• Peer Review
  • Non-profit
  • Global Open Access
  • Green Channel for Rising Stars
Volume 10 Issue 1
Jan.  2023
Turn off MathJax
Article Contents

LI Xiuhao, LIU Huaixi, ZHANG Zhiyong, ZHANG Min, WU Di, MIAO Desheng. Very Short-Term Wind Direction Multistep Forecast Based on VMD-LSTM[J]. SOUTHERN ENERGY CONSTRUCTION, 2023, 10(1): 29-38. doi: 10.16516/j.gedi.issn2095-8676.2023.01.004
Citation: LI Xiuhao, LIU Huaixi, ZHANG Zhiyong, ZHANG Min, WU Di, MIAO Desheng. Very Short-Term Wind Direction Multistep Forecast Based on VMD-LSTM[J]. SOUTHERN ENERGY CONSTRUCTION, 2023, 10(1): 29-38. doi: 10.16516/j.gedi.issn2095-8676.2023.01.004

Very Short-Term Wind Direction Multistep Forecast Based on VMD-LSTM

doi: 10.16516/j.gedi.issn2095-8676.2023.01.004
  • Received Date: 2022-06-23
  • Rev Recd Date: 2022-08-02
  • Available Online: 2022-12-01
  • Publish Date: 2023-01-11
  •   Introduction  In order to accurately forecast the wind direction in the next 4 hours, a very short-term wind direction multistep forecast algorithm based on VMD-LSTM(Variational Mode Decomposition-Long Short-Term Memory) is proposed.   Method  Wind direction sequence was collected from 3 wind turbines of a wind farm of Mingyang Smart Energy Group for preprocessing and analysis. The correlation of wind direction in different periods was calculated using the autocorrelation function (ACF) to select the characteristic length of wind direction sequence. Based on variational mode decomposition (VMD), the wind direction sequence was decomposed into relatively intrinsic mode functions, the number of which was determined by minimum sample entropy. Models were build for each intrinsic mode function to make very short-term wind direction 24-step forecast. Finally, the wind direction sequence was reconstructed from the forecasted intrinsic mode functions.   Result  The results obtained demonstrate that the average MAE(Mean Absolute Error), RMSE(Root Mean Square Error) and MAPE(Mean Absolute Percentage Error) of the 24-step wind direction forecast based on VMD-LSTM in 4 quarters are 8.430°, 16.870° and 9.155, respectively. The algorithm performs better than other common data modeling methods regarding each error evaluation index at different time scales in each quarter.   Conclusion  The proposed algorithm can optimize the control yaw angle in the actual production of wind farms.
  • [1] 张东东, 崔新维. BP神经网络在风力发电机风向预测中的应用 [J]. 太阳能, 2015(3): 47-49. DOI:  10.3969/j.issn.1003-0417.2015.03.011.

    ZHANG D D, CUI X W. Application of BP neural network in wind direction forecast of wind turbine [J]. Solar Energy, 2015(3): 47-49. DOI:  10.3969/j.issn.1003-0417.2015.03.011.
    [2] CHENG W Y, LIU Y B, BOURGEOIS A J, et al. Short-term wind forecast of a data assimilation/weather forecasting system with wind turbine anemometer measurement assimilation [J]. Renewable Energy, 2017, 107: 340-351. DOI:  10.1016/j.renene.2017.02.014.
    [3] ERDEM E, SHI J. ARMA based approaches for forecasting the tuple of wind speed and direction [J]. Applied Energy, 2011, 88(4): 1405-1414. DOI:  10.1016/j.apenergy.2010.10.031.
    [4] KAVASSERI R G, SEETHARAMAN K. Day-ahead wind speed forecasting using f-ARIMA models [J]. Renewable Energy, 2009, 34(5): 1388-1393. DOI:  10.1016/j.renene.2008.09.006.
    [5] AMBACH D, SCHMID W. A new high-dimensional time series approach for wind speed, wind direction and air pressure forecasting [J]. Energy, 2017, 135(17): 833-850. DOI:  10.1016/j.energy.2017.06.137.
    [6] WANG S X, ZHANG N, WU L, et al. Wind speed forecasting based on the hybrid ensemble empirical mode decomposition and GA-BP neural network method [J]. Renewable Energy, 2016, 94: 629-636. DOI:  10.1016/j.renene.2016.03.103.
    [7] KHOSRAVI A, KOURY R N N, MACHADO L, et al. Prediction of wind speed and wind direction using artificial neural network, support vector regression and adaptive neuro-fuzzy inference system [J]. Sustainable Energy Technologies and Assessments, 2018, 25: 146-160. DOI:  10.1016/j.seta.2018.01.001.
    [8] 张亚超, 刘开培, 秦亮. 基于VMD-SE和机器学习算法的短期风电功率多层级综合预测模型 [J]. 电网技术, 2016, 40(5): 1334-1340. DOI:  10.13335/j.1000-3673.pst.2016.05.007.

