Regression Based Reactive Power Sizing under Grid Code Requirements

This paper frames reactive power device sizing at the point of common coupling as a supervised regression problem grounded in physically consistent data. Operating points are sampled from voltage dependent capability curves and validated through power flow analysis; among options that meet the grid code, the minimum loss shunt bank rating is taken as the target. Several regressors are benchmarked, with a Gaussian Process model delivering the best overall accuracy and Bagged Trees performing competitively. Errors remain well below the device step size, indicating practical suitability. A subsequent feature selection stage reduces the input set without degrading accuracy, yielding a compact model that relies only on routine plant signals and a requirement curve lookup. Future work will develop dynamic models to test performance against broader grid code requirements, including transient conditions.