Ground-level Ozone Prediction Using Machine Learning Techniques: A Case Study in Amman, Jordan

Abstract

Air pollution is one of the most serious hazards to humans′ health nowadays, it is an invisible killer that takes many human lives every year. There are many pollutants existing in the atmosphere today, ozone being one of the most threatening pollutants. It can cause serious health damage such as wheezing, asthma, inflammation, and early mortality rates. Although air pollution could be forecasted using chemical and physical models, machine learning techniques showed promising results in this area, especially artificial neural networks. Despite its importance, there has not been any research on predicting ground-level ozone in Jordan. In this paper, we build a model for predicting ozone concentration for the next day in Amman, Jordan using a mixture of meteorological and seasonal variables of the previous day. We compare a multi-layer perceptron neural network (MLP), support vector regression (SVR), decision tree regression (DTR), and extreme gradient boosting (XGBoost) algorithms. We also explore the effect of applying various smoothing filters on the time-series data such as moving average, Holt-Winters smoothing and Savitzky-Golay filters. We find that MLP outperformed the other algorithms and that using Savitzky-Golay improved the results by 50% for coefficient of determination (R 2) and 80% for root mean square error (RMSE) and mean absolute error (MAE). Another point we focus on is the variables required to predict ozone concentration. In order to reduce the time required for prediction, we perform feature selection which greatly reduces the time by 91% as well as shrinking the number of features required for prediction to the previous day values of ozone, humidity, and temperature. The final model scored 98.653% for R 2 , 1.016 ppb for RMSE and 0.800 ppb for MAE.

A Comprehensive Review of Available Battery Datasets, RUL Prediction Approaches, and Advanced Battery Management

“…ANN is a well-recognised machine-learning algorithm which are widely used due to their ability to discover non-linear relationships between variables [231]. [232] proposed a way of approximating the RUL of Li-ion cells depending on the long short-term memory model, empirical mode decomposition, and deep neural network (DNN).…”

Makine Öğrenme Algoritmaları ile PM10 Konsantrasyon Tahmini

“…Üçüncü algoritma, grafiksel bir ters ağaç yapısına sahip, iyi bilinen bir makine öğrenme algoritması olan karar ağacıdır. Regresyon için bir karar ağacı kullanıldığında, buna karar ağacı regresyonu (KAR) denir (Aljanabi, 2020).…”

A Comparison of Machine Learning Methods to Forecast Tropospheric Ozone Levels in Delhi

“…The simplest method is linear regression, which computes the coefficients of a hyperplane to best fit the data. This method typically has the highest errors because it does not account for local variations and nonlinearities in the data, but it remains a baseline for comparison in many studies [12][13][14]. Instance-based algorithms create a database of specific instances and rely on local relations rather than global rules or generalizations [15], and they include K-Nearest Neighbor (kNN) and Support Vector Machines (SVM).…”

“…Ensemble methods have gained popularity in recent years, including in air pollution forecasting, because they often outperform single-algorithm methods of machine learning [14,30]. However, results vary by application and even data set, so it is best to test and compare a suite of ML methods for each new project [13,18]. When counting all variations, there are easily 70-100 ML algorithms [11].…”

“…Capilla [31] predicted ozone 1, 8, and 24 h in advance in an urban area on the eastern coast of the Iberian Peninsula and compared multiple linear regression with a Multi-Layer Perceptron network. Aljanabi et al [13] predicted ozone in Amman, Jordan, one day in advance, comparing a Multi-Layer Perceptron neural network (MLP), SVM, DTR, and the XGBoost algorithm, and found that MLP performed the best. They applied feature selection to reduce the run time by 91% by only using the previous day's ozone, humidity, and temperature.…”