Prediction Enhancement of Machine Learning Using Time Series Modeling in Gas Turbines

Abstract

Abstract Blade-path temperature can serve as a precursor of anomalies in combustion system and/or cooling system. Given observations from blade-path temperature sensors of a power plant, we consider prediction of the temperature for each sensor. The only extraneous predictor is the combustion turbine fuel flow, while measurements of other potential predictors are unavailable. Long-memory behavior and heterogeneous variance are observed from the residuals of the generalized additive model. Autoregressive Fractionally Integrated Moving Average (ARFIMA) and Generalized Autoregressive Conditional Heteroskedasticity (GARCH) models are employed to fit the residual process, which significantly improve the prediction. Rolling one-step-ahead forecast is studied for each of the sixteen univariate blade-path temperature sensors. Their conditional variances are also estimated. Numerical experiments are performed with manually generated perturbation to evaluate the specificity and sensitivity of the prediction. Abrupt changes in the temperature are considered in the numerical study with various jump sizes. We also consider slowly increasing trend in the blade-path temperature with different slopes. Our prediction is sensitive given reasonable signal-to-noise ratio. It also has a much lower false positive rate than the generalized additive model prediction from the combustion turbine fuel flow. Difference between the real-time forecast and observation can be deployed to test for anomalies.