How maxaragon/skygpt works
Project type and high‑level purpose
• This is a research‑oriented machine‑learning pipeline for ultra‑short‑term solar forecasting.
• It has two major stages:
– Stochastic sky‑video prediction (generating future sky frames)
– PV output forecasting (mapping images to power predictions)
File structure to extract components
• /codes/video_prediction/
– ConvLSTM (deterministic baseline)
– PhyDNetGAN (physics+GAN baseline)
– SkyGPT (VQ‑VAE + physics‑informed transformer)
– eval/ (quantitative evaluation notebooks)
• /codes/pv_output_prediction/
– SUNSET (AlexNet‑style CNN)
– UNet (modified for power regression)
• Data artifacts and dependencies
– video_prediction_dataset.hdf5, test_set_2019nov_dec.hdf5
– times_curr_*.npy
– requirements.txt (Python 3.6, PyTorch 1.8.1, TensorFlow 2.4.1)
• Figures and README describe framework and data flow
Sketch the main components
A. Data Storage
• Raw dataset (HDF5 files on local disk or Google Drive)
• Metadata (timestamps .npy)
B. Preprocessing
• reformat_input.py (SkyGPT input)
• any notebook cells that load, normalize, and batch images/PV
C. Video‐Prediction Models
• ConvLSTM module
• PhyDNet + GAN module
• SkyGPT module:
– VQ‑VAE (encoder/decoder)
– PhyCell (physics constraint)
– Transformer (auto‑regressive token model)
• Training scripts: training_vqvae.py, training_transformer.py
• Sample generation script: sample_gen.py
D. PV Output Prediction Models
• SUNSET training notebook (AlexNet CNN)
• U‑Net training notebook (modified UNet)
E. Evaluation & Comparison
• quant_eval_*.ipynb (compute MAE, MSE, VGG similarity)
• comp_all_models.ipynb (aggregate comparisons)
F. Hardware & Environment
• GPU cluster (NVIDIA A100)
• Python environments (TF, PyTorch)
Architectural patterns and technologies
• Modular ML pipeline (each model lives in its own subfolder + scripts)
• Hybrid deep‐learning frameworks (PyTorch for video, TensorFlow for PV)
• Discrete latent modelling (VQ‑VAE) + physics‐informed transformer
• Notebook‑driven orchestration for experiments and result visualization