Deep Learning Framework-Based Chaotic Hunger Games Search Optimization Algorithm for Prediction of Air Quality Index

"Air pollution" refers to the accumulation of damaging compounds in the atmosphere that can have negative effects on human well-being [1]. Air contamination is a major problem in China due to the country's rapidly emerging economy and the accompanying rise in the number of cars and industrial facilities [2]. Because of this, the Chinese government has assisted most major cities in setting up air quality monitoring networks [3]. Predicting air quality accurately, however, helps cities grow and protects people's health, which is critical for addressing the problem of air pollution [4]. Constantly worsening air quality has had devastating effects on China's economy and people's health. Pollutant levels in the atmosphere can be measured. It is derived. A summary of the AQI categorization criteria is provided in Table 1. To put it simply, studies reveal a link between air pollution and breathing illnesses [5]. The respiratory system is the main entry point for polluted air into the human body, where it can have devastating effects on health. When it comes to limiting the negative health effects of urbanisation, nothing is more important than having access to reliable early warnings about the air quality forecast for the next few days. As a result, it is crucial to keep an eye on air quality and issue alerts when necessary.

Table 1. Classification standards for AQI

In the 1980s, it became common to measure contaminants with the help of numerical studies and methods for making accurate and statistical predictions [6]. Standard statistical methods include time series analysis. Classical statistical models are employed in this area [7]. Predicting air quality, however, is difficult. Weather conditions (temperature, humidity, wind speed, precipitation, etc.), vehicular pollution, and factory releases are just a few of the things that can readily alter it. Particulate matter is deposited and dispersed differently depending on the temperature, humidity, and precipitation [8, 9]. Wind speed also plays a role in the diffusion of particles throughout the atmosphere.

This is especially true in the realms of false intellect and big data, where deep learning has recently become increasingly popular. The reason for this is that it can effectively learn feature illustrations from massive amounts of input statistics, allowing it to uncover the underlying, rich properties of the data. Therefore, interdisciplinary research has popularised the subject of predicting urban air quality concentrations using deep learning [9]. The two most common types of air quality prediction models used today are the time series estimate model and the spatial and temporal prediction classical [10, 11]. When it comes to sequence learning, RNNs have proven to be effective. Incorporating a LSTM or gated recurrent unit (GRU) into RNNs allows them to acquire knowledge of temporal dependencies over extended periods of time [12].

The present study contributes primarily in four ways: We propose a multi-output, multi-index supervised learning (MMS) model for (LSTM) networks, and we show how this model can be used to address two major challenges in the field: (1) integrating data on airborne particulate matter, atmospheric conditions, and gaseous pollutants. improving forecast accuracy by considering the interplay of many contaminants; CHCS is a technique for hyper-parameter tuning that was developed by fusing with chaotic maps present in the starting population. Animal behaviour motivated by hunger served as inspiration for the HGS algorithm. With the help of idea, this optimization technique mimics the impact of hunger on the exploring events of the animals. The presented CHGS algorithm successfully completed the optimization process and arrived at the best possible result. The process does not stall out at a regional plateau, but rather advances to the global level. The CHGS optimization approach is advantageous because of its quick and consistent convergence. The next section of the paper is the blueprint for the rest of the work: Related works are provided in Section 2, and the projected model is explained in Section 3. In Sections 4 and 5, we detail the validation analysis and its contribution.

• Discussion of the suggested mechanism's specific implementation follows. To begin with, the acquired Dataset was pre-processed to fill in any blanks and eliminate any unnecessary information. The AQI has the potential to serve as both a descriptive framework and a crucial means of risk communication. It informs the general public on the present state. Because of this, there is a potential threat to health, especially among the young, the poor, and the elderly who are more likely to have conditions such as respiratory and cardiovascular disease.
• For the common man, the AQI is "One Number- One Color-One Description" of the air quality in his area. Values of the (AQI) and concordant background concentrations of the eight contaminants are shown in Table 2. As a complex composition, the effects of exposure to outdoor air pollution on human health varied widely. Sulfur dioxide. The (Cleanness Mission). Eight contaminants are used to determine air quality:
• Particulate matter (PM_2.5),
• Ozone (O₃),
• Particulate matter (PM₁₀),
• Sulphur Dioxide (SO₂),
• Nitrogen Dioxide (NO₂),
• Carbon Monoxide (CO),
• Lead (Pb), and
• Ammonia (NH₃)

See Table 3 for a rundown of the AQI readings and the associated health effects.

