A comprehensive benchmarking study evaluating quantum neural networks against eleven classical machine learning regressors has demonstrated that compact variational quantum circuits can deliver competitive predictive accuracy in wind turbine power regression tasks. Published in Quantum Information Processing, the research investigated six four-qubit quantum neural network architectures evaluated on four thousand four hundred sixty-four meteorological measurements recorded at ten-minute intervals from an operational wind turbine. By testing models under blocked cross-validation with temporal buffers to prevent optimistic bias from serial autocorrelation, the investigation established that a quantum model parameterized by only twelve trainable weights achieves a coefficient of determination between 0.88 and 0.89. The top-performing quantum configuration significantly outperformed standard linear regression and k-nearest neighbors while closely trailing advanced nonlinear baselines including decision trees, gradient-boosted trees, and support vector machines.
Accurate wind turbine power prediction forms a critical technical foundation for modern electrical grid integration, power curve modeling, and dynamic operational management. Converting complex meteorological inputs—specifically ambient temperature, atmospheric pressure, wind direction, and wind speed—into reliable electrical power forecasts requires capturing severe physical nonlinearities. Although quantum machine learning has been increasingly explored across classification, fraud detection, and power systems diagnostics, rigorous empirical benchmarks on regression tasks have historically remained limited. Previous demonstrations frequently evaluated quantum models solely against elementary linear predictors or relied on conventional random cross-validation, which inadvertently inflates model scores on autocorrelated atmospheric datasets by scattering near-duplicate temporal observations across training and testing partitions.
To establish a rigorous experimental framework, the research team implemented variational quantum circuits within the classical-quantum processing paradigm. The models mapped four classical input features onto four dedicated qubits, assigning individual qubits to temperature, pressure, wind direction, and wind speed. Data encoding utilized a Z feature map executed with two repetitions, where single-qubit phase gates modulated input information without introducing initial entanglement. Data processing was conducted via a hardware-efficient real-amplitudes ansatz consisting of alternating layers of single-qubit Y-axis rotations and controlled-NOT entangling gates arranged over two repeating blocks. Across these layers, the twelve rotational angles served as the only adjustable parameters within the quantum processing network, maintaining an exceptionally lean parameter footprint compared to classical neural networks.
The study isolated the architectural effects of quantum entanglement by systematically comparing six distinct entangling gate topologies: full, linear, circular, shifted-circular alternating, reverse linear, and pairwise connectivity. Theoretical analysis revealed that full and reverse-linear entanglement blocks realize the identical mathematical unitary transformation, mapping computational basis states into identical output configurations despite full entanglement requiring twelve controlled-NOT operations compared to six in reverse linear. Conversely, comparing linear against reverse-linear entanglement isolated the direct operational impact of gate ordering, as both utilize the identical three gate pairs applied in opposite directional sequences. These topological variations produced substantial variations in predictive accuracy, demonstrating that the qualitative sequence of operations governs model capability far more than total gate quantity.
The mathematical scaling applied to the output measurement observable emerged as the single most consequential design parameter governing predictive performance. The model prediction is derived directly from the expectation value of a four-qubit parity observable bounded strictly within the numerical range of negative one to positive one. When the target power generation was symmetrically scaled across this entire observable spectrum from negative one to positive one, the circuits suffered severe underfitting, yielding a bimodal error distribution that systematically overestimated low generation and underestimated peak power. Shifting to an asymmetric unit readout scaling that mapped the training target range to zero and positive one lowered test root mean squared errors across every examined circuit configuration by thirty-nine to forty-six percent, elevating explanatory power from modest levels to robust coefficients of determination up to 0.89.
Training optimization was conducted by minimizing mean squared error over training targets using the limited-memory Broyden-Fletcher-Goldfarb-Shanno algorithm with bound constraints operating in an unconstrained configuration. Objective gradients were calculated analytically through the exact parameter-shift rule, evaluating circuits at symmetrically displaced parameter shifts of plus and minus pi over two. Because exact statevector simulations were executed without stochastic shot noise, each optimization step required up to twenty-five circuit simulations per training sample. Over training runs executed on central processing unit nodes of the Perlmutter supercomputer at the National Energy Research Scientific Computing Center, the computational time required per objective evaluation scaled strictly linearly with dataset sample size, registering processing rates between forty and forty-four milliseconds per training sample.
