期刊名称：International Journal of Hybrid Information Technology
印刷版ISSN：1738-9968
出版年度：2016
卷号：9
期号：7
页码：89-106
DOI：10.14257/ijhit.2016.9.7.10
出版社：SERSC
摘要：Selangor is an important river basin in adjacent to the city of Kuala Lumpur, the federal capital of Malaysia and it supplies about 70% of the water required for domestic and industrial use for the city. Selangor river basin is presently regulated by two water supply dams, namely the Tinggi dam and the Selangor dam. Water is abstracted at an intake located 21 and 42 km downstream of the Tinggi and Selangor dam respectively. In the wet season, when unregulated flows downstream of the dams are sufficient for abstraction, no releases from the dams are required. However, releases are required in the dry season when flows downstream fall below the normal level. The present practice in dam operation is to use recession analysis in low flow forecasting during prolonged dry periods. Recession constants were derived using stream flow data and future flows were forecasted using the current flow and the recession constants assuming that there is no rain for the coming period where forecasts were made. Decisions were then made for releases from the dams. The disadvantage of recession analysis in forecasting low flow is that the forecast is not accurate if rain falls during the period and over release will occur. This study reports the use of Artificial Neural Network (ANN) models to forecast one and two time steps ahead river flows at the Rantau Panjang gauging station near the water supply intake for different travel times from the dams to the intake point to help in determining the regulating releases from the dams for more efficient reservoir operation. Two different ANN models, the Multi -Layer Perceptron (MLP) and the General Regression Neural Network (GRNN), were developed and their performances were compared. Endogenous and exogenous input variables such as stream flow and rainfall with various lags were used and compared for their ability to make future flow predictions. The input variables required are decided considering statistical properties of the recorded rainfall and flow such as cross-correlation between flow and rainfall, auto and partial autocorrelation of the flows which are best in representing the catchment response. Results show that both methods perform well in terms of R2 but GRNN models generally give lower RME and MAE values indicating their superiority compared to MLP models.
