syeedalireza/transformer-load-forecast

Transformer / LSTM Load Forecast

Short-term electrical load forecasting using a sequence model (LSTM or Transformer). Predicts the next time steps from a window of past load (and optionally other features).

Problem

Utilities need short-term load forecasts for unit commitment and balancing. This project trains a PyTorch model on historical load series to predict future values (e.g. next 24 hours from the last 168 hours).

Data

• Use public load datasets (e.g. UCI Electricity Load, ISO New England, or similar) or the provided synthetic CSV in data/ for a quick run.
• Expected format: one column for load (and optionally date/time); script normalizes and builds sliding windows.

Model

• LSTM (default) consumes a fixed-length history and outputs a forecast horizon. Script args: --seq_len, --horizon, --epochs, --data_path.

Files

• model.py — LSTM; train.py — training and validation (synthetic series if no CSV); data/load_sample.csv — sample data.

Usage

pip install -r requirements.txt
python train.py

For custom data, place a CSV with a load column and set --data_path. A small sample is in data/load_sample.csv; if the path is missing, the script falls back to synthetic data.

Metrics

• Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) on the test set.
• Optional: MAPE if values are strictly positive.

Limitations / future work

• Single series / single region; multi-region or multivariate inputs would require a different setup.
• No exogenous variables (temperature, holidays) in the minimal version; these can be added for better accuracy.

Author

Alireza Aminzadeh

• Email: alireza.aminzadeh@hotmail.com
• Hugging Face: syeedalireza
• LinkedIn: alirezaaminzadeh