Early warning for critical transitions using machine-based predictability

Abstract

Detecting critical transitions before they occur is challenging, especially for complex dynamical systems. While some early-warning indicators have been suggested to capture the phenomenon of slowing down in the system's response near critical transitions, their applicability to real systems is yet limited. In this paper, we propose the concept of predictability based on machine learning methods, which leads to an alternative early-warning indicator. The predictability metric takes a black-box approach and assesses the impact of uncertainties itself in identifying abrupt transitions in time series. We have applied the proposed metric to the time series generated from different systems, including an ecological model and an electric power system. We show that the predictability changes noticeably before critical transitions occur, while other general indicators such as variance and autocorrelation fail to make any notable signals.