标题:
改善水文气象预报的统计后处理Statistical Post-Processing to Improve Hydrometeorological Forecasts
作者:
段青云, 叶爱中
关键字:
水文集合预报, 统计后处理方法, 集合前处理器, 广义线性模型Ensemble Hydrological Forecast; Statistical Post-Processing Methods; Ensemble Pre-Processor; Generalized Linear Regression Model (GLM)
期刊名称:
《Journal of Water Resources Research》, Vol.1 No.4, 2012-08-10
摘要:
水文预报存在由模型输入、初始与边界条件、模型参数与结构等因素带来的不确定性，这些不确定性的存在给基于模型输出结果而得出的预报带来预报均值和区间的偏差，在发布最终水文预报前我们必须要消掉这些偏差，统计水文气象后处理器是对水文气象模型直接输出结果进行后处理、从而达到消掉预报偏差的一个有效方法。本文简单介绍两个实用的统计后处理方法：1) 对降水或气温集合预报进行后处理的集合预报处理器(Ensemble Pre-Processor)和2) 适合于径流预报后处理的广义线性模型方法(Generalized Linear Model)，并将这些方法在我国淮河流域和美国French Broad河流域进行展示。研究结果表明，统计后处理方法能够很大程度提高原始降水与水文预报的精度，一个很有意义的结果是统计后处理改善原始水文预报的程度与水文模型参数率定对原始水文预报的改善程度相当，这意味着在受人类活动影响很大而不能率定水文模型的情况下，水文统计后处理方法能够起到模型率定相同的效果。
Hydrologic forecasts based on direct outputs from a hydrologic model contain significant uncertainty from various sources, including model inputs, initial/boundary conditions and model structure/model parameters. The uncertainty leads to various biases in the hydrologic forecasts. Before issuing final hydrologic forecasts to the forecast users, it is necessary to remove these biases. A statistical post-processor is an effective tool than can be used to remove various biases from the hydrometeorological forecasts. In this paper, we briefly describe two practical post-processing methods: 1) The ensemble pre-processor for post-process- ing quantitative precipitation and temperature forecasts; and 2) The generalized linear models for post-proc- essing streamflow forecasts. We demonstrated the effectiveness of these two methods in China’s Huai River basin and the French Broad River basin in the US. Results clearly show that post-processing can significantly improve the raw hydrometeorological forecasts. An interesting observation is that post-processing can achi- eve the same degree of improvement in streamflow simulation as model calibration. This suggests that, for basins where calibration cannot be done properly due to data issues (i.e., streamflow regulations), we can use post-processing to compensate for the lack of model calibration.