基于门控循环单元—支持向量回归组合模型的湖泊水位预测方法探索

为改善传统循环神经网络预测梯度消失的问题，准确预测水位变化，采用门控循环单元(gated recurrent unit, GRU)和支持向量回归(support vector regression, SVR)构建组合预测模型，对广州市猎德涌的源头西湖水位进行预测。选择了3种不同核函数下的GRU-SVR(多项式核、RBF核、Sigmoid核)模型，并确定了最佳核函数组合，探索了GRU组合模型在水文时序预测中的有效性。该组合模型通过GRU提取雨量与水位间时空特征，SVR增强整体的非线性预测能力。结果表明，GRU-SVR(多项式核)适用于湖泊降雨时期预测，与CNN-GRU及GRU、SVR相比，该模型整体预测精度分别提升了3.2%、10.3%和59.3%。

Exploration of lake level forecast method based on GRU-SVR model combination

the study was aimed to improve the problem of gradient disappearance predicted by traditional recurrent neural networks (RNN) and discuss the use of combination model in water level forecast. A combination forecast model on account of gated recurrent unit (GRU) and support vector regression (SVR) was proposed. It was used to predict the water level of West Lake at Headwaters of the Liede Chung Valley. GRU-SVR (polynomial kernel, RBF kernel, and Sigmoid kernel) models were compared. Through GRU, the spatio-temporal characteristics between rainfall and water level were extracted by the combination model, and the overall nonlinear prediction ability was enhanced by SVR. The results showed that GRU-SVR (polynomial kernel) model had better prediction accuracy in lake rainfall period compared with single GRU and SVR. Finally, the integration of model components was discussed and analyzed.