.

Table 3 demonstrates that using different methods can improve the accuracy of crop recommendation by enhancing classification accuracies. Empathy is an interpersonal measurement process that is used in rational assessment. Accurate positive correlation methods with actual values in the negative range can enhance sensitivity estimates compared to other methods.

Table 3
Classification accuracy.

METHODS/DATASET RECORDS	CLASSIFICATION ACCURACY IN %
	NFIM	MDC-SVR	CNN-LSTM	FF-RNN	UNET-RNN
105	86.4	89.3	92.13	92.19	93.26
507	88.07	92.56	93.26	93.78	94.64
1115	92.5	92.78	93.37	94.46	95.17
2289	93.8	93.5	94.27	94.82	95.24
3001	93.9	93.15	94.49	95.16	95.39

Figure 8 illustrates the sensitivity enhancement of the recorded outcomes across various datasets, with the function effectively classifying the results to achieve a superior rate compared to the commonly used combination method.

Figure 8
Analysis of sensitivity.

Table 4 outlines the sensitivity analysis of the proposed method and compares it with existing methods.

Table 4
Performance of sensitivity analysis.

METHODS/NO OF RECORDS	ANALYSIS OF SENSITIVITY %
	NFIM	MDC-SVR	CNN-LSTM	FF-RNN	UNET-RNN
105	85.17	87.02	88.14	89.24	90.28
507	87.56	89.26	89.76	91.23	92.7
1115	89.33	89.37	90.25	92.46	92.67
2289	89.84	90.78	91.54	92.57	93.14
3001	91.03	91.36	92.09	93.18	94.09

Figure 9 illustrates the specific performance of the proposed algorithm results. The x-axis shows the volume of data, and the y-axis shows efficiency in percentage. This approach outperforms existing methods.

Figure 9
Analysis of specificity evaluation.

Table 5 shows that the variation of parameters can be controlled by different data sets, creating different representations through other methods.

Table 5
Specificity evaluation.

METHODS/NO OF RECORDS	ANALYSIS OF SPECIFICITY IN %
	NFIM	MDC-SVR	CNN-LSTM	FF-RNN	UNET-RNN
105	82.3	87.3	89.3	91.3	91.4
507	83.8	87.6	91.2	91.8	92.1
1115	84.2	88.5	92.6	92.8	93.2
2289	85.3	88.9	92.8	93.2	94.6
3001	86.3	90.2	93.5	94.5	94.8

Figure 10 demonstrates that the false classification is lower for the difference between the error rates produced by the prediction method and the other methods. The proposed method archives the false classification results of 4.5, 4.8, 5.2, 5.3, and 6.2 for 105, 507, 1115, 2289, and 3001 records correspondingly. When records are gradually increased, the performance of false classification is improved by using the proposed methodology according to the number of records. While the existing methods have increased their false classification performance, the previous method gets a higher false classification result than the proposed work.

Figure 10
False classification evaluation.

Table 6 defines the relative proportion based on the number of actual positive values.

Table 6
False classification evaluation.

METHODS/NO OF RECORDS	ANALYSIS OF FALSE CLASSIFICATION IN %
	NFIM	MDC-SVR	CNN-LSTM	FF-RNN	UNET-RNN
105	5.7	6.2	5.9	5.4	4.5
507	6.8	6.9	6.4	5.8	4.8
1115	7.9	7.2	6.1	5.9	5.2
2289	8.3	7.8	6.4	6.1	5.3
3001	8.6	8.1	6.6	6.3	6.2

In Figure 11, we can see that there is a timing issue with the current implementation. It takes time to estimate the complexity of the implementation during feature selection and classification. The activity time is measured as the time spent filtering and classifying a feature out of the total time. The proposed system performs better than other methods when it comes to performance, as shown in Table 7.

Figure 11
Time complexity evaluation.

Table 7
Time complexity evaluation.

METHODS/NO OF RECORDS	ANALYSIS OF TIME COMPLEXITY (MS)
	NFIM	MDC-SVR	CNN-LSTM	FF-RNN	UNET-RNN
105	8.3	9.3	8.7	7.2	6.5
507	8.8	9.6	9.2	8.7	7.8
1115	10.3	9.8	9.1	8.8	8.1
2289	11.2	11.1	10.3	9.8	8.5
3001	11.8	11.4	10.2	9.5	8.7

The proposed UNET-RNN system’s time complexity representation performs better than existing methods. This hybrid model combines feature estimation and classification with O (n) time complexity metric verification, ensuring that the process is executed redundantly to be completed on time. This model is more efficient than other existing systems.

