Photovoltaic systems (PV) are having an increased importance in modern smart grids systems. Usually, in order to maximize the energy output of the PV arrays a maximum power point tracking (MPPT) algorithm is used. However, once deployed, weather conditions such as clouds can cause shades in the PV arrays affecting the dynamics of each panel differently. These conditions directly affect the available energy output of the arrays and in turn make the MPPT task extremely difficult. For these reasons, under partial shading conditions, it is necessary to have algorithms that are able to learn and adapt online to the changing state of the system. In this work we propose the use of deep reinforcement learning (DRL) techniques to address the MPPT problem of a PV array under partial shading conditions. We develop a model free RL algorithm to maximize the efficiency in MPPT control. The agent’s policy is parameterized by neural networks, which take the sensory information as input and directly output the control signal. Furthermore, a PV environment under shading conditions was developed in the open source OpenAI Gym platform and is made available in an open repository. Several tests are performed, using the developed simulated environment, to test the robustness of the proposed control strategies to different climate conditions. The obtained results show the feasibility of our proposal with a successful performance with fast responses and stable behaviors. The best results for the presented methodology show that the maximum operating power point achieved has a deviation less than 1% compared to the theoretical maximum power point.

光伏系统（PV）在现代智能电网系统中的重要性日益提高。通常，为了使PV阵列的能量输出最大化，使用了最大功率点跟踪（MPPT）算法。但是，一旦部署，天气条件（例如云）会在PV阵列中造成阴影，从而以不同方式影响每个面板的动态。这些条件直接影响阵列的可用能量输出，进而使MPPT任务极为困难。由于这些原因，在部分阴影条件下，必须具有能够在线学习并适应系统变化状态的算法。在这项工作中，我们建议使用深度强化学习（DRL）技术来解决部分阴影条件下PV阵列的MPPT问题。我们开发了无模型RL算法，以最大程度地提高MPPT控制的效率。代理的策略由神经网络参数化，该神经网络将感官信息作为输入并直接输出控制信号。此外，在开放源代码OpenAI Gym平台中开发了在遮光条件下的PV环境，并在开放存储库中提供了该环境。使用开发的模拟环境进行了一些测试，以测试所提出的控制策略对不同气候条件的鲁棒性。获得的结果显示了我们的建议的可行性，该建议具有快速响应和稳定行为的成功性能。所提出方法的最佳结果表明，与理论最大功率点相比，所达到的最大工作功率点的偏差小于1％。