Abstract
AbstractMany existing approaches for multisite weather generation try to capture several statistics of the observed data (such as pairwise correlations) in order to generate spatially and temporarily consistent series. In this work, we analyze the application of Bayesian networks to this problem, focusing on precipitation occurrence and considering a simple case study to illustrate the potential of this new approach. We use Bayesian networks to approximate the multivariate (multisite) probability distribution of observed gauge data, which is factorized according to the relevant (marginal and conditional) dependencies. This factorization allows the simulation of synthetic samples from the multivariate distribution, thus providing a sound and promising methodology for multisite precipitation series generation.
Highlights
Stochastic weather generators (WGs) produce synthetic time series of weather data of unlimited length for a location based on the statistical characteristics of observed weather at that location
In this work we describe the application of Bayesian networks (BNs) to stochastic weather generation of precipitation occurrence
BNs allow for a qualitative analysis of the dependencies and independencies codified in the graph, and this is not the purpose of this work, they can be used for answering questions such as “Is variable Xi independent of Xj given a set of variables Xk?” As we will see later, the density of the graph can vary depending on the complexity required for the particular problem/application, since a directed acyclic graph (DAG) with more arcs captures more dependence relationships but requires more parameters
Summary
Stochastic weather generators (WGs) produce synthetic time series of weather data of unlimited length for a location based on the statistical characteristics of observed weather at that location. BNs have gained widespread use in several fields (Niedermayer, 2008; Pourret et al, 2008), boosted by the availability of several commercial and open software packages allowing to efficiently learn them from data, such as the bnlearn popular implementation in R used in this work (Scutari, 2010; Scutari & Denis, 2014) Their application to environmental sciences is still limited (Aguilera et al, 2011; Borunda et al, 2016; Uusitalo, 2007), and only a few applications for water resource management have been described in the literature (Phan et al, 2016; Ropero et al, 2017).
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
