Keywords

1 Introduction

The amount of exhaust gas produced by several factories and automobiles continues to climb as industrialisation levels rise, substantially increasing air pollution. People's daily lives are significantly impacted by air quality. Accurate air quality forecasting has emerged as a key strategy for reducing pollution and raising air quality. Data on air quality has caused great worry throughout the world. For predicting air quality, time series data prediction techniques are frequently employed, along with time series prediction models and conventional machine learning techniques. Some methods imitate the temporal and geographical dependency of air quality data concurrently. However, commonly used machine learning techniques frequently exhibit considerable performance variability under various conditions. Numerous variables, including temperature, wind speed, and geographical arrangement have an impact on air quality. As a result, it is challenging to produce certain and precise prediction results using the popular single model prediction method. Our thoughts in this paper are based on a strategy that has recently been discussed in the literature: integrating various models to predict air quality. When compared with current models, the integrated model can greatly increase the ability to forecast. But there is still much to learn about how to combine the benefits of several models depending on the features of the data collection.

2 Literature Survey

Petr Hájek et al. in [1] genetic algorithms optimise the input variable sets for each forecast of an air pollutant. Based on information gathered by the Pardubice city monitoring station in the Czech Republic, models are developed to predict the specific air quality indices for each air pollutant. The results show that when the root mean squared error is taken into consideration, individual prediction model compositions outperform single forecasts of the common air quality index. As a result, these models can be used to produce air quality index predictions that are more accurate one day in advance.

In order to avoid air pollution in urban areas and improve the quality of life for city dwellers, Kang et al. [2] highlighted the importance of conducting work on city air quality forecasting. Following that, AQI prediction models based on back propagation (BP) neural networks, genetic algorithm optimisation, and genetic simulated annealing algorithm optimisation are established. Comparing and evaluating the prediction outcomes reveal that the BP neural network based on genetic simulated annealing method has a higher accuracy rate, excellent generalisation capacity, and global search ability.

According to Wang et al. [3], who found that air pollution was becoming more severe, the most significant air pollutant, PM2.5 in aerosols, had a negative impact on people's regular output, way of life, and employment, as well as their health. As a result, the forecasting of PM2.5 concentration has taken on significant practical importance. The study selects real-time air quality data that is released, collects historical monitoring data of air environmental contaminants, normalises the data, and then splits the sample data into the two sets in a suitable ratio to form the training dataset and test dataset.

A key component of a smart city is a system for measuring and forecasting air quality, Mahajan et al. [5]. Making a forecast system with great accuracy and a reasonable calculation time is one of the biggest challenges. In this study, we demonstrate that a variety of clustering algorithms may be used to forecast fine particulate matter (PM2.5) concentrations reliably and quickly. We cluster the monitoring stations depending on their geographic proximity using a grid-based methodology. Data from 557 stations that have been distributed throughout Taiwan's Airbox device network is used in the tests and evaluation. The accuracy and processing time of the various clustering algorithms are compared in a final study.

3 Existing System

Commonly used machine learning techniques frequently exhibit considerable performance variability under various conditions. Numerous variables, including temperature, wind speed, and geographical arrangement have an impact on air quality. As a result, it is challenging to produce certain and precise prediction results using the popular single model prediction method.

4 Proposed System

In this work, a prediction approach based on integrated dual long short-term memory (LSTM) models was created to handle the complexity of air quality prediction. First, a single-factor prediction model that can independently forecast the value of each component in air quality data is created using sequence to sequence (Seq2Seq) technology. The multi-factor prediction model is then the LSTM model plus the attention mechanism. The model takes into account the air quality parameters such as the data from nearby stations and the weather. The two models are then combined using the eXtreme Gradient Boosting (XGBoosting) tree.

5 System Architecture

See Fig. 1.

Fig. 1
A flowchart of a system architecture. It starts with raw input data, followed by normalizing the dataset based on air quality, single and multi-factor prediction models and results. This collective result and metrological data lead to an integrated dual-mode to provide final prediction results.

System architecture

Full size image

6 Flow Chart

See Fig. 2.

Fig. 2
A flowchart of the processes on a dataset. It starts with loading and verification. If verification is successful, it processes and splits the dataset, calculates values of R M S E, M A E, and M A P E based on different algorithms and finally gives the comparison graph. If not, it ends directly.

Flow chart

Full size image

7 Results

Single-factor model is subsequently improved using the ATTENTION layer to create a multi-factor (combination of LSTM, sequence 2 sequence, and attention). In order to combine both models and improve prediction accuracy, features from the multi-model are extracted and retrained using XGBOOST.

The screen below displays information from the air quality dataset, which was used to construct this project.

The first row of the dataset's screen in Fig. 3 shows its column names, while the following rows show its values. As the training features, we used PM values, the target variable, and others.

Fig. 3
A screenshot of an Edit plus window with air quality dataset of Ahmedabad in a c s v file.

Air quality dataset

Full size image

All three models—single-factor LSTM, multi-factor LSTM with attention, and multi-factor integrated with XGBOOST—have been coded by our team. Below are the code and output screens for all the models we coded in the Jupyter notebook. You can see BLUE colour comments in each screen to learn about code (Figs. 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18 and 19).

Fig. 4
A screenshot of a window has an untitled tab with Python code. It has a command that reads loading python require packages.

Above screen we are loading required Python classes, LSTM, attention and XGBOOST class

Full size image
Fig. 5
A screenshot of a window has an untitled tab with Python code. It has functions to calculate MAPE and to read and display dataset values. A table at the bottom has the dataset values of Ahmedabad.

