Examples of Bayesian probabilistic quantitative precipitation forecasts

This poster compares examples of Probabilistic Quantitative Precipitation Forecasts (PQPFs) output from two systems: the Bayesian Processor of Output (BPO) and Model Output Statistics (MOS). The probabilistic forecasts also are compared with the deterministic forecasts of precipitation amount output from the Global Spectral Model run under the code name AVN. The BPO is a theoretically-based technique for probabilistic forecasting of weather variates. It is currently being tested by producing PQPFs for a set of climatically diverse stations in the contiguous U.S.

Each system, the BPO and the MOS, processes output from a numerical weather prediction (NWP) model in order to quantify uncertainty about the future precipitation amount. The difference between the two systems lies in how the NWP model output is processed. The BPO optimally fuses a continuous prior distribution function of precipitation amount, estimated from climatic data, with NWP model output in order to produce a continuous posterior distribution function of precipitation amount, conditional on precipitation occurrence. MOS uses regression estimation of event probabilities on the NWP model output to estimate three, four, or five points on the distribution function of precipitation amount, conditional on precipitation occurrence.

The poster graphically depicts BPO PQPFs and MOS PQPFs quantifying various degrees of uncertainty about the precipitation amount along with the deterministic forecasts of precipitation amount output from the AVN model. The examples highlight the BPO's ability to update the climatic prior distribution function of precipitation amount with output from the NWP model and to produce forecasts which quantify the uncertainty about extreme amounts of precipitation (the BPO PQPF assigns a nonzero exceedance probability to all realizations of the precipitation amount). The examples demonstrate gross errors of the deterministic forecasts output from the AVN model and the inability of MOS's discrete representation of the conditional distribution function to characterize the uncertainty about extreme precipitation amounts (which lie in the tail of the prior distribution function).