Supplementary material for "Uncertainty-quantified, physics-informed TinyML for reliable real-time solar PV diagnostics"

It comprises: (i) the PyTorch implementation of the proposed uncertainty-quantified physics-informed autoencoder (UQ-PI-AE), including the two-channel inference procedure (deterministic detection channel plus a post-hoc Monte Carlo dropout copy for epistemic uncertainty) and symmetric per-tensor INT-8 post-training quantization; (ii) the dataset and generated script that produces the 86,400-sample, minute-level physics-grounded synthetic PV corpus from the five-parameter single-diode model, emulating the schema and fault taxonomy of the public Solar PV Anomaly Detection Dataset (Kaggle, CC BY 4.0); (iii) the four baselines evaluated in the paper (decision tree, one-dimensional CNN, data-only autoencoder, deterministic PI-AE) and the full evaluation harness; and (iv) the result files underlying the reported metrics (F1 = 0.702 at 0.53% FPR; ECE = 4.41%; Brier = 0.030; 2.12 KB flash and 0.60 mJ per inference projected on an ARM Cortex-M4F).

Requirements: Python 3.9+, PyTorch, NumPy, scikit-learn.