Purpose: To select a predictive model for daily output without being affected by specific patterns such as autocorrelation
Methods: Appropriate predictive models were selected by evaluating the flow of data that were experienced in historical observations. We used the autoregressive integrated moving average (ARIMA) and nonlinear autoregressive (NAR) neural network algorithms to predict time series data. To verify the autocorrelation for a given time series data, we carried out the autocorrelation function (ACF) and partial autocorrelation function (PACF) analyses. The ACF and PACF plots represent the visual information to determine the appropriate ARIMA model provide the necessary guidance for optimizing a number of hidden layers and neurons in the NAR. When quality assurance (QA) data are autocorrelated, we select an appropriate model of the ARIMA and NAR algorithms and compare the actual measured value to the value predicted by the selected models for the sampled daily output.
Results: Considering the mean absolute percentage error as the evaluation criteria, ARIMA model is 0.2651 and NAR model is 0.1385, respectively. Our findings that the ARIMA model can be used quickly and relatively easily without complex computational approach. We expected ARIMA model to have a similar effect as the NAR model when excluding autocorrelation structures and also recommended that they be used to build predictive models for short periods of time when QA data is less likely to be distorted. Our other finding in autocorrelated time series, the NAR outperformed the ARIMA developed on the basis of autocorrelation and that there has real clinical utilization as a prediction model.
Conclusion: We found that autocorrelation can have a significant impact on the accuracy of machine behavior predicted by that data obtained from the historical observations. It is therefore essential to analyze specific patterns and autocorrelated structures in the data for designing predictive monitoring.