Multi-Condition Magnetic Core Loss Prediction and Magnetic Component Performance Optimization Based on Improved Deep Forest

In recent years, power electronic technology has been widely applied in energy conversion in the fields of renewable energy and information and communication technology. Magnetic components play a crucial role in voltage stabilization, conversion, and filtering, conventional loss models struggle to accurately characterize the nonlinear loss characteristics under multi-physics coupling and complex operating conditions. To achieve high efficiency and high power density designs, it is essential to study core loss characteristics and optimize operating parameters. This paper proposes a core loss prediction model based on an improved deep forest algorithm and information entropy-enhanced genetic algorithm. First, a multi-factor analysis of variance (ANOVA) and Bayesian inference are employed to explore the effects of temperature, material, and excitation waveform on core losses, as well as their optimal combinations. Subsequently, a temperature correction factor is introduced into the conventional Steinmetz equation, and the ridge regression regularization technique is applied to improve the goodness-of-fit from 0.954 to 0.986. Next, the deep forest algorithm is employed to perform multi-granularity scanning for feature extraction from the samples, which are then fed into an enhanced cascade forest (with two random forests replaced by XGBoost and LightGBM), key parameters (maximum tree depth and number of trees) are optimized via grid search, constructing a highly accurate and robust core loss prediction model. The model's superiority is validated from multiple perspectives, accompanied by feature importance analysis using SHAP values. Finally, an information entropy and game theory-based genetic algorithm is utilized to optimize the operating conditions of magnetic components, yielding an optimal combination of low losses and high magnetic energy transfer efficiency.