Combined use of long short-term memory neural network and quantum computation for hierarchical forecasting of locational marginal prices

Abstract

Accurate locational marginal price forecasting (LMPF) is crucial for the efficient allocation of resources. Nevertheless, the sudden changes in LMP make it inadequate for many existing long short-term memory (LSTM) network-based prediction models to achieve the required accuracy for practical applications. This study adopts a hierarchical method of three layers based on double quantum-inspired grey wolf optimisation (QGWO) to improve the LSTM model (HD-QGWO-LSTM) for a one-step LMPF. The top layer completes the data processing. The middle layer is a QGWO-optimised support vector machine (SVM) for classifing whether LMPs are price spikes. The bottom laver is a double QGWO-improved LSTM (QGWO-LSTM) model for a real LMPF, where one QGWO-LSTM is for the spike LMPF and the other is for the non-spike LMPF. To address the issue of excessively long training times during the design of the LSTM network structure and parameter selection, a QGWO algorithm is proposed and used to optimise four LSTM parameters. The simulation results on the New England electricity market show that the HD-QGWO-LSTM method achieves similar prediction accuracy to other four LSTM-based methods. The results also validate that the QGWO algorithm significantly reduces time consumption while ensuring optimisation effectiveness when optimising SVM and LSTM.