A Change Point Prediction Algorithm for Time Series Data with Periodic Structural Changes

Abstract

We present an algorithm for predicting change points in time-series data with periodic structural changes. Energy demand profiles in buildings or factories often show abrupt increases and decreases between peak and off-peak times. Such abrupt changes may significantly affect the performance of a prediction algorithm since a slight error in change points results in a huge error between real and predicted values. We take a probabilistic approach that provides confidence intervals in which change points can fall. Numerical experiments show the effectiveness of our algorithm.