Wake Properties Prediction by means of machine learning algorithms

风力发电已成为最成功的可再生能源，是欧洲仅次于天然气的第二大发电能力（175吉瓦）。截至2019年底，英国累计风电装机容量为23，513 MW，发电量超过50 TWh。尾流转向是一种可以提高风电场性能的控制策略。最近和正在进行的多项研究表明，由于尾流不足的减少，偏航迎风涡轮机可以为顺风涡轮机提供更多的功率。类似的效果可以通过使用逆风涡轮机的降额或通过偏航控制和降额的组合来潜在地实现。为了稳健地应用尾流转向或风力涡轮机降额，有必要知道涡轮机是否实际上正在经历任何有害的尾流效应以及尾流扰动的相对位置，以便可以执行适当的控制措施。结果是有希望的，然而它们需要额外的装置（激光雷达），如果将其安装在每个涡轮机上，则增加了成本和复杂性。其他问题是数据可用性相对较低，以及对数据处理的要求。此外，优选的是，可以利用标准传感器设置来执行预测。风切变和偏航失准等风入流条件已通过安装在涡轮机上的传感器（如应变计）成功表征，并辅以加速度计数据和SCADA信号（如转速和功率）。由于尾流不足引起的风廓线与风切变以及转子上的风速和湍流的其他变化有些相似，因此可以预期尾流效应将导致负载谐波的类似变化。机器学习已成功应用于风力涡轮机的各种建模和检测目的，迄今为止最大的焦点是功率输出预测和故障检测。对于尾流特性预测，问题类似于故障检测，因为需要识别和评估信号中的模式是否与正常行为一致。因此，机器学习有望成为基于负载信号实现自动尾流特性预测的最佳方法。该项目的主要目标是设计一种基于来自受尾流影响的涡轮机的测量信号的风力涡轮机尾流特性预测方法。尾流特性预测能力将能够控制单个风力涡轮机，以最大限度地提高风电场的性能。这是通过以下具体目标来实现的：-使用尾流条件和风力涡轮机动力学的数值模拟来设计基于机器学习的尾流特性预测算法-分析特征重要性，使得相关输入被理解和包括;另外，可以基于个体输入贡献来降低维度，以实现更准确的预测。此外，还应了解涡轮机群中可用的输入，并考虑这些配置，并根据所考虑的模型进行评估。通过与其他预测能力（如机舱安装激光雷达）进行比较，对预测方法进行现场验证-建议一种控制方案，以利用尾流特性预测信息来改善风电场性能

Wind power has become the most successful renewable energy and represents the second-largest power generation capacity (175 GW) in Europe after natural gas. At the end of 2019, there was an installed accumulative wind power capacity in the UK of 23,513 MW of wind power, generating more than 50 TWh of energy.Wake steering is a control strategy that can increase wind farm performance. There are multiple recent and ongoing studies, showing that yawing upwind turbines can deliver more power for the downwind turbines due to reduction of wake deficits. Similar effect can potentially be achieved using derating of upwind turbines, or by a combination of yaw control and derating. In order to robustly apply wake steering or wind turbine derating, it is necessary to know whether the turbine is actually experiencing any detrimental wake effects and the relative location of the wake disturbance so that appropriate control action can be carried out.There are multiple examples of how wake characteristics have been identified using nacelle-mounted lidars and other remote sensing technologies. The results are promising however they require additional devices (lidars) which increase cost and complexity if it were to be installed on each turbine. Further issues are relatively low data availability and the requirements for data processing. Furthermore it would be preferable for prediction can be carried out with standard sensor setups. Wind inflow conditions such as wind shear and yaw misalignment have been successfully characterized with turbine-mounted sensors, e.g. strain gages, supplemented with accelerometer data and SCADA signals such as rpm and power. Since the wind profiles caused by wake deficits represent somewhat similar phenomena to wind shear and other variations of wind speed and turbulence over the rotor, it is expected that wake effects will cause similar variations in the load harmonics.Machine learning has been successfully applied for various modelling and detection purposes in wind turbines, with the greatest focus by far being power output prediction and detection of faults. For wake properties prediction, the problem is similar to fault detection since patterns in the signal need to be identified and evaluated as to whether or not they are consistent with normal behavior. Machine Learning is, thus, expected to be the best approach for achieving automated wake properties prediction based on load signals.The main goal of the project is to devise a method for wind turbine wake properties prediction based on the use of measurement signals from the turbine affected by the wake. The wake properties prediction capability will enable controlling individual wind turbines to maximize wind farm performance. This is achieved through the following specific objectives:- Devise a Machine Learning-based wake properties prediction algorithm using numerical simulations of wake conditions and wind turbine dynamics- Analyze feature importance so that relevant inputs are understood and included; additionally dimensionality may be reduced based on individual input contribution to a more accurate prediction. Furthermore understand which inputs are available across turbine fleet, and reflect on these configuration and evaluate in relation to considered models. Carry out field validation of the prediction method by comparison with other prediction capabilities such as nacelle-mounted lidars- Suggest a control scheme to utilize the wake properties prediction information for improving wind farm performance