THE COMPARISON OF LINEAR MODELS FOR PM₁₀ AND PM_2.5 FORECASTING

Abstract

Air pollution is a very serious problem in Poland and elsewhere, and it is a factor that significantly affects the quality of human life. However, people are not fully aware of the terrible air quality due to the insufficient number of monitoring stations. This means they have no access to information about the quality of the air they breathe. The aim of this paper is to present and compare some linear procedures for PM10 and PM2.5 forecasting. Herein, the simulations concerning investigated prediction algorithms are based on real data originating from the Airly company network of pollution measurement stations. Related data, including measurements, were gathered every hour for a period of about one year, moreover, for forecasting purposes, weather data from the Dark Sky portal was additionally used. In this study, several Machine Learning predictive methods are considered. Among these, the results of three are presented. These are: Multiple Linear Regression, Multiple Linear Regression with Regularisation and, finally, Linear Neural Networks. The task for each predictive algorithm was to predict the concentration of PMx dust in the following hours of the next day. As a measure of the prediction task evaluation, several types of error were considered, while, during the research, machine learning group algorithms were utilized as learning models. Via these advanced, efficient and convenient algorithms, a detailed air quality forecast for the next 24 hours is obtainable. The presented algorithms will be implemented into the common air condition prediction system.

PM10 forecasting through applying convolution neural network techniques

“…Finally, Section 5 contains concluding remarks and plans for future investigations. This work is a continuation of the research presented in [1].…”

“…In addition, the use of a heterogeneous model is proposed. This will include numerous neural networks of various types, from classical linear networks [1], through deep learning [6] to those of the probabilistic type [17,18], as the application of such a hybrid predictive model will ease the task of analysing the sensitivity of individual elements of the input vector [17,19].…”

“…The next, the CNN Model with time (yellow line), adds time coordinates. The last line in green is a plot of the linear regression that the prediction is based on (1). Figure 6 shows the change in the Pearson correlation coefficient depending on the time of prediction.…”

“…When reviewing literature related to this problem, it is possible to distinguish several groups of machine learning algorithms that allow for its solution, e.g. Linear Regression [1], Fuzzy Logic [2], Neural Network Procedures [3], Neural Networks with ARMA models [4], Deep Neural Networks [5,6] and Autoencoders [7].…”

Numerical analysis of factors, pace and intensity of the corona virus (COVID-19) epidemic in Poland

“…Indeed, for some measuring stations the value is approx. R = 0.96 ( Kowalski and Warchałowski, 2018 ). Due to the great correlation between PM10 and PM2.5 assessments and that of these with certain observations related to the analyses already performed in the first phase of the pandemic ( Kowalski and Konior, 2020 ), the inquiry used only the PM2.5 pollutant as a defining unit.…”