Model Predictive Control for Water Supply Networks with Storages

Balla, Krisztian Mark

4. term

Control and Automation, Master

2018

2018-06-06

144

Denne afhandling undersøger, hvordan pumpestationer i drikkevandssystemer med forhøjede reservoirer (vandtårne) kan styres, så driftsomkostningerne bliver så lave som muligt, samtidig med at forsyningssikkerhed og trykkrav overholdes. Første del beskriver drikkevandssystemer (WSS) matematisk for at udvikle en model til modelprædiktiv regulering (MPC). MPC er en reguleringsmetode, der forudsiger, hvordan systemet vil opføre sig, og vælger pumpehandlinger, der opfylder efterspørgslen til lavest mulig pris inden for givne grænser. Vi udvider en eksisterende modelleringsramme, som tidligere ikke tog højde for lagertanke i nettet, til en topologi med flere pumpestationer og forhøjede reservoirer. Til identifikation bruges neurale netværk (datadrevne modeller) til at beskrive vandstand i lagertanke og indløbstryk ved pumpestationer. Identifikationen udføres både på simulerede data fra EPANET (et simuleringsværktøj til vandnet) og på virkelige data fra et distributionsnet leveret af Verdo A/S. Anden del handler om at forfine NN-modellen og finde en passende modelkompleksitet til styring. En lineært tidsvarierende (LTV) model etableres ved at linearisere det neurale netværk i forskellige driftspunkter, baseret på netværkets flowbehov. Derefter designes en MPC, der antager, at energipriser og vandefterspørgsel kan prognosticeres. Sammenlignet med den nuværende tænd/sluk-styring i Verdos WSS viste MPC bedre performance ved at reducere omkostninger og udnytte lagringskapaciteten i de forhøjede reservoirer mere effektivt. Ud over simulationer er styringen implementeret og verificeret på et småskaligt WSS-testanlæg ved Aalborg Universitet.

This thesis studies how to operate pumps in drinking water supply systems with elevated reservoirs (water towers) in a cost‑optimal way, while meeting demand and pressure requirements. The first part develops a mathematical description of water supply systems (WSS) to build a model for Model Predictive Control (MPC). MPC is a control method that predicts future behavior and chooses pump actions that meet demand at the lowest cost within constraints. We extend an existing modeling framework—which previously did not include storage—to a topology with multiple pumping stations and elevated reservoirs. For system identification, neural networks (data‑driven models) are used to describe water levels in storage tanks and inlet pressure at pumping stations. The identification is carried out using both simulation data from EPANET (a water network simulation tool) and real data from a distribution grid provided by Verdo A/S. The second part focuses on refining the neural‑network model and selecting an appropriate complexity for control. A Linear Time‑Varying (LTV) model is obtained by linearizing the neural network at different operating points based on network flow demand. An MPC controller is then designed under the assumption that energy prices and flow demand can be forecast. Compared with the currently used ON/OFF control on Verdo’s WSS, the MPC algorithm showed better performance by reducing costs and making more effective use of the elevated reservoirs’ storage capacity. In addition to simulations, the controller was implemented and verified on a small‑scale WSS test setup at Aalborg University.

[This abstract was generated with the help of AI]

Model Predictive Control ; Neural Network ; Verdo ; Pressure management ; Water tank ; elevated reservoir ; linear ; non-linear ; MPC ; multiple inlet ; multiple tank ; storage ; water supply network ; water supply system ; pump ; pipe ; valve ; AAU ; Water Lab ; EPANET ; system identification ; radial basis function ; RBF ; optimal control ; cost optimization ; validation

Download
View record in AAU Student Projects