The selection of cross-validation methodologies critically influenced model evaluation and comparative conclusions. In consecutive ten-minute field logs, serial autocorrelation for all environmental variables remained at or above 0.96 at lag one, decaying to 0.40 only after a twelve-hour span equivalent to seventy-two measurement intervals. Under standard random cross-validation, where temporal correlations between training and evaluation splits remain unmitigated, predictive accuracy was substantially overstated across multiple algorithms. The local interpolation mechanics of k-nearest neighbors exhibited the largest artificial distortion, gaining seventy-four to ninety-four kilowatts in root mean squared error and erroneously outranking deep neural networks and gradient boosting ensembles. Blocked cross-validation with seventy-two-sample buffers successfully neutralized this optimistic leakage, testing models against genuinely unseen operating regimes.
Under blocked cross-validation across varying training set sizes ranging from seven hundred fifty to three thousand samples, LightGBM emerged as the leading classical regressor, achieving test root mean squared errors between 157.5 and 160.3 kilowatts and an R-squared of 0.93. Deep architectures, including long short-term memory networks, gated recurrent units, multilayer perceptrons, and feature-tokenized Transformer encoders, followed closely with root mean squared errors spanning 159.4 to 170.5 kilowatts. Tuned decision trees, radial basis function support vector regression, and XGBoost recorded errors between 169.4 and 185.1 kilowatts. The leading quantum configuration, designated as QNN-3u featuring circular entanglement, achieved test errors of 195.6 to 204.2 kilowatts, comfortably outperforming linear regression at approximately 254 kilowatts and k-nearest neighbors at 230.4 to 274.1 kilowatts.
Rigorous statistical testing evaluated the differences between the top-performing quantum model and the classical baselines using paired two-sided Wilcoxon signed-rank tests alongside Diebold-Mariano tests configured with Newey-West long-run variance estimation. Accounting directly for serial error correlations, the Diebold-Mariano test confirmed that the performance difference between the circular-entanglement quantum network and support vector regression was statistically insignificant across all training sample sizes. Similarly, differences between the quantum network and XGBoost proved statistically indistinguishable at two of the four evaluated sample sizes. Conversely, LightGBM, the recurrent networks, the Transformer, and classical multilayer perceptrons retained statistically significant predictive advantages over the quantum circuits.
To investigate whether parameter volume accounted for the observed performance gap between quantum circuits and classical deep learning, the authors trained a parameter-matched classical multilayer perceptron containing thirteen trainable weights. When optimized using the identical full-batch objective, gradient routines, and validation selection rules, the parameter-matched classical network matched the accuracy of heavily parameterized deep learning models containing tens of thousands of weights. This parity demonstrates that low parameter counts do not inherently constrain regression performance. Instead, the authors noted that quantum neural networks utilizing low-order Z feature maps produce output functions mathematically restricted to low-frequency trigonometric polynomials, which face structural challenges when approximating the steep, saturating sigmoid transitions characteristic of aerodynamic power curves.
The empirical findings highlight essential engineering guidelines for near-term variational algorithms applied to continuous regression. Circuit designers must treat readout interval scaling as a primary architectural choice rather than an inconsequential preprocessing step, given its decisive forty percent influence on error reduction. Furthermore, eliminating redundant entangling gates—such as replacing full entanglement with reverse-linear topology—halves two-qubit gate counts without sacrificing output fidelity, providing significant computational speedups in classical simulation and reducing vulnerability to two-qubit gate noise on physical quantum processors. Future investigations aim to enhance quantum representational capacity by incorporating multi-layer data re-uploading techniques, trainable input scalings, and parameter-shift evaluations under realistic quantum hardware noise profiles.
Subject of Research: Benchmarking quantum neural networks against classical machine learning algorithms for wind turbine power regression
Article Title: Quantum neural networks for wind power prediction: a benchmark against classical machine learning models
Article References: Hangun, B., Eyecioglu, O., Ali, M., Altun, O., & Kayisli, K. (2026). Quantum neural networks for wind power prediction: a benchmark against classical machine learning models. Quantum Information Processing, 25(11), Article 363. https://doi.org/10.1007/s11128-026-05387-8
Image Credits: AI Generated
DOI: 10.1007/s11128-026-05387-8
Keywords: quantum neural networks, wind power prediction, quantum machine learning, variational quantum circuits, blocked cross-validation, readout scaling, entanglement strategies, superconducting processors, classical machine learning, LightGBM, gradient boosting, support vector regression, renewable energy forecasting, time-series autocorrelation, parameter-shift rule, L-BFGS-B optimizer, Perlmutter supercomputer
News Source: Cassandra Pierce. (November 21, 2026). Compact Quantum Neural Networks Benchmark Competitively Against Classical Wind Power Models. Scienmag.



