Interpretable Seasonal Hidden Markov Model for spatio-temporal stochastic rain generation in France

• Main Authors: Gobet, Emmanuel, Métivier, David, Parey, Sylvie
• Other Authors: Centre de Mathématiques Appliquées de l'Ecole polytechnique (CMAP), École polytechnique (X), Institut Polytechnique de Paris (IP Paris)-Institut Polytechnique de Paris (IP Paris)-Centre National de la Recherche Scientifique (CNRS), Mathématiques, Informatique et STatistique pour l'Environnement et l'Agronomie (MISTEA), Institut National de Recherche pour l’Agriculture, l’Alimentation et l’Environnement (INRAE)-Institut Agro Montpellier, Institut national d'enseignement supérieur pour l'agriculture, l'alimentation et l'environnement (Institut Agro)-Institut national d'enseignement supérieur pour l'agriculture, l'alimentation et l'environnement (Institut Agro), EDF R&D (EDF R&D), EDF (EDF), Chaire Stress Test - BNP Paribas/Ecole polytechnique/Fondation de l'X
• Format: Report
• Language: English
• Published: HAL CCSD 2024
• Subjects: Hidden markov model; Stochastic weather generators; Interpretable Learning; Rainfall; Climate model evaluation; MSC2020: 62M05; 37H10; 62P12; [STAT.ME]Statistics [stat]/Methodology [stat.ME]; [STAT.AP]Statistics [stat]/Applications [stat.AP]; North Atlantic
• Online Access: https://hal.inrae.fr/hal-04621349; https://hal.inrae.fr/hal-04621349/document; https://hal.inrae.fr/hal-04621349/file/main.pdf

We present a Stochastic Weather Generator described based on a multisite Hidden Markov Model (HMM) and trained with French weather stations data. It generates correlated precipitation, with a special focus on seasonality and the correct reproduction of the distribution of dry and wet spells. The hidden states are viewed as global weather regimes, e.g., dry all over France, rainy in the north, etc. The resulting model is fully interpretable; it can even approximately recover large-scale structures such as North Atlantic Oscillations. The model achieves very good performances, specifically in terms of extremes. Its architecture allows easy integration of other weather variables. We show an application where the model is trained on future climate scenarios, allowing easy comparison and interpretation with the historical data in terms of parameters evolution and extremes.