Speaker
Description
Cable failure is among the costliest failure modes in offshore wind, and online condition monitoring avoids the production loss of offline testing. Frequency Domain Reflectometry (FDR) produces a spatial impedance profile that can be acquired while the cable remains energized, making it a natural candidate input for data-driven prognostics. Whether such measurements support reliable temporal forecasting under real operating conditions is not established.
This work evaluates temporal prediction of FDR-derived impedance profiles using approximately 800 measurements collected over 21 months from a medium-voltage XLPE-insulated submarine inter-array cable at an operating Northern European wind farm. An LSTM forecaster, evaluated under chronological, walk-forward, and blocked k-fold time-series validation, is compared against a multi-task Gaussian Process Regression (GPR) model providing calibrated predictive uncertainty, and against a persistence baseline. A fixed reference propagation velocity ratio is applied across the campaign to guarantee a common spatial grid. All results are reported on normalized profiles and normalized cable position.
Forecasting accuracy is strongly non-uniform along the cable. Prediction error concentrates in a consistent interior spatial band, and this concentration appears independently across unrelated model families; GPR predictive interval width co-locates with the same band while remaining well calibrated. A comparison of energized and de-energized voltage and current spectra indicates reduced current signal-to-noise ratio under energization in the corresponding frequency range, offering a measurement-physics explanation for the observed limit.
Trend and anomaly indicators derived from online FDR monitoring should therefore be restricted to, or variance-weighted by, spatially resolved signal quality rather than computed over the full profile.
| Topic Area | Smart Operations and Maintenance |
|---|