Artificial Neural Network and Fuzzy Inference System Models for Forecasting Suspended Sediment and Turbidity in Basins at Different Scales

Abstract

The monitoring of hydro-sedimentological processes is important for environmental control but depends on resources that are not always available. The estimation of sedimentological variables with mathematical models is often limited by the scarcity of data for a single basin. This research experiments with the simulation of suspended sediment concentration (SSC) and turbidity (T) using a regional model, with data from agricultural basins of different scales within the same hydrographic region, using hourly precipitation as one of the predictive variables, aggregated through the exponentially weighted moving average (EWMA) of past rainfall, in artificial neural network (ANN) and fuzzy inference system (FIS) models. The data monitoring was performed from January 2013 to February 2020 in four watersheds within the same region in southern Brazil, with areas ranging from 1.3 to 524.3 km2. For the turbidity estimation, the FIS model, which also made use of the discharge (Q) and area (A) of each basin as inputs, performed best, with a Nash-Sutcliffe efficiency (NS) of 0.860 for the verification samples. Several FIS and ANN models performed very well for SSC prediction (with NSs ranging from 0.950 to 0.977) due to the EWMA variable, including an FIS model that uses only this variable (NS 0.952). The results allow us to conclude that it is possible, with few data for the individual basin and a regional empirical model, to estimate SSC and turbidity, provided the aggregation of hourly precipitation by the EWMA, as long as the basins have similar physical and climatic characteristics.