85-Sequential Neural Network with Error Updating for Improved Higher Lead Time Flood Forecasts.

Abstract

A novel ANN architecture is proposed for forecasting river flows at higher lead times with greater accuracy. The paper predominantly demonstrates the potential in computing paradigm, through 'sequential ANN (SANN)', to extend the lead time of forecast. In SANN, a series of ANNs are connected sequentially„ each of them taking forecast value from an immediate preceding network as input. The output of each network modifies itself by adding an expected value of error so that residual variance of the forecast series is minimized. The efficacy of the developed model has been tested through a real case study for the data on hourly river flow forecasting for Kolar River, India. The binary-coded genetic algorithm is used to establish the weights among the neurons because of the dynamic nature of input layer in SANN model. The main objective function of the proposed model is to minimize the root mean square error. Our results demonstrate that the SANN is capable of providing accurate forecasts up to 8 hours ahead. The SANN model tends to preserve the performance at higher lead times compared to both ANN1 and ANN2 models.