Photovoltaic power forecasting using quantum machine learning

TL;DR

This study advances photovoltaic power forecasting by introducing three hybrid quantum–classical architectures (HQNN, HQLSTM, HQSeq2Seq) that leverage variational quantum circuits and quantum depth-infused layers to achieve higher accuracy and data efficiency than strong classical baselines. Using a Mediterranean PV dataset and rigorous cross-validation, the HQLSTM emerges as the top performer with substantially fewer parameters, while HQSeq2Seq enables horizon-flexible predictions without prior meteorological inputs. The authors also provide circuit-level analyses (ZX-calculus, Fisher information, Fourier) to characterize trainability and expressivity, and demonstrate the approach’s robustness when data are scarce and its generalization to an additional energy dataset. Together, these results suggest that hybrid quantum models can offer practical, data-efficient improvements for PV forecasting and grid integration tasks.

Abstract

Accurate forecasting of photovoltaic power is essential for reliable grid integration, yet remains difficult due to highly variable irradiance, complex meteorological drivers, site geography, and device-specific behavior. Although contemporary machine learning has achieved successes, it is not clear that these approaches are optimal: new model classes may further enhance performance and data efficiency. We investigate hybrid quantum neural networks for time-series forecasting of photovoltaic power and introduce two architectures. The first, a Hybrid Quantum Long Short-Term Memory model, reduces mean absolute error and mean squared error by more than 40% relative to the strongest baselines evaluated. The second, a Hybrid Quantum Sequence-to-Sequence model, once trained, it predicts power for arbitrary forecast horizons without requiring prior meteorological inputs and achieves a 16% lower mean absolute error than the best baseline on this task. Both hybrid models maintain superior accuracy when training data are limited, indicating improved data efficiency. These results show that hybrid quantum models address key challenges in photovoltaic power forecasting and offer a practical route to more reliable, data-efficient energy predictions.

Figures (7)

• Figure 1: The input to a hybrid quantum model is presented as a chronological data table, documenting hourly meteorological parameters including ambient temperature ($T_a$), module temperature ($T_m$), and solar irradiance ($I_3$, $I_{15}$), alongside the mean PV power output ($P$). The model is designed to leverage these data to generate predictions of PV power output for a short-term forecast aimed at near-future output, typically the next hour, and a long-term forecast that extends to a broader temporal horizon.
• Figure 2: (a) Mean and standard deviation of the PV power value for each hour of the day. (b) Mean and standard deviation of the PV power value for each month of the year. The plot shows the PV power reaching maximal values in June and July. (c) Correlation matrix of input features. (d) Joint distribution of features.
• Figure 3: The architectures of: (a) Hybrid Quantum Neural Network with a VVRQ layer, (b) Hybrid Quantum Long Short-Term Memory, (c) QDI layer used in HQLSTM model, (d) Hybrid Quantum Seq2Seq with QDI layer.
• Figure 4: Results of training and testing for the HQNN, MLP, HQLSTM, and LSTM models. (a-b) Training and testing history shown by the dotted and solid lines, respectively. The filled region shows the standard deviation of the models averaged over the different testing subsets. (c) Bar chart showing MAE and RMSE loss on the test dataset averaged over the results of 5 models trained on different subsets of training and test data. (d) The bar chart shows MSE loss on the test dataset for MLP, HQNN, LSTM and HQLSTM models for the reduced training dataset size. (e) The bar chart shows the RMSE loss on the test dataset for the Seq2Seq and HQSeq2Seq models for the reduced training dataset size. (f-g) The training and testing learning curves for the Seq2Seq and HQSeq2Seq models. (h) Example for the classical and hybrid Seq2Seq models inference on the testing data. The models receive $124$ hours of data as an input (before the dashed line) and forecast the "Power" up to $137$ hours ahead (after the dashed line). The solid black line represents the ground-truth value of the "Power" feature. The MSE/MAE errors in this particular example are $0.0052/0.0349$ for Seq2Seq and $0.0040/0.0292$ for HQSeq2Seq.
• Figure 5: Training dynamics of HQLSTM and LSTM models on the energy prediction task. (a) MSE loss on the training subset shows both models converging similarly. (b) MAE loss on the test subset indicates that HQLSTM achieves lower test loss values overall. Specifically, the lowest loss is $0.0513$ and $0.0535$ for the HQLSTM and LSTM models, respectively.