Neural Network Modelling of Magnetic Effects in Solid-State Transformer Performance Under Variable Magnetic Operating Conditions

Abstract

Solid State transformers (SSTs) are pivotal for renewable energy integration and smart grids, yet their performance is critically limited by non-linear magnetic effects such as core losses, flux- density saturation, and magnetising-inductance drift, that conventional analytical methods and prior artificial neural network (ANN) studies have struggled to model simultaneously across the electrical, thermal and magnetic domains. This study develops and evaluates a unified multi-output ANN framework that predicts six SST performance indicators which are efficiency, power density, total harmonic distortion, temperature rise, core loss and magnetising inductance from six operating variables (switching frequency, duty cycle, load current, ambient temperature, input voltage and maximum flux density). The synthetic dataset was expanded to 2000 operating conditions generated from physically grounded equations, the modified Steinmetz core-loss model and a flux-density-dependent inductance roll-off model, with heteroscedastic noise added to emulate measurement variability and an explicit first order thermal-coupling term. A baseline model that faithfully reproduces the original two-hidden layer, Z-score normalised architecture on a 500-sample subset is reported alongside an improved model that combines a deeper feed-forward architecture, log-scale normalisation of skewed outputs, a physics-informed auxiliary input derived from the analytical Steinmetz equation, and oversampling of the high-flux-density on the loss-cliff region. On a held-out test set, the improved model achieves R2 = 0.991 (efficiency), 0.996 (power density), 0.990 (THD), 0.995 (temperature rise), 0.993 (core loss) and 0.995 (magnetising inductance), resolving the core-loss and inductance R2 inconsistencies. A single slice sweep around the operating median identifies a favourable magnetic-flux-density region ( ≈ 1.0 - 1.2T) in which efficiency remains above the exponent. An illustrative comparison with a nano-crystalline core option is included to broaden material scope. Because the dataset remains synthetic, all quantitative operating recommendations are explicitly framed ANN-predicted trends requiring experimental confirmation, and a concrete roadmap toward physics informed neural networks (PINNs) and laboratory validation is proposed.