As one of the techniques for anomaly detection and future forecasts utilizing time-series operating data, there have been a growing number of applications of long short-term memory (LSTM), one form of deep learning. LSTM is expected to demonstrate high prediction accuracy when actual operating conditions are close to those simulated in training. If they are not, however, there is a possibility that the prediction accuracy could decrease significantly. Meanwhile, since there has been no means to know whether the predictive results are reliable in conventional LSTM, it has not been applied much to fields where a high level of reliability would be required such as plant operation control.
This report will describe the predictive results produced by a deep learning model utilizing LSTM developed jointly with Kyushu University, as well as a technique to evaluate its reliability.
