This project develops and evaluates machine learning models to forecast short-term regional electricity load within the PJM Interconnection, integrating real-time energy market data with external weather signals. The goal is to outperform PJM’s official forecasts using custom neural networks and autoregressive baselines. Data Sources:
Real-time and day-ahead load forecasts, actual instantaneous demand.
OpenWeatherMap API: Weather data for major cities across the PJM grid footprint.
Dual-stage attention-based Seq2Seq LSTM for multi-horizon load forecasting.
AR (autoregressive) baselines for comparison and ablation.
Forecasting 24 time steps (2 hours at 5-minute granularity).
Trained and deployed using AWS SageMaker, leveraging cloud resources for scalability.
Evaluation performed via rolling MAE and MAPE across all 24 horizons.
Achieved up to 65% lower MAE than PJM’s official 2-hour-ahead forecasts in out-of-sample testing.
Real-time leaderboard and visualizations comparing personal forecasts vs. PJM vs. actuals.
