Dual-branch LSTM-Transformer network for load duration prediction of governor oil pump in pumped storage power station

Abstract

With the development of grid energy storage systems, the stable operation of governor oil pumps has become increasingly important, and accurately predicting the load durations is essential for ensuring system reliability in a hydropower generation station. This paper proposes a dual-branch Long Short-Term Memory (LSTM)-Transformer network for load duration prediction of the governor oil pump. A hybrid LSTM-Transformer network is designed as baseline to combine the strengths of LSTM and Transformer. A two-branch structure is designed to respectively focus on the temporal correlation and cross-correlation factors. A Backpropagation (BP) neural network is designed as the fusion module to intelligently integrate the features from both branches, enabling a more powerful and non-linear combination than a fixed linear weighting. The model is trained and tested by using operational data from a real pumped storage power station. Comprehensive experiments, including ablation studies, comparisons with various baseline and ensemble models, robustness tests under different missing data mechanisms, are conducted. The results consistently demonstrate the superiority of our proposed model, validating the effectiveness of the dual-branch architecture and the BP fusion module. The proposed model provides support for the intelligent operation and maintenance of pumped storage power stations.