EVALUATION OF A WATER-LEVEL PREDICTION MODEL THAT EMPLOYS SUPPORT VECTOR REGRESSION IN WATER & DRAINAGE MANAGEMENT

Abstract

 In general, a deep-leaning neural network (DNN) poorly predicts without a large amount of data although DNN has been testedly implemented as a quick predictable model into pumping and drainage systems in lowland for management efficiency. Using support vector regression (SVR), we developed a water-level prediction model that can perform good predictions even with a small amount of data. We employed DNN to compare with SVR for an accuracy performance based on data size. The input data are the rainfall and water level, collected in two different lowland areas. The datasets are separated into the short-term-period data and the long-term-period data based on the number of seasonal variations. For the test of the short-term period, SVR predictions upto 6 h lead time were better than those of DNN by 6 to 28% improvement. For the long-term period test, both models had similar performannces in prediction accuracy. In addition, the accuracy of SVR was poor when quick changes of water level were simulated.