Hybrid Quantum-Classical Recurrent Neural Networks for Time Series Prediction

Renewable energy forecasting is one of the most demanding challenges in modern artificial intelligence. Photovoltaic production changes with the weather, follows seasonal patterns, and often presents fluctuations that are difficult to capture with classical deterministic models. Predicting the next day of solar power generation, therefore, requires more than learning a smooth curve: it requires a model capable of interpreting instability. The work described here explores this challenge through a hybrid quantum-classical recurrent neural network (QRNN). The idea is to combine the temporal memory of Long Short-Term Memory (LSTM) networks with the expressive power of a variational quantum layer, creating a model that processes time series through both classical recurrence and quantum feature transformation.

The architecture begins with two stacked LSTM layers, which learn the temporal dependencies embedded in the photovoltaic power sequence. These layers process the past behavior of the system and extract information about daily and seasonal dynamics. After this recurrent stage, a fully connected layer compresses the learned representation into a form suitable for the quantum component. The variational quantum layer then maps the input into a high-dimensional quantum state, using a feature map and a parametrised circuit to process the information through superposition and entanglement. This hybrid structure does not replace classical deep learning. Instead, it extends it. The LSTM layers learn how the signal evolves over time, while the quantum layer enriches the representation space in which the final prediction is produced.

The system is tested on real photovoltaic production data collected from the Oak Ridge National Laboratory plant in the United States. The task consists of predicting one full day of solar power production using historical hourly observations. The experiments consider different months of the year, including both more regular sunny periods and more unstable conditions. This choice is important because photovoltaic forecasting becomes more difficult when clouds, seasonal variability, and low irradiation alter the expected production curve. The hybrid model is compared with a standard LSTM network and a Random Forest regressor. The results show that the QRNN often achieves lower prediction errors, especially in conditions where the production profile is more irregular..

The central contribution lies in the role of the variational quantum layer. By embedding classical features into quantum states, the model can process information in a richer mathematical space than the one used by a purely classical layer of similar size. This does not mean that quantum computing is already a universal replacement for classical forecasting models. The current technological context is still constrained by noise, limited qubit availability, and the complexity of quantum circuit design. However, the hybrid approach shows a practical path forward: quantum components can be integrated into existing neural architectures as specialised processing layers. In this case, the quantum layer acts as a feature processor placed at the end of a recurrent neural network. It receives the temporal representation built by the LSTM layers and transforms it before producing the final output.

The broader significance of this approach is not limited to photovoltaic forecasting. It suggests a new design principle for time series prediction, where classical layers handle temporal structure and quantum layers contribute to representation power. For renewable energy systems, this direction is particularly promising. Energy data are noisy, intermittent, and influenced by many external factors. Models capable of better approximating stochastic fluctuations can support more reliable grid management, improved scheduling, and more effective integration of renewable resources. In this perspective, hybrid quantum-classical neural networks represent an early but meaningful step toward a new generation of forecasting tools, where quantum computation does not stand apart from artificial intelligence, but becomes one of its internal mechanisms.