Table 2. Value of the (AQI) and concordant background concentrations of the eight contaminants

Table 3. AQI and their related health influences

3.1 Dataset description

Chennai provided the data used in this analysis [20]. There are three different stations located at Manali, Velachery, and Alandur. RH, PM2.5 levels, BP, WS, and WS intensity are the wind-related variables that have been gathered from these locations for further study (WD). All stations combined produced a dataset with 35039 data rows, for a grand total of 105117 data rows, and the composed data are displayed in 15-minute intervals corresponding to the time period 2020. There was almost a 78% gap in PM2.5 data. After processing, the data was cleansed to get rid of any rows with blank columns and filtered to only include information on rows with PM 2.5 levels below 250 g/m³. Because not all rows contained the same information, we had to narrow the remaining set down to just 22827 rows. The dataset is depicted in detail in Figures 1 and 2.

11.png

Figure 1. Sample input dataset

Every 15 minutes, data on atmospheric components like, CO, and Ozone was collected from three stations spread over Chennai. Each station contributes 35039 rows of data, for a total of 490546 rows in the Dataset, covering the period of time between May 1, 2019, and April 30, 2020. Data sets are implemented to store data that may be accessed by an application. Source code, macro libraries, system variables, and parameters are all examples of data sets utilised by applications and the operating system.

2.png

Figure 2. Details of the dataset

3.2 Pre-processing

In the first step, the data are taken from the Dataset and pre-processed to get rid of extraneous information. As a result, the input data is filtered using a pre-processing procedure to remove extraneous information. The normalisation strategy is utilised for the pre-processing stage as it is more efficient at identifying outliers and filling in missing data. The primary benefit of these methods is based on the assumption that aggregating forecasts from classifiers may allow for more accurate detection of class noise in input data. As a result, the raw data is pre-processed to obtain silent data for subsequent processing, which occasionally suppresses valuable info or leads to the info loss.

3.3 Air quality forecast perfect based on LSTM

3.3.1 Time series forecast

Both meteorological elements and the mutual limitations between pollutants have an impact on the concentration of air pollutants, in addition to the concentration of pollutants in the past. Regarding time series data, LSTM shines. To better anticipate MMS messages, we developed an LSTM-based prediction model. Here is how we characterise time series forecasting:

The concentration of particulate matter in the air is signified by P=p1, p2, ..., pT, the information of meteorological factors is signified by W=w1, w2, ..., wT, and the concentration of gaseous pollutants in the air is denoted by A=a1, a2, ..., aT; hence, R=PWA. The 15 features that make up each element r i in the set R are as follows: pm2.5.

The target prediction time series is denoted by z1=(t,t + 1,t + 2,...,t + N), whereas O z1=o1, o2, ..., os and Fz1=f1, f2, ..., fs are the experiential and forecast values of s air pollutant pointers.

Predicting the concentration of the target pollutant for the next N hours requires looking at the time series before time t, which is represented by z2=t-1, t-2, ..., t-D. O pz2, O wz2, and O az2 stand for the air quality observation consequences of particle pollutants, atmospheric factors, and gaseous pollutants, respectively, over the past D hours. The same holds true for z2 [1, T] and O pz2, O wz2, O az2, R. were utilised to evaluate the results of the predictions. That example, RMSE and MAE have their limits, in that the same algorithm model can be used to solve multiple issues, and they cannot capture the benefits and drawbacks of this model for diverse problems. Because the data dimensions are unique across different practical applications, it is difficult to determine for which problem a given model is most suited for making predictions. As a result, the accuracy of the predictions is measured on a scale from 0 to 1. R² is the regression, and is used to assess. Eq. (1) is provided to help with the RMSE computation. Below are the formulas for determining both MAE and R²:

$M A E=\frac{1}{m} \sum_{i=1}^m\left|y^i-p^i\right|$                       (1)

$R^2=1-\frac{\sum_{i=1}^m\left(y^i-p^i\right)^2}{\sum_{i=1}^m\left(y^i-\hat{y}\right)^2}$                       (2)

where, yⁱ, pⁱ, and y^b stand for the expected, measured, and mean levels of air pollution. This is the number of samples used in the experiment, denoted by m. When the Mean Absolute Error and (RMSE) are both modest and the R² value is high, the model error is small and the prediction presentation is good.

