Improving ARIMA Forecasting Accuracy Using Decomposed Signal on PH and Turbidity at SCADA Based Water Treatment

In industrial plants, accurate forecasting is critical for decision making. Autoregressive Integrated Moving Average (ARIMA) is a statistical analysis model used widely in time series forecasting. A suitable forecasting methodology must accurately predict future values. In the testing or validation process, the model should relatively follow the pattern of the actual signal. Most studies about ARIMA use directly observed signals in modeling and forecasting. The lack of this method, the predicted signal produces a straight line instead of following the actual signal when the time series data does not have strong seasonality. In this paper, we propose a customized forecasting methodology. First, the observed signal is decomposed into trend, seasonal, and residual component. Then decomposed components are modeled and forecasted independently. Finally, the forecasted components are recomposed to achieve the forecasted observed signal. In this study's experiment, the proposed method can reduce MSE of turbidity forecast 90.021% lower than the direct forecasting method. Meanwhile, the MSE reduction of pH forecast reaches 97.062% lower than the direct forecasting method. The average MSE reduction reaches 42.597%.