Biased wavelet neural network and its application to streamflow forecast

Long leading-time streamflow forecast is a complex non-linear procedure. Traditional methods are easy to get slow convergence and low efficiency. The biased wavelet neural network (BWNN) based on BP learning algorithm is proposed and used to forecast monthly streamlfow. It inherits the multiresolution capability of wavelets analysis and the nonlinear input-output mapping trait of artificial neural networks. With the new set of biased wavelets, BWNN can effectively cut down the redundancy from multiresolution calculation. The learning rate and momentum coefficient are employed in BP algorithm to accelerate convergence and avoid falling into local minimum. BWNN is applied to Fengtan reservoir as case study. Its simulation performance is compared with the results obtained from autoregressive integrated moving average, genetic algorithm, feedforward neural network and traditional wavelet neural network models. It is shown that BWNN has high model efficiency index, low computing redundancy and provides satisfying forecast precision.