CFD–ANN dataset and surrogate-optimisation code for a Savonius VAWT: power coefficient and torque ripple across the TSR–overlap-ratio design space

Description

This dataset contains 170 CFD-simulated operating points for a two-bladed Savonius vertical-axis wind turbine, sampled across the tip-speed-ratio and overlap-ratio design space using Latin hypercube sampling, together with the Python code used to train ensemble ANN surrogate models and perform constrained multi-objective optimisation. It supports the associated manuscript on CFD–ANN surrogate optimisation of a Savonius VAWT.

Steps to reproduce

Place DATASET.csv in the working directory (or /content/ on Google Colab). Open savonius_cfd_ann_pipeline.ipynb in Colab with a GPU runtime (T4). Install dependencies: numpy, pandas, matplotlib, scipy, scikit-learn, tensorflow, joblib. Set FAST_MODE = False for the full publication run (FAST_MODE = True runs a quick reduced-scale test). Run all cells. The pipeline performs the stratified hold-out split, 5-fold cross-validation architecture search, single-model and 10-member ensemble training, constrained SLSQP optimisation, constraint-sensitivity sweep, response-surface and Pareto-front generation, residual diagnostics, and stability analysis. Outputs (600 dpi figures and result CSVs) are written to the configured output directory. Random seeds are fixed (SEED = 42) for reproducibility, and cross-validation scalers are fitted inside each fold to avoid data leakage.

Institutions

• University of Greater Manchester

  England, Bolton

Categories

Licence