The loss in magnetic core components is closely linked to their microstructure and operating conditions. Accurate prediction of this loss is crucial for improving the efficiency and power density of power electronics. At present, there are few universally applicable and highly accurate models, making it difficult for the industry to accurately assess magnetic core losses, which in turn affects the evaluation of power converter efficiency. To enhance the accuracy of magnetic core loss prediction models, this paper introduces a new model that combines the Northern Goshawk Optimization (NGO) algorithm with a Gated Recurrent Unit (GRU) network. Initially, historical data related to magnetic core loss are preprocessed, addressing missing values using interpolation and handling outliers using the Ryder’s Criterion ( \(3\sigma \) rule). During the optimization process, we introduce adaptive weighting factors to enhance search efficiency in the exploration phase of the Northern Goshawk algorithm and incorporate nonlinear convergence factors to balance the capabilities of global search and local development in the development phase. By constructing a GRU model and optimizing key hyperparameters of the network using the improved NGO algorithm, including the number of hidden layer units, learning rate, and training epochs. Finally, comparative analysis of the models shows that the NGO-GRU model outperforms others in terms of prediction accuracy and adaptability, highlighting its strong potential for application in magnetic core loss prediction.