FELINES is a research project aiming to design a preventive protection concept for electrical infrastructures by sensing electromagnetic fields generated during the early phases of lightning inception—in particular the Preliminary Breakdown Pulses (PBP)—and using them to predict whether the upcoming Return Stroke (RS) could be dangerous, enabling timely disconnection of vulnerable equipment.
The work combines PBP/RS modeling, electromagnetic field & coupling simulations, and machine learning for early classification (“dangerous RS” vs “not dangerous”).
More context is available on the project website and public deliverables (links below).

git clone https://github.com/ShayanDodge/FELINES-Lightning-Forecast.git
cd FELINES-Lightning-Forecast

• 📦 What the Repository Includes

• 📌 Phase 1 — First Case Study

  • 🌩 Project Vision
  • 📦 Repository Content
  • 📊 Datasets
  • 🧠 Implemented Methods
  • 📈 Performance — New Scenario (MTDM)
  • 🚀 Getting Started

• 🔮 Next Steps — Phase 2

• 👥 Project Team

• 🤝 Acknowledgments

• Developing deep learning models trained on electromagnetic signals
• Estimating lightning strike location
• Regressing channel-based peak current

• This phase provides the physical and data-driven foundation for risk assessment and protection strategies
• Machine learning–based classification of dangerous vs non-dangerous events
• PBP and RS electromagnetic modeling
• Field-to-line coupling simulations

⚠️ This repository currently contains the implementation of Phase 1 (Geolocation and Peak Current Estimation). Phase 2 (Early Classification and Protection Strategy) will be integrated in future releases.

This release corresponds to Phase 1 of the FELINES project, dedicated to lightning geolocation and peak current estimation from lightning-induced voltages on overhead lines.

The scientific foundations of this work are presented in:

Dodge, S.; Nicora, M.; Barmada, S.; Brignone, M.; Procopio, R.; Tucci, M.
“A deep learning based lightning location system.”
Electric Power Systems Research, 2025, https://doi.org/10.1016/j.epsr.2025.111437.
Read the article

📌 This repository implements, reproduces, and extends the methodology described in the above peer-reviewed publication. If you use this code, dataset, or results in your research, please cite the original article.

Rather than relying solely on conventional far-field Lightning Location Systems (LLS), FELINES introduces a line-integrated approach that infers lightning characteristics directly from voltage waveforms induced on overhead power lines.

The framework enables the estimation of:

• Lightning strike coordinates (x, y)
• Channel-base peak current (kA)

At its core, the system learns the physical-to-data mapping:

Induced Voltage Waveforms  →  {Strike Location (x, y), Peak Current}

This repository provides the complete implementation of Phase 1, including datasets, models, and reproducible experimentation pipelines.

• Simulation Datasets

  • Old Scenario — A benchmark 10-km single-conductor transmission line used to implement and compare the three proposed Deep Learning approaches (FFDM, FTDM, MTDM) under controlled conditions.
  • New Scenario — A realistic 2-km three-phase medium-voltage distribution system including surge arresters and nonlinear components. In this configuration, only the best-performing model (MTDM) is evaluated, following its superior benchmark results.

• Deep Learning Architectures
  Full implementations of the best regression models:

  • MTDM — Modified Time Domain Method
    with MTDM selected as the reference model for the high-fidelity scenario.

• Preprocessing and Feature Engineering Pipeline
  Signal conditioning, dimensionality reduction, and adaptive time-domain compression routines.

• Training and Evaluation Framework
  10-fold cross-validation, performance metrics, and robustness analysis utilities ensuring full reproducibility of the published results.

The repository includes two simulated scenarios used for model development and validation:

Three Deep Learning approaches are implemented and evaluated:

The Modified Time Domain Method (MTDM) demonstrated the best overall accuracy and robustness.

Key characteristics:

• Leading-zero removal (arrival-time normalization)
• Adaptive non-uniform time-domain compression
• Fully connected neural network
• 3 hidden layers (74 neurons each)
• 10-fold cross-validation

The Modified Time Domain Method (MTDM) achieves the following results on the realistic distribution line scenario:

The model maintains stable performance under realistic perturbations:

✔ Additive White Gaussian Noise (AWGN) — tested down to SNR = 25 dB
✔ Variations in soil conductivity (limited sensitivity observed)

1. Navigate to:
Phase 1 — Lightning Geolocation and Peak Current Estimation/
2. Choose:
- Old scenario
- New scenario
3. Follow the README inside each folder.

• Preliminary Breakdown Pulse (PBP) analysis
• Early classification of dangerous return strokes
• Protection-triggering strategy design
• Experimental validation on real networks

FELINES is developed by a multidisciplinary research team from University of Genoa (DITEN), University of Pisa (DESTEC), and University of Campania “Luigi Vanvitelli”.

Department of Naval, Electrical, Electronic and Telecommunication Engineering

Department of Energy, Systems, Territory and Construction Engineering

Department of Engineering

Supported by: Italian Ministry for Education, University, and Research (PRIN 2022) European Union — NextGenerationEU