The Power Line That Diagnoses Itself: A New Math for Microgrid Faults

Inside a DC power line strung with a dozen or more converters — the nervous system of a modern microgrid, where solar arrays, batteries, and electric loads plug in at scattered points along a single conductor — a fault can be almost invisible. A converter starts silently drawing less power than it should. A short drifts slowly. A sensor goes quietly wrong. Nothing trips a circuit breaker, because nothing dramatic happens. But the problem is real, and it is sitting somewhere along a line that can stretch for hundreds of meters with only a handful of measurement points watching it.

The hard part is not that the signal is small. It is that the signal is everywhere. A disturbance at one point on a conductor doesn't stay at that point — it spreads out along the whole line, riding the electromagnetic dynamics of the cable, arriving at your sensors mixed together with the line's own behavior and with the jitter of every other converter. You cannot tell, from a single measurement, whether the anomaly came from a plant shutting down at position $x=10$ m or a failing sensor at $x=85$ m. The information is there, but it is folded together.

A team of engineers from the University of Almería and the research institute SINTEF in Norway has built a mathematical tool that unfolds it. Their trick, reported in a 2026 preprint, is to give time and space separate imaginary directions — two independent geometric "imaginary units," one for the temporal phase and one for the spatial phase — so that a fault's signature factors apart like $\dots$. The temporal character of the fault stays in one factor, and its location becomes a readable angle in the other. In the researchers' own words, the position of a fault "can be read as a geometric argument" of a single transformed object that also encodes what the fault is (Arrabal-Campos et al., 2026). It is, to use their phrase, a genuinely two-dimensional answer to a genuinely two-dimensional problem.

The Science

The paper, Multivariable Geometric Laplace Transform and Fault Detection in Distributed-Converter Lines, is aimed at a specific but increasingly common physical setup: a DC distribution line — the kind that links renewable generators, storage, and electronic loads in a microgrid — with many power converters attached at distinct positions along its length. Voltage and current along such a line depend on both time $t$ and the spatial coordinate $x$ simultaneously. This is a "spatio-temporal" system, and the paper's central claim is that both dimensions deserve their own phase.

To understand the mathematics, it helps to recall what the ordinary Laplace transform does. A classical Laplace transform takes a time signal $f(t)$ and turns it into a function $F(s)$ using an exponential kernel $e^{-st}$, where the complex variable $s=\sigma +j\omega$ and $j^2=-1$. The imaginary unit $j$ encodes the phase — the wobble — of the oscillatory parts of the signal. The transform is the workhorse of control and signal-processing theory: it turns messy differential equations into algebra.

Here is the crucial subtlety the authors exploit. There is nothing sacred about the specific choice of $j$. It can be replaced by any element $B$ of an algebra that squares to $-1$. If $B$ is a bivector — a geometric object representing an oriented plane, like a signed patch of area spanned by two vectors — then the whole calculus still works, because everything rests on the single relation $B^2=-1$. With one independent variable, swapping $j$ for a bivector gives you a geometric interpretation of phase but no new power.

The situation changes with two independent variables, time and space. Now a temporal phase (tied to frequency $\omega$) and a spatial phase (tied to wavenumber $k$) coexist. The authors introduce two bivectors, ${\mathbf{B}}_t$ and ${\mathbf{B}}_x$, each squaring to $-1$, living in algebraically distinguishable planes, and commuting with each other. Their product ${\mathbf{B}}_{tx}={\mathbf{B}}_t{\mathbf{B}}_x$ squares to $+1$. These four objects $\{1,{\mathbf{B}}_t,{\mathbf{B}}_x,{\mathbf{B}}_{tx}\}$ generate a four-dimensional commutative algebra, a small slice of the larger geometric algebra $\mathrm{Cl}(4,0)$. This subalgebra turns out to be isomorphic to Segre's bicomplex numbers — a number system that is essentially "complex numbers made of complex numbers," where an element is $a+bi+cj+dij$ with $i^2=j^2=-1$ and $ij=ji$.