    ZHANG Y C, LIU K P, QIN L. Short-term wind power multi-leveled combined forecasting model based on Variational Mode Decomposition-Sample entropy and machine learning algorithms [J]. Power System Technology, 2016, 40(5): 1334-1340. DOI:  10.13335/j.1000-3673.pst.2016.05.007.
    [9] 唐振浩, 赵赓楠, 曹生现, 等. 一种基于数据解析的混合风向预测算法 [J]. 太阳能学报, 2021, 42(9): 349-356. DOI:  10.19912/j.0254-0096.tynxb.2020-0119.

    TANG Z H, ZHAO G N, CAO S X, et al. A data analystic based hybrid wind direction prediction algorithm [J]. Acta Energiae Solaris Sinica, 2021, 42(9): 349-356. DOI:  10.19912/j.0254-0096.tynxb.2020-0119.
    [10] 林涛, 王建君, 张达. 基于VMD-BA-LSTM的短期风向预测研究 [J]. 高技术通讯, 2021, 31(6): 653-659. DOI:  10.3772/j.issn.1002-0470.2021.06.010.

    LIN T, WANG J J, ZHANG D. Short-term wind direction prediction research based on VMD-BA-LSTM [J]. Chinese High Technology Letters, 2021, 31(6): 653-659. DOI:  10.3772/j.issn.1002-0470.2021.06.010.
    [11] 向玲, 李京蓄, 王朋鹤, 等. 基于VMD-FIG和参数优化GRU的风速多步区间预测 [J]. 太阳能学报, 2021, 42(10): 237-242. DOI:  10.19912/j.0254-0096.tynxb.2019-1083.

    XIANG L, LI J X, WANG P H, et al. Wind speed multistep interval forecasting based on VMD-FIG and parameter-optimized GRU [J]. Acta Energiae Solaris Sinica, 2021, 42(10): 237-242. DOI:  10.19912/j.0254-0096.tynxb.2019-1083.
    [12] 吴迪, 刘怀西, 苗得胜. 尾流算法与风向变化对海上风机排布影响研究 [J]. 南方能源建设, 2019, 6(2): 54-58. DOI:  10.16516/j.gedi.issn2095-8676.2019.02.010.

    WU D, LIU H X, MIAO D S. Research on offshore wind farm units layout considering the algorithm of wake model and the change of wind direction [J]. Southern Energy Construction, 2019, 6(2): 54-58. DOI:  10.16516/j.gedi.issn2095-8676.2019.02.010.
    [13] DRAGOMIRETSKIY K, ZOSSO D. Variational mode decomposition [J]. IEEE Transactions on Signal Processing, 2014, 62(3): 531-544. DOI:  10.1109/tsp.2013.2288675.
    [14] RICHMAN J S, LAKE D E, MOORMAN J R. Sample entropy [J]. Methods in Enzymology, 2004, 384: 172-184. DOI:  10.1016/S0076-6879(04)84011-4.
    [15] ZHANG Y G, PAN G F. A hybrid prediction model for forecasting wind energy resources [J]. Environmental Science and Pollution Research, 2020, 27(16): 19428-19446. DOI:  10.1007/s11356-020-08452-6.
    [16] GRAVES A. Long short-term memory [M]//GRAVES A. Supervised Sequence Labelling with Recurrent Neural Networks. Berlin, Heidelberg: Springer, 2012: 37-45. DOI: 10.1007/978-3-642-24797-2_4.
    [17] 国家市场监督管理总局, 国家标准化管理委员会. 风电场气象观测资料审核、插补与订正技术规范: GB/T 37523—2019 [S]. 北京: 中国标准出版社, 2019.

