An extreme learning machine model for the simulation of monthly mean streamflow water level in eastern Queensland

An extreme learning machine model for the simulation of monthly mean streamflow water level in eastern Queensland

A predictive model for streamflow has practical implications for understanding drought hydrology, environmental monitoring and agriculture, ecosystem and resource management. In this study state-or-art Extreme Learning Machine (ELM) model was utilized to simulate the mean streamflow water level (QWL) for three hydrological sites in eastern Queensland (Gowrie Creek, Albert River & Mary River). The performance of ELM model was benchmarked with the Artificial Neural Network (ANN) model. The ELM model was a fast three step simulation method designed using the Single Layer Feedforward Neural Network (SLFNs) and randomly determined hidden neurons that learned the historical patterns embedded in the input variables related to QWL. A set of nine predictors with the month (to consider seasonality of QWL), monthly rainfall, Southern Oscillation Index (SOI), Pacific Decadal Oscillation (PDO) Index, ENSO Modoki Index (EMI), Indian Ocean Dipole (IOD) Index and Nino 3.0SST, Nino 3.4SST and Nino 4.0SSTs were utilized. A pre-selection of variables was performed using cross correlation analysis with QWL, yielding the best inputs defined by (month; P; Nino 3.0SST; Nino4.0SST; SOI; EMI) for Gowrie Creek, (month; P; SOI; PDO; IOD; EMI) for Albert River and (month; P; Nino3.4 SST; Nino4.0 SST; SOI; EMI) for Mary River site. A three layer neuronal structure was employed to trial activation functions defined by the sigmoid, logarithmic, tangent sigmoid, sine, hardlim and triangular and radial basis equations for feature extraction that utilized between 2 to 150 hidden neurons. This resulted in the optimum ELM model executed by hard-limit function with a neuronal architecture 6-106-1 (Gowrie Creek), 6-74-1 (Albert River) and 6-146-1 (Mary River). The simulations were also performed with two inputs (month & rainfall) and all nine inputs. The model performance was evaluated using mean absolute error (MAE), coefficient of determination (r2), Willmott index (d), peak percentage deviation (Pdv) and Nash-Sutcliffe coefficient (ENS). Results found that the ELM was more accurate than the ANN model for simulation of QWL. Inputting the best six input variables improved the performance of both models where the optimum model ELM yielded R2  (0.964, 0.957 & 0.997), d  (0.968, 0.982 & 0.986), MAE  (0.053, 0.023 & 0.079 for Gowrie Creek, Albert River and Mary River, respectively, and the optimum ANN model yielded smaller R2 and d and larger simulation errors. When all nine inputs were utilised, the simulations were consistently worse for all stations with R2 (0.732, 0.859 & 0.932 (Gowrie Creek); d (0.802, 0.876 & 0.903 (Albert River) and MAE (0.144, 0.049 & 0.222 (Mary River) although they were relatively better than using only the month and rainfall as inputs. Also, with best input combinations, the frequency of simulation errors fell in smallest error bracket. Therefore, it is ascertained that the ELM algorithm offers an efficient soft-computing approach for simulation of streamflow, and therefore, can be explored further for its practicality in hydrological modeling.

Files associated with this item cannot be displayed due to copyright restrictions.