Real-Time Operational Strategy for Ice Thermal Storage District Cooling Systems via Model-Free Safe Deep Reinforcement Learning

District cooling system (DCS) offers centralized cooling services for building groups that significantly reduce energy consumption and carbon emissions. However, existing model-based methods struggle with the large-scale, stochastic, and nonlinear characteristics of the DCS optimization problem under strict time limits. To fill this research gap, this paper proposes a model-free safe deep reinforcement learning (MFS-DRL) algorithm for the DCS to deliver real-time scheduling strategies of multiple chillers and ice thermal storage units under varying operating conditions. The problem is modeled as a constrained Markov decision process (CMDP) to systematically decouple the objective from the constraint, therefore overcoming the confusion between reward and penalty caused by manual penalty coefficients in traditional DRL methods. Moreover, prior expert knowledge is acquired through the interior point optimizer (IPOPT) solver, and the agent is pre-trained using imitation learning (IL) to mitigate exploration disorder and local convergence in the early training stage. Ultimately, the Lagrangian relaxation approach is applied to modify the optimality criteria and gradient update rules of the classic deep deterministic policy gradient (DDPG) algorithm, thereby achieving a trade-off between feasibility and optimality. Numerical experiments on 12-node and 619-node DCSs demonstrate that the proposed method outperforms representative model-based and data-driven baselines, delivering high-quality solutions within only milliseconds and maintaining robustness across hyperparameter variations.