リカレントニューラルネットワークによる多変数プログラム制御の管路系シミュレーションへの適用(流体工学,流体機械)

概要

論文の詳細を見る
The aim of this study is to control pressure and flow rate in a circulating pipeline system to conform to various target cyclic waveforms by operating a pump at rotational speed n and valve opening θ. In a first step, the pipeline system is operated at arbitrarily selected cyclic waveforms n and θ. Then, a recurrent neural-network (RNN) computer program learns the response characteristics of the system from the input and output data. At the next step, suitable values of n and θ are obtained by using the two-dimensional steepest descent method so that the sum of root-mean-square residuals between target values and the respective outputs of the RNN is minimized. Thereby, the RNN can generate values of n and 6 that can produce a pressure and flow rate that conform to the target waveforms. This method was investigated by numerical simulation of the pipeline system as a preliminary step toward applying it to a real experimental apparatus. By using the RNN which has learned the response characteristics obtained by measuring the controlled variable time-series of the system without investigating the individual characteristics of each device, it was possible to derive operating variables time-series that produced pressure and flow rate conforming to target waveforms with high accuracy.
2007-05-25