人工神经网络结构对径流预报精度的影响分析

建立了基于径流形成机理的以时段降水量与前期径流量为预报因子的前向多层人工神经网络径流预报模型；分析了网络结构对月径流预报精度的影响，发现随网络结构的复杂化，网络训练误差减小，模型评定的确定性系数增大，并均趋于稳定，预报检验的确定性系数总趋势是减小；发现影响模型精度的决定因素是网络输入单元数，亦即径流影响因素；提出了以模型评定与预报检验共同高效或等效的模型选择的折衷方法，以及按模型适宜预报域进行多模型组合预报的最佳预报域组合法。

AN ANALYSIS OF EFFECTS OF ARTIFICIAL NEURALNETWORK STRUCTURES ON PRECISION OFSTREAM FLOW FORECASTING

A stream flow forecasting model of feed forward multi layer artificial neural network(ANN), in which current precipitation and antecedent flow are considered as the model inputs according to runoff generation mechanism, is introduced. The deterministic coefficient is adopted as a norm to control ANN training error and precision of model calibration and verification. It is shown through the study that ANN training error is decreased and the coefficient of model calibration is increased, and meanwhile the coefficient of model verification is persistently decreased, with increase of complexity of ANN structures. It is also recognized that the key factor affecting the model precision is the number of neurons in the input layer, i e., the number of flow effecting factors. A method to select models for operational application, and to combine optimal forecasting ranges is proposed.