Detected Latent Seasonal Precipitation Patterns with Genetic Search and Fourier Compression

Abstract Book of the 6th World Conference on Climate Change and Global Warming

Year: 2025

[PDF]

Detected Latent Seasonal Precipitation Patterns with Genetic Search and Fourier Compression

Alina Barbulescu, Youssef Saliba, Cristian Stefan Dumitriu

ABSTRACT:

Change point detection in climate time series is a complex task due to inherent seasonality, trends, and serial correlation within the data. We present an approach that integrates regression modeling with autoregressive error structures to effectively identify abrupt regime shifts in climate datasets. The proposed model decomposes the observed time series into seasonal means, a linear trend component, and regime shift offsets, while capturing serial correlations through a periodic autoregressive (PAR) process that varies by season. To balance model fit and complexity, the Minimum Description Length (MDL) principle is used as the objective function, incorporating penalties for the number of change points, their positions, and the autoregressive order. Given the combinatorial nature of selecting change points and AR orders, a multi-island genetic algorithm (GA) is utilized to explore the model space. The GA employs chromosome representations that encode both change point locations and AR orders, and utilizes genetic operators such as crossover, mutation, and migration to optimize the MDL score. The methodology is validated through synthetic data generation that mirrors real climate series characteristics, including induced change points and controlled shift magnitudes. Performance assessments indicate the GA-based model selection effectively identifies true change points under varying conditions of shift magnitude and autocorrelations.

Keywords: Ga, Mdl, Par, Precipitation, Residual