AI4ALL Final Project: Hybrid LSTM-Embedding Neural Network for AQI Prediction

This project focuses on predicting the Air Quality Index (AQI) by integrating temporal and categorical data using a hybrid deep learning approach. The model combines LSTM layers to process sequential data (e.g., pollutant levels over time) and embedding layers to handle categorical features (e.g., state, county, and city). By leveraging this multi-input architecture, the model provides accurate AQI predictions based on historical pollutant levels and meteorological data.

Skills Applied

• Time-series analysis with LSTM
• Categorical data representation with embedding layers
• Data preprocessing, feature engineering, and scaling
• Deep learning model design and optimization
• Visualization of trends and results

Program Context

Developed as part of AI4ALL, this project showcases the application of cutting-edge AI techniques to address critical environmental challenges, specifically air quality forecasting.

Problem Statement

Air pollution significantly impacts public health, the environment, and the economy. Accurate AQI forecasting is essential for enabling informed decisions, mitigating health risks, and guiding policymaking.

Relevance

• Provides a predictive tool to monitor air quality in real-time.
• Supports public health efforts by forecasting poor air quality days.
• Assists policymakers in designing effective interventions for air pollution control.

Key Results

1. Processed and merged data from multiple sources into a unified dataset with over 40,000 observations.
2. Built and trained a hybrid LSTM-Embedding Neural Network:
   • Combines temporal patterns and categorical location features for robust AQI prediction.
3. Model Performance:
   • Mean Absolute Error (MAE): 8.172
   • Mean Squared Error (MSE): 11.675

Methodologies

1. Data Collection and Preprocessing

• Downloaded 10 separate feature attribute CSV files from AQS Data Mart.
• Merged multiple datasets, condensing 10 features into 3 categories:
  • Criteria Gases: Ozone, SO2, CO, NO2
  • Particulates: Fine particles (PM2.5), Coarse particles (PM10)
  • Meteorological Features: Wind, Temperature, Barometric Pressure, Relative Humidity, and Dewpoint
• Connected with output feature AQI values to create a complete dataframe.

2. Feature Engineering

• Scaled features using MinMaxScaler.
• Created time-series input sequences with a 30-day rolling window for temporal data.

3. Model Architecture

• Temporal Input:
  • Stacked LSTM layers to capture sequential patterns in AQI and pollutant data.
• Categorical Input:
  • Embedding layers to represent state, county, and city as dense vectors.
• Fusion:
  • Combined temporal and categorical features through concatenation.
• Dense Layers:
  • Fully connected layers refine features for AQI prediction.

4. Training

• Trained the hybrid model using TensorFlow and Keras.

5. Evaluation

• Evaluated the model on a separate test dataset.
• Visualized actual vs. predicted AQI values for validation.

Data Sources

• AQS Data Mart (EPA): Comprehensive pollutant and meteorological data for air quality monitoring.

Technologies Used

• Programming Languages: Python
• Deep Learning Frameworks: TensorFlow and Keras
• Data Processing: Pandas, NumPy, scikit-learn
• Visualization: matplotlib, seaborn, folium
• Geospatial Analysis: folium (interactive maps)

Authors

This project was completed in collaboration with:

• Uyen Nguyen (nguyen_u1@denison.edu)
• Neha Janardhan (njanardhan@umass.edu)
• Nurbu Lama (laman1@montclair.edu)
• Ignacio Ruiz (iruiz047@fiu.edu)