Indoor Air Quality Prediction

Predicting indoor air quality pollutants is a crucial step in ensuring the health and safety of building occupants. By identifying potential sources of pollutants, such as mold or volatile organic compounds, measures can be taken to mitigate their impact and create a healthier indoor environment. Additionally, predicting indoor air quality can prevent long-term health issues caused by exposure to pollutants, such as respiratory problems or even cancer. As a part of my Master studies subject, I have developed a project aimed at predicting indoor air quality pollutants, which will contribute to the development of more effective mitigation strategies and healthier indoor environments for all.

Table of Contents

• Installation
• Usage
• Results
• License

Installation

To install the required packages for this project, switch to a new environment. clone the repository using

git clone https://github.com/en-ashay/Airquality-predction

## switch to the folder 
cd Airquality-prediction

then run the following command:

pip install -r requirements.txt

Usage

The repository contains a python script file alongwith three Jupyter Notebooks: EDA.ipynb, feature_engineering.ipynb, and time_series_forecasting.ipynband a folder for the PLots.

1. The EDA.ipynb notebook contains our exploratory data analysis (EDA), where we examine the data to uncover patterns and relationships. We use visualizations to present our findings and gain insights into the data.

2. In the feature_engineering.ipynb notebook, we generate new features to improve our predictive models. These include window features, which capture trends and patterns over a specified time window; time lag features, which incorporate past observations into our models; and seasonality and cyclical features, which account for recurring patterns in the data.

3. The time_series_forecasting.ipynb notebook contains our analysis of eight different time series forecasting models. We use grid search to systematically explore different combinations of model parameters and identify the best-performing models for each feature. The training history is saved as a CSV file for future reference. Additionally, we generate plots comparing real versus predicted values for all models and all features, which are stored in a plots folder for easy access.

4. Instead of the time_series_forecasting.ipynb , the main.py file can be used to run all the models and save the plots and the model results in the results.json file.

5.The Plots folder stores all the plots made during the main.py for all pollutants and also during the EDA process. To run all models on the preprocessed data -

python main.py

Results

The best performing models for each feature are:

Pollutant	Model	MSE	MAPE
CO2	LightGBM	0.0003834291266491532	0.0960325396742516 %
Humidity	SVR	5.930342021103925e-05	0.01851282401967276 %
PM10	XGBoost	2.649203592035359e-05	0.041401718434739304 %
PM25	XGBoost	6.058677648190005e-05	0.03985426224063079 %
Temperature	SVR	0.0002899542450183089	0.026371552642149455 %
VOC	SVR	0.00043322552369807647	3.195298841037611 %

The top three performing models for different pollutant values are:

Feature	Best Performing Models
CO2	LightGBM, RandomForest, XGBoost
Humidity	SVR, LightGBM, XGBoost
PM10	LightGBM, RandomForest, XGBoost
PM2.5	XGBoost, LightGBM, RandomForest
Temperature	SVR, RandomForest, LightGBM
VOC	SVR, LightGBM , RandomForest

for example here is the prediction on CO2 values by LightGBM model

License

This project is licensed under the MIT License - see the LICENSE file for detail