Abstract
一般而言動態系統大多以聯立微分方程組來求解該系統之動態。因此若要求解某輸入條件下系統動態,則需較長之時間來建立此系統聯立方程式,而且常需相當長之計算時間, 在考慮整體計算時間的需求下, 一些暫態模式常省略一般系統動態計算, 本研究以類神經網路來建立一般系統的動態模式, 如此可大量縮減計算時間。 本論文採用對角遞迴類神經網路,僅使用很少數目的類神經元即可達到模擬動態系統的目的, 並且此網路收歛快速, 可縮短訓練所需時間。本論文採用核研所之核三廠精緻模擬器以產生所需之訓練數據對。將一般系統分為高壓汽機, 汽水分離再熱器, 低壓汽機, 冷凝器, 低壓飼水加熱器,高壓飼水加熱器等次系統, 分別建立網路模式。俟所有之次系統網路模式依次建立後，加以傳遞聯結即可模擬計算核能電廠一般系統之動態反應。Generally, dynamic system's responses were described by thedifferential equations .Therefore, to model the system dynamicsof a complicated system, it requires rather long time to setup the system governing equations. In addition, the computationeffort is also very huge. In order to reduce the computingtime, the calaulation of dynamic response of balance of plantis usually neglected in the large system codes. This researchis to set up the dynamic model of balance of plant using theneural networks so that the computing time can be very short.In this research, the diagonal recurrent neural networks wereadopted, which need rather few neurons and also achieve goodperformance. In addition, it converged very fast and thenreduced the training time. The data used for training weregenerated by the compact simulator of Mannashan Nuclear PowerPlant which was developed by Institute of Nuclear EnergyResearch and Institute for Information Industry. The wholesystem was divided into six subsystems:high pressure turbine,moisture seperator and reheater, low pressure turbine,condenser, low pressure feedwater heater, and high pressurefeedwater heater. The neural network models were trained forsuch subsystems. When the training was finished, the neuralnetworks of subsystem were connected to simulate the dynamicbehavior of balance of plant.