灌区降水量影响因子分析与集合降水预测

Analysis of Influencing Factors of Rainfall in Irrigation Area and Combining Rainfall Forecasting

针对降水随机性较强、影响因素复杂、单一模型预测精度低的特点,采用集对分析法,研究降水量与气象影响因子的关系。将基于密度参数的径向基函数人工神经网络模型与灰色模型相结合,利用信息熵权重法计算2个单一模型的权重,构建基于信息熵的集合模型(Combing model based on information entropy,IE-CM),用于三江平原友谊农场的降水量预测。研究结果表明,与单一模型相比,IE-CM模型预测结果的决定系数、平均相对误差及均方根误差较单一模型均有所提高,分别为0.99、10.655%和3.03 mm,预测结果的合格率为83.3%,均满足水文预测要求。

Sanjiang Plain is located in the east of Heilongjiang Province, which belongs to the humid climate area. In recent years, Sanjiang Plain＇s natural state has changed due to several factors, such as the warming climate and human activity. Precipitation is one of the major sources of agricultural irrigation in the irrigation area. Due to the strongly stochastic characteristic of precipitation which was influenced by many factors and the lower accuracy of single forecasting model, set pair analysis was introduced which could discuss the relation between rainfall and meteorological factors. In order to improve the training speed of the radial basis function neural network, the K - means algorithm based on density parameter was applied. In this way, the sensitivity of conventional K - means algorithm to initial clustering center was also removed. A combing model based on information entropy （IE -CM） was built, which combined the radial basis function artificial neural network based on density parameter with the grey model, and the weight of each single model was calculated by using the information entropy weight method. The constructed model was applied to forecast the rainfall over the Youyi Farm in Sanjiang Plain. The case study showed that the determination coefficient, average relative error and root mean square error of IE - CM were better than those of single models, which were demonstrated to be 0. 99, 10. 655% and 3.03 mm, respectively. The qualification rate of the forecasted result was 83.3%, which satisfied the requirements of hydrologic prediction. In conclusion, the built combining model could provide a new method for forecasting precipitation.