    State Administration for Market Regulation, Standardization Administration. Specification for data inspection and correction of wind power plant meteorological observation: GB/T 37523—2019 [S]. Beijing: Standards Press of China, 2019.
  • 通讯作者: 陈斌, bchen63@163.com
    • 1. 

      沈阳化工大学材料科学与工程学院 沈阳 110142

    1. 本站搜索
    2. 百度学术搜索
    3. 万方数据库搜索
    4. CNKI搜索

Figures(9)  / Tables(4)

Article Metrics

Article views(710) PDF downloads(109) Cited by()

Related

Very Short-Term Wind Direction Multistep Forecast Based on VMD-LSTM

doi: 10.16516/j.gedi.issn2095-8676.2023.01.004

Abstract:   Introduction  In order to accurately forecast the wind direction in the next 4 hours, a very short-term wind direction multistep forecast algorithm based on VMD-LSTM(Variational Mode Decomposition-Long Short-Term Memory) is proposed.   Method  Wind direction sequence was collected from 3 wind turbines of a wind farm of Mingyang Smart Energy Group for preprocessing and analysis. The correlation of wind direction in different periods was calculated using the autocorrelation function (ACF) to select the characteristic length of wind direction sequence. Based on variational mode decomposition (VMD), the wind direction sequence was decomposed into relatively intrinsic mode functions, the number of which was determined by minimum sample entropy. Models were build for each intrinsic mode function to make very short-term wind direction 24-step forecast. Finally, the wind direction sequence was reconstructed from the forecasted intrinsic mode functions.   Result  The results obtained demonstrate that the average MAE(Mean Absolute Error), RMSE(Root Mean Square Error) and MAPE(Mean Absolute Percentage Error) of the 24-step wind direction forecast based on VMD-LSTM in 4 quarters are 8.430°, 16.870° and 9.155, respectively. The algorithm performs better than other common data modeling methods regarding each error evaluation index at different time scales in each quarter.   Conclusion  The proposed algorithm can optimize the control yaw angle in the actual production of wind farms.

LI Xiuhao, LIU Huaixi, ZHANG Zhiyong, ZHANG Min, WU Di, MIAO Desheng. Very Short-Term Wind Direction Multistep Forecast Based on VMD-LSTM[J]. SOUTHERN ENERGY CONSTRUCTION, 2023, 10(1): 29-38. doi: 10.16516/j.gedi.issn2095-8676.2023.01.004
Citation: LI Xiuhao, LIU Huaixi, ZHANG Zhiyong, ZHANG Min, WU Di, MIAO Desheng. Very Short-Term Wind Direction Multistep Forecast Based on VMD-LSTM[J]. SOUTHERN ENERGY CONSTRUCTION, 2023, 10(1): 29-38. doi: 10.16516/j.gedi.issn2095-8676.2023.01.004
    • 在风力发电中,偏航系统依据风向进行对风操作,准确对风可以降低风力发电机组的载荷、提高风能利用率,进而提高发电量,直接影响机组的经济性和安全性[1]。而风向却是随机的、不稳定的,这使得偏航系统的对风存在一定的误差和滞后性,继而导致偏航系统偏航动作频繁。因此,风向的准确预测可以提高偏航系统的对风精度、减小偏航滞后性,有利于风电场结合风向变化趋势制定偏航控制策略,进而对偏航优化提供保障,对风电企业经济、安全生产具有较大的工程实践指导意义。

