Convolutional quantum extreme learning machines for time series forecasting: a case study on solar energy generation

Abstract Accurate forecasting of solar power generation is critical for maintaining grid stability, yet it remains a significant challenge due to the inherent volatility of solar resources. While classical machine learning models have advanced the field, they often struggle to capture the complex, non-linear dynamics underlying rapid generation changes. This paper introduces a novel hybrid framework, Convolutional Quantum Reservoir Computing (C-QRC), designed to enhance time series forecasting by creating rich, high-dimensional features. Our method employs a gate-based quantum circuit as a reservoir, which is convolved over input time series to augment the feature space for classical predictive models, including LSTM, Transformer, and XGBoost. We conduct a rigorous evaluation using a rolling-forecast methodology on three distinct datasets, comparing the performance of QRC-enhanced models against classical and baseline counterparts. Our results reveal that QRC does not act as a universal accuracy enhancer for point forecasts. Instead, its primary advantage lies in a specialized capability: the improved modeling of system dynamics, particularly during high-volatility periods and significant ramp events. We identify strong architectural synergies, where the QRC-LSTM combination excels at predicting system dynamics, while the QRC-Transformer pairing demonstrates superior robustness in the resulting point forecast. We conclude that QRC is not a general replacement for classical pre-processing but a strategic component for advanced hybrid systems, designed to mitigate critical failure modes by providing a more profound understanding of system dynamics.