Geïntegreerde niet-lineaire modelgebaseerde voorspellende regeling van een kleinschalig vierde generatie warmtenet

The residential heating sector accounts for approximately one-fifth of the European Union's (EU's) final energy use, making its decarbonisation essential for achieving the EU's climate goals. This necessitates the transition to a more efficient and renewable heat supply. In areas with a low heat demand density, individual heat pumps are suitable, while district heating (DH) networks will play a significant role in urban areas. This dissertation studies fourth generation DH (4GDH), which is characterised by lower network temperatures, a high share of renewable and residual energy sources, and (seasonal) thermal energy storage ((S)TES). 4GDH networks offer a significant amount of (energy) flexibility through (S)TES systems, network capacity, and the thermal inertia of the buildings. This flexibility can be utilised by a model predictive controller (MPC) to match heat supply and demand while minimising a cost function. MPC uses a mathematical model of the system (the controller model) and predictions of the system disturbances to determine the optimal set of control inputs. To fully exploit the flexibility sources (TES systems, network capacity, and thermal building inertia), their characteristics must be incorporated into the controller model. Moreover, DH networks are subject to non-linear physics which should be incorporated as well. These challenges lead to the main goal of this dissertation: to develop an integrated non-linear MPC for a 4GDH network. The term 'integrated' emphasises that the entire DH system (heat supply, distribution, and demand side) is accounted for with all its interactions, enabling the MPC to fully exploit the available flexibility. Additionally, this dissertation addresses the challenges posed by integer control inputs and STES. The first part of the dissertation focuses on developing an integrated (mixed-integer) non-linear MPC for the small-scale 4GDH network of De Schipjes. Firstly, an integrated non-linear MPC (referred to as MPC-NLP) is formulated using TACO, which only supports continuous optimisation variables. Since the De Schipjes DH system includes several binary control inputs, a post-processing step is developed. MPC-NLP is then tested and evaluated in a co-simulation environment against a conventional rule-based controller (RBC). The simulation results indicate that MPC-NLP outperforms RBC in terms of electrical energy use and thermal comfort, due to its ability to anticipate, the integrated approach, and the inclusion of important non-linearities and all flexibility sources in the controller model. Secondly, an integrated mixed-integer non-linear MPC (referred to as MPC-MINLP) is developed using the combinatorial integral approximation decomposition algorithm. This approach eliminates the need for a post-processing step and enhances model accuracy. Simulation results demonstrate that MPC-MINLP more accurately predicts the system behaviour of the emulator compared to MPC-NLP, but the performance (electrical energy use and thermal comfort) of both MPCs is similar. However, both MPCs assume perfect predictions of the system's disturbances and a perfect state observer. The latter immediately detects and accounts for mismatches between the MPC's predicted behaviour and the emulator's system behaviour. To address this, the dissertation also examines the impact of imperfect predictions and an imperfect state observer. This analysis quantifies the effect of imperfect predictions and/or an imperfect state observer on the MPC performance and provides a more realistic comparison between the different controllers (RBC, MPC-NLP, and MPC-MINLP). The simulation results reveal that the combined effect of prediction uncertainties and an imperfect state observer significantly impacts MPC performance, although MPC still outperforms RBC. Additionally, the results demonstrate that MPC-NLP and MPC-MINLP again perform similarly regarding electrical energy use and thermal comfort. The second part of the dissertation investigates the application of MPC to 4GDH systems with STES. Firstly, a fictitious solar-based DH network with a seasonal storage tank is described and modelled. Secondly, a shadow-cost-based MPC is developed to manage the long-term effects of STES without extending the MPC's prediction horizon (which could jeopardise real-time operation). The long-term MPC (LTMPC) is then tested and assessed using an open-loop optimisation setup. The results show that LTMPC outperforms a short-term MPC and approaches the performance of a full-year optimal control problem (the theoretical optimum, referred to as FY-OCP). Finally, a parameter study examines the impact of the tank size and the shadow-cost weighting factor on the performance of LTMPC. This study reveals that, while the tank size affects the optimal control behaviour of FY-OCP, the performance gap of LTMPC compared to FY-OCP remains consistent. The shadow-cost weighting factor has a minimal effect on the performance of LTMPC if the factor is kept constant. However, it is shown that a well-tuned, time-varying shadow-cost weighting factor can adapt the control behaviour of LTMPC, allowing it to closely replicate that of FY-OCP, thereby enhancing the overall performance of LTMPC.