基于BP神经网络的降水量估算模型在自动气象站降水量质量控制中的应用

Application of Precipitation Estimation Model Based on BP Neural Network in Precipitation Quality Control of Automatic Weather Station

为提高自动气象站降水量质量控制效果和效率,针对传统空间一致性检查方法在对流性降水质量控制中效果不佳,且参数估计复杂、应用不便的问题,提出了基于BP神经网络的自动气象站降水量估算模型,模仿人脑神经系统结构和信息处理方式对自动气象站降水量、位置数据进行学习,利用邻近站小时降水量、邻近站与待检站距离估算出待检站降水量,进而通过对比估算值与实测值的差异,完成实测降水量质量控制。模型采用三层网络结构,综合4种经验公式确定隐层节点数的区间,使用试凑法和对比法确定最优节点数和最优训练函数。模型通过训练,达到较好的估算效果,训练、检验、预测样本估算值与实测值R2均超过0.9;与空间插值法估算值相比,神经网络模型估算准确性更高。利用模型对宁河镇、汉沽街降水量异常实例的分析结果表明,模型估算值能准确反映出实测降水量偏低的情况。通过比较站点实测降水量与神经网络模型估算结果,可及时发现异常降水数据,提升质量控制工作的效果和效率,提高自动气象站维修保障的针对性。

In order to improve the effect and efficiency of precipitation quality control of automatic weather station, aiming at the problems of poor effect of traditional spatial consistency check method in convective precipitation quality control, and complex parameter estimation and inconvenient application, this paper proposes a precipitation estimation model for automatic weather station based on the BP neural network. The proposed model imitates the structure of human brain neural system and its information processing mode and manages to learn the precipitation and location data of automatic weather stations. The hourly precipitation of adjacent stations and the distance between the adjacent stations and the stations which are to be tested are used to estimate the precipitation of test stations. Then the quality control of the measured precipitation is completed by comparing the difference between the estimated value and the measured value. The model adopts three-layer network structure, integrates four empirical formulas to determine the interval of hidden layer nodes, and uses trial and error methods as well as comparison method to determine the optimal node number and the optimal training function. Through training, the model achieves good estimation effect. The R~2 of training, testing, prediction samples and the measured value all passes over 0.9. The estimation accuracy of neural network model is much higher than that of spatial interpolation method. In addition, the model is used to analyze the abnormal precipitation cases in Ninghe Town and Hangu Street, and the analysis result shows that the estimated value of the model can accurately reflect the low measured precipitation. The comparison between the measured precipitation and the estimation results of neural network model indicate that the abnormal precipitation data can be found in time, the effect and efficiency of quality control work can be improved, and the pertinence of maintenance support for automatic weather stations can be improved.