Does quantum machine learning actually help on the smart grid?

Most papers claiming a "quantum advantage" in machine learning are comparing a carefully tuned quantum model against a classical baseline that nobody tuned at all. That is not a fair fight, and it is not a useful result.

So when we built QLID-Net, a hybrid quantum–classical network for smart-grid load identification, we made ourselves a rule: the classical baseline gets the same hyperparameter budget as the quantum model. Same search, same compute, same patience.

What we were actually predicting#

Non-intrusive load monitoring (NILM) asks: given only the aggregate power signal at a building's meter, which appliances are running? It is a multi-label classification problem over a signal that is noisy, non-stationary, and — this is the important part — expensive to label.

That last property is why it is a reasonable place to look for a quantum advantage. The theoretical arguments for QML rarely promise faster training. What some of them promise is better generalisation from few samples.

The architecture#

The model is a variational circuit sandwiched between two classical layers:

import pennylane as qml
from pennylane import numpy as np

n_qubits = 8
dev = qml.device("default.qubit", wires=n_qubits)

@qml.qnode(dev, interface="torch", diff_method="backprop")
def circuit(inputs, weights):
    # Angle embedding: each feature becomes a rotation angle.
    qml.AngleEmbedding(inputs, wires=range(n_qubits), rotation="Y")

    # Strongly-entangling ansatz — the part that is actually learned.
    qml.StronglyEntanglingLayers(weights, wires=range(n_qubits))

    return [qml.expval(qml.PauliZ(w)) for w in range(n_qubits)]

The classical encoder compresses 128 spectral features down to the 8 the circuit can accept; the decoder maps the 8 expectation values back out to appliance labels.

The result that mattered#

With the full training set, the classical baseline matched the hybrid model. No advantage. If we had stopped there — as a lot of papers effectively do, in the other direction — the story would have been "quantum wins," and it would have been wrong.

The interesting behaviour only appears when you starve both models of data:

Training samples	Classical $F_1$	Hybrid $F_1$	Δ
5,000	0.913	0.916	+0.003
1,000	0.871	0.894	+0.023
500	0.804	0.858	+0.054
100	0.612	0.731	+0.119

The gap widens as the data shrinks. At 100 labelled samples the hybrid model is nearly 12 $F_1$ points ahead.

Why this might be happening#

The usual explanation is expressivity: the feature map

$$\phi(x) = U(x)\,|0\rangle^{\otimes n}$$

lifts the input into a $2^n$-dimensional Hilbert space, and the induced kernel

$$k(x, x') = \left| \langle 0 |\, U^{\dagger}(x)\, U(x')\, |0 \rangle \right|^{2}$$

can separate points that a classical kernel of comparable parameter count cannot.

I want to be careful here. That is a plausible mechanism, not a demonstrated one. We measured an effect; we did not prove its cause. The honest summary is:

On this problem, under a matched hyperparameter budget, the hybrid model generalises better from small labelled datasets. We do not yet know how far that finding travels.

What we are doing next#

• Testing whether the advantage survives on real meter data rather than simulated aggregates.
• Running on actual quantum hardware, where noise may eat the entire margin.
• Checking whether a classical model with a comparable inductive bias — not just comparable parameter count — closes the gap.

If the third experiment closes the gap, then what we found is a statement about inductive bias, not about quantum mechanics. That would still be worth knowing.