A Predictive Data Feature Exploration-Based Air Quality Prediction Approach

Abstract

In recent years, people have been paying more and more attention to air quality because it directly affects people's health and daily life. Effective air quality prediction has become one of the hot research issues. However, this paper is suffering many challenges, such as the instability of data sources and the variation of pollutant concentration along time series. Aiming at this problem, we propose an improved air quality prediction method based on the LightGBM model to predict the PM2.5 concentration at the 35 air quality monitoring stations in Beijing over the next 24 h. In this paper, we resolve the issue of processing the high-dimensional large-scale data by employing the LightGBM model and innovatively take the forecasting data as one of the data sources for predicting the air quality. With exploring the forecasting data feature, we could improve the prediction accuracy with making full use of the available spatial data. Given the lack of data, we employ the sliding window mechanism to deeply mine the high-dimensional temporal features for increasing the training dimensions to millions. We compare the predicted data with the actual data collected at the 35 air quality monitoring stations in Beijing. The experimental results show that the proposed method is superior to other schemes and prove the advantage of integrating the forecasting data and building up the high-dimensional statistical analysis. INDEX TERMS Predictive data fusion, high dimensional statistical features, air quality prediction, machine learning.

Air Quality Prediction in Smart Cities Using Machine Learning Technologies Based on Sensor Data: A Review

“…Using only the historic air pollution data. [52] AQ, MET, WFD, Spatial NO Faster training rate, higher accuracy. N/S [14] AQ, MET YES To obtain a sequence pattern.…”

“…A predictive data feature exploration-based air quality prediction approach [52]: Zhang et al proposed Light Gradient Boosting Machine (LightGBM) model and combining predictive and historical data executed prediction of the PM 2.5 concentration over the next 24 h. This method helped to process the high-dimensional large-scale data and support parallel learning. The problem of the lack of data was solved by applying the sliding window mechanism, which increases the training dimensions to millions.…”

“…Ma et al [13] mentioned that LSTM based algorithms performed better than RNN, and because of bidirectional modelling concept BLSTM provided better results compared to LSTM, integration of IDW improved BLSTM by 5.6% because the spatial factor was taken into consideration. Zhang et al [52] pointed out the advantage of LightGBM, being a histogram-based algorithm, processes high-dimensional big data better compared to the other boost algorithms. Tao et al [14] also mentioned the superiority of LSTM compared to RNN and confirmed the advantage of the bidirectional model.…”

A Smart Rig for Calibration of Gas Sensor Nodes

“…The process of gas sensor calibration is tedious and includes (i) collection of sensor response at different operating points, (ii) data processing and determination of an appropriate mathematical model linking the real gas concentration to the sensor response, and (iii) storing of the calibration model or look-up tables in a memory associated with the sensor to correct its readings during subsequent deployments. To make the low-cost gas sensor usable, it is important to increase its accuracy by developing a real-time adaptive calibration [36][37][38][39][40][41].…”

Robust Analysis of the Information Obtained From a Set of 12 Years of SO₂ Concentration Measurements

“…In [32], to process high-dimensional data with the aim of predicting the PM 2.5 concentration at 35 air quality monitoring stations in Beijing, China over the subsequent 24 hours, a LightGBM model [33] was proposed. Additionally, to capture the trend of the PM 2.5 concentration in a time series and reduce its dimensions, correlation analysis and PCA were used.…”