Estimation of hydrological parameters in Apa Dam basin by means of machine learning

In recent years, especially after the industrial revolution, humankind's need for water has increased. The preservation of the future and current amounts of this water, whose vital importance for living things is indisputable, has gone beyond becoming an issue that concerns all human beings and has become debatable what everyone can do about this issue in recent days. Drought, defined as a decrease in the average rainfall over a certain period, is among the factors that affect the availability of water resources. Drought is one of the most insidious natural disasters. When it occurs, people may not be aware of it. Drought indices are generally used to monitor and evaluate drought events. In this study, Standardized Precipitation Index (SPI) and Streamflow Drought Index (SFI) were used as drought index. In addition, hydrological and meteorological data affecting droughts and used in their calculations were also utilized. Monthly total precipitation, monthly average temperature, monthly maximum temperature, monthly minimum temperature, monthly average dam lake level elevation, monthly average dam volume data, monthly average flow data covering the widest range of 1955-2022 were used in this study. The Apa Dam Region, located in the middle parts of the Konya Closed Basin which meets a significant part of Turkey's grain needs, was chosen as the study area. From past to present, human beings have been struggling with environmental problems. In this struggle, scientists are looking for alternative solutions for the drought phenomena. As with almost every problem, one of the best ways to find a solution to a problem is to make future predictions by taking advantage of our past experiences. In hydrology, future actions can be predicted by using past data - assuming that past behavior will not change - and studies are carried out on topics such as water cycle, water resources planning, basin management, drought analysis and mathematical modeling. Inspired by all these, in this study machine learning algorithms, Support Vector Machines, Artificial Neural Networks, Decision Trees, Random Forest, Multiple Support Vectors were used to predict the hydrological parameters affecting the Apa Dam Region, Konya. Subsequently, 3 different models were established. These models are drought prediction model, rainfall-runoff prediction model, dam lake level prediction model. Wavelet transform was used to improve these models and increase their predictive power. Both the established models were compared with each other according to their performance, and the machine learning algorithms were compared with each other according to statistical criteria. For this purpose, Nash-SutCliffe Efficiency(NS), Root Mean Square Error (RMSE), Correlation Coefficient (r) were used to compare model performances. As a result of the study, it has been determined that SVM generally gives better results than other algorithms in drought prediction models. In drought prediction models, it has been determined that models whose model input structure is created with both SPI and monthly total precipitation data perform more effectively than other models. The most successful model in drought prediction models is M05, which was determined in the wavelet transform of SVM (with performance values of NS:0.9970, RMSE:0.0498, r:0.9986). In rainfall-runoff prediction models; in analyzes performed without wavelet transform, SVM and RO generally gave better results than other algorithms. Here, increasing the data diversity in the model structure positively affected the model performance parameters. One of the striking findings in the rainfall-runoff modeling in this study is that using only flow data in the input structure after wavelet transformation positively affects model performances, unlike the findings in previous models without wavelet transformation. M26 (NS: 0.9781, RMSE: 0.2989, r: 0.9893) gave the most effective result in this class. In dam lake level prediction models; RO and SVM without wavelet transform are superior to other algorithms. Using dam volume and dam lake level data in the model input structure increased model performances. After wavelet transformation, M18 (NS: 0.9851, RMSE: 0.6942, r: 0.9926), which was created from flow data and elevation data in the model input structure, showed superior performance compared to other models.