Mixture of Regression Models for Precipitation Prediction

One of the key challenges of climate science today is the understanding and prediction of rainfall and other precipitation events [6]. It holds high values for both scientists and policymakers in terms of obtaining novel scientific insight into the processes of precipitation formation, as well as predictions and associated uncertainties which decide future policy decisions based on them. One of the foremost aspects of this understanding is to model the influence of atmospheric variables on the intensity, duration and frequency of precipitation. In recent times, physics based models of the atmosphere have been used for simulations of the atmosphere for both historical verification and future projections [5]. Although they provide partial understanding of the influence of climate variables on precipitation, they have been found to be insufficient for predictive modeling with sufficient accuracy [6]. The high complexity and nonlinearity of the system calls for novel advancements through statistical methods to complement and support the physical models for accurate predictive modeling. In this work, we propose a non-parametric mixture of regression model for precipitation prediction, which considers temporal measurements of local atmospheric variables and some of their interactions as covariates. Our model is learned individually for different rainfall ‘regimes’, determined through percentiles or peak over threshold values. Prediction occurs in two stages: prediction of rainfall regime, followed by the prediction of rainfall distribution, which is a non-parametric mixture of a normal and an extreme-value distributions. The mixture model is suitable for simultaneously modeling normal and extreme precipitation values.