ABSTRACT

Various hybrid approaches combined the different deep learning and machine learning models with evolutionary optimization algorithms and have improved the accuracy of streamflow forecasting problem. In this article, three deep learning models were investigated for streamflow forecasting with various lag times at both stations (i.e. Sidi Aich and Ponteba Defluent), Algeria. Also, a machine learning [i.e. feedforward neural network (FFNN)] model was implemented to compare the forecasting accuracy of deep learning models. The particle swarm optimization algorithm was combined to determine the hyperparameters (i.e. model structure) automatically based on adaptive moment estimation algorithm.

The addressed two-stage hybrid models were assessed and evaluated by root mean square error (RMSE), signal-to-noise ratio (SNR), and Nash–Sutcliffe efficiency (NSE) statistical indices. Evaluating all models explained that the GRU II two-stage hybrid model (RMSE = 35.241 m³/s, SNR = 0.5159, and NSE = 0.7337 at Sidi Aich and RMSE = 11.074 m³/s, SNR = 0.3600, and NSE = 0.8703 at Ponteba Defluent) was found to produce more accurate results compared to the Elman recurrent neural network, long short-term memory, and FFNN two-stage hybrid models during testing phase for forecasting streamflow.

摘要

各种混合方法将不同的深度学习和机器学习模型与进化优化算法相结合，提高了流量预测问题的准确性。在本文中，研究了三个深度学习模型，用于在阿尔及利亚两个站点（即 Sidi Aich 和 Ponteba Defluent）进行具有不同滞后时间的流量预测。此外，还实施了机器学习 [即前馈神经网络 (FFNN)] 模型来比较深度学习模型的预测准确性。结合粒子群优化算法，基于自适应矩估计算法自动确定超参数（即模型结构）。

通过均方根误差 (RMSE)、信噪比 (SNR) 和 Nash–Sutcliffe 效率 (NSE) 统计指标对已解决的两阶段混合模型进行评估和评价。评估所有模型解释了 GRU II 两级混合模型（RMSE = 35.241 m ³ /s，SNR = 0.5159，NSE = 0.7337 在 Sidi Aich 和 RMSE = 11.074 m ³ /s，SNR = 0.3600，NSE = 0.8703在 Ponteba Defluent）被发现在预测流量的测试阶段与 Elman 递归神经网络、长短期记忆和 FFNN 两阶段混合模型相比产生更准确的结果。