Prediction Of Atmospheric Carbon Monoxide Concentration Utilizing Different Machine Learning Algorithms: A Case study in Kuala Lumpur, Malaysia

Insidious toxin carbon monoxide (CO) can imitate a wide range of different disease states. Clinicians have, and will continue to have, serious concerns about the impact of CO imbalances on the human body. Carbon monoxide concentration has been exceeding the allowable levels in Malaysia. Owing to this, the main objective of this research is to propose a carbon monoxide (CO) prediction model based on machine learning techniques. Three years of historical data were used as input to develop the proposed models to predict carbon monoxide concentrations on a 12-hour and 24-hour basis. Four different machine learning technique models were used for the prediction which are Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Automated Neural Network – Multi-Layer Perceptron (ANN-MLP). The input parameters used are wind speed, humidity, Ozone (O3), Nitric oxide (NOx), Sulfur dioxide (SO2), and Nitrogen Dioxide (NO2). For each location, in this study, the uncertainty of the models utilized has been implemented to ensure the robustness of the performance. Furthermore, Taylor Diagram has been constructed to distinguish the performance of each model. The results indicate that ANN-MLP outperformed the all-other models involved in this study and showed efficiency in predicting Carbone monoxide concentration. By using ANN-MLP, the highest determination coefficient R2 were achieved which are 0.7190, 0.8914 and 0.7441 for the first station (S1), second station (S2) and the third station (S3) respectively by using 24-hour dataset. Meanwhile, by using a 12-hour dataset, 0.7490 for S1, 0.8942 for S2 and 0.8127 for S3. The uncertainty analysis of the ANN-MLP has 0.99 of confidence level and the lowest d-factor achieved, at S2 by using 12-hour dataset, is 0.000250455. These results ensure the effectiveness and robustness of ANN-MLP to predict carbon monoxide in the tropospheric layer. Not applicable.