基于不同输入层的BP人工神经网络径流模拟研究

摘要

In order to study the influence of different input layers of BP artificial neural network on runoff simulation results, based on the long-term daily precipitation of 8 precipitation monitoring stations in Binjiang River Basin, the comparative analysis of monthly runoff differences under different input layers, namely the original precipitation, the average precipitation, the com-posite antecedent precipitation, the watershed area weighted precipitation and the composite runoff area weighted precipitation, is conducted. The studies showed that the correlation coefficient and uncertainty factor of the runoff simulated by using the water-shed area weighted precipitation data as the input layer is relatively optimal, while the relative error of the original precipitation data input is the smallest. The results calculated by using the average precipitation as the input layer is the most concentrated and stable. The prediction accuracy of composite input layer is relatively higher, but the results are not necessarily better by adopting more complex input layer. The watershed area weighted precipitation obtained the best runoff prediction results.%为定量研究BP人工神经网络不同输入层对径流模拟的影响,以滨江流域8个雨量监测站长系列逐日降水径流资料为基础,对比分析原始降水、算术平均降水、复合前期径流降水、流域面积加权降水和复合径流面积加权降水作为输入层时,BP人工神经网络月径流量模拟的结果差异。研究表明：采用流域面积加权降水模拟的径流量,具有最优相关系数和确定性系数,以原始降水作为输入层所得结果相对误差最小,由算术平均降水模拟出的结果分布最集中,网络模拟效果稳定。复合输入层的模拟精度相对较高,但输入层并非越复杂越好,基于面积加权降水得出的模拟径流量综合评价最高。