Reservoir water level forecasting using normalization and multiple regression

Many non-parametric techniques such as Neural Network (NN) are used to forecast current reservoir water level (RWLt). However, modelling using these techniques can be established without knowledge of the mathematical relationship between the inputs and the corresponding outputs. Another important issue to be considered which is related to forecasting is the preprocessing stage where most non-parametric techniques normalize data into discretized data. Data normalization can influence the the results of forecasting. This paper presents reservoir water level (RWL) forecasting using normalization and multiple regression. In this study, continuous data of rainfall (RF) and changes of reservoir water level (WC) are normalized using two different normalization methods, Min-Max and Z-Score techniques. Its comparative studies and forecasting process are carried out using multiple regression. Three input scenarios for multiple regression were designed which comprise of temporal patterns of WC and RF, in which the sliding window technique has been applied. The experimental results showed that the best input scenario for forecasting the RWLt employs both the RF and the WC, in which the best predictors are three day’s delay of WC and two days’ delay of RF. The findings also suggested that the performance of the RWL forecasting model using multiple regression was dependent on the normalization methods