Abstract
A prediction method of plant status using neural network has been proposed for operation support systems in plants. The change of time series data from plant is complex because the data contain many frequency components. It is difficult to improve prediction performance of neural network by using the time series data directly. The proposed method resolves the data into some components based on their change frequencies and a prediction model using a neural network is generated for each components individually.
To evaluate the proposed method, two kinds of time series data are trained and predicted. One is the data containing several periodic components and a non-periodic component generated by an equation. The other is the simulation data of a fossile power plant under an abnormal condition. As a result, time averaged prediction error was reduced to 1/4-1/3 compared with a method using time series data directly. Thus, the proposed method is effective to predict plant status precisely.
