Evolving Connectionist Systems Versus Neuro-Fuzzy System for Estimating Total Dissolved Gas at Forebay and Tailwater of Dams Reservoirs

Abstract

AbstractThis chapter proposes new method to estimate total dissolved gas (TDG) concentration, which is a critical factor causing gas bubble trauma in fish. Two kinds of data-driven approaches were applied: evolving connectionist systems (ECoS) and neuro-fuzzy systems (NFs). For the first group, we selected two ECoS models, namely (i) the off-line dynamic evolving neural-fuzzy inference system called DENFIS_OF and (ii) the on-line dynamic evolving neural-fuzzy inference system called DENFIS_ON. For the second group, three NFs models were selected, namely (i) adaptive neuro-fuzzy inference system (ANFIS) with fuzzy c-mean clustering (FC) algorithm called ANFIS_FC, (ii) adaptive neuro-fuzzy inference system with grid partition (GP) method called ANFIS_GP, and (iii) ANFIS with subtractive clustering (SC) called ANFIS_SC. In addition, results using the standard multiple linear regression (MLR) were provided for comparison. The proposed models were developed using several inputs variables, e.g., water temperature, barometric pressure, spill from dam, and discharge. Several inputs combinations were considered and compared to find the best inputs variables for estimating TDG, and several scenarios were developed and tested. Firstly, the proposed models were applied and compared for predicting TDG measured at the Tailwater of the dams using 70% of the data for training and 30% for validation (scenario 1). Secondly, using the same splitting ratio, the models were applied and compared for predicting TDG measured at the Forebay (scenario 2). Thirdly, the best models for the first two scenarios were selected and trained using validation data set and tested with the training data set (scenario 3). Fourthly, and finally, TDG is predicted without the well-known inputs variables, but rather, using the component of the Gregorian calendar as inputs variables (scenario 4). All the four scenarios were achieved using data collected from US Army Corps of Engineers and measured at hourly time step. The accuracy of the models was evaluated using coefficient of correlation (R), Nash-Sutcliffe efficiency (NSE), root mean squared error (RMSE), and mean absolute error (MAE). The applications, at Forebay and Tailwater of dam’s reservoirs, revealed that the proposed methods could be successfully utilized for estimation of TDG concentration using the component of the Gregorian calendar as input variable.KeywordsMachine learningInput variabilityTotal dissolved gasFuzzy set