3.3.2 MMS prediction model

Through a process of layer-by-layer feature translation, deep learning is able to take the features of a sample in one space and convert them to a new feature space, which improves prediction, leverages massive data to learn features, and better represents the data's rich internal information. Gathering relevant information is the first stage. Using time-space analysis, we may extract the air quality time series data set R. The next step. Next, come the data entry and anticipated outcomes. The D-hours-earlier time series would be transformed into numerous 2-by-2 matrices. To facilitate LSTM's processing of input data, a computational architecture with multiple layers and an adequate number of neuron nodes was crafted. Through the network's learning and tuning, we determined which input and output layers would be most beneficial, and we shaped the function association from input to output to ensure that the resulting association was as accurate as possible. Decoding the LSTM output and obtaining the final prediction result required a completely linked layer, which was analogous to the model's output layer.

The first step is to obtain the raw data from each illustrative station in Chennai (S1, S23, S29, and S31); the second step is to apply the aforementioned MMS model to process the data in order to obtain a sequence at times T, T+1, ..., T+N (the attentiveness sequence of SO₂); and the third step is to calculate the average value of each pollutant index (PM2.5)

3.3.3 Steps of MMS prediction model

Processing the data beforehand is the first stage. The raw dataset should undergo outlier removal and missing value stuffing before being fed into the model. This will reduce the likelihood of unwanted effects from the data dimension on the prediction outcome and speed up the rate at which the data converges. The range of the data has been standardised to (0, 1).

After that, you must change the file type. Synchronizing input and output sequences, as is done in supervised learning, from the original data sequence. A multivariate time series is used for analysis. The multivariate time series can be put to use by utilising.

Third, initialise the LSTM's parameters by splitting the data into a training and testing set. Limiting the number of generations, configuring the sum of neurons, and determining the size of the fully linked network all require setting the epoch. The process of training an LSTM model involves selecting an appropriate, loss function, and optimization method.

Fourth, anticipate using the perfect. Investigating how well the model does on a test set.

Output: Prediction model parameters such as maximum epoch, sum of neurons, learning rate, minibatch size, L2 regularisation should be saved.

Several hyperparameters in this study were determined with the aid of CHGSA. These hyperparameters included the number of LSTM layers, the sum of neurons in each LSTM layer, the number of fully connected layers, time lag D. In order to find the optimal parameter, we hold all other variables constant and examine how they affect the model's ability to make predictions.

3.4 Hyper-parameter optimization using Hunger Games Search (HGS) optimization algorithm

The Human Genome Project (HGS) is an example of a optimization technique. Used to tackle optimization problems subject to restrictions, it is easy to implement, stable, and competitive [21]. The HGS algorithm is predicated on the idea that animals would act in ways that best satisfy their hunger. The animals' social behaviours serve as inspiration, while their hunger dictates the extent of their foraging. Using the idea of "Hunger" as a driving force behind all human endeavours, this dynamic optimization method is based. The hunger is simulated in the HGS optimization algorithm through the use of weights that are meant to reflect how the hunger impacts the search stages. Animal logic is employed as the basis for the algorithm's operation. Animals engage in these behaviours, which are viewed as adaptively evolving in nature, because they do so in an effort to increase their chances of obtaining food.

3.4.1 The logic of search, games

Animals follow social norms that are shaped by their surroundings. The behaviour and appearance of animals are subject to rules. An animal's hunger will influence its decisions and actions. Animals' nervousness and the hunters' uneasiness both increase when they're hungry. When their energy reserves are low, animals will search for food. In order to make a clean transition between exploration and defence, they need to look for food in addition to moving across habitats. The ability to flee from predators and discover new food sources is bolstered by animals' social lives. Living in groups increases an animal's chances of survival. Animals who are in better health have a greater chance of survival in the wild because they are better equipped to find food. Nature provides its own version of "Hunger Games," where the stakes are high and each false move could be fatal. Animals' actions can be influenced not just by physical needs like hunger but also by psychological ones like fear of humans acting as predators. There is a direct correlation between the intensity of hunger and the intensity of the quest for food. This means the animal is making a greater effort to find food in the days leading up to his death. Species migrations and logical alternatives underpin the proposed optimization strategy.

