Conditional transmission section limits (C-TSLs) depend on the actual operational state of the power grid, such as commitment status and reserve capacity, exacerbating the computational complexity of the unit commitment problem. To overcome this complexity, this paper proposes a data-model hybrid-driven approach for solving unit commitment problems. Furthermore, a unit commitment prediction algorithm is proposed to reduce the problem scale by fixing a subset of unit commitment variables, thereby eliminating inactive C-TSL intervals. Specifically, the proposed algorithm incorporates a convolutional attention module to perform deep feature extraction and enhancement on time series data, complemented by a multi-head cross-attention mechanism designed to synthesize temporal and non-temporal features. The introduction of attention mechanisms enhances the prediction performance of the proposed algorithm. Numerical results demonstrate that the proposed approach significantly reduces the number of variables and constraints in the unit commitment model, thereby expediting the solution process while maintaining high solution accuracy.