Figure 12 illustrates the comparison of the performance of various metrics of existing and proposed methods for air quality index prediction. The metrics are F-measure, precision, and recall. The proposed method attains an F-measure result of 94.03%, a precision result of 92.14%, and a recall result of 96.1% for AQI prediction. However, the conventional methods obtained less performance compared to the proposed system.

Figure 12
Comparison of existing and proposed method performance.

5. CONCLUSION

A novel model that predicts air pollution by integrating the Multi-Scaling Adaptive-Network-Based Fuzzy Inference System (MS-ANFIS-FS) and Unet Optimized Recurrent Neural Network (Unet-RNN) was proposed and evaluated. The model very well captures variance dependencies by applying the Successive Feature Defect Scaling Rate (SFDSR) and ARIMA ratios methods, which leads to the selection of the optimal features. In multi-objective continuous feature indexing, the model selects the features that have the largest contribution to the predictive performance. The Unet-RNN modeling which predicts pollutant variance through logistic activation functions with softmax has a very high precision score. Therefore, we can achieve lower RMSEs and thus, higher classification accuracy. The strategy, therefore, points out the significance of the material properties in the impact on pollution behavior and it is a very useful instrument for the quick solution of air quality problems and the environmental management. This model serves as a means for the further applications in pollution control and mitigation, thereby making the decision-making process for the sustainable environmental strategies more effective.

