Comparison of Different Artificial Neural Network Methods in Determining Reservoir Capacity

Abstract

Making estimation of river flow with hydrological and methodological data of the past period and using in water resources project studies have been used for a long time in many studies in our country and in the world. In addition to this, along with frequent drought problems in recent years, it is also important to store and use the water in the reservoirs correctly. This study aims to research how two separate Artificial-Neural-Network-Functions, which are generated based on the study of neural networks in the human brain, can be used in areas with a certain reservoir capacity such as dams or ponds and to provide an example of the most appropriate ANN model for predicting the level changes that will occur depending on the next years. In this context, two separate functions have been evaluated. These are the Gradient-Descent-with-Momentum (GDM) and Levenberg-Marquardt-(LM) training functions. Here, instead of the best results obtained from the models, the best three model outputs were evaluated and the models were considered from a wide frame. Gökçe Dam Basin, which is located in the Marmara region was chosen as the study area. With the different architectures prepared for the period of 31 December 2019 -1 January 2000, monthly basin capacities for 2019 have been tried to be estimated. 95% of the data belonging to the selected period was used for training and the remaining 5% for testing purposes. The flow rates entering and leaving the dam basin and the average precipitation, evaporation, and dam seepage amount were used as model inputs. While evaluating the model performances, Mean squared error (MSE), Mean absolute percent error (MAPE), Mean absolute error (MAE), and coefficients of determination were taken into consideration. As a result; It can be stated that the LM training models are more successful in estimating the dam basin levels and converge to the real values.