Development a new hydraulic method for prediction of river flood discharge

The correct estimation of flood flow in rivers is an important issue and plays a significant role in the optimal use of water resources, operation of dam reservoirs, and the design and planning of water projects.

In this research, a simple and conceptual method based on Manning's formula in real flow conditions is used to estimate the flood flow discharge. In this method, firstly, for the combined effect of friction slope and Manning's roughness coefficient, the alpha parameter (α) was defined and calculated for 12 hydrometric stations located in three main rivers of Golestan province (including Gorganrood, Atrak, and Qarasoo).

The results showed that the value of this parameter decreases continuously with the increase of the flow depth and finally asymptotically reaches a constant value. This behavior shows that the value of α is nearly constant for the upper flow depths which indicate the occurrence of floods, and hence using this constant value and the Manning formula, the river flood discharge can be estimated. In the next step, we tried to provide a regression model between the Alpha parameter and the flow depth. The regression modeling results showed that for most of the hydrometric stations, the coefficients of determination (R2) of the presented equations are smaller than 0.3 which demonstrates its low efficiency. For this reason, machine learning models were used and the parameter was modeled by the Artificial Neural Networks (ANN), Decision Tree (M5tree), and Support Vector Regression (SVR) models.

The modeling results showed that the decision tree model with a mean absolute error of 0.35, determination coefficient of 0.88, and root mean square error of 0.86 has the best accuracy in the test phase. After determining the parameter α, the amount of flood discharge was predicted. The best performance among the models was the decision tree in predicting the flow rate in rivers. After comparing the observed values, the decision tree model has an average absolute error of 1.32, a determination coefficient of 0.89, and an average square root error of 63. 3. It has the best accuracy in the test phase.