Modelling Stochastic Inflow Patterns to a Reservoir with a Hidden Phase-Type Markov Model

TL;DR

This paper introduces a novel Hidden Markov Model with phase-type distributions to better capture the complex, non-Markovian rainfall and inflow patterns in a specific region, improving water resource management insights.

Contribution

It develops a non-Markovian HMM using phase-type distributions to model climate-related inflow patterns, extending traditional models for enhanced accuracy.

Findings

Effective modeling of dry and wet periods in the climate system.

Improved forecasting of reservoir inflows based on latent states.

Enhanced understanding of the temporal structure of local rainfall.

Abstract

This paper presents a novel methodology for modelling precipitation patterns in a specific geographical region using Hidden Markov Models (HMMs). Departing from conventional HMMs, where the hidden state process is assumed to be Markovian, we introduce non-Markovian behaviour by incorporating phase-type distributions to model state durations. The primary objective is to capture the alternating sequences of dry and wet periods that characterize the local climate, providing deeper insight into its temporal structure. Building on this foundation, we extend the model to represent reservoir inflow patterns, which are then used to explain the observed water storage levels via a Moran model. The dataset includes historical rainfall and inflow records, where the latter is influenced by latent conditions governed by the hidden states. Direct modelling based solely on observed rainfall is insufficient due to the complexity of the system, hence the use of HMMs to infer these unobserved dynamics. This approach facilitates more accurate characterization of the underlying climatic processes and enables forecasting of future inflows based on historical data, supporting improved water resource management in the region.