AIEG Forecasting — Electric Load & Solar Production Forecasting

Project summary

This repository implements forecasting for electrical signals: both consumption (load) and photovoltaic (solar) production. Models include classical ML and deep learning approaches to predict short-term horizons.

Problem statement and change of scope

The initial goal was to forecast the next timestep (t+15 minutes) and the next 8 timesteps (t+2 hours). During development we found data are not collected in real time but consolidated once per day (before midnight). Therefore the problem was reframed: use the previous day's consolidated data as input to predict the required horizons.

Key features

• Data loaders and preprocessing for time-series (windowing, scaling).
• Multiple model families: XGBoost, KNN, MLP, GRU, CNN-GRU, Transformer.
• Training, early stopping, learning rate scheduling.
• Rolling/expanding window cross-validation and k-fold helpers.
• Prediction, denormalization and standardized metrics (MAE, MAPE, RMSE, NRMSE).
• Visualization scripts for predictions and evaluation.

Quick start

1. Clone the repository: git clone https://github.com/BricePetit/AIEG_Forecasting
2. Create and activate a Python 3.10+ venv: python -m venv .venv source .venv/bin/activate
3. Install dependencies: pip install -r requirements.txt

Project layout

• src/ : source code (data, models, training, evaluation).
• src/saved_models/ : saved model checkpoints (ignored by git).
• data/ : raw and processed datasets (ignored by git).
• notebooks/ : exploratory notebooks.
• plots/ : figures and diagnostics (ignored by git).

Usage examples

• Prepare dataset with src/data loaders or provided scripts.
• Train a model using training scripts in src/ (e.g. GRU, Transformer).
• Evaluate using expanding window or k-fold utilities.

Evaluation & metrics

Standard metrics are provided: MAE, MAPE, MSE, RMSE, NRMSE. Predictions are denormalized before scoring when scalers are available.

Notes & recommendations

• Because input uses daily consolidated snapshots, align modeling and operational expectations with the daily cadence.
• Consider additional exogenous features (weather forecasts, calendar) to improve day-ahead predictions.
• Use rolling/expanding validation to reflect temporal data leakage constraints.

Contributing

• Open issues for bugs or feature requests.
• Submit PRs with clear descriptions and tests.
• Follow code formatting (black/flake8 recommended).

License

MIT

Contact

See repository metadata for authors and contact details.