A polynomial neural network for river’s water-level prediction

Abstract

Flood risk prediction is based on collecting long time-series data using in situ sensors. As such, the consistency and quality of the hydrological flood models is subsequent to information gaps included in the aforementioned time series, which may cause issues that negatively affect the efficiency of the prediction. Herein, to cope with that problem, a novel polynomial neural network is developed, which utilizes Legendre polynomial activation functions and it is able to fill those gaps by predicting the river’s water level by intertwining observations coming from different areas of the river, and weather data. The network is trained by a modified evolutionary computation algorithm, which is based on the well-known artificial bee colony method. The experimental case study concerns the Kifisos river basin (Attica, Greece), where a number of sensors have been deployed to collect data. The network was compared to a feed-forward neural network in terms of the root mean square error performance index. Results indicated a superior performance of the proposed network obtaining predictions from 1 to 2 cm.