This paper proposes novel approaches to jointly design user equipment (UE) association and power control (PC) in a downlink user-centric cell-free massive multiple-input multiple-output (CFmMIMO) network. The goal is to maximize the sum spectral efficiency (SE) of UEs while considering constraints on per-AP transmit power, quality-of-service (QoS) rate requirements, maximum fronthaul signaling load, and maximum number of UEs served by each AP. For small-scale CFmMIMO systems, a successive convex approximation (SCA) method is proposed to solve the mixed-integer nonconvex optimization problem, and a deep learning-based method called JointCFNet is developed to reduce computational complexity. For large-scale CFmMIMO systems, an accelerated projected gradient (APG) technique is introduced to achieve low-complexity suboptimal solutions. Numerical results show that JointCFNet can achieve similar performance to SCA in small-scale systems with significantly reduced computational complexity, while the APG approach offers faster performance and higher SE compared to SCA in large-scale systems.This paper proposes novel approaches to jointly design user equipment (UE) association and power control (PC) in a downlink user-centric cell-free massive multiple-input multiple-output (CFmMIMO) network. The goal is to maximize the sum spectral efficiency (SE) of UEs while considering constraints on per-AP transmit power, quality-of-service (QoS) rate requirements, maximum fronthaul signaling load, and maximum number of UEs served by each AP. For small-scale CFmMIMO systems, a successive convex approximation (SCA) method is proposed to solve the mixed-integer nonconvex optimization problem, and a deep learning-based method called JointCFNet is developed to reduce computational complexity. For large-scale CFmMIMO systems, an accelerated projected gradient (APG) technique is introduced to achieve low-complexity suboptimal solutions. Numerical results show that JointCFNet can achieve similar performance to SCA in small-scale systems with significantly reduced computational complexity, while the APG approach offers faster performance and higher SE compared to SCA in large-scale systems.