Automated CFD Optimisation of a Small Hydro Turbine for Water Distribution Networks

Abstract

Leakages in water distribution networks is a global problem, with leakage rates up to 50% of the flow [1]. One way to find leakages is to adopt more sensors in the network. Such sensors need electric power that is commonly not distributed with the network. An option is to generate the electricity where it is needed, using small hydro turbines that extract energy from the flowing water and produces just enough electric power for the sensor at each site. The present work provides an automated optimisation procedure for the design of such turbines, and employs it to optimise a turbine with respect to efficiency and kinematic pressure drop at a desired operating point. The automated optimisation loop handles the entire procedure of geometry creation, meshing, simulation, post-processing and optimisation using open source software. The geometry generation and solution procedure is based on a the work by Bergqvist [2], who implemented a Ruby code to generate a parameterised turbine geometry, generated the mesh with cfMesh, and used OpenFOAM to simulate the flow. The present work extends the method by using Dakota to create an automated loop for optimisation, and applies it for water network turbines. The optimisation is done by varying the pitch, number of rotor and stator blades, as well as the hub radius. A steady-state frozen-rotor approach with multiple reference frames (MRF) is used in the simulations. The turbulence is modelled with the k −! SST turbulence model. After convergence studies of the pilot design, the simulations are carried out on unstructured grids with 2 − 3 million cells for the varying geometries. To reduce the computational burden that direct optimisation with CFD simulations implies, a surrogate modelling technique is implemented using the Efficient Global Optimisation (EGO) algorithm [3] [4]. For this, Kringing modelling is used to create a response surface, and division of rectangles (DIRECT) to optimise it and find infill points for further evaluation. 51 sample points, generated with latin hypercube sampling, are simulated before the first abstraction of the objective function is created. After the learning period, 61 additional design iterations are simulated before the algorithm reaches the stopping criteria. In between each of the 61 optimisation designs, a surrogate optimisation is carried out to select the next infill point. The completed optimisation provides a new design that performs at 4.4 percentage points higher efficiency and with 72% lower kinematic pressure drop. The new design features a bigger flow through area and a lower work output that is more adapted to the application with a reduction of 70%.