The streamflow prediction of Kurkursar river using hybrid artificial intelligence models

Streamflow prediction is a challenging and critical task for water resource management. Streamflow prediction is one of the essential steps for reliable and robust water resources planning and management. On the other hand, streamflow prediction plays a crucial role in water resources systems planning and mitigating hydrological extremes such as floods and droughts. Since a variety of uncertainties exist in streamflow prediction, it is necessary to enhance our efforts to robustly address uncertainties and their interactions for improving the reliability of streamflow prediction. It is highly vital for hydropower operation, agricultural planning, and flood control. Physical process-based models are developed based on the understanding of the runoff generation processes, transport in channels, and mathematical formulations or parameterization of these physical processes. Data-driven models have the advantages of low demand for model input and ease of use. Due to the influence of various factors, including climate change, human activities, and socio-economic development, the hydrological time series leads to its complicated stochastic characteristics. Statistical models are the typical examples of this category and have been widely used in streamflow predictions over different regions. Different statistical models, machine learning methods are effective in describing the nonlinear characteristics of the observations and offer an alternative approach for streamflow prediction. Over the last two decades, a variety of machine learning techniques, including the artificial neural network (ANN) and the support vector machine (SVM), have been extensively applied in water resource management and hydrological prediction.

In this research, from the data of precipitation (Pt), precipitation with a delay of one day (Pt-1) until precipitation with a delay of three days (Pt-3) and discharge with a delay of one day (Qt-1) until discharge with a delay of 3 days (Qt- 3) are used as input variables and discharge (Qt) is used as output variable to predict the flow of Kurkursar river in Nowshahr. The Kurkursar river basin in the north of Iran is considered a sub-basin of the North Sea, its area is 75.495 km2, the average height of the basin is about 860 m, and the overall average slope is about 80.21%. This basin has an oblique shape and is limited to the Caspian Sea from the north, the Meshlak river and part of the Chalus watershed from the south, and the Chalus watershed from the west. Various geomorphological forms such as floodplains, alluvial cones, and sediment dams have been identified in the river basin area. The time series is daily and 70% and 30% of the data are respectively used for the training and test processes. The models used in this research include three individual models (random forest model - artificial neural network and regression support vector machine) and three hybrid models including bagging tree model - random forest (BA-RF), neural network - creative rifleman (ANN-AIG) and Support Vector Machine Regression-Crow Search (SVR-CSA) was used. In order to evaluate the model, the Mean square error (MAE), root mean squared error (RMSE), Nash–Sutcliffe model efficiency coefficient (NSE), and the PSR were used. Pearson's correlation coefficient (PCC) is used for the relationship between input and output variables. Therefore, we calculate the correlation coefficients between input and output variables, then we evaluate the composition of the input model based on different scenarios. The model combination that has the highest correlation is selected as the selected model. Then, the internal coefficients of each model are optimized with the help of meta-heuristic optimization algorithms. Therefore, the prediction results and observational data of the model are compared with each other. Finally, time series graphs, data dispersion, and box plots will be drawn and we will compare different evaluation indicators quantitatively and qualitatively. Finally, we compare the results of all models with each other and choose the model that has the best result as the best model. In the final step, we will compare the accuracy increase of the hybrid model (error reduction) with the standalone model to determine the error reduction percentage.

The Koukursar Nowshahr river flow was predicted using seven model with combinations of predictive variables such as R (t), Q (t-1), Q (t-2), R (t-1), Q (t -3), R (t-2) and R (t-3) as input variables and flow rate (Qt) as output variable. The results show that the precipitation variable (Rt) with a value of 0.563 has the highest correlation with the output variable, followed by Q (t-1) with 0.463, Q (t-2) with 0.297, R (t-1) with 0.281, Q (t-3) with 0.251, R (t-2) with 0.124, and R (t-3) with 0.072. Variable R(t) with 0.563 has the highest correlation, and variable R(t-3) with 0.072 has the lowest correlation with the output variable. Based on this, the relationship between the input and output variables was shown based on the correlation coefficient using the radar chart. In this regard, in this research, three individual models and three hybrid models were used to predict the river flow. Among the different scenarios and combinations of the input model, model 7 was used as the final model for forecasting. The evaluation results show that the RF model in the training phase has R2= 0. 957 and in the test, a phase has R2=0.717. The RF model has the highest correlation coefficient among all research models in the training phase. Also, in the test phase, it has the third rank among all models. ANN-AIG model improved the error of the single model by 32.94 %, the single SVR-CSA model by 23.17%, and the BA-RF model by 17.74%.

The qualitative results of the models show that all the models performed very well in predicting the model. Among the research models, the ANN-AIG model has performed best in forecasting.