Multistate Prediction for In-Service Gas Turbine via Adaptive Diffusion Graph Network

Abstract

State estimation is integral to the continuous monitoring and management of the energy consumption in gas turbines. Recent research about multistate prediction adopts separate predictions and uses few connections between signals. The complex physical mechanism of the gas turbine makes it difficult to extract and obtain the spatial information between states as prior knowledge, and the spatial information has irregularities as the operating conditions change. To overcome these limitations, a graph structure with a fusion strategy was adopted to solve the irregularity of multisource sensor data, and then the adaptive adjacency-matrix diffusion graph network (AdapDGN) was proposed. First, an adaptive adjacency matrix module learned through a query and key and based only on the input data can capture the temporal trends of spatial information. Second, the diffusion graph recurrent module can remember the history sequences. The two aforementioned modules are combined in a unified framework, and the entire framework is learned end to end. The performance of AdapDGN for multistate prediction is demonstrated on a real-world sensor network of a gas turbine control system under variable working conditions. The results show a 2% improvement in prediction accuracy of the fusion strategy to the separate prediction, and the learned connections between signals under diverse working conditions are also visualized.