SEEDTrans is a Transformer-based framework for interpretable day-ahead photovoltaic (PV) power forecasting, incorporating multi-level series decomposition, expert-driven variable grouping, and cross-scale semantic fusion. It achieves state-of-the-art accuracy while enhancing interpretability in extreme weather scenarios like El Niño.
- 📊 Interpretable Forecasting: Integrates learnable wavelet transforms and seasonal-trend decomposition.
- 🎯 Adaptive Grouping: Groups meteorological variables dynamically, guided by expert priors.
- ⛅ Extreme Weather Robustness: Tested under El Niño conditions with adaptive representation.
- 🧠 Transformer Core: Encoder-decoder structure with multi-level cross-fusion and full attention.
The framework includes:
-
Adaptive Variable Grouping (AVG)
Learns to identify and reweight variables critical to short-term fluctuation and long-term trend. -
Wavelet-based Decomposition (WTDU)
Extracts fine-grained frequency-aware features. -
Seasonal-Trend Decomposition (STDU)
Disentangles seasonal and long-term trends. -
Cross-Fusion Strategy
Fuses features across scales and decomposition levels.
SEEDTrans significantly outperforms baselines like ARIMA, LSTM, ConvLSTM, and iTransformer:
| Model | RMSE ↓ | MAE ↓ | MAPE ↓ |
|---|---|---|---|
| ARIMA | 1.161 | 0.816 | 0.442 |
| FC-LSTM | 1.049 | 0.842 | 0.405 |
| CNN-BiLSTM | 0.502 | 0.271 | 0.177 |
| SEEDTrans | 0.439 | 0.223 | 0.129 |
(See paper for full benchmark and ablation results)
We use:
- 6 PV stations in Hebei, China (15-min resolution, 1 year)
- Stanford PV Plant, USA (30-min resolution, 2 years)
Each includes NWP data (irradiance, temperature, humidity, etc.) and historical PV output.
git clone https://github.com/AI4SClab/SEEDTrans.git
cd SEEDTrans
pip install -r requirements.txt
# Training
python train.py --config configs/seedtrans_stanford.yaml
# Inference
python predict.py --checkpoint checkpoints/best_model.pth