Quantum neural networks for wind power prediction: a benchmark against classical machine learning models
Quantum Information Processing, cilt.25, sa.11, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 25 Sayı: 11
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s11128-026-05387-8
- Dergi Adı: Quantum Information Processing
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Compendex, INSPEC, MathSciNet, zbMATH, Academic Search Ultimate (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Benchmarking, Blocked cross-validation, Entanglement, Quantum machine learning, Quantum neural network, Readout scaling, Wind power prediction
- Açık Arşiv Koleksiyonu: AVESİS Açık Erişim Koleksiyonu
- Yıldız Teknik Üniversitesi Adresli: Evet
Özet
Quantum neural networks (QNNs) are frequently proposed for regression on small datasets, yet the design and evaluation choices that govern their accuracy are rarely isolated. This study benchmarks six QNN circuits, built from a Z feature map and a real-amplitudes ansatz with six entanglement strategies and each trained with two readout scalings, against 11 classical regressors, including gradient boosting, support vector regression, and recurrent and attention-based networks, for predicting wind turbine power from four meteorological measurements. Because consecutive samples are strongly autocorrelated, all models are evaluated by blocked cross-validation with buffers, and conventional random cross-validation is analyzed for comparison. The readout scaling has the largest effect: mapping the target to [0, 1] instead of [-1,1] lowers the test error of every configuration by 39–46%, and the best configuration, with circular entanglement, reaches R2 of 0.88–0.89 with 12 trainable parameters. The best QNN selected by validation significantly outperforms linear regression and k-nearest neighbors and trails support vector regression, a decision tree, and XGBoost by 16–40 kW, while LightGBM and the neural networks remain ahead (R2 of 0.92–0.93). The ordering of the entangling gates alone changes the error by up to 17%, and the time per objective evaluation grows linearly with the training size and increases with the gate count. Random cross-validation overstates accuracy most strongly for k-nearest neighbors and for QNNs with the [-1,1] readout.