Abstract: Due to the strong intermittency of renewable energy output and the multi-dimensional dynamic characteristics of terminal loads, the smart grid is facing the predicament of system stability. The stability diagnosis algorithm, as the "nerve center" for the safe operation of the smart grid, not only provides decision-making basis for risk feedforward control, but also becomes a key technical support for optimizing power flow distribution and tapping the potential of flexible resources. By integrating Recurrent Neural Network (RNN) and Bi-directional Long Short-Term Memory. The stability diagnosis model RNN-BiLSTM-CNN of the smart grid was constructed by using BiLSTM and Convolutional Neural Network (CNN). Process time series data using RNN to capture temporal dependencies; BiLSTM bidirectional encoding enhances the feature extraction ability and captures long-term and short-term dependencies. CNN automatically extracts local features through convolutional layers to enhance the ability of feature learning and prediction. The experimental results show that in 20 traversals, the accuracy rate, precision rate, recall rate and F1 score of RNN-BiLSTM-CNN for the stability diagnosis of smart grids reached up to 99.43%, 99.21%, 99.41% and 0.9859 respectively at the highest.