2020年超长梅汛期降水概率预报应用与检验

基于欧洲中期天气预报中心（European Center for Medium-range Weather Forecasts，ECMWF）集合预报资料及浙江全省自动站降水观测资料，采用贝叶斯模型平均（Bayesian Model Average, BMA）方法对2020年浙江超长梅汛期开展降水概率预报订正试验。采用平均绝对误差、连续等级概率评分、布莱尔评分B_S、Talagrand、概率积分变换(Probability Integral Transform, PIT)直方图及属性图检验方法对本次过程BMA订正前后的概率预报进行对比分析，结果表明：①50 d为适用于浙江梅汛期ECMWF集合预报订正的BMA最优训练期，经最优训练期的BMA订正后，预报离散度有所增加，预报误差有所下降；②BMA对0.1 mm、10.0 mm和25.0 mm阈值降水的订正效果显著，经BMA订正后3个阈值的降水预报B_S下降率分别为25.92%、19.29%、4.76%，但对超过50.0 mm的降水订正效果不明显，且随着降水阈值增加，BMA的订正效果减弱；③在强降水个例中，BMA能有效减少各阈值降水预报概率大值落区偏差，使订正后的降水预报概率大值区与观测落区更一致。

Application and Verification of Probabilistic Forecast of Precipitation During Super Long Meiyu Season in 2020

Based on the ensemble forecast data derived from European Centre for Medium-range Weather Forecasts (ECMWF) ensemble forecast system and observation data derived from automatic observation stations in Zhejiang region, the Bayesian Model Averaging (BMA) method is used to calibrate the probabilistic forecasts of precipitation during the super long Meiyu season in 2020. In this paper, we verify the raw ensemble probabilistic forecast and BMA calibrated probabilistic forecast from 1 June to 15 July, 2020, by Mean Absolute Error (MAE), Continuous Ranked Probability Score (CRPS), Brier Score (BS), Talagrand, Probability Integral Transform (PIT) histogram, and attribute diagram. The verification results before and after calibration are compared. The analysis results are listed as follows. (1) In 8 different training periods (10 days to 80 days), 50 days correspond to smaller MAE and CRPS score values. So we set 50 days as the optimal BMA training period for ECMWF ensemble forecast calibration in the Meiyu season in Zhejiang Province. After BMA calibration in the optimal training period, the spread of ensemble forecast increases and the forecast error decreases. Analysing from the quantitative verification indicators, BMA can effectively calibrate the overall precipitation in the test stage, but it cannot calibrate the daily precipitation in the test stage. (2) For forecasting of different threshold precipitation, BMA has different calibration performance. For the thresholds of 0.1 mm, 10.0 mm, and 25.0 mm, BMA has a significant calibration effect. After BMA calibration, the CRPS of precipitation probabilistic forecast for these three thresholds (0.1 mm, 10.0 mm, and 25.0 mm) decreases by 25.92%, 19.29%, and 4.76%, respectively. However, the calibration effect of BMA weakens with the increase of precipitation threshold. For the events with total precipitation exceeding 50.0 mm, the BMA calibration effect is not as significant as that of the smaller threshold. In addition, BMA can effectively improve the forecast skills of 0.1 mm, 10.0 mm and 25.0 mm threshold precipitation and make the forecast probability more closely match the observation. (3) In the case of heavy rain, the high probability range of the raw ensemble probabilistic forecast is always wider than that of the observation. BMA has the ability to slightly calibrate the raw ensemble forecast probability. After BMA calibration, the high probability range of precipitation forecast at each threshold effectively reduces the deviation. The empty message information and the probability of empty message events also reduce after calibration. So BMA can make the calibrated high probability range of precipitation forecast more consistent with the observed range. But unfortunately, BMA cannot adjust the spatial distribution of precipitation forecast probability.