Electric-Traction Parameter Identification from Instantaneous Power (Mishchenko, 2012)

T. M. Mishchenko's 2012 paper asks how a complicated traction device can be replaced by a tractable terminal model without reproducing its internal circuit. The proposed answer is to observe only terminal voltage $u(t)$ and current $i(t)$, represent the device as a passive one-port, and infer a parallel resistive-reactive equivalent from instantaneous power. The application is specific to AC locomotives, but the identification problem is the general one behind waveform-based RLC identification.

The proposed one-port decomposition #

The paper separates current into an active component with the same waveform as voltage and an orthogonal residual:

$$i(t)=i_a(t)+i_r(t), \qquad i_a(t)=\frac{P}{U^2}u(t), \qquad i_r(t)=i(t)-\frac{P}{U^2}u(t),$$

where $P$ is average active power over the observation interval and $U$ is RMS voltage over that interval. This makes the static parallel resistance

$$R=\frac{U^2}{P},$$

and assigns the remaining instantaneous power to the reactive branch:

$$q(t)=u(t)i_r(t)=u(t)i(t)-\frac{P}{U^2}u^2(t).$$

That construction is a projection: $i_a(t)$ is the component of current parallel to $u(t)$ under the interval inner product, while $i_r(t)$ is orthogonal to it. The terminology is therefore broader than the sinusoidal active/reactive split and belongs with instantaneous power in nonsinusoidal circuits.

From power to equivalent L and C #

For a parallel $R$-$L$ model, the method equates $q(t)$ to the derivative of magnetic energy $L i_r^2/2$. For a parallel $R$-$C$ model, it equates $q(t)$ to the derivative of electric energy $C u^2/2$. Closed-form expressions are given when $L$ or $C$ is treated as constant; first-order differential equations are derived when the equivalent parameter itself varies with time.

This distinction turns out to be decisive. Applying the constant-$L$ expression produces values above 150 H, which the paper itself rejects as physically implausible for the modeled locomotives. Solving the time-varying equations produces parameter trajectories that reproduce the assumed power balance much more closely, but some inferred $L(t)$ and $C(t)$ values become negative. The author explicitly treats those values as computational parameters without direct physical meaning, suggesting that a negative inductance may stand in for capacitive behavior and vice versa.

Numerical results reported #

The calculations reuse measured input waveforms for DS3 and 2ES5K locomotives from a cited 2009 study. The paper reports these current-dependent static resistances:

Substitution of the inferred $R(I)$, $L(t)$, and $C(t)$ into the paper's differential equations is reported to leave a 0.3–0.4% equation residual. That is an internal consistency check against the same formulation, not independent validation against withheld waveforms, another operating regime, or a physical measurement of stored energy.

What survives and what does not #

The durable contribution is the modeling workflow: reduce a difficult device to terminal behavior, make the assumed equivalent topology explicit, derive parameters from an energy or power identity, and test whether the result remains physically interpretable. The paper is also a compact example of why equivalent one-port topology and identifiability cannot be separated.

The method needs more evidence before it can be treated as a general identification procedure:

• differentiating measured voltage and current amplifies noise, but the paper gives no filtering, regularization, sampling-rate, or uncertainty analysis;
• the numerical study reuses waveforms from another publication and does not describe a train/test split or independent operating-condition validation;
• a small residual in the defining differential equation shows algebraic closure, not uniqueness, passivity, causality, or predictive power;
• negative equivalent storage parameters are a sign that the selected parallel $R$-$L$ or $R$-$C$ topology is representing behavior outside the physical meaning of its components.

Use the paper as a worked case in system identification, especially for the difference between fitting terminal behavior and recovering a physically meaningful internal model.

Reproducible projection check #

The active-current formula is the least-squares projection of current onto voltage over the observation interval. Choose a scalar $a$ to minimize

$$J(a)=\langle(i-au)^2\rangle.$$

Setting $dJ/da=0$ gives $a=\langle ui\rangle/\langle u^2\rangle=P/U^2$, exactly the paper's coefficient. Consequently $\langle u(i-au)\rangle=0$. This derivation establishes the decomposition without assuming sinusoidal waveforms, but it establishes only orthogonality—not a unique reactive-element topology.

Evidence classification #

The paper is therefore strong as a modeling caution and weak as evidence of a generally validated estimator. Its numerical closure should be reproduced only from archived waveforms and followed by out-of-sample tests, constraint checks, and uncertainty analysis.