Designing Machine Learning Solution for Real-World Air Quality Predition

Demo in YouTube: https://www.youtube.com/watch?v=T06xX9X0MH0

Overview

This project investigates the relationship between industrial facilities, greenhouse gas emissions, population density, and air quality in various US counties using machine learning models and data analysis techniques.

Data Preparation

1. We cleaned and merged datasets from the EPA's Greenhouse Gas Reporting Program (GHGRP) and Air Quality System (AQS), as well as population estimates for US counties. This process included renaming columns for readability, filtering top air quality measures, synchronizing datasets by common years and states, and merging them based on common identifiers such as State and County. The final result was two comprehensive DataFrames: facility_gas (for Model 1) and county_gas_air_quality (for Model 2).
2. We randomly splitted both DataFrames into training and validations datasets using the same approach: 70% for training sets and 30% for validation set (using train_test_split function from sklearn.model_selection).

Feature Engineering

Key steps included aggregating emissions data by facility and year, merging these aggregates with county population data, and creating a pivot table for air quality measures.

• For model 1, we applied standarization of features by subtracting the mean and scaling to unit variance (preprocessing.StandardScaler from sklearn).
• For model 2, we performed one-hot encoding (pandas.get_dummies()) on major industry types and categorized the percentage of days with PM2.5 levels into three severity levels: Good, Moderate, and Poor. We also standardized the non-one-hot encoded features.

These features were crucial for building predictive models.

Machine Learning Model Selection

Two machine learning models were constructed to address our research questions

• Model 1: The first model aimed to predict the industry type of facilities based on their gas emissions. We used the standarized training data to train the first logistic regression model (usingLogisticRegression from sklearn) and evaluated the model performance using training and validation datasets.
• Model 2: aimed to predict county air quality levels based on the sum of gas emissions, population density, and major industry types. We used one-hot coding data, non-one-hot coding training dataset to train the second logistic regression model and to determine which level of air quality they are mostly associated with. We evaluated the model performance using training and validation datasets.

Performance Evaluation

Hyperparameter tuning was performed to optimize the performance of both models. For the logistic regression models, we standardized features and evaluated performance using precision and recall metrics.

Model 1 Performance

Training Accuracy: 0.8133481646273637
Validation Accuracy: 0.7612869745718733

Class	Precision	Recall	F1-Score	Support
Crude Petroleum and Natural Gas Extraction	0.91	0.06	0.11	356
Fossil Fuel Electric Power Generation	0.70	1.00	0.82	830
Solid Waste Landfill	0.86	0.83	0.85	741
Accuracy			0.76	1927
Macro Avg	0.82	0.63	0.59	1927
Weighted Avg	0.80	0.76	0.70	1927

Model 2 Performance

Training Accuracy: 0.7881308929561841
Validation Accuracy: 0.7868217054263565

Class	Precision	Recall	F1-Score	Support
Good	0.80	0.98	0.88	600
Moderate	0.25	0.03	0.05	103
Poor	0.57	0.24	0.34	71
Accuracy			0.79	774
Macro Avg	0.54	0.42	0.42	774
Weighted Avg	0.71	0.79	0.72	774

Conclusion

Additionally, we explored the corrlations between different features related to air quality and gas emission using corrlation analysis. We visualized teh relationship between variables in the datasets. There is a certain correlation betwen air quality measurements, and there is also a certain correlation between population and air quality indicators.

To better understand air quality in the US, we created a choropleth map of the Annual Average PM2.5 level for each county in 2010. It is worh noting that the air quality in the estern and western regisons of the central US was poor.

Overall, the study highlights the significant impact of industrial activities and population density on air quality. The two maching learning models we established from this project can be used to predict air quality based on different factors, based on the data being provided in the future. The models developed provide valuable insights for policymakers to mitigate air pollution through targeted industrial regulation and management.

Libraries utilized:

pandas, seaborn, matplotlib, numpy, sklearn (including linear_model, model_selection, metrics, and preprocessing)

Dataset Used

1. EPA's Greenhouse Gas Reporting Program (GHGRP): Includes datasets us_greenhouse_gas_emissions_direct_emitter_facilities.csv and us_greenhouse_gas_emission_direct_emitter_gas_type.csv, detailing greenhouse gas emissions reported by facilities.
2. EPA's Air Quality System (AQS): Contains us_air_quality_measures.csv, providing air quality measurements on a county level from approximately 4000 monitoring stations across the US.
3. Population Estimates: The dataset PopulationEstimates.csv provides population estimates and Federal Information Processing Standards (FIPS) codes for US counties in 2010.