A novel non-homogeneous hidden Markov model for simulating and predicting monthly rainfall

Abstract

Monthly-to-seasonal precipitation forecasts are important in water resource management. Hidden Markov models (HMM) are widely applied in precipitation simulation due to its simplicity and advancements in associated computing techniques. HMMs, however, lack the flexibility to accommodate external factors in the dynamics. In this study, we consider a novel semiparametric Bayesian non-homogeneous hidden Markov model (BNHMM) that can explicitly incorporate the effects of observed covariates, e.g., climate index, on the state transition probabilities as well as the observation emission distributions. The proposed approach is tested for three rainfall stations in the service area of Tampa Bay Water, a regional water supply agency in the Southeastern United States. The BNHMM is first examined to simulate historical monthly rainfall data from the period 1979 to 2016. It is then evaluated in a retrospective mode to generate 3-month-ahead precipitation forecasts. Results indicate that the proposed BNHMM can capture historical rainfall properties well, and it is a promising alternative in providing operational rainfall forecasts. Although large-scale atmospheric and oceanic teleconnections, e.g., ENSO, have a strong influence in regional rainfall at the seasonal time scale, especially for the winter and early spring; its influence on monthly scale rainfall is limited. Potential improvement in forecasting performance is also discussed.