This paper presents a synergistic integration of machine learning (ML) and mathematical optimization for unit commitment (UC) in power systems. UC is a critical problem in power system operations, involving the determination of unit commitment statuses and generation levels to meet system demand while minimizing operating costs. Traditional mathematical optimization methods, such as Mixed-Integer Linear Programming (MILP), are computationally intensive and face challenges due to increasing intermittent renewables and intra-hour net load variability. ML offers a promising alternative, but directly learning good solutions is difficult due to the combinatorial nature of UC. The paper proposes a method that integrates ML within the Surrogate Lagrangian Relaxation (SLR) framework to learn "good enough" subproblem solutions for deterministic hourly UC. SLR decomposes the problem into subproblems, which are easier to solve, and uses multipliers to coordinate the solutions. To improve learning, the paper introduces dimensionality reduction by aggregating multipliers, novel multiplier distributions based on "jumps" in binary decisions, and innovative loss functions to enhance prediction quality. Ordinal Optimization (OO) and branch-and-cut (B&C) are used as backups for unfamiliar cases. Online self-learning is seamlessly integrated with offline learning to exploit solutions from daily operations. Results on the IEEE 118-bus and Polish 2383-bus systems demonstrate that continual learning improves the subproblem-solving process while maintaining near-optimality of overall solutions. The method opens a new direction for solving complex UC problems by combining ML with mathematical optimization. The paper also discusses the formulation of UC, the SLR framework, and the integration of ML for subproblem solutions. It highlights the challenges of learning good-enough solutions, the design of loss functions, and the use of OO and B&C as backups. The method is tested on various examples, showing its effectiveness in solving both small and large UC problems. The results indicate that the proposed method can achieve near-optimal solutions with significantly reduced computational time compared to traditional methods.This paper presents a synergistic integration of machine learning (ML) and mathematical optimization for unit commitment (UC) in power systems. UC is a critical problem in power system operations, involving the determination of unit commitment statuses and generation levels to meet system demand while minimizing operating costs. Traditional mathematical optimization methods, such as Mixed-Integer Linear Programming (MILP), are computationally intensive and face challenges due to increasing intermittent renewables and intra-hour net load variability. ML offers a promising alternative, but directly learning good solutions is difficult due to the combinatorial nature of UC. The paper proposes a method that integrates ML within the Surrogate Lagrangian Relaxation (SLR) framework to learn "good enough" subproblem solutions for deterministic hourly UC. SLR decomposes the problem into subproblems, which are easier to solve, and uses multipliers to coordinate the solutions. To improve learning, the paper introduces dimensionality reduction by aggregating multipliers, novel multiplier distributions based on "jumps" in binary decisions, and innovative loss functions to enhance prediction quality. Ordinal Optimization (OO) and branch-and-cut (B&C) are used as backups for unfamiliar cases. Online self-learning is seamlessly integrated with offline learning to exploit solutions from daily operations. Results on the IEEE 118-bus and Polish 2383-bus systems demonstrate that continual learning improves the subproblem-solving process while maintaining near-optimality of overall solutions. The method opens a new direction for solving complex UC problems by combining ML with mathematical optimization. The paper also discusses the formulation of UC, the SLR framework, and the integration of ML for subproblem solutions. It highlights the challenges of learning good-enough solutions, the design of loss functions, and the use of OO and B&C as backups. The method is tested on various examples, showing its effectiveness in solving both small and large UC problems. The results indicate that the proposed method can achieve near-optimal solutions with significantly reduced computational time compared to traditional methods.