3.4.2 Mathematical model

In this subsection, we present the fundamental exact equations underlying the HGS optimization procedure. As such, the mathematical model is constructed around behaviours that are driven by hunger.

Getting ready to dig in: Everyone is believed to get along and aid one another. Eq. (3) [21] provides the central equation of the proposed HGS optimization method, and it is this equation that embodies the individuals' cooperative communication:

$\begin{aligned} & X \overrightarrow{(t+1)} \\ & =\left\{\begin{array}{lr}\text { Game }_1: \overrightarrow{X(t)}\left(1+\operatorname{randn}(1)\right) & r_1<l \\ \text { Game }_2: \vec{W}_1 \vec{X}_b+R \rightarrow \vec{W}_2\left|\vec{X}_b-\overrightarrow{X(t)}\right|, & r_1>l, r_2>E \\ \text { Game }_3: \vec{W}_1 \vec{X}_b-R \rightarrow \overrightarrow{W_2}\left|\overrightarrow{X_b}-\overrightarrow{X(t)}\right|, & r_1>l, r_2<E\end{array}\right.\end{aligned}$                         (3)

In Eq. (3), people are searching for both sites close to the optimal one and other locations far away from it. This guarantees that the search space is fully covered during the exploration process.

Hunger's part: Mathematical expressions of the population starving characteristics in the exploration field are provided, and W₁ is determined using the formula (4) [21].

$\overrightarrow{W_1}(i)= \begin{cases}\text { hungry }(i) \frac{N}{\text { SHungry }} \times r_4 & r_2<l \\ 1, & r_3<l\end{cases}$                         (4)

Meanwhile, $\vec{W}_2$  in Eq. (3) is calculated as shown in Eq. (5):

$\vec{W}_2(i)=2\left(1-e^{-\mid \text {hungry }(i)-\text { SHungry } \mid}\right)_{r 5}$                         (5)

where, hungry is the hungry of the populace. r₃, r₄, and r₅ are random values among 0 and 1. N is the population size. SHungry describes the synopsis of the populaces’ hungry feelings.

The hungry (i) can be signified exactly as shadows Eq. (6) [21]:

$\begin{aligned} & \operatorname{hungry}(i) \\ & =\left\{\begin{array}{lr}0, & \forall \operatorname{AllFitness}(i)==B F \\ \text { hungry }(i)+H, \forall \operatorname{AllFitness}(i)=B F\end{array}\right.\end{aligned}$                         (6)

where, AllFitness(i) is the suitability of the present iteration. The optimal population has their hunger reduced to zero at each repetition. Meanwhile, the old population's hunger is multiplied by the new hunger (H). H values for each population are distinct from one another. Optimization using the HGS method is designated in pseudo-code form in Procedure 1.

3.4.3 Chaotic hunger games search optimization algorithm (CHGS)

The initial population in meta-heuristic optimisation techniques is chosen at random between two given boundaries. The effectiveness of optimization algorithms relies heavily on the state of the agents at the outset. The quality of the starting population greatly affects the final product. In this study, we use chaotic maps to enhance the founding population. The idea of using chaotic maps to alter the seed population in metaheuristic algorithms was introduced [22, 23].

Based on the discussion [23], it has been shown that the logistic chaotic maps are the most computationally efficient of the modern chaotic maps; this is because they use a random initialization of statistics close to 0 and 1. Mathematical representations of this sympathetic of chaotic mapping are provided in Eq. (7).

$\begin{gathered}y_1=\text { rand }, \\ y_{i+1}=4 y_i\left(1-y_i\right), \forall i \in N\end{gathered}$                               (7)

In this case, rand is a random vector between 0 and 1. As a result, the proposed CHGS optimization method uses the values produced from this kind of chaotic mapping to define its initial population, rather than the values provided by the HGS. The HGS simulation performance is enhanced by this replacement of the starting population.