Neural network extended state-observer for energy system monitoring

Abstract

Due to the complexity of industrial energy systems such as the thermal power plants, renewable plants and batteries, energy system monitoring is gaining more and more attentions. Energy system monitoring methods give the estimations of internal process variables or parameters that can be adopted as the possible indicators for health conditions, which not only founds the basis for diagnosis and prognostics but also strengthens the situation awareness of system operators. In this paper, a neural network extended state-observer (NN-ESO) is newly proposed for general energy systems with measurements of key process variables, where the neural network adopted is the multi-layer perception (MLP). The advanced characteristics of this NN-ESO lies in the following three aspects: (1) No prior knowledge about the process dynamics is used in the observer design. (2) The globally bounded estimations of the exterior disturbances and their changing rates in the measurement channels can be provided. (3) The MLP in the NN-ESO is trained and deployed online without any offline training samples. By adopting the estimations of the exterior disturbances and their varying rates as the health indicators, the NN-ESO is applied for the operation monitoring of a modular high temperature gas-cooled reactor which is a typical small modular nuclear reactor. Simulation results show not only the satisfactory estimation performance but also the feasibility of applying NN-ESO for system monitoring.