 |
 หัวเรื่อง:ไม่มีชื่อไทย (ชื่ออังกฤษ : Application of Artificial Neural Networks for Reservoir Inflow Forecasting) ผู้เขียน: ดร.วราวุธ วุฒิวณิชย์, รองศาสตราจารย์ ,  Santi Thongpumnak, นายนิมิตร เฉิดฉันท์พิพัฒน์, ผู้ช่วยศาสตราจารย์ , อารียา ฤทธิมา, Nattaphun Kasempun สื่อสิ่งพิมพ์:pdf AbstractThis study showed the application of the Artificial Neural Networks in forecasting the reservoir inflow. Two cases were studied, (1) single reservoir inflow forecasting and (2) multi-reservoir inflow forecasting. The problems were formulated as daily, weekly and monthly inflow forecast. There were 4 types of model namely A, B, C and D according to the levels of data used as the input variables to the ANNs. Model A used all available data of that reservoir. Model B used the data having relatively high correlation with the reservoir inflow such as the first 3 lags of reservoir inflow, stream flow, rainfall and some meteorological data. Model C used only the first 3 lags of the reservoir inflow and stream flow data. Model D used the first 3 lags of reservoir inflow, stream flow and rainfall data. The 4 reservoirs namely Mun Bon, Lam Chae, Lam Phra Phloeng and Lam Takong reservoirs in Upper Mun basin, Nakhon Ratchasima province, were selected as the case study. Feed forwards back propagation algorithm was selected for the study. One to 3 hidden layers with different ANNs parameters were experimented. Two to 3 hidden layers were suitable for single reservoir problem while 1 to 2 hidden layers were suitable for multi-reservoir problem. Sigmoid transfer function was used in all the models. The initial weight, learning rate and momentum were in the ranges of 0.80-0.90. However they were not sensitive to prediction performance. For single reservoir forecasting, models A and B showed better performance (R2) than models C and D. The monthly model showed the better result than the weekly and daily models. For multi-reservoir forecasting, the performance of the 4 models was not different. Model C was recommended since it required less data. The training and testing performance of daily, weekly and monthly models were not much different in case of multi-reservoir. |
 |
|
|