Forecasting PM 2.5 Pollution Using Weather Data

Introduction

• Presenter: Karthik
• Partner: Jack
• Focus: Forecasting PM 2.5 pollution using weather data

Background

• Air pollution is a significant global issue:
  • 7 million deaths globally attributed to air pollution (World Bank estimate)
  • Costs the global economy $225 billion annually
• PM 2.5 defined as particulate matter with a diameter of less than 2.5 microns
• Need for data-driven solutions to combat air pollution
• A well-trained algorithm can provide robust predictions and granular data

Data Collection

• Meteorological parameters linked to PM 2.5 levels:
  • Role in dispersion and dilution
• Data sets collected:
  • Weather data: Relative humidity, temperature, air pressure, wind speed, wind run, precipitation
  • PM 2.5 data
• Locations:
  • Training sites: Sebastopol, San Rafael, Santa Cruz
  • Test sites: Oakland, Richmond, Napa
  • Test sites significantly separated from training sites to assess model efficacy

Methods and Results

• Model architecture: LSTM (Long Short-Term Memory)
  • A recurrent neural network suitable for sequential data like air pollution
• Loss function: Mean Squared Error (MSE)
  • Penalizes wrong predictions to train faster
• Experiment:
  • Saved four versions of the model after tuning hyperparameters and structure
• Evaluation metric: Root Mean Squared Error (RMSE)

Version Analysis

• Version 2:
  • Oakland predictions showed a trend but were biased (predicted line shifted from true values)
• Need for model improvement:
  • Either add more data or deepen model architecture
• Version 4:
  • Deepened model architecture, reducing bias
  • Applied final model to other test sites with similar results

Future Work

• Explore capabilities of using more data to expand prediction area
• Include other features such as emission source data
• Develop a prediction map for better visualization and analysis

Conclusion

• Emphasis on the importance of data-driven approaches in tackling air pollution
• Acknowledgment of potential improvements and next steps for the model
• Thank you for your attention!