---
license: apache-2.0
pipeline_tag: time-series-forecasting
tags:
- weather
- precipitation-nowcasting
- climate
---

# PW-FouCast: Pangu-Weather-guided Fourier-domain foreCast

[![Paper](https://img.shields.io/badge/arXiv-2603.21768-B31B1B.svg)](https://arxiv.org/abs/2603.21768)
[![GitHub](https://img.shields.io/badge/GitHub-Repository-181717?logo=github)](https://github.com/Onemissed/PW-FouCast)
[![Conference](https://img.shields.io/badge/IJCNN-2026-blue.svg)](https://attend.ieee.org/wcci-2026/)

This is the official Hugging Face repository for **PW-FouCast**, a novel frequency-domain fusion framework designed to extend precipitation nowcasting horizons by integrating weather foundation model priors with radar observations. 

The model was introduced in the paper [Extending Precipitation Nowcasting Horizons via Spectral Fusion of Radar Observations and Foundation Model Priors](https://huggingface.co/papers/2603.21768).

## 🌟 Model Overview

**PW-FouCast** addresses the challenge of representational heterogeneities between high-resolution radar imagery and large-scale meteorological data. By leveraging Pangu-Weather forecasts as spectral priors within a Fourier-based backbone, the model effectively bridges the gap between atmospheric dynamics and local convective patterns.

### Key Features

* **Pangu-Weather-guided Frequency Modulation (PFM):** Aligning spectral magnitudes and phases with physical meteorological priors to ensure physically consistent forecasts.
* **Frequency Memory (FM):** A learned repository of ground-truth spectral patterns that dynamically corrects phase discrepancies and preserves complex temporal evolutions (e.g., expansion/contraction).
* **Inverted Frequency Attention (IFA):** A residual-reinjection mechanism designed to recover high-frequency details typically lost during spectral filtering, maintaining sharp structural fidelity in long-term predictions.
* **Extended Horizon:** Demonstrates superior performance on **SEVIR** and **MeteoNet** benchmarks, significantly mitigating performance decay in long-lead nowcasting.

## 🚀 How to Use

You can load the model weights for inference or fine-tuning as follows:

```python
import torch
from pw_foucast import PW_FouCast
from safetensors.torch import load_model
from huggingface_hub import hf_hub_download

MODEL_REGISTRY = {
    'pw_foucast': PW_FouCast,
}
ModelClass = MODEL_REGISTRY.get(args.model.lower())

model = ModelClass(**model_kwargs).to(args.device)
model = torch.nn.DataParallel(model)

# Load the model from Hugging Face
weights_path = hf_hub_download(repo_id=f"Onemiss/PW-FouCast", filename=f"{args.model}/{args.dataset}/model.safetensors")
load_model(model, weights_path)

# Eval
model.eval()
……
```

## ✍️ Citation

If you find this work or code useful for your research, please consider citing:

```bibtex
@article{qin2026extending,
  title={Extending Precipitation Nowcasting Horizons via Spectral Fusion of Radar Observations and Foundation Model Priors},
  author={Yuze Qin, Qingyong Li, Zhiqing Guo, Wen Wang, Yan Liu, Yangli-ao Geng},
  journal={arXiv preprint arXiv:2603.21768},
  year={2026}
}
```