Novel SABO-optimized LSTM and BiLSTM models for enhancing engine coolant bleed flow prediction

Abstract This paper proposes a turbine engine coolant bleed flow prediction method based on improved long short-term memory (LSTM) algorithms using subtraction average-based optimizer (SABO), specifically aiming at high-precision modeling for high pressure turbine (HPT) and low pressure turbine (LPT). Firstly, from the 21 sensor detection data in the C-MAPSS dataset, variables with high correlation to coolant bleed air flow are selected to construct the objective parameters of the model. Secondly, to enhance the prediction accuracy of the LSTM and bidirectional LSTM (BiLSTM) networks, the particle swarm optimization (PSO) and SABO are employed to optimize the parameters of the two memory neural networks. This leads to the proposal of coolant bleed air flow prediction models based on the frameworks of PSO-LSTM, PSO-BiLSTM, SABO-LSTM and SABO-BiLSTM. Through the training and testing of 33,991 datasets of coolant bleed air flow data measured from HPT and LPT, the results demonstrate that the SABO-BiLSTM model exhibits the best emission prediction accuracy and fitting performance, which provides technical support for evaluating the synergistic effects of aero-engine internal structures on coolant bleed air emissions and formulating effective emission reduction strategies.