Quantile loss Long Short-Term Memory network for anomaly detection in PM10 air pollution data

Abstract

Particulate matter (PM) is a prevalent air pollutant globally with significant health risks. Regulatory standards for particulate matter are monitored by air quality monitoring stations, which provide accurate data but have limited coverage. Recent advancements in sensor technology offer a way to increase the spatial coverage of the air quality monitoring network. However, the use of sensors is associated with decreased accuracy and precision, as well as limited possibilities for on-site maintenance. This study focused on developing an anomaly detection method for PM10 monitoring, aiming to enhance the reliability of air quality data measured with air quality sensors. Using modern deep learning method, this research sought to identify anomalies in PM10 measurements, ultimately contributing to more effective management of the air quality monitoring network. The study had three objectives: firstly, to perform data analysis to determine which parameters can be used to predict PM10 concentrations; secondly, to develop a Long Short-Term Memory (LSTM) neural net-work and compare it’s performance against multiple linear regression (MLR) model in predicting PM10 concentrations. Lastly, to evaluate the effectiveness of the quantile LSTM neural network in detecting malfunctioning sensors. The results indicated that the LSTM quantile loss neural network is efficient at detecting invalid measurements in the air quality sensor monitoring network. It appears to be a useful tool for issuing warnings to maintenance personnel, serving as a data preprocessing tool, and automating data cleaning before presenting results to the general public