DeepCAPE applies CNN to predict Convective Available Potential Energy using ERA5 meteorological data over Morocco. Complete pipeline from NetCDF processing to model evaluation with temporal validation strategies.
Data Pipeline: Processes NetCDF meteorological data using Xarray library
CNN Architecture: Implements 1D convolutional neural networks for regression
Temporal Validation: Uses mixed-date splitting strategy to avoid seasonal bias
Comprehensive Evaluation: Includes cross-validation, hyperparameter tuning, and seasonal performance analysis
Operational Focus: Designed with meteorological forecasting applications in mind
Data Processing: Xarray, Pandas, NumPy
Machine Learning: TensorFlow/Keras, scikit-learn
Visualization: Matplotlib, Seaborn
Data Format: NetCDF, CSV
The repository contains implementations of all project phases:
Data preprocessing and conversion from NetCDF to DataFrame
Exploratory data analysis and feature engineering
CNN model development with hyperparameter tuning
Cross-validation and performance evaluation
Seasonal and temporal analysis of predictions
The model achieved:
R² score of 0.5974 on test data
Mean Absolute Error of 37.27 J/kg
Significant improvement over traditional temporal splitting approaches
Best performance during summer conditions