Forecasting and Failure Prediction
A 16-day solar power forecasting platform that compares deep-learning models against gradient-boosted baselines. I also wrote a custom accuracy metric for dawn and dusk, where normal percentage-error metrics break down.
- Role
- ML Engineer
- Period
- 2024 to 2025
- Status
- Production
How the system fits together.
The calls that shaped it.
- 01
Authored a custom accuracy metric (MSCA, Mean Squared Capacity Accuracy) that solves the sunrise problem of dividing by near-zero capacity at dawn and dusk. A per-site seasonal bias-correction layer added roughly 17% accuracy on internal benchmarks at v0.13a.
- 02
Wrote a 12-step preparation pipeline with BSRN physics-based radiation cleaning, adaptive LOESS smoothing, operational features derived from work-order history, and a strict data-leakage gate that runs before every training experiment.
- 03
Companion XGBoost early-warning system predicts inverter failures 7, 14, and 30 days ahead across 189 sites and roughly 2,000 hardware units, with the work formalized in a technical report.
The interesting work isn't the stack. It's the boundaries.
What it runs on.
- 01 PyTorch + Lightning + PyTorch Forecasting for TFT
- 02 NeuralForecast for TiDE
- 03 Darts for the multi-model tournament
- 04 Optuna for hardware-aware hyperparameter search
- 05 XGBoost for the failure-prediction track