Predicting Eddy current losses in large generator rotor using data-driven short connection deep Gaussian process

In response to the increased prediction errors observed when applying deep Gaussian process (DGP) models to the eddy current loss data of large generator rotors, an explainable artificial intelligence(AI)-based DGP prediction method incorporating short connections design is proposed. This method reparameterizes multivariate normal random variables between layers and concatenates processed input-layer data to construct new inter-layer inputs for the subsequent layer, thereby establishing a short-connected DGP model. This approach effectively mitigates the accumulation of prediction errors due to increased model depth, significantly enhancing the model's adaptability and predictive accuracy in complex data structures. Experimental results demonstrate that the short-connected DGP model achieved exceptional predictive performance for eddy current loss estimation, as evidenced by a coefficient of determination of 0.9886 and a mean squared error of 0.0026. Furthermore, as the number of layers increases, the short-connected DGP model outperforms both existing DGP models and alternative approaches in prediction accuracy. The proposed AI method not only robustly supports high-precision prediction of eddy current losses in the rotor slot filling components of large generators, but also serves as an efficient and reliable modeling tool for their design and optimization, thereby possessing significant theoretical and practical value.