Session: 19-01: Turbines and Energy recovery.

Submission Number: 177307

LSTM-Based Modeling of Turbocharger Turbine Response to Pulsating Flow 

LSTM-Based Modeling of Turbocharger Turbine Response to Pulsating Flow

 

Turbocharger turbines, in both light- and heavy-duty engines, operate under highly pulsating flow conditions generated by the reciprocating motion of the cylinders. These unsteady conditions invalidate the quasi-steady assumption traditionally used to describe turbine behavior. Indeed, the turbine response exhibits hysteresis behavior, where the instantaneous performance deviates from the steady-flow characteristics measured in hot gas-stand continuous-flow conditions. The magnitude of these deviations is governed by the pulsation dynamics—primarily the amplitude and frequency of the pressure waves—and by the finite time required for inflow variations to propagate through the internal volumes of the system. To account for these effects in engine–turbine interaction studies, reduced-order modeling approaches are commonly applied. Zero-dimensional models typically rely on empirical corrections of steady performance maps, while one-dimensional formulations represent the hot side of the turbine using networks of elementary flow elements, such as ducts and junctions.

This paper presents a novel data-driven approach based on a neural network model trained on pulsating flow data obtained from an experimentally calibrated URANS model of a twin-scroll turbocharger turbine for off-highway vehicles. A range of operating conditions, generated by varying the amplitude and frequency of the inlet pulsation together with the turbine rotational speed, is used to build the training and validation datasets. The model is based on a Long Short-Term Memory (LSTM) architecture, suitable for capturing the time-dependent behavior of the system. It uses as inputs the time series of total pressure, total temperature, and mass flow rate at the inlet of the cold-side system to predict the instantaneous turbine torque as output. The LSTM hyperparameters (number of layers, neurons, time-series length, and learning rate) are first optimized using a Bayesian optimization algorithm. Then, the model’s accuracy is evaluated as a function of the number of pulsating-flow simulations used for training, providing insight into the computational cost required to achieve a reliable predictive model.

This data-driven method offers a promising alternative to conventional 0D and 1D turbine models, as it achieves higher predictive accuracy by directly learning from high-fidelity flow data without relying on the geometric simplifications and empirical corrections typically required by other reduced-order approaches.

Presenting Author: Roberto Mosca Accelleron Switzerland Ltd

Presenting Author Biography: Roberto is a CFD engineer working at Accelleron, Switzerland, and he got his PhD at KTH, the Royal Institute of Technology in Stockholm, Sweden.

Authors:

Roberto Mosca Accelleron Switzerland Ltd
Alex Kennedy Trinity College Dublin, The University of Dublin