6. ACKNOWLEDGMENTS

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

7. BIBLIOGRAPHY

• ^[1] SETHI, J.K., MITTAL, M., “A new feature selection method based on machine learning technique for air quality dataset”, Journal Of Statistics and Management Systems, v. 22, n. 4, pp. 697–705, 2019. doi: http://doi.org/10.1080/09720510.2019.1609726.
  » https://doi.org/10.1080/09720510.2019.1609726
• ^[2] SHU, Y., DING, C., TAO, L., et al, “Air pollution prediction based on discrete wavelets and deep learning”, Sustainability (Basel), v. 15, n. 9, pp. 7367, 2023. doi: http://doi.org/10.3390/su15097367.
  » https://doi.org/10.3390/su15097367
• ^[3] LIANG, Y.-C., MAIMURY, Y., CHEN, A.H.-L., et al, “Machine learning-based prediction of air quality”, Applied Sciences (Basel, Switzerland), v. 10, n. 24, pp. 9151, 2020. doi: http://doi.org/10.3390/app10249151.
  » https://doi.org/10.3390/app10249151
• ^[4] KOTHANDARAMAN, D., PRAVEENA, N., VARADARAJKUMAR, K., et al, “Intelligent forecasting of air quality and pollution prediction using machine learning”, Adsorption Science and Technology, v. 2022, pp. 5086622, 2022. doi: http://doi.org/10.1155/2022/5086622.
  » https://doi.org/10.1155/2022/5086622
• ^[5] LIN, X., WANG, H., GUO, J., et al, “A deep learning approach using graph neural networks for anomaly detection in air quality data considering spatiotemporal correlations”, IEEE Access: Practical Innovations, Open Solutions, v. 10, pp. 94074–94088, 2022. doi: http://doi.org/10.1109/ACCESS.2022.3204284.
  » https://doi.org/10.1109/ACCESS.2022.3204284
• ^[6] KUMAR, V., SINGH, S., AHMED, Z., et al, “Air pollution prediction using machine learning algorithms: a systematic review”, International Journal of Engineering Research & Technology (Ahmedabad), v. 11, n. 12, pp. 84–93, 2022.
• ^[7] MOKHTARI, I., BECHKIT, W., RIVANO, H., et al, “Uncertainty-aware deep learning architectures for highly dynamic air quality prediction”, IEEE Access: Practical Innovations, Open Solutions, v. 9, pp. 14765–14778, 2021. doi: http://doi.org/10.1109/ACCESS.2021.3052429.
  » https://doi.org/10.1109/ACCESS.2021.3052429
• ^[8] LIU, B., YAN, S., LI, J., et al, “A sequence-to-sequence air quality predictor based on the n-step recurrent prediction”, IEEE Access: Practical Innovations, Open Solutions, v. 7, pp. 43331–43345, 2019. doi: http://doi.org/10.1109/ACCESS.2019.2908081.
  » https://doi.org/10.1109/ACCESS.2019.2908081
• ^[9] HOTA, L., DASH, P.K., SAHOO, K.S., et al, “Air quality index analysis of indian cities during COVID-19 using machine learning models: a comparative study”, In: 8th International Conference on Soft Computing & Machine Intelligence (ISCMI), pp. 27–31, 2021. doi: http://doi.org/10.1109/ISCMI53840.2021.9654925.
• ^[10] BAROT, V., KAPADIA, V., “Air quality monitoring systems using IoT: a review”, In: 2020 International Conference on Computational Performance Evaluation (ComPE), pp. 226–231, 2020. doi: http://doi.org/10.1109/ComPE49325.2020.9200053.
• ^[11] GU, K., QIAO, J., LIN, W., “Recurrent air quality predictor based on meteorology- and pollution-related factors”, IEEE Transactions on Industrial Informatics, v. 14, n. 9, pp. 3946–3955, Sep. 2018. doi: http://doi.org/10.1109/TII.2018.2793950.
  » https://doi.org/10.1109/TII.2018.2793950
• ^[12] MOGARAJU, J.K., “Prediction of air quality index using auto machine learning tools over Andhra Pradesh state of India”, 2023. doi: http://doi.org/10.21203/rs.3.rs-3202656/v1.
  » https://doi.org/10.21203/rs.3.rs-3202656/v1
• ^[13] YANG, Y., MEI, G., IZZO, S., “Revealing influence of meteorological conditions on air quality prediction using explainable deep learning”, IEEE Access: Practical Innovations, Open Solutions, v. 10, pp. 50755–50773, 2022. doi: http://doi.org/10.1109/ACCESS.2022.3173734.
  » https://doi.org/10.1109/ACCESS.2022.3173734
• ^[14] SAMSAMI, M., SHOJAEE, N., SAVAR, S., et al, “Classification of the air quality level based on analysis of the sky images”, In: 27th Iranian Conference on Electrical Engineering (ICEE), pp. 1492–1497, 2019. doi: http://doi.org/10.1109/IranianCEE.2019.8786738.
