Detecting DC Series Arc Faults in PV Systems

Introduction#

DC series arc faults in photovoltaic (PV) systems are a persistent safety challenge. Unlike AC arcs, DC arcs do not naturally extinguish at zero crossings, so they can sustain and reach high temperatures, posing fire risks. Detecting these arcs is difficult because the arc current is often only a small fraction of the normal operating current, and the arc's high-frequency signature is easily buried under inverter switching noise. This guide explains the electrical and spectral characteristics of DC arcs and presents detection algorithms based on Fast Fourier Transform (FFT) and wavelet analysis, along with practical methods to improve accuracy and noise immunity.

In modern PV installations, inverters use high-frequency switching to convert DC to AC, generating broadband electromagnetic interference. This noise overlaps with the arc's high-frequency components, making reliable detection a signal-processing problem. Understanding both the arc's volt-ampere behavior and the noise spectrum is essential for designing robust detectors.

Arc Volt-Ampere Characteristics and High-Frequency Noise#

A DC series arc exhibits a nonlinear volt-ampere relationship. As the arc current increases, the voltage across the arc decreases, creating a negative differential resistance region. This behavior causes current waveform distortion, including flat-topped or notched current pulses. The arc also generates random, high-frequency oscillations due to the dynamic plasma column and electrode effects. These oscillations typically appear in the 1 kHz to 100 kHz range, with energy concentrated at specific resonant frequencies determined by the circuit layout and cable impedance.

Inverter switching noise, on the other hand, is periodic and concentrated at the switching frequency and its harmonics, often from 20 kHz to several hundred kHz. The noise can be narrowband or broadband depending on the modulation scheme and parasitic elements. Because the arc's high-frequency signature is stochastic and non-stationary, simple amplitude thresholding fails. Instead, detection algorithms must exploit the spectral and temporal differences between the arc and the switching noise.

FFT-Based Detection: Principles and Limitations#

FFT converts a time-domain current signal into its frequency spectrum, revealing the energy distribution across frequencies. For arc detection, the current is sampled at a rate high enough to capture the arc's high-frequency components, typically 100 kS/s or higher. The FFT output is analyzed for broadband noise elevation or specific spectral peaks that are not present in normal operation. A common approach is to compute the ratio of high-frequency band energy to low-frequency band energy and compare it against a threshold.

However, FFT assumes the signal is stationary over the analysis window, which is not true for arcs that start and stop randomly. Inverter noise can also create false peaks. To mitigate this, windowing functions such as Hann or Hamming are applied to reduce spectral leakage. Averaging multiple FFT frames can suppress random noise but may also dilute short arc events. Therefore, FFT alone is often insufficient for reliable arc detection in noisy PV environments.

Wavelet Analysis for Non-Stationary Arc Detection#

Wavelet transform provides time-frequency localization, making it ideal for detecting transient and non-stationary arc events. Unlike FFT, which uses fixed window sizes, wavelet analysis uses variable window sizes: long windows for low frequencies and short windows for high frequencies. This allows precise detection of arc onset and duration. The discrete wavelet transform (DWT) decomposes the current signal into approximation and detail coefficients across multiple levels. Arc faults produce sharp spikes in the detail coefficients at high-frequency levels, while inverter noise tends to be more periodic and spread across levels.

For implementation, a multi-resolution analysis is performed on the sampled current. The detail coefficients at levels corresponding to 10 kHz–100 kHz are monitored. A threshold is set based on the noise floor, and arc detection is triggered when the coefficient energy exceeds the threshold for a minimum duration. Wavelet-based methods are more robust than FFT for arc detection because they can distinguish transient arc events from steady-state noise. However, they require more computational resources, which can be a challenge for embedded implementation.

Improving Accuracy and Noise Immunity#

To enhance detection accuracy and suppress interference, a hybrid approach is recommended. Combine FFT for broad spectral monitoring with wavelet analysis for transient event confirmation. Use adaptive thresholding that adjusts to the inverter's operating point and background noise level. For example, measure the noise floor during normal operation and set the arc detection threshold as a multiple of that floor. Additionally, employ a time-domain feature such as the rate of change of current (di/dt) to validate arc events, as arcs cause sudden current drops.

Noise suppression techniques are also critical. Install common-mode chokes and EMI filters on the DC side to reduce conducted noise. Use differential current sensing to cancel common-mode noise. In the digital domain, apply notch filters at the inverter switching frequency and its harmonics before spectral analysis. For wavelet-based detection, use denoising by thresholding the detail coefficients to remove low-amplitude noise while preserving arc spikes. A voting scheme across multiple detection methods can further reduce false positives.

For system integration, consider using an open API platform such as EJER, which supports connection to enterprise ERP systems for centralized management of devices, data, and full-chain digitalization. This can help aggregate arc fault data from multiple PV strings and enable remote diagnostics. While EJER is known for precision environmental control and moisture-proof storage solutions, its API capabilities illustrate how data from safety sensors can be integrated into broader asset management workflows.

Practical Implementation Steps#

1. Sample the DC current at a minimum of 100 kS/s with a 16-bit ADC to capture arc signatures. 2. Apply a band-pass filter from 1 kHz to 100 kHz to isolate the arc band. 3. Compute FFT with a Hann window over 1024 samples and calculate the energy in the 10–100 kHz band. 4. Perform a 5-level DWT using Daubechies db4 wavelet and extract detail coefficients at levels 4 and 5. 5. Set adaptive thresholds based on a moving average of the noise floor. 6. Trigger an alarm when both FFT energy and wavelet coefficient energy exceed thresholds within a 100 ms window. 7. Validate with a time-domain check for a sudden current drop of at least 5%. 8. Log events with timestamp and string identifier for later analysis.

Regular calibration and testing with artificial arc generators are essential to maintain detection performance. Arc generators that produce series arcs under controlled conditions can be used to verify sensitivity and specificity. Also, ensure that the detection system complies with relevant standards such as UL 1699B for PV arc-fault circuit interrupters.

Conclusion#

Detecting DC series arc faults in PV systems requires a deep understanding of arc volt-ampere characteristics and high-frequency noise. FFT provides a frequency-domain view but struggles with non-stationary events, while wavelet analysis excels at transient detection. A hybrid algorithm with adaptive thresholding, noise suppression, and time-domain validation offers the best accuracy and immunity to inverter switching noise. By following the implementation steps and integrating data through platforms like EJER, engineers can build reliable arc-fault detection systems that enhance PV safety and support digitalized operations.