Daily reservoir inflow forecasting using fuzzy inference systems

Abstract

This paper presents the application of a methodology for daily reservoir inflow forecasting in Brazilian hydroelectric plants. The methodology is based on Fuzzy Inference Systems (FIS) and the technique used for adjusting of the model parameters is an offline version of the Expectation Maximization (EM) algorithm. In order to automate the application of the methodology and facilitate the analysis of the results, a tool that allows managing streamflow forecasting studies and visualizing their information in graphical form was developed. A case study was applied to the data from three Brazilian hydroelectric plants whose operation is under the coordination of the Electric System National Operator. They are located in the Grande basin, a part of the Parana basin with two main rivers: the Grande and the Pardo. The benefits of the model are analyzed using statistics calculations, such as: root mean square error, mean absolute percentage error, mean absolute error and mass curve coefficient. Besides that, graphics that compare the registered and predicted streamflow are presented. The results show an adequate performance of the model, leading to a promising alternative for daily streamflow forecasting.

A Gradient Boosting Model Optimized by a Genetic Algorithm for Short-Term Riverflow Forecast

“…Considering the importance of water resources in the Brazilian energy matrix and the fact that the hydrological events of rainfall and flow are naturally random, the studies on flow forecasting are fundamental mechanisms to subsidize the planning and coordination of activities of energy production and dispatch of energy, thus reducing the risks of not meeting the country's energy demand. However, streamflow prediction remains one of the most problematic tasks in the hydrological field since the system is characterized by a nonlinear dynamic process that depends on hydrological and geographic variables, as well as the river morphology (HUAMANI et al, 2011). In some cases, the collection and sharing of hydrometeorological data are not direct or they are of low quality (BOU-FAKHREDDINE et al, 2018).…”

Daily river flow prediction based on Two-Phase Constructive Fuzzy Systems Modeling: A case of hydrological – meteorological measurements asymmetry

“…Accurate prediction of river flow is of vital importance for efficient reservoir water management and control. However, forecasting river flow remains one of the very difficult issues in hydrological sciences because it is characterized by a dynamic, uncertain and nonlinear problem (Huamani et al, 2011). This problem deals with a system that receives thousands of inputs interacting in a complex and noisy environment.…”

“…In fact, (Porporato and Ridolf, 2001) dealt with local linear models with time-dependent parameters, whereas (Dibike and Solomatine, 2001) and (Pulido-Calvo and Portela, 2007) have considered data-driven nonlinear models based on Artificial Neural Network (ANN) or on Wavelet Neural Network (WNN) as in (Cuia et al, 2015). In (Huamani et al, 2011), the followed methodology was based on Fuzzy Inference Systems (FIS). However, (Coulibaly and Baldwin, 2005), (Firat, 2008) and (Kisi et al, 2012) have 1 discussed extensively neuro-fuzzy hybrid models that have the capability of preserving the learning abilities of ANNs and the reasoning of fuzzy systems.…”

“…Indeed, the presented method is based on a Constructive-Fuzzy System Modeling (C-FSM) and it is formed of two steps: First, the model is initialized by applying the Subtractive Clustering (SC) algorithm on the available historical data to determine the initial structure of the system (Huamani et al, 2011). In fact, this procedure had provided an extra tool to divide the heterogeneous data into more homogeneous sub-populations which in turn improves the forecasting accuracy (Asadia et al, 2013).…”

“…It aims to analyze and discuss stochastic modeling of river flow time series using FIS coupled with traditional correlation analysis. (Huamani et al, 2011) have presumed the normality of the gathered data, and in the course of the C-FSM training process they didn't state a clear approach for choosing the appropriate cluster radius. In this work, and to boost the performance of the C-FSM, the collected data had been brought to near a normal distribution using a suitable transformation.…”