Climate-Change Disaster Forecasting Tool

Overview

A predictive web application designed to forecast the severity of climate-related disasters—including expected deaths, injuries, property damage, and disaster types—based on historical global climate and disaster data. This tool empowers policymakers, researchers, and emergency responders with forward-looking insights into potential climate change impacts in specific regions and timeframes.

As the team lead, I guided the design and development of the entire system—from data ingestion to prediction interface. Our approach uses a Temporal Fusion Transformer (TFT), a state-of-the-art deep learning model for time series forecasting, to make informed predictions about future disasters based on multi-dimensional climate and disaster data.

Features

• Predicts key disaster metrics (deaths, injuries, damages, disaster type)
• Region-specific and time-specific forecasting interface
• Interactive frontend to input parameters and view predictions
• Scalable backend built for real-time model interaction
• Designed for use by scientists, policy analysts, and disaster management teams

Implementation

• Machine Learning Model: Temporal Fusion Transformer trained on historical datasets including global climate patterns, past disaster records, geographical features, and more
• Backend: Built in Python using PyTorch Lightning, FastAPI for model-serving, and APIs to ingest real-world climate/disaster datasets
• Frontend: JavaScript-based interface for users to input target regions and forecast times, and visualize model predictions
• Team Workflow: Led collaborative development through Git, Agile sprints, and regular code/design reviews

Challenges & Solutions

• Data Complexity: Managed high-dimensional, sparse, and heterogeneous data by pre-processing with normalization and feature engineering
• Model Interpretability: Incorporated attention visualizations from the TFT to give users insight into what factors influenced each prediction
• Real-Time Interaction: Optimized model inference and backend design to support responsive prediction generation

What I Learned

This project deepened my knowledge of advanced time series models and how to deploy them in real-world applications. I gained hands-on experience in MLOps, full-stack development, and leading a cross-functional team. Most importantly, it showed how machine learning can contribute to addressing one of the most pressing global challenges—climate change.

GitHub