PJME Hourly Load Forecasting

This project implements and compares advanced time series forecasting models for PJM's hourly energy consumption, including a hybrid Prophet-XGBoost approach.

📦 Project Structure

• notebooks/

  • 1. analysis.ipynb: Exploratory Data Analysis (EDA) and feature engineering.
  • 2. prophet II.ipynb: Prophet baseline model with cross-validation and hyperparameter tuning.
  • 3. hybrid III.ipynb: Hybrid Prophet + XGBoost model, evaluation, and results.

• data/: Raw PJME hourly load CSV file.

• utils/: Project utility scripts for data loading, feature engineering, and evaluation.

• output_hybrid_v2/, output_hybrid_v3/, output_prophet_II_v2/: Model outputs, artifacts, and results.

• requirements.txt: Python dependencies.

🚀 How to Run

1. Install dependencies:

   pip install -r requirements.txt

2. Run notebooks in order:

   • Start with EDA (1. analysis.ipynb)
   • Train Prophet baseline (2. prophet II.ipynb)
   • Run hybrid modeling (3. hybrid III.ipynb)

3. Data:
   Make sure PJME_hourly.csv is in the data/ folder.

VS Code Settings

This repository includes a .vscode/settings.json file with recommended settings for Visual Studio Code. These settings help ensure a consistent development environment by:

• Setting the Python interpreter to .venv/bin/python
• Adding ./utils to the Python analysis path
• Enabling basic type checking and linting with pylint
• Formatting code on save
• Auto-saving files after a short delay

Using these settings is optional, but recommended for a smoother experience and consistent code style. If you use a different editor, you can ignore this file.

📝 Features & Methods

• EDA:
  Visualizes seasonal, weekly, and holiday effects in PJM load data.

• Feature Engineering:
  Adds time-based, cyclical, and holiday features for modeling.

• Prophet Model:
  Captures trend, seasonality, and holidays. Hyperparameters tuned with Optuna.

• Hybrid Prophet-XGBoost:
  XGBoost learns Prophet residuals using engineered features, improving short-term and local accuracy.

• Evaluation:
  MAPE and visual comparisons across monthly and rolling windows.

📊 Results

• The hybrid model reduces MAPE by ~40% compared to Prophet alone.
• Hybrid forecasts better capture local fluctuations and the true shape of the load curve.

📁 Other Notes

• archive/: Place experimental or deprecated files here.
• utils/: Contains reusable code for data processing and evaluation.
• Notebooks are the main workflow; scripts can be modularized as needed.

📚 References

• PJM Hourly Load Data
• Facebook Prophet
• XGBoost

Acknowledgement:
AI tools were used to assist with markdown formatting, code refactoring, and improvement suggestions. All final implementation, analysis, and decisions are my own.