USA AIR QUALITY: FORECASTING PM2.5 CONCENTRATIONS USING MACHINE LEARNING MODELS

Emissions have been a serious topic worldwide and every authority in each country needs to set regulations and penalties to control emissions. These gas emissions have an impact on many sectors like society’s health, countries’ economies, and ecosystems. For that, it is always needed to build any new or updated regulation to be based on strong evidence and that is machine learning comes to fulfil the need. Machine learning algorithms can read big numbers of records and train them then build prediction models. This step allows the entry of historic data, and the algorithm will read and recognize the pattern of the PM2.5 concentrations. As a result of the prediction models, the forecasting will take a part and predict future values of PM2.5 concentrations in the provided meteorological condition available in the dataset. Using the dataset from Kaggle, the steps started with understanding the dataset and handling the errors that appeared like outlier removal, missing values imputation and eliminating the duplicated records. Then four regression machine learning models were applied which are: linear regression, random forest, neural network, and polynomial regression. These four models have been tested to predict the PM2.5 concentrations and tuned the hyperparameters to attain the best accuracy score measured by the value of R-squared and mean squared error MSE. As a result, the four models achieved a good score where linear regression R-squared value= 75.6%, Random Forest R-squared value= 74.9%, neural network R-squared value=75.4% and Polynomial regression accomplished the best model accuracy with R-squared= 100% and an MSE=0%. Taking one further step after defining the best-fitted model, is to forecast future values of the concentrations of PM2.5. Using forecaster code in python and a polynomial regression model to produce forecasting values by reading 500 backward steps of data and then generating the forecasted values.