Overview

This repository contains the official open-source implementation of PhyDiffNet and RaPVFormer, the two-stage framework proposed in our paper "High-fidelity full-sky video prediction for photovoltaic ramp event forecasting".

The project integrates physics-informed video prediction, generative diffusion modeling, and transformer-based ramp-aware PV forecasting for ultra-short-term solar forecasting.

File Descriptions

File Name	Function Overview
PhyDNet.py	Training and evaluation CLI for the paper-aligned PhyDNet stage. Builds the physics/residual recurrent model, applies scheduled teacher forcing and differential-moment regularization, and writes start-index-aligned coarse future-frame predictions.
RaPVFormer.py	Defines RaPVFormer, its PV regression and auxiliary ramp-classification losses, equal-weight real/generated dual-view training, and generated-view-only validation and testing.
constrain_moments.py	Implements the kernel-to-moment (K2M) transform used to constrain PhyCell's learned physical convolution kernels.
diff_modules.py	Provides diffusion-specific layers, linear self/cross-attention, conditional encoders, the Gaussian diffusion process, and the diffusion trainer.
generate_phydiffnet.py	Loads a trained diffusion checkpoint, refines PhyDNet coarse forecasts, and saves the generated 16-frame sequences with their eval_stidx and split metadata.
pipeline.py	Exposes the leakage-safe end-to-end PVRampPipeline, connecting PhyDNet, conditional diffusion, and RaPVFormer without accepting future observed RGB frames.
prepare_diffusion_data.py	Creates the train-only conditional-diffusion HDF5 dataset containing observed futures, coarse predictions, four historical frames, start indices, and split-manifest metadata.
rnn_models.py	Defines PhyCell, ConvLSTM, the dual-branch PhyDNet encoder/decoder, channel attention, SSIM, and the composite frame loss.
utilities.py	Centralizes dataset classes, synchronized spatial augmentation, day-level split manifests, normalization, device transfer, prediction HDF5 I/O, checkpoint metadata, logging, and reproducibility helpers.
video_conditional_diffusion.py	Defines the conditional 3D U-Net and the paper/legacy diffusion architectures, and provides the diffusion-training CLI.
requirements.txt	Lists the Python runtime and test dependencies, including PyTorch 2.6, torchvision, NumPy, pandas, h5py, einops, and pytest.
supplementary-file.pdf	Supplementary document containing detailed model hyperparameters, training/inference protocols, video-quality metric definitions, and dataset links.
LICENSE.txt	Repository license terms.
README.md	Project overview, complete file inventory, dataset resources, environment requirements, training workflow, and method-to-code correspondence notes.

Dataset Resources

Sky Image and Photovoltaic Power Generation Dataset (SKIPP'D):

• 2017: https://purl.stanford.edu/sm043zf7254
• 2018: https://purl.stanford.edu/fb002mq9407
• 2019: https://purl.stanford.edu/jj716hx9049

Requirements

• Python 3.12
• PyTorch 2.6

Install the remaining dependencies from requirements.txt.