We develop a flexible spline-based Bayesian hidden Markov model stochastic weather generator to statistically model daily precipitation over time by season at individual locations. The model naturally accounts for missing data (considered missing at random), avoiding potential sensitivity from systematic missingness patterns or from using arbitrary cutoffs to deal with missingness when computing metrics on daily precipitation data. The fitted model can then be used for inference about trends in arbitrary measures of precipitation behavior, either by multiple imputation of the missing data followed by frequentist analysis or by simulation from the Bayesian posterior predictive distribution. We show that the model fits the data well, including a variety of multi-day characteristics, indicating fidelity to the autocorrelation structure of the data. Using three stations from the western United States, we develop case studies in which we assess trends in various aspects of precipitation (such as dry spell length and precipitation intensity), finding only limited evidence of trends in certain seasons based on the use of Sen's slope as a nonparametric measure of trend. In future work, we plan to apply the method to the complete set of GHCN stations in selected regions to systematically assess the evidence for trends.
翻译:我们开发了一个灵活的、基于样板的贝叶西亚隐藏的马尔科夫模型随机天气生成器,以便按季节对各个地点的每日降水进行统计性模型。模型自然地记录了缺失数据(随机考虑缺失),避免了系统性缺失模式的潜在敏感度,或者在计算每日降水数据时使用任意截断处理缺失情况的方法。随后,可使用该适当模型来推断任意测量降水行为的趋势,或者对缺失数据进行多次估算,随后进行经常分析,或者对贝叶西亚后方预测分布进行模拟。我们显示该模型非常适合数据,包括多种多日特性,表明数据对自动调节结构的忠诚性。我们利用来自美国西部的三个站点,开展案例研究,评估降水的各个方面的趋势(如干法长度和降水强度),根据Sen的斜度作为非对趋势的计量,只在某些季节找到有限的趋势证据。我们计划对选定区域的全部GHCN站采用这种方法,以便系统评估趋势。