Optimal Prediction of Air Quality Index in Metropolitan Cities Using Fuzzy Time Series with Deep Learning Approach

Abstract

Predicting the air quality index (AQI) with high accuracy is just as crucial as predicting the weather. The research selected a few potential meteorological parameters and historical data after taking into account a variety of complex factors to accurately anticipate AQI. The dataset was gathered, pre-processed to substitute missing values (MV) and eliminate redundant information, and before being applied to predict the AQI. The data was collected from 2019 to 2022 to analyse the AQI founded on time series forecasting (TSF). Many AQI parameters, including accumulated precipitation, the daily normal temperature, and prevailing winds, are lacking in this study. To preserve the characteristics of the time series, kNN classification was implemented to fill in the MV and integrate Principal Component Analysis (PCA) to decrease the noise of data to recover the accuracy of AQI prediction. However, the majority of research is limited due to a lack of panel data, which means that characteristics such as seasonal behaviour cannot be taken into account. Consequently, the research introduced a TSF based on seasonal autoregressive integrated moving average (SARIMA) and stochastic fuzzy time series (SFTS). The stacked dilated convolution technique (SDCT) which effectively extracts the time autocorrelation, while the time attention module focuses on the time intervals that were significantly linked with each instant. To control the strongly connected features in the data set, the Spearman rank correlation coefficient (SRCC) was utilised. The selected features included SO2, CO and O3, NO2, PM10 and PM2.5, temperature, pressure, humidity, wind speed and weather, as well as rainfall. Additionally, to estimate the AQI and SO2, PM10, PM2.5, NO2, CO, and O3 concentration from 2019 to 2022, the data of climatological elements after PCA and historical AQI were input into the multiple linear regression (MLR) techniques with a temporal convolution network (TCN) built deep learning model (DLM). The proposed DLM springs a correct and detailed assessment for AQI prediction. The experimental results confirm that the expected background yields a stable forecasting result, that the pollutant concentration of the surrounding areas affects the AQI of a place, and that the planned model outperforms existing state-of-the-art models in terms of prediction of consequences. Consequently, utilising this presented innovative approach integrates fuzzy time series with deep learning, addressing missing values and noise reduction, incorporating seasonal behaviour, utilising the SRCC for feature control, employing a comprehensive set of meteorological parameters, and presenting a hybrid model that outperforms existing models. These aspects collectively contribute to the advancement of air quality prediction methodologies, particularly in metropolitan cities. However, this hybrid approach leverages the strengths of both traditional statistical methods and deep learning techniques, resulting in a robust and accurate assessment for AQI prediction as well as providing more stable and accurate forecasting results.