SIMULATION OF RAIN FLOODS OF THE STRYI RIVER BY AN ARTIFICIAL NEURAL NETWORK

Timely forecasting of rain floods on rivers allows you to avoid the negative consequences that they can cause by destroying structures and communications located in their channels or on floodplains. It is relevant for the Stryi River, as catastrophic rain floods occur in its basin from time to time. The article is devoted to the use of an artificial neural network (ANN) for modeling the rain runoff of the Stryi River near the Verkhnie Synyovydne village for the period 2005-2012. Using the 'nnet' package in RStudio (version 2024.12.0 Build 467), a direct-connection ANN model was developed. In addition, the classical linear multiple regression model (CLMRM) was also used. The model based on ANN has an advantage over the CLMRM, as its statistical indicators of modeling quality are higher. Thus, the Nash-Sutcliffe efficiency coefficient of the ANN model for the training sample was 91.6%, and for the test sample was 92.5%, which classifies it as excellent. However, for CLMRM these indicators were 81.5% and 89.4%, respectively. Graphical analysis also demonstrated the advantage of the ANN model, since it was for it that a better match between the simulated and historical values was obtained, which is confirmed by higher of determination coefficients (0.92 ANN and 0.82 CLMRM for the training sample and 0.93 ANN and 0.89 CLMRM for the test sample). The statistic indicators RMSE of the CLMRM model turned out to be greater than the ANN model for both the training (29.8 m3/s and 20.1 m3/s) and test (25.2 m3/s and 21.1 m3/s) samples. The advantage of the ANN model over the CLMRM model is that it takes into account the nonlinearity of 'rainfall-runoff' relationship due to the parallelism of its architecture. In Ukraine, river flow modeling by ANNs was caried out for the first time. This approach may be particularly relevant for transboundary rivers of the country, for which there are significant problems with the availability of observation data.