High-Precision Real-Time Flow Prediction in a Multi-tributary River System: A Bio-inspired Dynamic Neural Network Model

Floods are among the most severe natural disasters globally, particularly in densely populated areas with extensive agriculture, concentrated rivers, and abundant rainfall. In recent years, human activities have altered river confluence conditions, exacerbating the frequency and severity of floods. To address the limitations of existing multi-tributary stream flow prediction models, which suffer from poor real-time performance and low prediction accuracy, we developed a bio-inspired neural network (Bio-NN) model motivated by a cooperative regulation mechanism in biological systems. Considering the problem that there is less feedback information in existing neural networks, the proposed model combines a biohormone multi-level nonlinear feedback regulation mechanism with a neural network. This enhances traditional neural networks by improving network structure and dynamically incorporating feedback information, allowing real-time optimization and improving optimization speed and precision over time. We tested the Bio-NN model by applying it to predict river flow along the lower Murray River in Australia. To obtain deeper insights into the performance of Bio-NN, indicators such as NSE, RSR, PCC, and KGE, were determined in the basin. The simulation demonstrated its superior performance, achieving a Nash-Sutcliffe efficiency coefficient (NSE) of 0.991, root mean squared to standard deviation ratio (RSR) of 0.096, a Pearson's correlation coefficient (PCC) of 0.996, and a Kling-Gupta efficiency coefficient (KGE) of 0.995. Compared to a back propagation neural network (BP-NN), a dynamic learning BP-NN, and a self-feedback BP-NN, the Bio-NN showed significant improvements in prediction performance: improved by 8-65% (NSE), 4-28% (PCC), 67-85% (RSR), 9-27% (KGE). The results underscore Bio-NN's capability to significantly enhance the accuracy and stability of flood prediction models.