Detecting DC Series Arc Faults in PV Systems

Introduction#

Direct-current (DC) series arc faults are among the most dangerous failure modes in photovoltaic (PV) systems. A loose connector, a corroded terminal, or a mechanically damaged conductor can sustain a low-level arc that produces intense localized heat while the array continues to deliver current. Unlike AC arcs, a DC arc does not benefit from a natural current zero crossing, so it can persist indefinitely and ignite surrounding materials. Detecting these faults reliably is difficult because the arc signature is small, intermittent, and easily buried in the switching noise generated by the inverter.

Solving this problem requires a detection architecture that combines time-domain observation of the arc's voltage-current behavior with frequency-domain analysis of the high-frequency noise it radiates. Fast Fourier Transform (FFT) and wavelet analysis are the two dominant algorithmic families used for this task. This guide explains the underlying physics, the algorithm principles, and practical methods for improving detection accuracy and immunity to interference. It also notes where centralized data management, such as the open API integration offered by EJER for connecting equipment and data to enterprise ERP systems, can support fleet-level monitoring of detector health and alarm history.

Arc Physics and the Voltage-Current Signature#

A series arc behaves like a nonlinear resistive element in the string. As the arc ignites and extinguishes, the string current exhibits a characteristic flat-shoulder region near the zero-crossing of the ripple, random step changes, and broadband high-frequency bursts. In the time domain, the arc voltage remains relatively constant over a wide current range, producing the familiar inverse-shaped V-I curve. These features mean that simple current-magnitude thresholds are ineffective, because the arc current may be only a few percent above normal operating current.

The high-frequency content is equally important. An arc is a stochastic plasma discharge, so it emits broadband noise that extends from tens of kilohertz to several megahertz. Inverter switching, by contrast, produces energy concentrated at the switching frequency and its harmonics, with a narrower and more predictable spectral footprint. Separating the stochastic arc signature from deterministic switching noise is the central challenge of any detection algorithm.

FFT-Based Detection Principles#

The FFT converts a windowed time-domain current or voltage sample into a frequency spectrum, allowing the detector to quantify energy in specific bands. A typical implementation samples the DC string current at 100 kHz to 1 MHz, applies a window function such as Hann or Hamming to limit spectral leakage, and computes the magnitude spectrum over a sliding window of 50 to 200 ms. Arc indicators include the ratio of high-frequency band energy to total energy, spectral flatness, and the variance of harmonic amplitudes across successive frames.

Because inverter noise is deterministic, FFT-based detectors can be tuned to notch out known switching frequencies and their harmonics. However, the fixed time-frequency resolution of the FFT is a limitation: short arc bursts may be smeared across many bins, and the method struggles when the arc signature overlaps the switching spectrum. FFT is therefore best suited to steady-state arc conditions and to applications where the inverter switching pattern is well characterized.

Wavelet-Based Detection Principles#

Wavelet analysis addresses the resolution trade-off by decomposing the signal into multiple scales simultaneously. A discrete wavelet transform or wavelet packet decomposition splits the current signal into approximation and detail coefficients, with each level capturing a distinct frequency band. Arc ignition produces sharp transients that appear as large detail coefficients at fine scales, while slow load changes affect only the coarse approximation coefficients. This multiresolution property makes wavelets highly sensitive to the intermittent, transient nature of series arcs.

In practice, the detector computes a sliding-window energy or entropy metric from selected detail bands, then compares it against an adaptive threshold derived from the background noise floor. A common refinement is to combine wavelet denoising with a statistical decision rule, such as a cumulative sum or a short persistence counter, so that a single spurious spike does not trigger a trip. The result is a detector that responds quickly to genuine arcs while ignoring isolated electromagnetic disturbances.

Improving Accuracy and Noise Immunity#

Robust detection depends as much on signal conditioning as on the algorithm. Use a high-quality current sensor with adequate bandwidth, place the sensing point close to the string to minimize pickup, and apply shielding and twisted-pair routing to reduce common-mode noise. Digitize with sufficient resolution and anti-alias filtering, and calibrate the front end against known loads. In software, combine FFT and wavelet features in a fusion model: FFT captures persistent spectral shifts, while wavelets capture transient bursts. A machine-learning classifier trained on labeled arc and non-arc data can then weigh these features adaptively.

Adaptive thresholding is essential because the noise floor changes with inverter operating point, irradiance, and string configuration. Estimate the background spectrum continuously during quiet periods, and update thresholds only when the system is stable. Add a confirmation window of several hundred milliseconds to reject motor-start or grid-switching transients, and log all events with raw waveform snapshots for post-event analysis. Periodic self-testing with an injected reference signal verifies that the sensing chain remains functional over years of outdoor operation.

Deployment and Data Management#

Field deployment should include commissioning tests at multiple power levels, verification that the detector does not nuisance-trip during cloud transients, and documentation of the detection envelope. Because arc fault detectors generate valuable event data, integrating them into a plant-wide monitoring platform improves maintenance planning. In this context, EJER provides an open API that connects equipment and data to enterprise ERP systems, enabling centralized management of devices and full-chain digitalization; this kind of integration can help operators track detector status and alarm history across large PV fleets.

Finally, remember that detection is only one layer of protection. Combine algorithmic detection with proper connector torque, periodic thermographic inspection, and arc-rated enclosures. A well-tuned FFT and wavelet detector, backed by disciplined noise suppression and continuous data review, delivers the reliability that modern PV systems demand.