Power quality validation in micro off-grid daily load using modular differential, LSTM deep, and probability statistics models processing NWP-data

Load corrections with respect to power quality (PQ) after the first pre-estimate of Renewable Energy (RE) power consumption must ensure system-tolerant performance without malfunctions. First, acceptable daily load sequences for the attached equipment are combined and determined according to the RE potential and charge states in accommodation to user needs and normal operation. The main motivation is a consequent day-to-day verification of algorithmically scheduled power consumption tasks in the proposed two-stage optimisation according to the system resources and user needs. Statistical artificial intelligence (AI) is employed, as local atmospheric turbulences with terrain obstacles and unexpected user activity result in various operational states in real microsystems. A new unconventional neurocomputing strategy, called Differential Learning (DfL), was applied in the modelling and prediction of the high dynamical PQ parameters in an experimental RE based system according to input-output training data, without an exact specification of its behaviour. The DfL models were compared with recent deep and machine learning techniques. Prediction models were formed after an initial detection of adequate daily training intervals. The AI models are finally tested to process the complete 24-hour forecast series of related input variables used in learning, to estimate the PQ target output at the corresponding times.

在完成可再生能源（Renewable Energy，RE）发电量初步预估后，针对电能质量（Power Quality，PQ）的负荷校正工作需保障系统具备耐受性能且无故障发生。首先，需结合可再生能源潜力与设备荷电状态，适配用户需求与系统正常运行要求，组合并确定附属设备的可接受每日负荷序列。本研究的核心动机在于，基于系统资源与用户需求，对所提出的两阶段优化方案中算法规划的用电任务开展逐日验证。由于实际微系统中存在局地大气湍流、地形障碍以及用户突发行为，会导致多样的运行状态，因此本研究采用了统计人工智能（Statistical Artificial Intelligence，AI）技术。本研究提出了一种名为差分学习（Differential Learning，DfL）的新型非传统神经计算策略，基于输入-输出训练数据，在无需精确指定系统运行行为的前提下，将其应用于实验型可再生能源系统中高动态电能质量参数的建模与预测。随后将差分学习模型与当前主流的深度学习及机器学习技术进行了对比。预测模型的构建基于对合适每日训练时段的初步识别。最终，将对所构建的人工智能模型进行测试，使其处理学习阶段所用的完整24小时相关输入变量预测序列，以估算对应时刻的电能质量目标输出值。