Physics-Guided Latent Diffusion for Synthetic Solar Irradiance Generation

SOLAD generates arbitrarily long, 10-minute resolution GHI sequences given only a site's latitude, longitude, and date range, without any concurrent meteorological observations.

👉

🌍 No installation needed — generate synthetic data directly in your browser.

💤 If asleep, click "Yes, get this app back up!" to wake it

---

Key Classes

Class	File	Description
SolarVAE	vae.py	Physics-conditioned 1-D conv autoencoder. Encodes daily clear sky index K* into z_flat + z_var via FiLM conditioning on solar geometry.
SolarDenoiser	denoiser.py	Transformer DDPM on windows of day-latents conditioned on location, solar geometry, and four-class sky-regime with CFG.
LatentTauTransform	optical_depth.py	Beer--Lambert reparameterisation K* = K_max · exp(−τ). Fit once on the training latent cache.
TauNoiseSchedule	optical_depth.py	Cosine noise schedule with v-prediction, used by training and inference.
generate_sequence	generate.py	Full inference pipeline: solar geometry -> reverse diffusion -> VAE decode -> K* output.
IntraDayNormStats	physics_utils.py	Loads normalisation stats and assembles intraday physics tensors.
DayFeatureNormStats	physics_utils.py	Loads normalisation stats and assembles day-level feature tensors.

---

Installation

git clone https://github.com/frimane/SOLAD.git
cd SOLAD
pip install -r requirements.txt

---

Running the App

streamlit run app.py

---

Training Data

This version of the model was trained on SURFRAD stations (BON, DRA, FPK, GWN, PSU, SXF, TBL), years 2020–2023, covering a wide range of North American climate regimes. No auxiliary meteorological variables are used -- only solar irradiance records and their solar-geometry-derived features.

---

Retraining or Fine-Tuning (Optional)

The model is two-stage and trained sequentially. All hyperparameters are in config.yaml. Refer to the paper for full architectural and loss details.

Stage 1 — Train SolarVAE

Step	Action
1. Data	Prepare daily K* profiles at your needed resolution. Compute solar geometry (zenith, ETR, clear-sky GHI) via pvlib using physics_utils.py.
2. Normalisation	Fit z-score statistics on your training split -> save to data/norm_intraday.json and data/norm_day_feat.json.
3. Regime labels	Fit a 4-component GMM on per-day K* statistics. Labels: 0=clear, 1=partly-cloudy, 2=cloudy, 3=overcast (sorted by descending mean K*) -> save to data/regime_gmm.json.
4. Training	Train SolarVAE with the losses in vae.py: reconstruction_loss, spectral_loss, regime_separation_loss, latent_variance_penalty.

Stage 2 — Train SolarDenoiser

Step	Action
1. Latent cache	Encode the full training set with frozen SolarVAE -> save z_full, physics tensors, and regime labels.
2. τ-transform	Fit LatentTauTransform on z_flat = z_full[:, :d_z] -> save to inference_bundle/latent_tau_stats.json.
3. Training	Train SolarDenoiser on W-day windows in τ-space using v-prediction and TauNoiseSchedule. CFG dropout probability is set in config.yaml.

---

Citation

@article{frimane2026solad,
  title   = {Physics-Guided Latent Diffusion for Synthetic Solar Irradiance Generation},
  author  = {Frimane, Azeddine and others},
  journal = {under review},
  year    = {2026},
}

---

License

This project is licensed under the GNU General Public License v3.0 — see the file for details.

Contact

Azeddine Frimane — Azeddine.frimane@yahoo.com