Multiple changepoint detection for periodic autoregressive models with an application to river flow analysis

TL;DR

This paper introduces an automatic method for detecting multiple changepoints in periodic autoregressive models, enhancing river flow forecasting accuracy by accounting for structural changes and seasonality.

Contribution

It develops a novel changepoint detection procedure for P-AR models using genetic algorithms, allowing for structural changes beyond mean or variance shifts.

Findings

Improved river flow forecasting accuracy.

Effective detection of multiple changepoints.

Validation on real river flow data shows superior performance.

Abstract

In river flow analysis and forecasting there are some key elements to consider in order to obtain reliable results. For example, seasonality is often accounted for in statistical models because climatic oscillations occurring every year have an obvious impact on river flow. Further sources of alteration could be caused by changes in reservoir management, instrumentation or even unexpected shifts in climatic conditions. When these changes are ignored the statistical results can be strongly misleading. This paper develops an automatic procedure to estimate number and locations of changepoints in Periodic AutoRegressive models. These latter have been extensively used for modelling seasonality in hydrology, climatology, economics and electrical engineering, but there are very few papers devoted also to changepoints detection, moreover being limited to changes in mean or variance. In our proposal we allow the model structure as a whole to change, and estimation is performed by optimizing an objective function derived from the Information Criterion using Genetic Algorithms. The proposed methodology is brought out through the example of three river flows, for which we built models with possible changepoints and evaluated their forecasting accuracy by means of Root Mean Square Error, Mean Absolute Error and Mean Absolute Percentage Error. The last years of data sets have been omitted from the selection and estimation procedure and were then used to forecast. Comparisons with literature on river flow forecasting confirms the efficiency of our proposal.