A GTU Model Based on a Recurrent Neural Network: Features of Elaboration and Application

Abstract

The application of artificial neural networks (ANN) based on various types of the recurrent architecture for simulation of the dynamics of development of complex process power-generating facilities, in particular gas-turbine units (GTUs), is examined. The challenges are outlined that arise in constructing classical mathematical models of such units, including a need to have reliable information on physical regularities, weight and size of equipment, correction factors, and to perform labor-intensive verification. The proposed approach eliminates these restrictions since the model uses only archived data on signals from the automatic process-control system. As a result, direct specification of physical parameters is no longer required. The author focuses on the application of ANNs to construct models that can reproduce nonlinear relationships among control actions, ambient conditions, and technical state of equipment. The proposed neural network models were trained and validated against actual operating data acquired at a 6-MW gas-turbine unit. Archival time series longer than 2 years were employed. These models have been demonstrated to offer good generalization service. A studied probabilistic extension of the model can form confidence intervals of predictions, thereby improving the reliability of diagnostics and detection of abnormal behavior. The architectural features of the applied solutions, including long short-term memory and sequential encoder-decoder models with an “attention” mechanism are analyzed. Besides numerical metrics, the model response is studied during individual characteristic regimes, including startups, shutdowns, and transients. The presented results confirm the practical significance of the proposed approach in solving the problems of developing digital twins, monitoring systems, and training simulators for operational personnel.