Hey Zindians!

• I wanted to share my solution to the recently concluded Hydropower Climate Optimisation Challenge. Even though I only started submitting on the last day,, I’m proud of the score I achieved and believe with more time, this solution could have made the Top 25. 💪
• Data Preprocessing: -Utilized Polars for fast data handling (cleaning, filtering, and extracting date, device, and user ID). -Reduced data from ~30 million records to ~16,295 by focusing on the relevant time range, devices, and users.
• Feature Engineering: -Aggregated hydropower data daily by device and user. -Transformed climate data with daily aggregations, lagged rainfall, rolling statistics, wind/heat computations, event flags, and cyclical encodings.
• Model Training: -Trained three base models: CatBoostRegressor, LightGBM, and XGBoost. - Used Optuna for hyperparameter tuning with GPU support and applied 10-fold TimeSeriesSplit for validation.
• Ensembling: -Stacked base model predictions to train an XGBoost meta-learner. -Further optimized blending via grid search and constrained weight optimization. -The ensembler, however, underperformed (Private Score: 5.000983378; Public Score: 8.847746558). With further tuning, I believe the RMSE could be lowered to around 4.0—or even 3.0.
• Results: -XGBoost: 4.92389054 -LightGBM: 5.583192817 -CatBoost: 5.385226374
• Final Thoughts: -Inspired by Team Central_Park’s second solution from the Inundata challenge.
• - Despite submitting on the last day, I had fun and encourage others to share their improvements and ideas for a more competitive and collaborative community.
• **GitHub:** [View my solution] https://github.com/ZambiaKuchalo/Zindi_Solutions.git
• Inspiration & Final Thoughts:
• My approach was inspired by Team Central_Park’s second solution from the Inundata: Mapping Floods in South Africa challenge. Shoutout to them for sharing their brilliant ideas! 💡
• Though I submitted on the last day, I had so much fun building this pipeline and truly believe in the power of open collaboration.