      风向建模方法主要分为物理建模法、统计建模法和数据驱动建模法。物理建模法通常利用数值天气预报(Numerical Weather Prediction,NWP)数据进行建模[2],模型计算量大,主要用于长期预测。统计建模法一般使用时间序列建模,主要包括自回归移动平均模型(Autoregressive Moving Average model, ARMA)[3]、自回归差分移动平均模型(Autoregressive Integrated Moving Average model, ARIMA)[4]、自回归条件异方差模型(Autoregressive Conditional Heteroskedasticity model, ARCH)[5]等,相比于物理模型,统计建模法不用求解复杂的物理模型,计算效率高,常用于短期预测,但无法准确地描述非线性数据。由于风向数据具有非线性、非平稳性、随机性等特点,数据驱动模型可有效地处理复杂时序预测,BP神经网络[6]、支持向量机[7]等模型都取得不错的效果。张亚超[8]等提出一种基于VMD-SE和基模型的自适应多层级综合预测模型,可实现短期风电功率3步预测。但传统的浅层神经网络难以充分挖掘数据的深层特征,导致模型预测精度受限。深度学习以其强大的特征提取能力,在时序预测等任务上取得巨大的成功。唐振浩[9]等提出一种基于数据解析的混合建模算法(Data Analytics based Hybrid Algorithm,DAHA),可实现风向的单步预测。林涛[10]等提出1种变分模态分解(VMD)和蝙蝠算法(Bat Algorithm,BA)优化长短期记忆(LSTM)神经网络的短期风向预测模型,可提前预测未来5 min、15 min和30 min的风向。向玲[11]等提出一种变分模态分解(VMD)-模糊信息粒化(Fuzzy Information Granulation, FIG)和参数优化门控循环单元(Gate Recurrent Unit,GRU)的风速多步区间预测方法,可实现风速3步区间预测。由于风向具有强波动性和随机性等特点,但现有文献进行超短期风向预测时多为单步预测,且风向多步预测中预测步长较短,难以满足风电场的实际生产需求。

      为此,针对超短期风向多步准确预测,本文提出一种基于VMD-LSTM的风向多步预测算法,准确预测未来4 h的风向。首先,通过自相关函数计算风向不同时期之间的相关性,以选取模型最佳的风向序列输入长度;然后,针对风向数据波动性、随机性的特点,采用变分模态分解法将风向序列分解为相对稳定的模态信号,通过最小样本熵确定分解的子模态数,并对分解后的模态信号分别建立超短期风向预测模型进行预测;最后,重构风向序列,叠加各分量预测结果,实现超短期风向24步预测。

    • 明阳智能某风电场3个风电机组(1#、2#、3#)1年风向数据的风玫瑰图如图1所示。其中,扇形区域的颜色表示不同的风速大小,长度表示风向的频率。

      Figure 1.  Year-round wind rose map

      图1所示及统计可知,1#、2#号风机的主导风向为东,3#号风机的主导风向为东南偏东;1#、2#、3#号风机的盛行风向区间为(67.5°,112.5°),盛行风频率分别为40.31%、39.96%、40.61%。虽然1#、2#、3#号风机同处一风场,但受风机排布位置及尾流的影响[12],同一风场不同风机的风向也有所差别。由此可见,风向存在较大的随机性与不稳定性,进一步增大了风向预测的难度。

    • 不同的序列输入长度对模型的预测结果有一定的影响。模型输入的序列长度过短,难以表征序列特征,降低模型预测的准确率;输入序列过长导致信息冗余,降低模型的建模效率。本文采用自相关函数计算前$ k $个时刻($ k > 0 $)的风向$ {x}_{t-k} $与当前时刻的风向$ {x}_{t} $之间的相关性,以选取最佳的风向序列输入长度。

      自相关函数(Autocorrelation Function, ACF)用于度量同一事件在不同时期之间的相关性程度。对于时间序列$ \left\{ {{X_t}} \right\} $,ACF度量时间序列中每隔$ k $个时间单位($ {x}_{t} $$ {x}_{t-k} $)的观测值之间的相关性,计算公式如下:

      $$ {\rm{ACF}} = \dfrac{{\displaystyle \sum\limits_{t = k + 1}^n {\left( {{x_t} - \bar x } \right)\left( {{x_{t - k}} - \bar x } \right)} }}{{\displaystyle \sum\limits_{i = 1}^n {{{\left( {{x_t} - \bar x } \right)}^2}} }} $$ (1)

      式中:

