IEC-Net

IEC-Net is a time-series forecasting codebase built on top of TSLib-style training loops, focused on long-term forecasting with large time-series models and memory-bank retrieval.

This repository currently centers on:

• Chronos2
• TiRex
• TimesFM

Key Features

• Unified training entry: run_membank.py
• Long-term forecasting pipeline with train/val/test flow
• Optional adapter-style tuning (train_mode=adapter)
• Optional retrieval augmentation via --use_membank
• Supports univariate (S) and multivariate (M, MS) settings
• NPU-oriented training path (Ascend environment variables supported)

Project Layout

.
├── run_membank.py               # main training/testing entry
├── exp/                         # experiment loops
├── models/                      # Chronos2 / TiRex / TimesFM wrappers
├── data_provider/               # dataset loaders and provider
├── scripts/                     # example shell scripts
│   ├── solar/irr/mem_irr.sh
│   ├── water/flowdb/mem_flowdb.sh
│   └── wind/{mem_kag.sh,mem_opsd.sh}
├── utils/                       # metrics, losses, tools
└── tutorial/                    # notebook and figures

Quick Start

1) Environment

• Python >=3.9
• PyTorch compatible with your device (NPU/GPU/CPU)

Install minimal dependencies:

pip install torch numpy pandas scikit-learn scipy matplotlib

If you use NPU, install your vendor-specific PyTorch/NPU stack first.

2) Prepare Data

For data_provider/data_load_other.py datasets (Wind, Water, Solar):

• CSV must contain timestamp column.
• features=S: must include target column (or pass your target via --target).
• features=M/MS: all non-timestamp columns are used as features.

For data_provider/data_load_stanford.py style datasets:

• provide two files:
  • trainval_timestamp_target.csv
  • test_timestamp_target.csv

3) Run Training (with MemBank)

Example (single-file solar dataset):

sh scripts\solar\irr\mem_irr.sh

Results

Detailed site-wise results are available in Site_wise_results.md.

License

MIT (see LICENSE).