• ^[15] KUMAR, A., KUMARI, M., GUPTA, H., “Design and Analysis of IoT based Air Quality Monitoring System”, In: 2020 International Conference on Power Electronics & IoT Applications in Renewable Energy and its Control, pp. 242–245, 2020. doi: http://doi.org/10.1109/PARC49193.2020.236600.
• ^[16] MAHAJAN, M., KUMAR, S., PANT, B., et al, “Feature selection and analysis in air quality data,” In: 11th International Conference on Cloud Computing, Data Science & Engineering, pp. 280–285, 2021. doi: http://doi.org/10.1109/Confluence51648.2021.9376882.
• ^[17] LIU, B.C., BINAYKIA, A., CHANG, P.C., et al, “Urban air quality forecasting based on multi-dimensional collaborative Support Vector Regression (SVR): a case study of Beijing-Tianjin-Shijiazhuang”, PLoS One, v. 12, n. 7, pp. e0179763, 2017. doi: http://doi.org/10.1371/journal.pone.0179763. PubMed PMID: 28708836.
  » https://doi.org/10.1371/journal.pone.0179763
• ^[18] SINGHAI, S., BARANWAL, A., JAHAGIRDAR, A.C., et al, “Air quality monitoring and analysis network”, In: IEEE 7th Uttar Pradesh Section International Conference on Electrical, Electronics and Computer Engineering, pp. 1–4, 2020. doi: http://doi.org/10.1109/UPCON50219.2020.9376497.
• ^[19] CROFT, D.P., ZHANG, W., LIN, S., et al, “The association between respiratory infection and air pollution in the setting of air quality policy and economic change”, Annals of the American Thoracic Society, v. 16, n. 3, pp. 321–330, Mar. 2019. doi: http://doi.org/10.1513/AnnalsATS.201810-691OC. PubMed PMID: 30398895.
  » https://doi.org/10.1513/AnnalsATS.201810-691OC
• ^[20] HUANG, C.J., KUO, P.H., “A deep CNN-LSTM model for Particulate Matter (PM 25) forecasting in smart cities”, Sensors (Basel), v. 18, n. 7, pp. 2220, 2018. doi: http://doi.org/10.3390/s18072220. PubMed PMID: 29996546.
  » https://doi.org/10.3390/s18072220
• ^[21] XU, X., REN, W., “Application of a hybrid model based on echo state network and improved particle swarm optimization in PM2. 5 concentration forecasting: a case study of Beijing, China”, Sustainability (Basel), v. 11, n. 11, pp. 3096, 2019. doi: http://doi.org/10.3390/su11113096.
  » https://doi.org/10.3390/su11113096
• ^[22] OTMANI, A., BENCHRIF, A., TAHRI, M., et al, “Impact of Covid-19 lockdown on PM10, SO2 and NO2 concentrations in Salé City (Morocco)”, The Science of the Total Environment, v. 735, pp. 139541, 2020. doi: http://doi.org/10.1016/j.scitotenv.2020.139541. PubMed PMID: 32445829.
  » https://doi.org/10.1016/j.scitotenv.2020.139541
• ^[23] WU, H., YANG, T., LI, H., et al, “Air quality prediction model based on mRMR-RF feature selection and ISSA-LSTM”, Scientific Reports, v. 13, n. 1, pp. 12825, 2023. doi: http://doi.org/10.1038/s41598-023-39838-4. PubMed PMID: 37550459.
  » https://doi.org/10.1038/s41598-023-39838-4
• ^[24] UL-SAUFIE, A.Z., HAMZAN, N.H., ZAHARI, Z., et al, “Improving air pollution prediction modelling using wrapper feature selection”, Sustainability (Basel), v. 14, n. 18, pp. 11403, 2022. doi: http://doi.org/10.3390/su141811403.
  » https://doi.org/10.3390/su141811403
• ^[25] YAN, R., LIAO, J., YANG, J., et al, “Multi-hour and multi-site air quality index forecasting in Beijing using CNN, LSTM, CNN-LSTM, and spatiotemporal clustering”, Expert Systems with Applications, v. 169, pp. 114513, 2021. doi: http://doi.org/10.1016/j.eswa.2020.114513.
  » https://doi.org/10.1016/j.eswa.2020.114513
• ^[26] GUO, C., LIU, G., CHEN, C.H., “Air pollution concentration forecast method based on the deep ensemble neural network”, Wireless Communications and Mobile Computing, v. 2020, n. 1, pp. 8854649, 2020. doi: http://doi.org/10.1155/2020/8854649.
  » https://doi.org/10.1155/2020/8854649
• ^[27] KAYA, K., GÜNDÜZ ÖĞÜDÜCÜ, Ş., “Deep Flexible Sequential (DFS) model for air pollution forecasting”, Scientific Reports, v. 10, n. 1, pp. 3346, 2020. doi: http://doi.org/10.1038/s41598-020-60102-6. PubMed PMID: 32098977.
  » https://doi.org/10.1038/s41598-020-60102-6
• ^[28] GUPTA, N.S., MOHTA, Y., HEDA, K., et al, “Prediction of air quality index using machine learning techniques: a comparative analysis”, Journal of Environmental and Public Health, v. 2023, n. 1, pp. 4916267, 2023. doi: http://doi.org/10.1155/2023/4916267.
  » https://doi.org/10.1155/2023/4916267
• ^29] [29] SUD, A., “Air quality prediction using feature selection and deep learning model”, International Journal of Novel Research and Development, v. 8, n. 6, pp. 339–346, 2023.
