Data mining techniques for predicting values of a faulty sensor at a refinery

Refineries rely heavily on the sensor data, decisions making in critical situation when a sensor failure happens is therefore essential. This paper proposes a method of predicting sensor values based on its historical data captured as time series. Main forecasting techniques such as linear regression, moving average, autoregressive integrated moving average; and exponential smoothing were used to predict the value of failed sensors. A comparison of the models based on their mean squared error is presented in order to simplify the selection of forecasting models. The proposed model assists engineers and experts at a refinery to make critical decision at critical moments.