Trustworthy interval prediction method with uncertainty estimation based on evidence neural networks

Developing accurate and reliable prediction models is critical to ensuring the safety of the system. However, traditional deep learning only provides point predictions is not enough. For some high-risk systems, such as aerospace and autonomous driving, the reliability of model predictions needs to be assessed. This requires quantifying the uncertainty of model predictions and constructing trustworthy prediction intervals. Thus, a new trustworthy interval prediction method based on evidence neural network (TIENN) is proposed. Firstly, evidence theory and the Dirichlet distribution are integrated into deep neural networks to quantify prediction uncertainty. Secondly, modified expected utility theory is used to construct trustworthy prediction intervals. Moreover, a new loss function is designed to achieve both accurate point predictions and high-quality prediction intervals. Finally, taking the lithium-ion battery interval capacity prediction as an example to verify the effectiveness of the TIENN. The output results of the TIENN can not only be explained in clear language semantics, but also are consistent with the degradation process of lithium-ion batteries in actual engineering, thereby improving decision makers' trust in the model.