Demand Response of Interruptible Loads Using Deep Reinforcement Learning

Interruptible loads (ILs) are important flexible resources which can be controlled to shave peak load demand and balance power in real-time when power system is under stress or emergency. However, physics based control of ILs requires complete and explicit information of distribution network, which is difficult to gather in practice. To address the above issue, in this article, a data-driven method is proposed to control ILs automatically under time of use (ToU) tariff. To minimize long-time accumulative operation costs, ILs control strategy is developed as a Markov decision process (MDP), where states, actions and costs are carefully formulated. To overcome dimensionality curses, an adapted deep deterministic policy gradient (DDPG) reinforcement learning with relaxation-prediction-correction procedures is proposed to solve it. Extensive numerical tests show that the proposed approach can reduce both peak power of loads and operation costs while regulate voltages magnitudes to the rated range. Moreover, it can achieve (near) optimum, using only minor current slot information of distribution systems. Further, the computing speed is significantly faster than traditional multi-slot physical conic model.