      $ \bar{x} $——变量$ x $的平均值。

      图2所示,横坐标表示时间滞后长度,纵坐标表示滞后为$ k $个时刻的风向序列与当前序列的相关性。本文计算了72个样本点的相关性,随着$ k $的增大,历史风向与当前风向的相关性逐渐减弱,当$k > 35 \nwarrow $时其自相关系数小于0.6。为了兼顾模型的预测精度与效率,选择自相关系数大于0.6的序列。此外,输入序列长度应不少于预测序列长度,本文中预测未来4 h的风向,即输入序列长度应不少于24个样本点。因此,本文选取自相关系数大于0.6的前24个采样点的风向作为模型的输入序列。

      Figure 2.  The autocorrelation coefficient of wind direction

    • 变分模态分解(Variational Mode Decomposition, VMD)是一种完全内在、自适应、非递归的信号分解技术[13],通过求解约束变分问题,将信号转换到频域内分解为K个有限带宽的本征模函数(Intrinsic Mode Function, IMF),这种方法可以有效避免经验模态分解(Empirical Mode Decomposition, EMD)和局部均值分解(Local Mean Decomposition, LMD),在分解中由于递归分解模式所造成的包络线估计误差,具备强大的非线性和非平稳性信号处理能力,相比于EMD和集合经验模态分解(Ensemble Empirical Mode Decomposition, EEMD)等信号分解方法,它在解决信号噪声和避免模态混叠的问题上有显著优势。但VMD分解的IMF子模态数K对分解结果有很大的影响:当K太小时,分解后的序列会丢失过多信息从而导致模态混叠;当K太大时,会出现过度分解的问题。

      为了评估序列数据的复杂性,Richman和Moornan等[14]提出样本熵(Sample Entropy, SE),通过度量信号中产生新模式的概率大小来衡量时间序列的复杂性,时间序列越复杂,样本熵的值越大。

      针对风向数据波动性、随机性的特点,本文采用VMD对风向序列进行分解,得到多个稳定的子信号,通过最小SE值对VMD进行优化[15],以获取合适的K值。

    • 长短期记忆(Long Short-Term Memory, LSTM)是一种特殊的循环神经网络(Recurrent Neural Network, RNN),主要为了解决长序列训练过程中的梯度消失和梯度爆炸问题[16]。相比于RNN,LSTM能够在更长的序列中有更好的表现。LSTM由多个单元组成,每个LSTM单元包括3个门控系统和1个记忆单元,具体为:

      遗忘门:

      $$ {F_t} = {\rm{sigmoid}}({w_{{\rm{xf}}}}{x_t} + {w_{{\rm{hf}}}}{h_{t - 1}} + {b_{\rm{f}}}) $$ (2)

      输入门:

      $$ {I_t} = {\rm{sigmoid}}({w_{{\rm{xi}}}}{x_t} + {w_{{\rm{hi}}}}{h_{t - 1}} + {b_{\rm{i}}}) $$ (3)

      输出门:

      $$ {O_t} = {\rm{sigmoid}}({w_{{\rm{xo}}}}{x_t} + {w_{{\rm{ho}}}}{h_{t - 1}} + {b_{\rm{o}}}) $$ (4)

      记忆单元状态值:

      $$ {C_t} = {C_{t - 1}} \otimes {F_t} + {I_t} \otimes \tan h \left( {{w_{{\rm{xc}}}}{x_t} + {w_{{\rm{hc}}}}{h_{t - 1}} + {b_{\rm{c}}}} \right) $$ (5)

      LSTM记忆单元在$t$时刻的输出:

      $$ {h_t} = {O_t} \otimes \tan h ({C_t}) $$ (6)

      式(2)~式(6)中,$w\left( {{F_t},{I_t},{O_t},{C_t}} \right)$$b\left( {{F_t},{I_t},{O_t},{C_t}} \right)$为三个门控单元和记忆单元的权重和偏置值。

      本文采用两层的LSTM网络挖掘风向时序序列的深层特征,每层的神经元个数为100,学习率为0.001,优化器为Adam,迭代次数为10,激活函数为ReLU,如图3所示。

