A survey on outlier detection methods applied on air quality data

This paper presents a study on the impact of various time series prediction algorithms applied on air quality data. This data is obtained from several sensors measurements, at every passing minute. The current research is concerned about finding a solution for a prediction algorithm based on fit functions. Traditional statistics models such as ARIMA (AutoRegressive Integrated Moving Average Model) and modern ones, like Facebook Prophet, were used for a comparative approach. Moreover, our proposed method has also been tested using different types of regression: Linear, Polynomial and Spline. After having made all the possible analogies between the selected algorithms for the given time series, regression spline has been found as the most accurate model. The purpose of this paper is to explain and to convince that results behave in a different manner depending on the used algorithm. The research has been done by studying air quality measurements received from various sensors, such as: PM2.5, PM1, PM10, O-3, CH2O, temperature, pressure and CO2. The study analyses sensors' values over a period of several months, obtaining over 43000 measurements per month for each sensor. The paper discusses the data obtained and its accuracy is tested using various metrics of evaluation.