On top of this algebra they define a two-dimensional geometric Laplace transform:

$${\mathcal{L}}_G\{f\}=F(s_t,s_x)={\int }_0^{\infty }{\int }_0^{\infty }f(t,x)e^{-s_{t}t}{e}^{-s_{x}x}\mathrm{d}x\mathrm{d}t,$$

with geometric variables $s_t={\rho }_t+{\mathbf{B}}_t\omega$ and $s_x={\rho }_x+{\mathbf{B}}_{x}k$. The factor $e^{-s_{t}t}$ carries temporal attenuation and phase in the ${\mathbf{B}}_t$ plane; $e^{-s_{x}x}$ carries spatial attenuation and phase in the ${\mathbf{B}}_x$ plane. Because the two phases live in distinct planes, they are never conflated. The transform inherits a full suite of operational properties — linearity, a rule for how derivatives transform, regions of convergence — that mirror the classical one-sided complex Laplace transform. When you drop the spatial variable, it reduces exactly to the familiar single-variable transform. The mathematics is not new as an integral object; the contribution is the geometric embedding that makes the two phases legible and the operational use the authors put it to.

From this foundation they build a complete toolbox: a transmission-line model of the converter network with its space–time "dispersion relation" $\Delta (s_t,s_x)=0$, a distributed control law by "admittance shaping," an exact treatment of discrete converter sites including spatial sampling and aliasing, and — the heart of the paper — a fault-diagnosis chain.

Figure 1 (above) shows the dispersion relation of the distributed line. For an ideal lossless line the relation is exact — $k=\sqrt{\ell c}\omega$, a nondispersive line with constant phase velocity — while the dashed lossy curve is shown for reference. The mathematics separates the temporal and spatial dynamics cleanly, which is precisely what enables the diagnosis scheme that follows.

What They Found

The centerpiece is the signature of a localized fault. Suppose a fault occurs at a position $x_f$ along the line — a converter losing its injection, a shunt short developing, a sensor drifting. The researchers show that, in the transformed domain, the residual left by such a localized fault factorizes:

$$F_f(s_t)e^{-s_x{x}_f}.$$

The first factor, $F_f(s_t)$, encodes the temporal nature of the fault — what kind of fault it is. The second factor, $e^{-s_x{x}_f}$, carries a spatial phase that depends only on where the fault is. Because the two phases live in distinguishable bivector planes, you can read the location directly as an "argument" in the ${\mathbf{B}}_x$ plane, through an operator ${\mathrm{Arg}}_{{\mathbf{B}}_x}$ — in exact analogy to how the temporal phase is read with ${\mathrm{Arg}}_{{\mathbf{B}}_t}$. This factorization is the engine of the whole diagnosis framework: the temporal character and the spatial location are separated on a single transformed object.

On a finite line, where the boundaries impose standing spatial modes ${\varphi }_n(x)=\sin (k_{n}x)$ with $k_n=n\pi /L$ (Figure 2), the same idea works through modal weights. Each spatial mode $n$ has its own temporal denominator, whose roots are its modal poles — the "natural frequencies" that govern its dynamics. A fault at $x_f$ weights the modes in a characteristic way, and an estimator recovers the position from the relative weights. Figure 6 shows the cost function of this estimator against candidate positions; its sharp minimum marks the true fault location. A second, phase-based estimator (Figure 7) confirms the result independently.

The diagnosis chain distinguishes four fault classes: injection loss (a converter failing to inject power, whose signature is proportional to $1/s$), shunt faults (resistive or capacitive shorts to ground), sensor faults, and local-controller faults. Crucially, it can also tell apart a plant outage from a cable defect — two failures that look similar at first glance but have different operational consequences. And it extends to multiple simultaneous faults, with automatic order selection to decide how many faults are present and where each one is.

The control layer works by "admittance shaping" — adding a distributed virtual conductance that shapes the line's dynamics in the $(s_t,s_x)$ domain. Its effect is dramatic. In open loop, the fundamental spatial mode of the modeled line is unstable: it carries a pole with real part $+8.240$. Under the distributed virtual-conductance control, all the tested modal poles shift to $\mathrm{Re}=-175.000$. In plain terms: a line that would blow up on its own is stabilized firmly.

Distributed control stabilizes every mode

Label	Value
Mode 1 (open loop)	8.24
Mode 1 (closed loop)	-175
Mode 2 (closed loop)	-175
Mode 3 (closed loop)	-175
Mode 4 (closed loop)	-175
Mode 5 (closed loop)	-175

Source: Arrabal-Campos et al. (2026). Modal poles of the first five modes, open loop vs. closed loop (Fig. 3). Positive real part = unstable growth; negative = decaying response.

The closed-loop control pulls every modal pole far into the left half-plane — the region where disturbances decay instead of growing. The open-loop instability (positive real part, $+8.240$) is replaced by a strongly damped response at $-175.000$ across all five modes shown.

