Controlling drum level is a major and crucial control objective in thermal power plant steam boilers. The drum level as a controlled variable is highly characterized by complex non-linear process dynamics as well as measurement noise and long-time delays. Developing a data-driven process model is particularly advantageous as it could be built from ongoing operational data. Such a model could be used to assist existing controllers by providing predictions regarding the drum level. The aim of this paper is to develop such a model and to propose a control architecture that can be easily integrated into existing control hardware. For that purpose, different neural networks are used, Multilayer Perceptron (MLP), Nonlinear Autoregressive Exogenous (NARX), and Long Short Term (LSTM) neural networks. LSTM and MLP were able to capture the dynamics of the process, but LSTM showed superior performance. The results demonstrate that the use of traditional machine learning criteria to evaluate a process model is not necessarily adequate. Using the model in an open-loop and a closed-loop simulation is more suitable to test its ability to capture the dynamics of the process. A novel architecture that integrates the process model within an existing closed-loop controller is proposed. The architecture uses adaptive weights to ensure that a good model is given more influence than a bad model on the controller’s output.