Abstract: In the presence of unmodeled dynamics, the parameters' drift even instability may occur in the identification system of neural networks. Input-to-state stability(ISS) approach is applied to achieving the error backpropagation-like time-varying learning algorithm of weight matrix and wavelet scaling parameters in wavelet neural networks, of which the robust stability is guaranteed without robust modification. Simulations showed that the stable learning algorithm outperforms conventional error backpropagation ones, and the application of the algorithm to the prediction of COD in wastewater treatment process gets good results.