Above screen we define function to calculate MAPE and to normalise values and then reading and displaying dataset values

Full size image
Fig. 6
A screenshot of a window has an untitled tab with Python code. It has a function to extract x features and y labels from the dataset, to normalize the features, and to split the dataset into train and test.

Above screen we are processing dataset and then removing irrelevant columns and then splitting dataset into train and test and for testing we used 50 values which considers 1 test data per second for next one hour. These test values you can see in blue colour text

Full size image
Fig. 7
A screenshot of a window has an untitled tab with Python code. It has a function to calculate R M S E, M A E, and MAPE for each algorithm.

Above screen we are defining function to calculate RMSE, MAE, and MAPE values and then plot TEST data air quality and predicted air quality graph

Full size image
Fig. 8
A screenshot of a window has an untitled tab with Python code. The algorithm has a function to train single-factor L S T M model without attention.

Above screen we are training single-factor LSTM model and below is the LSTM model summary

Full size image
Fig. 9
A screenshot of a window has an untitled tab. A table for a sequential model has layer type, output shape, and param number. The total number of params is 71051, which is the same as the number of trainable params.

Above summary you can see that in single-factor LSTM there is no attention layer and in below screen we can see single-factor model prediction output

Full size image
Fig. 10
A screenshot of a window has an untitled tab. A table lists 17 rows of values of actual test data and predicted air quality.

Above screen we can see single-factor test data quality and predicted air quality for next 50 records

Full size image
Fig. 11
A screenshot of a window has an untitled tab. A single factor L S T M air quality prediction graph plots predicted air quality versus test data quality. Both trends overlap and fluctuate.

Above screen with single factor we got RMSE as 27 and MAE as 17, and MAPE as 39 and then in graph x-axis represents 50 min and y-axis represents air quality values, red line represents TEST data actual values, and green line represents predicted air quality and we can see both lines are overlapping with little difference. In below screen we can see LSTM with attention

Full size image
Fig. 12
A screenshot of a window has an untitled tab. A Python code for training with L S T M algorithm with attention as multifactor is given.

Above screen we are defining multi-model by combining LSTM and attention layer and below is the multi-model summary

Full size image
Fig. 13
A screenshot of a window has an untitled tab. A table for a sequential model has layer type, output shape, and param number. The total number of params is 71109, which is the same as the number of trainable params.

Above screen LSTM is combined with attention and below is the multi-model predicted output

Full size image
Fig. 14
A screenshot of a window has an untitled tab. A table lists 16 rows of values of actual test data and predicted air quality.

Above screen we can see test data and multi-model predicted air quality for next 50 s

Full size image
Fig. 15
A screenshot of a window has an untitled tab. A multifactor L S T M air quality prediction graph plots predicted air quality versus test data quality. Both trends overlap and fluctuate.

Above screen with multi-model we got RMSE as 23 and MAE as 13 and MAPE as 37, and we can see both test data and predicted air quality in graph. Above model RMSE, MAE, and MAPE is lesser than single model. In below screen showing integrated model with XGBOOST

Full size image
Fig. 16
A screenshot of a window has an untitled tab with Python code for an integrated model using Xgboost and multimodel attention L S T M. A table below lists 10 rows of values of actual test data and predicted air quality.

Above screen showing result of integrated XGBOOST

Full size image
Fig. 17
A screenshot of a window has an untitled tab. An integrated model air quality prediction graph plots predicted air quality versus test data quality. Both trends overlap and fluctuate.

Above screen with integrated model we got RMSE as 24 and MAE as 12, and MAPE as 31 and we can see predicted and actual test values in graph. In integrated model we got RMSE as high but MAE and MAPE is lesser than single and multi-model. In below screen showing graph of all algorithms

Full size image
Fig. 18
A screenshot of a window has an untitled tab. It has a grouped bar graph of performance in percentage versus algorithms such as multi-factor, single-factor, and X G boost. 3 output bars of M A E, MAPE, and R M S E are plotted for each algorithm. Out of which MAPE has higher values in all.

Represents values where each different colour bar represents different metric such as RMSE, MAE, and MAPE and in above graph we can see integrated XGBOOST got less MAE and MAPE compared with all other algorithms and same output we can see in below tabular format

Full size image
Fig. 19
A screenshot of a window has an untitled tab. It has a Python code and a table with columns for algorithm name, R M S E, M A E, and MAPE.

Above table we can see metric values of all algorithms and integrated XGBOOST got better result with low error rate

Full size image

8 Conclusion

We suggested a prediction model based on integrated dual LSTM model method to increase the precision of air quality data prediction. The integrated model's realisation procedure and impact can be summed up as follows. The air quality characteristics in the model are taken into consideration, together with meteorological information and data from surrounding stations. The method tree is then used to integrate the two models. First, single-factor models for each factor in the temporal dimension were made. To obtain the forecasted outcomes, the temporal dimension's attributes are employed. The projected value and weight of each leaf node are put together to provide the ideal expected value. Since the technique outlined in this study is based on analysing the experimental data using five evaluation indicators, it can result in predictions that are more accurate.

In order to improve the accuracy of by integrating the advantages of various models, the integrated dual LSTM model technique will be expanded in the next phase of the study. Although our model's outputs have very low probability, we have also found certain prediction results with outlier values. The examination of this sort of outlier value is one of the concerns that has to be addressed in the feature scope.