Neural network models and shapley additive explanations for a beam-ring structure

A combination of neural network modeling, SHapley Additive exPlanations (SHAP) and model simplification is proposed and applied to the system of a beam-ring structure in this paper. Based on the long short-term memory based encoder-decoder (LSTM ED), a new framework (P_LSTM ED) is presented to improve prediction accuracy. Several iterative methods are selected to optimize the neural network model under the new framework with LSTM EDs. The normalized root mean square error (NRMSE) is regarded as an evaluation metric for the prediction accuracy of neural network models. The SHAP method provides interpretability for the optimal neural network model and there are two applications through the SHAP analysis. One is a modification strategy for inputs of the neural network model, the other is a model simplification method for state equations of the nonlinear system. Numerical simulation shows that the P_LSTM ED has higher prediction accuracy while considering efficiency. The neural network model with the modification strategy applied achieves a maximum accuracy improvement of 24.15 % with less training time. The effectiveness and generality of model simplification are verified by numerical simulation comparison between simplified and original equations. •A new network framework is proposed to improve the prediction accuracy.•A beam-ring structural system of a circular truss antenna is our research object.•Shapley additive explanations provide interpretability for neural networks.•Interpretability analysis helps to modify neural network models.•An effective model simplification method is raised through interpretability analysis.