Quantum Kernel, Leveraging 5 Qubits and Fourier Transform, Improves Solar Irradiance Forecasting Accuracy

Predicting solar irradiance accurately is crucial for efficient energy grid management and reliable renewable energy forecasting, and a new approach to this challenge leverages the power of quantum computing. Nawfel Mechiche-Alami from Ecole Polytechnique Fédérale de Lausanne, alongside Eduardo Rodriguez and Jose M. Cardemil, demonstrate a significant advance in short-term forecasting by introducing a quantum kernel based on the Quantum Fourier Transform. Their method encodes time-series data into quantum states, processes it with a protective rotation layer, and then applies it to kernel ridge regression, consistently outperforming traditional forecasting models across diverse climate conditions. This research not only improves prediction accuracy and reduces forecasting bias, but also establishes a promising pathway towards integrating quantum techniques into practical renewable energy applications, with Enrique López Droguett from the University of California, Los Angeles also contributing to this important work.

Quantum Kernel for Enhanced Time-Series Forecasting

Scientists developed a novel quantum kernel to enhance short-term time-series forecasting, beginning with a process of windowing each input signal and encoding its amplitude into a quantum state. To prevent signal cancellation, the team incorporated a protective rotation layer following the transformation, ensuring signal integrity throughout the process. The resulting kernel feeds into a kernel ridge regression (KRR) model, a machine learning technique used for forecasting future values based on patterns in the data.

To incorporate external factors, scientists convexly fused feature-specific kernels, allowing the model to consider exogenous predictors alongside the time-series data itself. Experiments conducted using a noiseless simulator with five qubits and a window length of 32 provided a controlled environment for evaluating the kernel’s performance. The team meticulously tuned only the feature-mixing weights and the KRR ridge alpha parameter, maintaining fixed classical hyperparameters to ensure a fair comparison against baseline models. To assess the kernel’s effectiveness, researchers analyzed multi-station solar irradiance data spanning various Koppen climate classes.

The results demonstrate consistent improvements in median R2 and normalized root mean squared error (nRMSE) compared to reference classical kernels, including radial basis function and polynomial models. Furthermore, the new kernel reduced bias, as measured by normalized mean bias error (nMBE), and tighter average errors were observed, indicating enhanced forecasting accuracy. Complementary analyses using mean absolute error (MAE) and maximum error (ERMAX) confirmed these improvements, revealing headroom for further refinement even during sharp transients in the data. This work pioneers a method for leveraging quantum computation to improve the accuracy and reliability of short-term time-series forecasting.

Quantum Kernel Improves Solar Forecasting Accuracy

This research introduces a quantum kernel designed to improve short-term time-series forecasting, specifically applied to multi-station solar irradiance data. Experiments consistently demonstrate improvements in forecasting accuracy across various Koppen climate classes. The proposed kernel achieves a higher median R2 score and lower normalized root mean squared error (nRMSE) compared to classical radial basis function (RBF) and polynomial kernels.

Furthermore, the new kernel reduces bias, as measured by the normalized mean bias error (nMBE). The team incorporated exogenous predictors through convexly fused feature-specific kernels, enhancing the model’s ability to capture complex relationships within the data. Detailed analysis reveals tighter average errors, quantified by mean absolute error (MAE), and reduced maximum error (ERMAX), indicating improved stability and reliability of the forecasts. The new kernel yielded lower normalized root mean squared error and higher R-squared values, alongside a near-zero mean bias error, indicating both improved accuracy and reduced systematic deviations in forecasts. These findings suggest that representing time-series data within a frequency-aware quantum feature space effectively captures underlying diurnal and seasonal patterns, as well as short-term variability. The methodology is broadly applicable, extending beyond solar irradiance to other periodic or quasi-periodic time-series tasks such as wind nowcasting, load forecasting, and prediction of air quality, wave patterns, and traffic flows. Furthermore, the approach can be adapted for multi-step forecasting, extended to spatio-temporal settings, and repurposed for anomaly detection or classification tasks.