Predicting the Unpredictable
Team coursework exploring four storm-prediction tasks across 800 example storms, supported by a shared Python package for data, metrics, preprocessing, plotting and submission workflows.
Role
- Team contributor
- MSc group coursework — Team Sally
Stack
- Python
- PyTorch
- scikit-learn
- Hugging Face
- NumPy
- Sphinx
Problem
The FEMA challenge combines temporal frame forecasting, cross-sensor reconstruction, event classification and lightning localisation in one multi-task storm dataset.
Constraints
- The dataset contains 800 example storms and four tasks with different targets and evaluation needs.
- The work was completed by a nine-person MSc team, so project-level outcomes must not be presented as individual achievements.
- The portfolio repository is private, limiting which implementation links can be shown publicly.
My contribution
- Contributed as one of nine members of Team Sally to the shared storm-prediction coursework project.
- This case study deliberately attributes the package and four-task artefact to the team because the repository documentation does not separate individual ownership by component.
Architecture / methodology
- Organise shared data, I/O, metrics, preprocessing, splitting, plotting and submission utilities as an installable Python package.
- Address VIL frame prediction, VIL reconstruction, eight-class event classification and lightning location/time prediction as distinct tasks over one dataset.
- Generate API documentation from docstrings with Sphinx and define development dependencies for Pytest and Flake8.
Evaluation
- Use task-specific notebooks and shared metric utilities for experimental work.
- The repository defines documentation and code-quality tooling; no benchmark score is reproduced or claimed in this portfolio entry.
Results
- The team coursework artefact covers all four FEMA challenge tasks and packages reusable project utilities.
- No individual model result or team benchmark value is claimed without a separately attributable report.
Evidence and links
The README identifies the work as 2025–2026 Imperial College London MSc group coursework and names Zhifeng Li among nine Team Sally members.
Private portfolio repository READMEThe package metadata lists PyTorch, scikit-learn, Hugging Face, NumPy, Pandas, HDF5, Matplotlib and OpenCV dependencies.
Private repository pyproject.toml