Open-loop instability concentrated in the fundamental mode

Label	Value
Mode 1 (open loop)	8.24
Mode 2 (open loop)	0
Mode 3 (open loop)	0
Mode 4 (open loop)	0
Mode 5 (open loop)	0

Source: Arrabal-Campos et al. (2026).

Figure 4 in the paper shows the practical consequence: the modal time response, which diverges in open loop, settles cleanly under control.

Why This Changes Things

Fault location in DC lines is not a solved problem, and the existing toolkit has real gaps. The established techniques — traveling-wave methods, differential protection, active impedance estimation — mostly target the large-signal transient of a hard short circuit. They require pronounced transients, or dedicated injection hardware, or both, and they trip protections that are looking for something big.

The niche this work addresses is complementary and arguably more common in practice: incipient faults that never grow big enough to trip the large-signal protection. An injection that slowly degrades. An admittance that drifts. A sensor that lies. A local controller that quietly misbehaves. These are the faults a monitoring system should catch early, before they cascade — and they are precisely the ones existing large-signal tools are blind to. The paper's framework works in the small-signal incremental regime (the authors state this explicitly as Assumption 1.1: deviations from a steady operating point are small enough for a linear description), detects the fault, locates it, and classifies it, all within one model-based framework.

Equally important is what the work doesn't overclaim. The researchers are careful to note that the geometric separation of phases is an aid to reading fault positions, not a magic eraser of the intrinsic limits of any spatial monitoring problem. Detection and localization still hinge on the number and placement of sensors, on measurement noise, on boundary conditions, and on the quality of the linear incremental model. They repeat Remark 1.4 like a mantra of intellectual honesty: the results "should not be understood as a universal guarantee of diagnosis." That is the right frame — and it's part of why the claims that remain are worth taking seriously.

The modular design matters too. Because the transient properties of the transform carry over cleanly from the classical case, the framework is not a one-off mathematical curiosity. It sits on a rigorous foundation (bicomplex analysis is a mature field) and is backed by a fully reproducible open implementation — the authors state plainly that "all results are reproduced by an accompanying open implementation," a standard that makes the work testable rather than merely plausible.

The authors are also candid about what they did not invent. The two-dimensional Laplace transform as an integral object is known in the bicomplex-analysis literature; with one variable it reduces to the complex Laplace transform. Their contribution, delimited carefully in Section 1.3, is the geometric embedding into $\mathrm{Cl}(4,0)$ that makes the two phases algebraically distinguishable, and the operational exploitation of that distinction for fault diagnosis. This honesty about novelty is rare and valuable — it lets a reader trust the parts that are claimed as new.

What's Next

The most exciting open thread — flagged by the authors in their conclusions — is the move to noncommutative bivectors. The current algebra works because ${\mathbf{B}}_t$ and ${\mathbf{B}}_x$ commute; that commutativity is what preserves the clean classical operational calculus. But real physical systems do not always cooperate with such clean separations. Extending the geometric embedding to cases where the two phases interact noncommutatively would open the door to richer models — and to problems the commutative framework cannot touch. That is a natural next frontier.

So too is the question of where this generalizes. The paper demonstrates the framework on a DC line with distributed converters — a niche within microgrids. But the underlying mathematics is about any spatio-temporal system: a transmission line of any kind, a diffusion process, a distributed sensor network, even certain classes of partial differential equations. The transform is a tool; the fault-diagnosis chain is the first concrete application. A reader comes away suspecting the tool has other jobs waiting.

For the practical world of microgrids — which are multiplying as renewable generation, storage, and electronic loads cluster on shared DC buses — the promise is a monitoring layer that catches trouble while it is still small, tells you where it is, what it is, and even whether multiple problems are brewing at once. The gap between a fault appearing and a fault being understood is where blackouts, fires, and cascading equipment failures are born. A framework that shrinks that gap — reproducibly, honestly, and with the underlying math available to be checked — is a meaningful step toward grids that diagnose themselves instead of merely failing.

This is early-stage theory with a working implementation, not a deployed product with field trials. But it marks a genuinely different way of seeing the problem: give time and space their own separate imaginary directions, and the tangled story of a sick power line starts to unravel into clean, readable parts.

Distinct spectral fingerprints for each fault class

Label	Value
Injection loss	1
Resistive shunt	1
Capacitive shunt	2
Inductive shunt	3

Source: Arrabal-Campos et al. (2026). Spectral signatures of fault models (Fig. 8). Values indicate the asymptotic frequency scaling family of each fault signature.