Abstract

Today, in many complex real-world systems, physics-based simulation models often provide sufficient precision but are computationally intensive. Machine learning surrogates, once trained, can achieve simulations by orders of magnitude faster than their original physical model without sacrificing much accuracy. In this paper, we present a surrogate model in form of a neural network that is fitted to a set of different time series. The time series data are generated partly by a physical model and partly by measurement. This is because a physical model is available only for a part of the entire state space that is to be modeled. This method is used to predict the flue gas temperature at the output of the evaporator in the heat recovery steam generator of a combined cycle power plant. For simulation we use a specialized in house tool for transient power plant processes, called 'Dynaplant'. The generated surrogate model is fast and captures the major dynamics. Consequently, the model can be used in applications where fast evaluation is required, e.g., in parallel to operation. One form of such usage is virtual sensors, whereby, physical detectors can be omitted, and thus costs are reduced. With this, we demonstrate a method that beneficially merges physical insight from simulation with reallife data and machine learning. Our findings are of interest to applications where either simulated or measured time series data or both of different operating points are available and a fast simulation model is required.