基于多种递归神经网络的海岛水库径流预报

海岛地区径流量偏小甚至出现断流，会极大影响短期径流预报精度。对海岛地区供水水库多组入库径流时间序列，基于三种递归神经网络（RNN）建立了不同预报因子组合和预见期的径流预报模型，探讨了RNN模型在海岛地区短期水文预报中的适用性。以舟山岛水库群为例，说明研究方法的有效性。结果表明仅考虑径流时间序列信息的预报精度最差，而耦合气象预报信息可提高径流预报准确性；随着预见期的增加，简单RNN模型的信息融合能力有限，而具有复杂神经元结构的基本长短时记忆神经网络和门控循环单元预报效果稳定；RNN模型对于平稳时间序列数据模拟效果优于非平稳序列，而气象信息的引入和参数优选能够改善其在处理非平稳时间序列中的缺陷。

Reservoir inflow forecasting for island areas based on multiple recurrent neural networks

Accuracy of short-term runoff forecasting is often lowered due to low or even zero river flows in island areas. This study adopts three different recurrent neural networks (RNNs) to forecast several runoff series at different lead times and input combinations. Application to a case study of the Zhoushan Island shows that runoff forecasting coupled with future meteorological forecasting information achieves better performance than that based on runoff information only. As forecasting lead time increases, the long short-term memory network and gated recurrent unit are becoming better than the simple RNN model. RNN models perform better on stationary runoff series than non-stationary ones, and their stability and reliability in calculation of non-stationary runoff series can be improved by coupling meteorological information and through parameter optimization.