A Gradient-Sampling-based Algorithm for Change Point Detection in Piecewise Linear Model

Change point detection, as an important technique in artificial intelligence, aims to identify abrupt changes in complex systems. In this paper, we propose a novel gradient-sampling-based approach for change point detection in piecewise linear model. The convergence to a point satisfying the first-order optimality condition is guaranteed. Through extensive numerical experiments, we compare the proposed algorithm with the well known method of Muggeo's segmentation by dynamic programming. By computing the change points on the dataset concerning the relationship between the residential electricity consumption and temperature in Fujian Province, we demonstrate that the proposed algorithm outperforms Muggeo's method. Moreover, when using the change points for power load forecasting, the change points from the proposed algorithm can significantly improve the predictive performance of the Long Short-Term Memory (LSTM) model.