Tie-line scheduling in an interconnected power grid is difficult to perform optimization by establishing a unified physical model due to the different scheduling methods in different dispatching centers. It may cause repeated coordination and iteration in multi-level dispatching centers. Therefore, a data-driven tie-line scheduling method with self-learning ability is proposed. First, perform cluster pre-processing of historical dispatching data with K-means algorithm; secondly, establish a deep learning model of tie-line scheduling based on long short-term memory (LSTM), and build a mapping model among system load, clean energy output and tie-line schedule through historical data training; After that, the all above process is used as a foundation to do the tie-line scheduling; finally, continuously revise the model by accumulating historical data, so that it has the ability of self-evolution and self-learning. We make a case analysis based on the actual power grid data, and the calculation results show the effectiveness of the proposed method.