This study presents a data-driven Deep Reinforcement Learning (DRL) framework to minimize power losses and mitigate voltage fluctuations in distribution networks with high photovoltaic (PV) penetration, without curtailing active power generation from PV systems. The voltage regulation problem is formulated as a Markov Decision Process (MDP) with a multi-objective reward function, and two advanced DRL algorithms—Soft Actor-Critic (SAC) and Trust Region Policy Optimization (TRPO)—are trained and evaluated on the IEEE 37-bus and 123-bus distribution test systems. Using real-world solar radiation, temperature, and residential load data for two representative days, the DRL agents learn optimal coordinated Volt-Var control policies for multiple smart PV inverters through continuous interaction with a power flow environment simulated in OpenDSS. Results demonstrate that both SAC and TRPO effectively regulate voltage profiles and coordinate reactive power support while preserving full PV active power output, achieving substantial reductions in network energy losses compared to no-control and conventional Volt-Var droop methods, all while maintaining voltages within standard limits. SAC exhibits faster convergence owing to higher sample efficiency, whereas TRPO produces marginally lower total losses. This work introduces a model-free, multi-tasking DRL approach that enables scalable and adaptive coordinated control across distributed inverters, offering a practical solution for modern active distribution grids facing dynamic PV and load variations.