Amirkabir Dam Inflow Prediction Using Teleconnection Patterns and Machine Learning Models

The demand for freshwater is increasing, while the limited water resources are subject to over-harvesting, pollution, and climate change, which require improving water resource management to distribute it equitably and achieve It highlights the goals of sustainable development. A low-cost option to support better water management strategies is to develop models capable of predicting available water amounts, especially amounts related to precipitation and river flow. Climatic diversity and climate changes are basic assumptions for hydro climatological predictions. One of the remarkable aspects of this issue is the correlation between large-scale atmospheric-oceanic phenomena or Teleconnection patterns with hydrological processes on a local scale, and these patterns can also affect the inflow to the dams. This study uses three machine learning models, an artificial neural network, a Bayesian neural network, and an adaptive neuro-fuzzy inference system to predict dam inflow and evaluate their efficiency. For this purpose, 12 scenarios consisting of rainfall variables, inflow to the dam, and nine climatic indicators with a delay of up to six-time steps were designed to investigate the effect of using long-term models as predictive variables of the flow one month later in Amirkabir Dam. to be placed The analysis of the results of this research showed that the use of the Nino3.4 index with one-time step delay as well as the PDO index with two-time step delays can increase the accuracy of the model compared to the scenarios in which only station variables are used. to be According to the results, the Nino 3.4 index was found to be the most effective index on the inflow to Amirkabir Dam, and the scenario in which the mentioned index along with the rainfall and flow data of one and two months before was used as input, in all three The model recorded the highest accuracy. Also, the performance of the ANFIS model for the mentioned scenario (scenario 9), with RMSE and R2 values, equal to 5.69 and 0.79 cubic meters per second, respectively, was better than the ANN and BNN models, so the value of the R2 index for the best scenario consisting of station variables (scenario 5), it increased by 0.15 and the value of RMSE index decreased by 0.78 cubic meters.