Anomaly symptom detection using ensemble prediction based on new weighting method

In this paper, anomaly symptom detection using ensemble prediction based on newly developed weighting method is presented for time series. This weighting method is characterized that weights are determined in proportion to the indices defined by the prediction errors in a certain time period in the past. Next, alarm prediction based on this prediction method is proposed. Prediction accuracy is considered to be stable even under the fluctuation of plant behavior and it is shown that long margin time (the period between the time of anomaly symptom detection and actual anomaly event) was obtained by simulation. This means that operators have enough time to cope with anomaly events before their occurrence by the predicted alarms.