The electricity industry has witnessed increasing challenges in power system operation and rapid developments of artificial intelligence technologies in the last decades. In this context, studying the approach of security-constrained unit commitment (SCUC) decision-making with high adaptability and precision is of great importance. This paper proposes an improved data-driven deep learning (DL) approach, following the sample coding and Sequence to Sequence (Seq2Seq) technique. First, an encoding and decoding strategy is utilized for high-dimensional sample matrix dimension compression. A DL SCUC decision model based on a Seq2Seq network with gated recurrent units as neurons is then constructed, and the mapping between load and unit on/off scheme is established through massive data from historical scheduling. Numerical simulation results based on the IEEE 118-bus test system demonstrate the correctness and effectiveness of the proposed approach.