Author(s): Donghao Miao
Linked Author(s):
Keywords: Catchment; Pollutant load; Neural network; Weight initialization; SOBOL sensitivity analysis; Stability
Abstract: Within a river catchment, the relationship between pollutant load accumulation and its related factors is nonlinear. When neural network models were used to identify the nonlinear relationship, data scarcity and random weight initialization might result in overfitting and instability. In this paper, we proposed an averaged weight initialization neural network (AWINN) to carry out multi-index integrated prediction of the pollutants loads under data scarcity. The results showed that: (1) Compared with PSONN and AdaboostR models that prevent overfitting, AWINN improved simulation accuracy significantly. The R2 in test sets of different pollutants’ models reached 0.51-0.80. (2) AWINN is effective in overcoming instability. With more hidden layers, the stability of the models’ outputs is stronger. (3) The study explains the variations of main factors across seasons. The algorithm proposed in this paper can realize stably integrated prediction of pollutant load in the catchment under data scarcity and has a good application prospect.
Year: 2024