基于静态与动态神经网络的运河水位预报
Water Level Forecasting by Static and Dynamic Neural Networks

以杭州市上塘河运河为例, 分别建立自回归、静态时延神经网络及动态反馈神经网络模型预报受闸门及泵站影响的城市河道未来1～3 h的水位变化, 并对模型预报误差、不同水位区间下的误差-频数关系及训练数据的不确定性进行分析．研究结果表明：在预见期1～3 h内, 时延神经网络预报效率系数均达到0.9以上, 比自回归模型分别提高1.34% 、5.57% 和6.86% , 比动态反馈网络分别提高0.21% 、1.97% 和1.98% ; 动态网络模型在人为调控的影响下仍能通过训练模拟出降雨径流关系, 对数据具有更好的自学习与调适能力; 时延神经网络模型随数据减少预测精度也减少, 在仅保留1个洪水场次下效率系数最大降低8.05% , 而动态反馈网络效率系数随训练数据量变化基本在0.11% 内波动, 因此在数据量较少的情况下宜建构动态模型．
The auto-regression(AR)model，static time-delay neural network(TDNN)and dynamic recurrent neural network(RNN)models were constructed for 1―3 h ahead water level forecasting at the complex Shangtang Canal system where water level was affected by sluices and pumping stations. Analyses were conducted of the forecast bias and error-frequency relationship at different water levels，as well as the uncertainty of training data. The results indicated that: the coefficient of efficiency obtained from TDNN was higher than 0.9，which resulted in 1.34% ，5.57% and 6.86% improvements as compared to AR，and is 0.21% ，1.97% and 1.98% improvements as compared to RNN for 1―3 h forecasting; the RNN model could still be trained to simulate the rainfall-runoff relationship even though the training data contained man-made noise，indicating that the RNN model is capable of self-learning and adaptability; the accuracy of TDNN model，based on a single flood event，was 8.05% lower than the original performance as the training size decreased in terms of coefficient of efficiency，whereas the forecasted error obtained from RNN model was basically within 0.11% with training size changing. Therefore，it is suggested to adopt RNN model in the case of insufficient training data