A 5-day wind speed & power forecasts using a layer recurrent neural network (LRNN)
A 5-day wind speed & power forecasts using a layer recurrent neural network (LRNN)
复制标题
DOI：
10.1016/j.seta.2013.12.001
复制
发表时间：
2014-06
影响因子：
8
通讯作者：
Zaccheus Olaofe Olaofe-Zaccheus-Olaofe-Olaofe-9364743
中科院分区：
文献类型：
--
作者：
Zaccheus Olaofe Olaofe-Zaccheus-Olaofe-Olaofe-9364743
以一台40 kW风力涡轮机为研究对象，采用分层递归神经网络作为预测器，对风力涡轮机的长期风速和功率输出进行了预测。预测模型采用了带有抽头延迟的Levenberg马夸特反向传播（BP）算法，以5分钟的步长预测A站5天前的风速和发电量。此外，BP算法被认为是预测站B在相同的塔高度使用10分钟的样本的风势。为了进行准确性比较，10分钟合成样本是从A站采样的5分钟测量值中产生的;风预测与5分钟预测进行了比较。为了准备预报模型，在两个风站的20 m塔高处获得了一个月的天气样本。第一天的数据用于训练模型，并在第二天开始预测，最长时间为5天。对于站点A和B，分别预测了30天期间使用采样5分钟测量的1322.61 kWh和使用采样10分钟测量的4485.56 kWh的可用总发电量。使用A站生成的合成样本，预测可用总发电量为1320.55 kWh。风预报显示，使用5分钟测量值和A站10分钟合成样本之间的偏差非常小。此外，对预测模型进行了评估，以测试LRNN在所选网络参数下的表现。一个新的天气样本是从一个20米高的塔远程站获得的，以测试预测模型的准确性。估计的误差被用来确定两个站的风预测与其可接受值或实际值的接近程度。使用独立样本的准确性测试结果与使用A站天气样本的验证结果密切相关。
This article presents the long term wind speed and power output of a 40 kW wind turbine based on a layer recurrent neural network as the predictor. The forecast model utilized the levenberg marquardt back propagation (BP) algorithm with a tap delay for prediction of the wind speed and power generation at 5-min steps of up to 5 days ahead at station A. In addition, the BP algorithm was considered for prediction of the wind potential at station B using 10-min samples at the same tower height. For accuracy comparisons, the 10-min synthetic samples were generated from the sampled 5-min measurements at station A; and the wind predictions were compared with the 5-min predictions. To prepare the forecast model, a one month weather samples were obtained at the 20 m tower height on both wind stations. The first day data was used to train the model and forecast began at the second day for maximum period of 5 days. A usable total electricity generation of 1322.61 kWh using the sampled 5-min measurements, and 4485.56 kWh using the sampled 10-min measurements were predicted for the period of 30 days for the stations A and B, respectively. Using the generated synthetic samples at station A, a usable total electricity generation of 1320.55 kWh was predicted. The wind forecast shows a very small deviation between the use of the 5-min measurements, and the 10-min synthetic samples at station A. Furthermore, the forecast model was assessed to test how well the LRNN performed with the selected network parameters. A new weather sample was obtained from a remote station at a 20 m tower height to test the forecast model accuracy. The estimated errors were used to determine the closeness of the wind predictions to its acceptable or actual value at both stations. Accuracy test results using independent samples show close relationship with the validation results using the weather samples at station A.