A Bayesian data assimilation method to enhance the time sequence prediction ability of data-driven models

Abstract

The generalization ability and robustness of data-driven models deteriorate when samples are scarce or input features are outside the training sample space. This research focuses on building a data assimilation (DA) framework that embeds the time sequence prediction model and improves the prediction results by constructing an enhanced system state vector that incorporates model bias parameters and new observation data to achieve the deviation correction and improve the robustness of the model. The DA framework proposed in this study includes three steps: prediction model inference, ensemble Kalman filter (EnKF) online deployment, and model parameter update. Wind tunnel experiments of a supersonic cascade are performed to obtain the original data for training the prediction model and optimizing the DA framework. Long short-term memory network (LSTM) was chosen as the prediction model. The testing set is distributed outside the sample space of the training set to verify the effectiveness of the DA framework for enhancing the time sequence prediction model. The improvement effects of the EnKF-enhanced LSTM model on the time sequence prediction results of the wall pressure in the oscillating flow field and the non-oscillating flow field are studied. The optimization results indicate that the LSTM model embedded in the DA framework can significantly improve the wall pressure prediction results. Thus, the results of this study lay a solid foundation for the condition monitoring and margin determination of the supersonic cascade flow field.