Laura Clemente-Harding, Luca Delle Monache, Guido Cervone, Martina Calovi, Weiming Hu, Mehdi Shahriari.

Probabilistic forecasts provide a distribution of the potential future states of the atmosphere. There are two main methods to generate probabilistic forecasts: one is to run multiple model realizations or multiple models with initial perturbations, and the second is to use analogs. The Analog Ensemble (AnEn) technique discussed in this tutorial generates probabilistic predictions using a single deterministic NWP, a set of past forecast predictions, and their corresponding observations. The AnEn technique compensates for the model bias by taking past errors into account. The main assumption is that if similar past forecasts can be found, the model error can be estimated. Specifically, the AnEn seeks to estimate the probability distribution of the observed future value of the predictand variable given a model prediction, which can be represented as p(y ~ f) where, at a given time and location, y is the unknown observed future value of the predictand variable and f the values of the predictors from the deterministic model prediction at the same location and over a time window centered over the same time. Advantages of the AnEn including the use of higher resolution forecasts and no need for initial condition perturbations, running multiple model instances, or post processing requirements. The AnEn is able to capture the flow-dependent error characteristics and show superior skill in predicting rare events when compared to state-of-the-art post processing methods. This tutorial will provide a guided discussion of the Analog Ensemble technique, one means of generating an ensemble prediction. It will discuss the AnEn technique, its local implementation and the parallel implementation on HPC resources, and case examples.