Multivariate Time Series Prediction Based On Quantum-Enhanced LSTM Models

The discussion of time series prediction has a long history,among which multivariate time series prediction is ubiquitous in practical application scenarios and is of great significance for guiding human production and life and assisting scientific research.With the development of artificial intelligence technology,it has become a mainstream method to solve the time series prediction problem based on deep learning technology such as long short-term memory neural network(LSTM).However,the sharp increase in model complexity and computational requirements has prompted people to explore new techniques to assist in timing prediction.The potential advantages of quantum computing in information processing make it possible to further optimize classical time series prediction.Quantum deep learning(QDL)combines the advantages of quantum computation and classical deep learning.It tries to improve the existing deep learning by constructing a quantum neural network(QNN)which is analogous to the classical neural network,so that it can be used to efficiently solve the time series prediction problem.In this work we attempt to construct a quantum-enhanced LSTM model(QLSTM)by the well-designed quantum neural network layers.Also,the proposed QLSTM model is used to solve the multivariate time series prediction problem.The main works of this study include:(1)Discussing the construction of the QNNs in detail,including designing a class of variational quantum circuits(VQCs)with well expressibility.Based on the hybrid quantum-classical neural network architecture,we construct the quantum-enhanced LSTM model.(2)Using the constructed QLSTM model,a series of experiments are conducted on the quantum software development environment py Qpanda to show the feasibility of proposed QLSTM model.(3)Considering a series of methods for optimizing the proposed QLSTM models,including using classical deep learning experience to optimize and constructing a series of quantum-enhanced models corresponding to LSTM variants.In addition,a series of general regularization methods for QNNs are designed in our work for further improving the models’ performance.The experimental results show that the constructed QLSTM model has certain feasibility and potential advantages in solving multivariate time series prediction problems,including simpler model structure and fewer parameters.Also,above optimization experiments show the possibility of optimizing the quantum deep learning model based on classical deep learning experience and demonstrate the availability of the proposed quantum regularization method designed for QNN in optimizing the performance of the QLSTM model and improving the quantum noise robustness.This work provides a reusable reference for the construction of quantum deep learning models based on QNNs and provides feasible ideas and methods for optimizing the performance of quantum deep learning models.