A data-driven approach is an alternative to extract general models for wind energy applications. A spatial sensitivity analysis is achieved using a probabilistic model to quantitatively identify the variability in performance due to individual parameters and visualize spatial distributions. Proper orthogonal decomposition results are combined with linear discriminant analysis under the clustering framework to present low-dimensional classifiers. Using the decomposition enables the system to be far away from ill-conditioned states. The optimal sensor locations are explicitly distributed in the transition region, where the velocity and Reynolds stresses relax toward a wake recovered state. With the optimal sensors, the cluster assignment and flow dynamics are obtained. There is an advantage in including more features in the reconstruction process to capture the slow and fast dynamics. Assessing the differences in the wake response and establishing the importance of spatial sensitivities are provided here for seeking accurate models. The bidirectional neural network is used to predict the fluctuating velocity of the considered sensors. The result of long–short term memory shows correlations of 92% between the real and predicted fluctuating velocities.

中文翻译：

---

聚类的稀疏传感器位置识别和基于深度学习的风轮机尾流预测

数据驱动方法是提取用于风能应用的通用模型的替代方法。使用概率模型来实现空间敏感性分析，以定量地识别由于各个参数导致的性能变化并可视化空间分布。在聚类框架下，将正确的正交分解结果与线性判别分析相结合，以提供低维分类器。使用分解可使系统远离病态。最佳传感器位置明确地分布在过渡区域中，在该过渡区域中，速度和雷诺应力向着唤醒恢复状态松弛。使用最佳传感器，可以获得群集分配和流动力学。在重建过程中包含更多功能以捕获慢速和快速动态效果是有利的。此处提供了评估唤醒响应的差异并确定空间敏感性的重要性，以寻求精确的模型。双向神经网络用于预测所考虑传感器的波动速度。长期记忆的结果显示，实际波动速度和预测波动速度之间的相关性为92％。