In this paper, we propose a policy function approximation (PFA) algorithm using machine learning to effectively control photovoltaic (PV)-storage systems. The algorithm uses an offline policy planning stage and an online policy execution stage. In the planning stage, a suitable machine learning technique is used to generate models that map states (inputs) and decisions (outputs) using training data. The training dataset is generated by solving a deterministic smart home energy management problem using a suitable optimization technique [e.g., mathematical programming or dynamic programming (DP)]. In the execution stage, the model generated by the machine learning algorithm is then used to generate fast real-time decisions. Since the decisions can be made in real-time, the policy can rely on up-to-date information on PV output, electrical demand, and battery state of charge. Moreover, we can use PFA models over a long period of time (i.e., months) without having to update them but still obtain similar quality solutions. Our results show that the solutions from the PFAs are close to the best solutions obtained using DP and approximate DP, which have the drawback of requiring an optimization problem to be solved before the beginning of each day or as new information on demand or PV becomes available.

中文翻译：

---

光伏存储系统的能源管理：使用机器学习的策略近似

在本文中，我们提出了一种使用机器学习来有效控制光伏（PV）存储系统的策略函数近似（PFA）算法。该算法使用离线策略计划阶段和在线策略执行阶段。在计划阶段，将使用一种合适的机器学习技术来生成模型，这些模型使用训练数据来映射状态（输入）和决策（输出）。通过使用适当的优化技术[例如，数学编程或动态编程（DP）]解决确定性的智能家居能源管理问题来生成训练数据集。在执行阶段，由机器学习算法生成的模型随后用于生成快速的实时决策。由于决策可以实时做出，因此该政策可以依赖于有关光伏输出，电力需求，和电池的充电状态。而且，我们可以在很长一段时间（即数月）内使用PFA模型，而不必更新它们，但仍然可以获得类似的质量解决方案。我们的结果表明，来自PFA的解决方案接近于使用DP和近似DP获得的最佳解决方案，其缺点是需要在每天开始之前或随着新的按需信息或PV可用而解决优化问题。 。