Seasonal Inflow Forecasts Based On Non-Homogeneous Hidden Markov Models: The Case Of Oros Reservoir, Northeastern Brazil

This study explores the use of non-homogeneous hidden Markov models (NHMMs) to forecast streamflows into the Oros Reservoir, located in the State of Ceara, Northeastern Brazil. These models employ hydro-climatic information from a given year to predict the next year inflow. The Oros Reservoir is a key infrastructure in the water resources system of the state, where most rivers present very high temporal variability, which is a challenge for water managers. Streamflow forecast systems at the climate scale (months ahead) can be very useful for the challenging process of water allocation in the region. This paper explores the use of NHMMs to provide a probabilistic one-year ahead forecast of the mean annual flow for the Oros Reservoir. Results based on a forecast verification study based on 60 years of independent data show that NHMMs with two or three hidden states, employing the NINO3 climate index as exogenous variable, provides forecasts with an average value of the continuous ranked probability skill score of about 0.2-0.25, meaning that they are better than climatology. It is also evident that those models performed much better during dry years than during wet years, which may be related to lack of other climate indices.