README_HF_SPACE.md

metadata

title: FBMC Chronos-2 Zero-Shot Forecasting
emoji: ⚡
colorFrom: blue
colorTo: green
sdk: jupyterlab
sdk_version: 4.0.0
app_file: inference_smoke_test.ipynb
pinned: false
license: mit
hardware: a10g-small

FBMC Flow-Based Market Coupling Forecasting

Zero-shot electricity cross-border flow forecasting for 38 European FBMC borders using Amazon Chronos 2.

🚀 Quick Start

This HuggingFace Space provides interactive Jupyter notebooks for running zero-shot forecasts on GPU.

Available Notebooks

1. inference_smoke_test.ipynb - Quick validation (1 border × 7 days, ~1 min)
2. inference_full_14day.ipynb - Production forecast (38 borders × 14 days, ~5 min)
3. evaluation.ipynb - Performance analysis vs actuals

How to Use

1. Open any notebook in JupyterLab
2. Run all cells (Cell → Run All)
3. View results and visualizations inline

📊 Dataset

Source: evgueni-p/fbmc-features-24month

• Rows: 17,880 hourly observations

• Date Range: Oct 1, 2023 - Oct 14, 2025

• Features: 2,553 engineered features

  • Weather: 375 features (52 grid points)
  • ENTSO-E: ~1,863 features (generation, demand, prices, outages)
  • JAO: 276 features (CNEC binding, RAM, utilization, LTA, net positions)
  • Temporal: 39 features (hour, day, month, etc.)

• Targets: 38 FBMC cross-border flows (MW)

🔬 Model

Amazon Chronos 2 Large (710M parameters)

• Pre-trained foundation model for time series
• Zero-shot inference (no fine-tuning)
• Multivariate forecasting with future covariates
• Dynamic time-aware data extraction (prevents leakage)

⚡ Hardware

GPU: NVIDIA A10G (24GB VRAM)

• Model inference: ~5 minutes for complete 14-day forecast
• Recommended for production workloads

📈 Performance Target

D+1 MAE Goal: <150 MW per border

This is a zero-shot baseline. Fine-tuning (Phase 2) expected to improve accuracy by 20-40%.

🔐 Requirements

Set HF_TOKEN in Space secrets to access the private dataset.

🛠️ Technical Details

Feature Availability Windows

The system implements time-aware forecasting to prevent data leakage:

• Full-horizon D+14 (603 features): Weather, CNEC outages, LTA
• Partial D+1 (12 features): Load forecasts (masked D+2-D+14)
• Historical only (1,899 features): Prices, generation, demand

Dynamic Forecast System

Uses DynamicForecast module to extract context and future covariates based on run date:

• Context window: 512 hours (historical data)
• Forecast horizon: 336 hours (14 days)
• Automatic masking for partial availability

📚 Documentation

• Project Repository
• Activity Log
• Feature Engineering Details

🔄 Phase 2 Roadmap

Future improvements (not included in zero-shot MVP):

• Fine-tuning on FBMC data
• Ensemble methods
• Probabilistic forecasting
• Real-time data pipeline
• Production API

👤 Author

Evgueni Poloukarov

📄 License

MIT License - See LICENSE file for details

---

Last Updated: 2025-11-14 Version: 1.0.0 (Zero-Shot MVP)