Energy Forecasting with Deep Learning ⚡️

---

Table of Contents

• Project Overview
• Repository Structure
• Quick Start
• Data
• Notebooks & Models
• Results & Visualizations
• Reproducibility (How to run)
• Contributing
• License & Contact

---

Project Overview 🔍

Energy Forecasting with Deep Learning is a compact, reproducible project that demonstrates multiple approaches for short-term (hourly) load forecasting using time-series modeling. The repository includes baseline models (naive and dense feed-forward), feature engineering, exploratory data analysis, and an LSTM memory-based forecasting model that predicts the next 24 hours using the previous 168 hours.

This work is designed to be educational, reproducible, and easily extensible for productionization or further research.

---

Repository Structure 📂

000_naive_baseline_models/
001_baseline_dense_model/
002_LSTM_model/
EDA/
Evaluation/
raw/                  # raw PJME_hourly.csv dataset
visuals/              # evaluation & result plots (included)
README.md

Key notebooks:

• EDA/EDA.ipynb — exploratory data analysis and transformation steps
• 000_naive_baseline_models/000_naive_baseline_model.ipynb — naive baseline approach
• 001_baseline_dense_model/baseline_dense_model_predicition.ipynb — dense network baseline (note: filename contains a typo: "predicition")
• 002_LSTM_model/LSTM_memorybased_forecasting_model.ipynb — LSTM forecasting model

---

Quick Start ⚡️

1. Clone the repo:

git clone <your-repo-url>
cd <repo-dir>

2. Create a Python virtual environment and install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Tip: If you don't have requirements.txt, install common packages used in the notebooks: pandas, numpy, scikit-learn, matplotlib, tensorflow (tested with 2.15.x), jupyter.

3. Start Jupyter and run notebooks interactively:

jupyter lab

---

Data 🗂️

• Original raw data: raw/PJME_hourly.csv (hourly load data).
• Preprocessed / feature-engineered file used by notebooks: 001_baseline_dense_model/feature_engineered_load.csv.

The notebooks include the full pipeline for loading, cleaning, feature engineering (hour, dayofweek, month, year), scaling, windowing (168-hour input → 24-hour horizon), training, and evaluation.

---

Notebooks & Models 🧠

• Baselines:
  • Naive: 000_naive_baseline_models/ (simple naive forecasts)
  • Dense model: 001_baseline_dense_model/baseline_dense_model_predicition.ipynb
• LSTM memory-based model: 002_LSTM_model/LSTM_memorybased_forecasting_model.ipynb
  • Input shape: last 168 hours (7 days)
  • Output (horizon): next 24 hours
  • Uses MinMaxScaler for scaling and Mean Absolute Error (MAE) for primary evaluation

---

Results & Visualizations 📈

The repository already contains evaluation visuals in visuals/. Examples:

LSTM vs Dense model (24-hour horizon)

Dense model evaluation

Dense model forecast vs True

Training loss curve (Dense)

Predicted dense model results (sample)

These images illustrate model predictions and training metrics — they were generated by the notebooks and are ready for inclusion in README or reports.

---

Reproducibility — How to Run (Step-by-step) 🔁

1. Ensure your environment has the required packages (see Quick Start).
2. Open EDA/EDA.ipynb and run cells to reproduce the feature engineering and save feature_engineered_load.csv.
3. Run 000_naive_baseline_models/000_naive_baseline_model.ipynb to generate naive baseline forecasts and evaluation metrics.
4. Run 001_baseline_dense_model/baseline_dense_model_predicition.ipynb to train and evaluate the dense baseline. Save dense_model_predictions.csv to compare with LSTM.
5. Run 002_LSTM_model/LSTM_memorybased_forecasting_model.ipynb to train the LSTM and produce comparison plots. Use the included early stopping callback for robust training.

Tips:

• Use a GPU to speed up LSTM training when available.
• If you want deterministic results, set seeds for numpy, tensorflow, and python's random at the start of notebooks.

---

Evaluation Metrics ✅

• Primary metric: Mean Absolute Error (MAE)
• Also consider: Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE)

The notebooks print MAE for sample forecasts and include visualization comparing true vs predicted loads.

---

Contributing & Extending 🔧

Contributions are welcome — suggestions:

• Add hyperparameter search (Optuna or Keras Tuner)
• Add probabilistic forecasting (quantile loss or Bayesian methods)
• Introduce exogenous variables (weather, holidays)
• Wrap models with a minimal inference API (FastAPI/Flask)

Available on request: a trimmed requirements.txt (core packages), a Makefile for reproducible commands, or a GitHub Actions CI workflow to run notebooks/tests.

---

License & Contact 📬

This repository is provided under the GNU General Public License v3.0 (GPL-3.0). See the LICENSE file for the full text.

For questions or collaboration, open an issue or contact: datatwinlabs.nl.