Modelling and adaptive control of industrial waste wood and residue boilers using neural networks

Abstract

Many real world processes are highly complex and not easily described by fundamental physical laws. This, along with the rising availability of data describing these processes, results in the modeling of these systems being increasingly performed using data-driven machine learning techniques. Additionally, the rise of computational processing power enables the use of larger datasets and more complex model architectures, leading to more powerful models in shorter training durations. In the process industry, these system models can be used in many different ways in order to raise efficiency, reduce undesirable process dynamics, increase process outputs, and decrease pollution. Possible applications include advanced control techniques such as model-based or adaptive control, predictive maintenance, process optimization, and fault detection. In this thesis, artificial neural networks are built to model industrial biomass combustion boilers using historical process data from two power plants. Specifically, two modeling approaches are tested, utilizing a static feedforward neural network on the one hand and a dynamic recurrent neural network on the other. Subsequently, the respective model performances are evaluated and compared to each other. After the evaluation of several model iterations the optimal model configuration is determined. Finally, different ways to further increase the modeling accuracy are discussed and implementation strategies into a conventional control setup are reviewed.