thingnario/Diffusion_SolRad

DDPM Solar Radiation model

A deep learning model for solar radiation nowcasting using modified MCVD model, a kind of DDPM model for video generation. The model predicts clearsky index and converts it to solar radiation for up to 6 or 36 time steps ahead.

Below is an example of DDPM generation process for 1-hour solar radiation prediction (6 time steps). The total iteration is 1000 steps, and every 50 steps are shown in the gif.

Overview

This repository contains two trained models (1hr & 6hr) for solar radiation forecasting:

• 1hr DDPM Model: Predicts solar radiation up to 1 hour ahead (6 time steps)
• 6hr DDPM Model: Predicts solar radiation up to 6 hours ahead (36 time steps).

The model uses multiple input sources:

• Himawari satellite data: Clearsky index calculated from Himawari satellite data
• WRF Prediction: Clearsky index from WRF's solar irradiation prediction
• Topography: Static topographical features

Installation

1. Clone the repository & install Git LFS:

git lfs install
git clone <repository-url>
cd Diffusion_SolRad
git lfs pull
git lfs ls-files # confirm whether models weights & sample data are downloaded

2. Install dependencies:

pip install -r requirements.txt

Requirements

• Python 3.x
• PyTorch 2.4.0
• NumPy 1.26.4
• einops 0.8.0

Usage

Basic Inference

Run solar radiation prediction using the pre-trained models:

python inference.py --pred-hr [1hr/6hr] --pred-mode [DDPM/DDIM] --basetime 202504131100

Command Line Arguments

• pred-mode: Choose between DDPM or DDIM sampling methods (default: DDPM)
• pred-hr: Choose between 1hr or 6hr prediction models (default: 1hr)
• --basetime: Timestamp for input data in format YYYYMMDDHHMM (default: 202504131100)

Example

# DDIM sampling method for 1-hour prediction
python inference.py --pred-hr 1hr --pred-mode DDIM --basetime 202507151200

Sample Data

The repository includes sample data files:

• sample_202504131100.npz
• sample_202504161200.npz
• sample_202507151200.npz

Model Weights

Pre-trained weights are available for both models:

• model_weights/ft06_01hr/weights.ckpt
• model_weights/ft36_06hr/weights.ckpt

License

This project is released under the MIT License.