基于贝叶斯原理降水订正的水文概率预报试验

The Experiment of Hydrologic Probabilistic Forecast Based on the Precipitation Forecast Calibrated by Bayesian Model Averaging

利用淮河流域加密站点2008年6月1日—8月31日逐日降水资料、对应的T213模式的24 h,48 h以及72 h集合预报,采用贝叶斯模型平均(Bayesian Model Averaging,BMA)方法对集合预报15个成员的降水预报进行了概率集成与偏差订正,采用排序概率评分(CRPS)、平均绝对误差(MAE)对BMA的订正结果进行检验,并将订正后的降水预报输入VIC(Variable Infiltration Capacity)水文模型中进行水文概率预报。结果表明:经BMA订正后的24 h,48 h,72 h降水预报精度较订正前有所提高;BMA模型给出的有效区间(第25百分位数至第75百分位数)预报将实况降水量包含在内的可能性比订正前更大;由水文概率预报检验指标分析可知,经BMA订正的降水集合预报,由VIC水文模型模拟得到的径流量变化趋势与实况较吻合。

Based on 24-h accumulated precipitation data of the Huaihe Basin from 1 June to 31 August in 2008 and the corresponding ensemble forecast of 24 h, 48 h, 72 h from T213 model, the method of Bayesian Model Averaging （BMA） is used to calibrate quantitative precipitation forecasts of 15 members from the ensemble forecast based on the training data of 30 days. The calibrated results are verified by continuous ranked probability score （CRPS） and mean absolute error （MAE）. Second, the Dapoling---Wangjiaha Catchment in the upper stream of the Huaihe River, which is subdivided into Dapoling--Xixian Catchment and Xixian-Wangjiaba Catchment, is investigated with the hydrological simulation experiment. The Xixian and Wangjiaba hydrologic stations in the upper stream of the Huaihe River are selected as representative stations. The rain process occurring from 23 June to 3 August in 2008 is investigated to simulate the runoff tendency. Then, the 25th and the 75th percentiles of 24-h, 48-h and 72-h precipitation ensemble forecast which are calibrated by BMA are used to force the Variable Infiltration Capacity （VIC） hydrological model respectively to obtain the corresponding runoff, and finally the simulate results are analyzed comparing with daily runoff observations. Results show that the precision of 24-h, 48-h, 72-h precipitation forecast of BMA model is improved after the calibration. The raw ensemble forecast of 24 h is calibrated well by BMA model. As the leading hours increase, the calibration of 48 h and 72 h is as good as that of 24 h. Although BMA calibrates on the raw ensemble forecast, the improvement of calibrated forecast depends on the accuracy of raw ensemble forecast. The valid interval given by BMA model, namely the interval from the 25th percentile to the 75th one of ensemble forecast, is more likely to contain the true value of observations according to the verifica- tion analysis of hydrological probabilistic forecast. From this aspect, the performance of BMA forecast outperform deterministic forecast. It can improve the accuracy of forecast and reduce the error by BMA, describing forecast uncertainty in the form of a probability distribution. Known from the analysis of verifi- cation index of hydrologic probabilistic forecast, it shows that the hydrological simulation forced by the calibrated precipitation is almost consistent with the runoff tendency of observations. It is effective to grasp the trend of runoff change. It indicates that the precipitation forecast calibrated by BMA can be es- tablished coupling with the VIC hydrological model as well as increasing the forecast accuracy significant- ly. It can meet the more and more objective, quantitative needs of decision-making service, and improve the benefit of weather forecast greatly.