Author(s): David Yaranga Lazaro; Kithner Espinoza Varillas; Jhon Zapana Arpasi

Linked Author(s): David Yaranga Lázaro

Keywords: Forecast; Neural networks; Water model; Wavelet

Abstract: Peru has been hit by at least twelve extreme El Niño events (EENE). Its presence is reflected in the increase of the water regime, greater transport of sediments, droughts and intense storms; characteristics of the climatic diversity observed along the Pacific and Titicaca slopes. One of the most devastating consequences of these events is flooding, causing economic losses, farmland and human lives. In view of this problem, the object of this research focuses all efforts on analyzing the forecast of daily flows using artificial neural network (ANN) and the wavelet transform. This investigation will be carried out in the Rio Chicama basin (7 ° 21 'and 7 ° 59' parallels and 78 ° 14 'and 79 ° 20' west longitude) Salinar hydrological station, Pacific slope, and the Huancane basin (14 29.75 'and 15 ° 21.36' south latitude and 69 ° 17.51 'and 70 ° 9.3' west longitude.) Huancane bridge hydrological station, Titicaca slope. The methodological approach developed was to compare the flow forecasts generated with artificial neural networks (ANN) and a hybrid that combines wavelet and ANN multiresolution analysis called wavelet neural network (WRN) model. In this respect, several simulations were carried out with different scenarios of univariate RNA and WRN models, in order to identify the optimal structuring of the neural network, an uncertainty analysis based on efficiency indicators (RMSE, MAE, MARE, NSE) and the best performance in the training and validation phase. The time series of daily flows for the Chicama basin, Salinar station (01 January 1975 – 06 December 2016) and the Huancane basin, Huancane bridge station (01 December 1975 – 31 August 2018). Later, flow forecasts were made for different horizons using the two approaches of ANN and WRN. The results achieved for the Huancane and Chicama river basin, showed a better performance in the water model WNN compared to the ANN; determining that in a maximum forecast horizon of 15 days, the Huancane basin obtained an RMSE = 10.8 m3/s and NSE=0.77, while in the Chicama basin an RMSE = 20.1 m3 / s and NSE = 0.62 was obtained, exceeding the ANN that had an RMSE = 16.93 m3 / s and NSE = 0.44 and RMSE=25.26 m3/s and NSE=0.41 for each basin respectively, in the validation period. Finally, the best forecast was obtained in a horizon of one day, observing an NSE=0.9967 and NSE=0.9964.

DOI: https://doi.org/10.3850/IAHR-39WC252171192022372

Year: 2022