vsidaarth/physics-mamba-zero-shot-forecasting

Physics-constrained state-space learning for geographical zero-shot solar irradiance forecasting. # Physics-Mamba Physics-constrained state-space learning for geographical zero-shot solar irradiance forecasting. This repository accompanies the manuscript **Physics-Mamba: Physics-Constrained State-Space Learning for Geographical Zero-Shot Solar Irradiance Forecasting**. It contains the reproducibility code, frozen protocol, station split metadata, trained checkpoints, sealed-test results, and paper figures. It does not contain BSRN observations. ## Main experiment One global model uses the latest 24 hours of 15-minute observations to forecast global horizontal irradiance at 15, 30, 60, 180, and 360 minutes. The location split is fully station-disjoint: - 33 stations for training and normalization - 7 geographically unseen stations for validation and model selection - 11 additional unseen stations for the sealed test Physics-Mamba combines a Mamba-2 encoder with deterministic target-time solar geometry. Its output is a bounded transmission-like coefficient multiplied by top-of-atmosphere horizontal irradiance. This gives nonnegative predictions and exact zero output during astronomical night. The primary sealed-test endpoint is daylight station-macro MAE. Mean MAE across seeds 42, 43, and 44 was: | Model | 15 min | 30 min | 1 h | 3 h | 6 h | |---|---:|---:|---:|---:|---:| | Physics-Mamba | **36.07** | **46.37** | **56.52** | **76.23** | **90.72** | | iTransformer | 36.64 | 47.08 | 57.58 | 77.85 | 91.45 | Values are W/m2. Paired station-bootstrap intervals support the difference through 3 hours. The 6-hour interval includes zero. ## Repository contents - `scripts/`: download, audit, preprocessing, training, evaluation, analysis, and figure-generation code - `configs/protocol_v1.json`: machine-readable forecasting protocol - `manifests/`: fixed station assignments, archive checksums, coverage audit, and preprocessing audit summaries - `freezes/`: frozen pre-test manifests and hashes - `checkpoints/`: trained model checkpoints, model metadata, and validati …