🚀 This is the official repo of "Non-stationary Diffusion For Probabilistic Time Series Forecasting"

The metrics in Table 6 about MAE/MSE is reversed. Please note this information when you are reproducing our results. I can change the arxiv paper, but I can not change publication text in the official ICML conference.

• Weiwei Ye (equal contribution): wwye155@gmail.com
• Zhuopeng Xu (equal contribution): xuzhuopeng@csu.edu.cn
• Ning gui (corresponding author): ninggui@gmail.com

📌 [2025年05月01日] 🎉🎉🎉🎉🎉 NsDiff is accepted as a Spotlight poster at ICML 2025 — Oral decision pending 🔎

NsDiff is a new diffusion-based theoretical framework for probalistic forecasting. Specifically designed for non-stationary scenarios.

pip install -r ./requirements.txt

⚠️⚠️⚠️⚠️The dataset will be downloaded automatically. Just run the following scripts.

see ./scripts/ for more examples.

1. pretrain and run

# pretraining
bash ./scripts/pretrain_F/ETTh1.sh
# run 
export PYTHONPATH=./
CUDA_DEVICE_ORDER=PCI_BUS_ID \
python3 ./src/experiments/NsDiff.py \
 --dataset_type="ETTh1" \
 --device="cuda:0" \
 --batch_size=32 \
 --horizon=1 \
 --pred_len=192 \
 --windows=168 \
 --load_pretrain=True \
 runs --seeds='[1232132, 3]'

2. run without pretraining

# run without pretraining
export PYTHONPATH=./
CUDA_DEVICE_ORDER=PCI_BUS_ID \
python3 ./src/experiments/NsDiff.py \
 --dataset_type="ETTh1" \
 --device="cuda:0" \
 --batch_size=32 \
 --horizon=1 \
 --pred_len=192 \
 --windows=168 \
 runs --seeds='[1232132, 3]'