To address the issue that traditional pure electromechanical transient simulation cannot accurately simulate the dynamic characteristics and strong nonlinear characteristics of power electronics devices, this paper establishes a hybrid electromechanical-electromagnetic transient simulation model for wind power grid-connected power systems. For the power system time series data, a long-short term memory (LSTM) network-based transient stability evaluation method is proposed, and a focal loss (FL) function is introduced to enhance the mining of the model for a few classes of samples. To improve the interpretability of the evaluation results, the Local Interpretable Model-agnostic Explanations (LIME) algorithm is used for predictive attribution analysis of example samples, which provides a reliable basis for online evaluation of grid transient stability.