Juan R. Rabuñal and Alejandro Pazos, Alberto Fernandez-Sanchez, Luis Cea-Gomez, Marcos Gestal and Daniel Rivero
In this paper, Artificial Intelligence techniques such as different types of Artificial Neural Network architectures and other regression models have been used to predict the inflow of water into a hydroelectric dam from rainfall records from different areas of the basin, a physical phenomenon known as "Rainfall-Runoff transformation".
The environment of this research work includes the Portodemouros dam, in northwestern Spain, and among the rainfall records used to predict the inflow to the dam, rainfall in different regions of influence have been processed. The training set contains samples of more than 3800 days from 2009 to 2020 and the validation and test sets contain more than 300 samples with data from 2021 and 2022 respectively.
Having predictions of the flow of water entering the dam at least one day in advance makes it possible to manage the dam's operation with greater security, as well as to prevent possible floods and flooding. In this research, real rainfall data from different areas of the basin are combined with meteorological estimates of rainfall predictions (using radar) as input to the Artificial Neural Network (ANN) that will produce as output the flow of water entering the dam. The ANN is adapted to work with real time series. The results are compared with other machine learning techniques such as SVM, etc.
The models created have obtained, in the test set, a correlation coefficient of the predicted output with the real output of 90%, which allows to provide a reliable prediction model to improve the operation of the hydraulic dam.