• ^[30] NATARAJAN, S.K., SHANMURTHY, P., AROCKIAM, D., et al, “Optimized machine learning model for air quality index prediction in major cities in India”, Scientific Reports, v. 14, n. 1, pp. 6795, 2024. doi: http://doi.org/10.1038/s41598-024-54807-1. PubMed PMID: 38514669.
  » https://doi.org/10.1038/s41598-024-54807-1
• ^[31] WANG, J., LI, J., LI, Z., “Prediction of air pollution interval based on data preprocessing and multi-objective dragonfly optimization algorithm”, Frontiers in Ecology and Evolution, v. 10, pp. 855606, 2022. doi: http://doi.org/10.3389/fevo.2022.855606.
  » https://doi.org/10.3389/fevo.2022.855606
• ^[32] DAO, T.H., VAN NHAT, H., TRUNG, H.Q.,et al, “Analysis and prediction for air quality using various machine learning models,” In: Proceedings of the Seventh International Conference on Research in Intelligent and Computing in Engineering, pp. 89–94, 2022. doi: http://doi.org/10.15439/2022R03.
• ^[33] PAULPANDI, C., CHINNASAMY., “Optimized forecasting air pollution model based on multi-objective staked feature selection approach using deep featured neural classifier”, International Journal on Recent and Innovation Trends in Computing and Communication, v. 11, n. 9, pp. 1506–1515, 2023. doi: http://doi.org/10.17762/ijritcc.v11i9.9132.
  » https://doi.org/10.17762/ijritcc.v11i9.9132
• ^[34] RAJAK, R., CHATTOPADHYAY, A., “Short and long-term exposure to ambient air pollution and impact on health in India: a systematic review”, International Journal of Environmental Health Research, 2019. PubMed PMID: 31070475.
• ^[35] GHADDAR, B., NAOUM-SAWAYA, J., “High dimensional data classification and feature selection using support vector machines”, European Journal of Operational Research, v. 265, n. 3, pp. 993–1004, Mar. 2018. doi: http://doi.org/10.1016/j.ejor.2017.08.040.
  » https://doi.org/10.1016/j.ejor.2017.08.040
• ^[36] QI, Z., WANG, T., SONG, G., et al, “Deep air learning: interpolation prediction and feature analysis of fine-grained air quality”, IEEE Transactions on Knowledge and Data Engineering, v. 30, n. 12, pp. 2285–2297, 2018. doi: http://doi.org/10.1109/TKDE.2018.2823740.
  » https://doi.org/10.1109/TKDE.2018.2823740
• ^[37] ADITYA, C.R., DESHMUKH, C.R., NAYANA, D.K., et al, “Detection and prediction of air pollution using machine learning models”, International Journal of Engineering Trends and Technology, v. 59, n. 4, pp. 204–207, 2018. doi: http://doi.org/10.14445/22315381/IJETT-V59P238.
  » https://doi.org/10.14445/22315381/IJETT-V59P238
• ^[38] ZHU, D., CAI, C., YANG, T., et al, “A machine learning approach for air quality prediction: model regularization and optimization”, Big Data and Cognitive Computing, v. 2, n. 1, pp. 5, 2018. doi: http://doi.org/10.3390/bdcc2010005.
  » https://doi.org/10.3390/bdcc2010005
• ^[39] LI, H., WANG, J., LI, R., et al, “Novel analysis-forecast system based on multi-objective optimization for air quality index”, Journal of Cleaner Production, v. 208, pp. 1365–1383, 2019. doi: http://doi.org/10.1016/j.jclepro.2018.10.129.
  » https://doi.org/10.1016/j.jclepro.2018.10.129
• ^[40] XAYASOUK, T., LEE, H., LEE, G., “Air pollution prediction using long short-term memory (LSTM) and deep autoencoder (DAE) models”, Sustainability (Basel), v. 12, n. 6, pp. 2570, 2020. doi: http://doi.org/10.3390/su12062570.
  » https://doi.org/10.3390/su12062570
• ^[41] OLIVEIRA, W.L., ULIANA, P.S., LINS, V.D.F.C., “Mechanical and chemical analysis of degraded plastic bags exposed in air, water, and soil”, Matéria (Rio de Janeiro), v. 29, n. 1, pp. e20230315, 2024. doi: http://doi.org/10.1590/1517-7076-rmat-2023-0315.
  » https://doi.org/10.1590/1517-7076-rmat-2023-0315
• ^[42] ALZABEN, N., MAASHI, M., ALAZWARI, S., et al, “Artificial neural network prediction of chloride diffusivity in concrete for sustainable development”, Matéria (Rio de Janeiro), v. 29, n. 4, pp. e20240423, 2024. doi: http://doi.org/10.1590/1517-7076-rmat-2024-0423.
  » https://doi.org/10.1590/1517-7076-rmat-2024-0423

Publication Dates

• Publication in this collection
  10 Feb 2025
• Date of issue
  2025

History

• Received
  13 Sept 2024
• Accepted
  07 Jan 2025