Abstract Micro-grid systems (MGS) are increasingly investigated for green and energy efficient buildings in order to reduce energy consumption while maintaining occupants’ comfort. It includes renewable energy sources for power production, storage devices for storing power excess, and control strategies for orchestrating all components and improving the system's efficiency. In fact, MGS can be seen as complex systems composed of different heterogeneous entities that interact dynamically and in collective manner in order to balance between energy efficiency and occupants’ comfort. However, the uncertainty and intermittency of energy production and consumption requires the development of real-time forecasting methods and predictive control strategies. The State-of-Charge (SoC) of batteries is one of the main parameters used in MGS predictive control algorithms. It indicates how much energy is stored and how long MGS can be relying on deployed storage devices. Several methods have been developed for SoC estimation, but little work, however, has been dedicated for SoC forecasting in MGS. In this paper, we focus on advancing MGS predictive control through near real-time embedded forecasting of batteries SoC. In fact, we have deployed, on two platforms, two forecasting methods, Long Short-Term Memory (LSTM) and Auto Regressive Integrated Moving Average (ARIMA). Their accuracy and performance have been evaluated in both classical batch mode and streaming mode. Extensive experiments have been conducted for different forecasting horizons and results are presented using two main metrics, the accuracy and the computational time. Obtained results show that LSTM outperforms ARIMA for real-time forecasting, it has the better tradeoff in terms of forecasting accuracy and performance.

中文翻译：

---

微电网系统中嵌入式实时电池充电状态预测

摘要 微电网系统 (MGS) 越来越多地被研究用于绿色和节能建筑，以减少能源消耗，同时保持居住者的舒适度。它包括用于发电的可再生能源、用于存储多余电力的存储设备以及用于协调所有组件和提高系统效率的控制策略。事实上，MGS 可以被看作是由不同的异构实体组成的复杂系统，这些实体以动态和集体的方式相互作用，以平衡能源效率和居住者的舒适度。然而，能源生产和消费的不确定性和间歇性需要开发实时预测方法和预测控制策略。电池的荷电状态 (SoC) 是 MGS 预测控制算法中使用的主要参数之一。它表明存储了多少能量以及 MGS 可以依赖部署的存储设备多长时间。已经开发了几种用于 SoC 估计的方法，但是很少有工作专门用于 MGS 中的 SoC 预测。在本文中，我们专注于通过电池 SoC 的近实时嵌入式预测来推进 MGS 预测控制。事实上，我们已经在两个平台上部署了两种预测方法，即长短期记忆 (LSTM) 和自回归综合移动平均 (ARIMA)。它们的准确性和性能已经在经典批处理模式和流模式下进行了评估。已经针对不同的预测范围进行了广泛的实验，结果使用两个主要指标呈现，精度和计算时间。获得的结果表明，LSTM 在实时预测方面优于 ARIMA，它在预测精度和性能方面具有更好的权衡。