DARC PV

This executive summary outlines the feasibility test of the DARC PV (Photovoltaic Arc Detection 
Device), conducted at the Smart Grid Interoperability Laboratory of the Joint Research Centre 
(JRC) in Ispra, Italy, from February 24 to March 7, 2025. Developed by Anastasios Kladas and 
Arman Ghaderi Baayeh from KU Leuven, the DARC PV leverages a Real-Time Digital Simula
tor (RTDS) to replay high-frequency arc time-series data, aiming to enhance arc fault detection 
in photovoltaic systems. The test successfully validated the device’s capability to identify arc 
faults with high precision, offering a promising solution for improving the safety and reliability of 
PV systems. 
Photovoltaic systems are a cornerstone of renewable energy, yet arc faults remain a critical 
challenge, posing risks of fires, equipment damage, and safety hazards. Traditional detection 
methods—relying on physical indicators, frequency domain techniques like Fast Fourier Trans
form (FFT), and even advanced wavelet transforms—have limitations, including sensitivity to 
noise and complexity in real-world applications. Emerging machine learning (ML) approaches, 
such as the Multi-Layer Perceptron (MLP) classifier used in DARC PV, offer a data-driven alter
native, capable of processing complex datasets to detect arc faults accurately. This feasibility 
test was designed to assess the DARC PV’s performance using simulated high-frequency data 
replayed via RTDS, addressing the need for reliable, real-time arc fault detection. 
The DARC PV setup involved a Raspberry Pi 4 paired with an ADALM2000K analog-to-digital 
converter, connected to an RTDS system that replayed simulated PV arc data at a 250 kHz 
sampling frequency. The test utilized a Simulink-designed PV system model with weather data 
from NIST’s Maryland station, simulating arcs under specific irradiance conditions. Data were 
processed into frequency bands (0–5 kHz, 5–50 kHz, 50–100 kHz, 100–250 kHz), with the MLP 
model trained on voltage signals from Bands 1, 2, and 3. Despite noise from the RTDS inverter, 
the optimized model—featuring a tanh activation function, 30 hidden neurons, and Adam solv
er—achieved 100% precision in detecting arc faults on both training and testing datasets. Tech
niques like Recursive Feature Elimination (RFE), Synthetic Minority Oversampling Technique 
(SMOTE), and hyperparameter tuning via GridSearchCV enhanced the model’s robustness 
against noise and class imbalance. 
The DARC PV demonstrated exceptional performance, accurately distinguishing arc faults from 
background noise in simulated conditions. The model’s 100% precision highlights its potential 
as a reliable detection tool. Key features included voltage signals across higher frequency 
bands, processed at a 0.5-second resolution due to computational constraints. While the test 
relied solely on simulated data due to time limitations, the methodology—combining prepro
cessing, feature engineering, and ML optimization—proved effective, with training loss conver
gence indicating a stable, scalable solution. 
Recommendations 
1. Expand Testing Scope: Conduct further tests with real-world measured data to validate 
the DARC PV’s performance beyond simulated conditions. 
2. Enhance Computational Capacity: Upgrade the Raspberry Pi or integrate a more power
ful processor to reduce data resolution below 0.5 seconds, enabling faster real-time de
tection. 
3. Field Deployment: Initiate pilot deployments in operational PV systems to assess practi
cal integration and long-term reliability.
These recommendations are justified by the test’s success with simulated data and the identi
fied limitations. Real-world validation is essential to confirm generalizability, while improved 
computational power would align the device with real-time application demands. Pilot deploy
ments would provide critical insights into scalability and operational challenges, accelerating 
adoption in the renewable energy sector. 
The feasibility test of the DARC PV at the JRC facilities marks a step toward reliable arc fault 
detection in PV systems. Achieving 100% precision with an MLP-based approach, the device 
showcases the power of machine learning in addressing a critical safety issue. While con
strained to simulated data, the results underscore its potential for real-time applications. With 
further testing and refinement, the DARC PV could enhance PV system safety and efficiency, 
supporting the broader transition to sustainable energy. This summary stands independently, 
offering decision-makers a clear path to advance this technology.