Multivariate and hybrid data-driven models to predict thermoelectric power plants fuel consumption.

Ensuring the reliable operation of diesel/Heavy fuel oil (HFO) engines in Thermoelectric Power Plants (TPPs) requires constant monitoring and control of operational parameters. Time series forecasting methods can predict physical system characteristics based on historical data, which can help regulate engine operational parameters. Despite the widespread use of these techniques in coal and natural gas powered TPPs, further examination of their application in large diesel/HFO engines in Brazilian TPPs operating under dispatch conditions is needed. This work investigated the fuel consumption generalization capacity of linear, nonlinear, and hybrid prediction models, both considering univariate and multivariate approaches, related to a TPP engine-driven generator in Pernambuco, Brazil. For multivariate model development, exogenous variables were selected based on change of causality overtime analysis, which sought not only to infer the relationship between signals but also to identify the influence regime of each causality during operations. The feature selection step initially identified a performance improvement when eight additional features were used. However, each feature causality throughout the operations indicated that only four signals helped significantly predict consumption in different ways. The exogenous variables introduction to ARIMA and NAR models resulted in significant improvements only to the nonlinear approach, NARX, which could recognize operational disturbances more quickly when compared to other analyzed models. In addition, the application of additive and multivariate hybrid models provided the ability to detect more complex variations and, at the same time, maintain model stability during full-load operation, benefiting from linear and nonlinear characteristics capturing ability related to the combined model. • Engines' FOC was analyzed based on univariate and multivariate prediction models. • Linear, nonlinear, and hybrid prediction models were used. • Exogenous variables were selected based on change of causality overtime analysis. • Exogenous variable's introduction to NAR resulted in significant improvements. • Hybrid prediction models detected complex variations and maintained model stability. [ABSTRACT FROM AUTHOR]