Hydrological Modeling – An Approach with Advanced Neural Network Models and their Sensitivity to Initial Solution

  • Unique Paper ID: 181888
  • Volume: 12
  • Issue: 2
  • PageNo: 66-73
  • Abstract:
  • This study introduces a novel approach to hydrological modeling through the application of a Generalized Neuron (GN), an advanced artificial neuron architecture. Unlike the conventional McCulloch and Pitts’ artificial neuron (MPAN) commonly utilized in artificial neural networks (ANNs), the GN is structured to handle complex nonlinearities in hydrological systems using a non-linear discriminant function. This compact model architecture eliminates the need for specifying hidden layers and neurons, offering a streamlined modeling process. Two neural system (NS) models were developed: (a) a standard multilayer perceptron (MLP) model trained using the back-propagation (BP) algorithm, and (b) the GN-based model. Both models were tested on rainfall and discharge data from the Kentucky River basin, using ten distinct initial weight configurations to assess sensitivity. Evaluation was carried out using multiple performance indicators. Results show that the GN model not only outperforms the traditional MLP model in terms of accuracy and training efficiency but also demonstrates robustness against initial weight sensitivity. This highlights the potential of GN as a powerful tool for modeling nonlinear hydrological processes.

Related Articles