      Figure 3.  The structure of LSTM

    • 本文采用明阳智能某风电场数据采集与监视控制系统(SCADA)提供的2021年风向数据。风向数据来源于3个风力发电机(1#、2#、3#),时间粒度为10 min,按季度分为4组。为保证时间样本的顺序性,使用样本总量前70%的序列作为训练集,后30%的时间序列作为测试集,具体如表1所示。

      数据集数据来源训练集数量测试集数量
      A2021/03/01 00:00-2021/03/31 23:503 1251 339
      B2021/06/01 00:00-2021/06/30 23:503 0241 296
      C2021/09/01 00:00-2021/09/30 23:503 0241 296
      D2021/12/01 00:00-2021/12/31 23:503 1251 339

      Table 1.  Dataset information

      根据GB/T 37523-2019[17]对风向数据进行合理范围筛选,剔除异常数据值,使用滑动平均法对缺失数据进行插补;采用最大-最小归一化方法对风向数据进行处理,将风向数据映射到[0,1]内,归一化计算如下:

      $$ {y_t}' = \dfrac{{{y_t} - {y_{\min }}}}{{{y_{\max }} - {y_{\min }}}} $$ (7)

      式中:

      ${y_t}'$ ——$t$时刻归一化后的风向数据;

      ${y_t}$ ——$t$时刻的原始风向数据;

      ${y_{\max }}$ ——风向序列的最大值;

      ${y_{\min }}$ ——风向序列的最小值。

    • 本文采用绝对平均误差(Mean Absolute Error, MAE)、均方根误差(Root Mean Square Error, RMSE)、平均绝对百分比误差(Mean Absolute Percentage Error, MAPE)作为评价指标,衡量预测值与真实值的偏离程度、评价数据的变化程度、定量评价预测模型,其表达式分别为:

      $$ {\rm{MAE}} = \dfrac{1}{n}\displaystyle \sum\limits_{i = 1}^n {\left| {{y_{{\rm{true}}}} - {y_{{\rm{pre}}}}} \right|} $$ (8)
      $$ {\rm{RMSE}} = \sqrt {\dfrac{1}{n}\displaystyle \sum\limits_{i = 1}^n {{{\left( {{y_{{\rm{true}}}} - {y_{{\rm{pre}}}}} \right)}^2}} } $$ (9)
      $$ {\rm{MAPE}} = \dfrac{1}{n}\displaystyle \sum\limits_{i = 1}^n {\left| {\dfrac{{{y_{{\rm{true}}}} - {y_{{\rm{pre}}}}}}{{{y_{{\rm{true}}}}}}} \right|} $$ (10)

      式中:

      ${y_{{\rm{true}}}}$ ——真实值;

      ${y_{{\rm{pre}}}}$ ——预测值;

      $n$ ——样本个数。

    • 本文算法流程图如图4所示。

      Figure 4.  Algorithm flowchart

      具体步骤如下:

      1) 采集风向序列并对其进行预处理。

      2) 绘制风向玫瑰图,分析风向特征。

      3) 基于ACF计算风向不同时期之间的相关性,选取自相关系数大于0.6的前24个采样点的风向作为模型的输入序列。

      4) 采用VMD将风向序列分解为相对稳定的模态信号,通过最小样本熵确定分解的子模态数K

      5) 对分解后的K个模态信号分别建立预测模型,进行超短期风向多步预测。

      6) 重构风向序列,叠加各分量预测结果。

    • VMD分解的子模态数K直接决定了风向信号分解的好坏,对预测结果有一定的影响。本文利用VMD对风向数据进行分解得到K个IMF,计算每个IMF的SE值,以获得具有最小SE的序列作为趋势项,通过对比不同K值的最小SE值以确定子模态数。本文令K取值为2~10,如图5显示了在取不同K值时最小SE值和预测绝对平均误差的变化趋势。

      Figure 5.  Variation trend of minimum sample entropy and absolute mean error

      图5可知,随着VMD分解的子模态数$K$的增加,最小样本熵的值减小,风向预测模型的绝对平均误差MAE减小,并且两者衰减的变化趋势基本保持一致,表明最小样本熵的值能够有效地表征信号分解能力。当子模态数$K$较小时,原始风向信号分解不足,序列趋势项中混入了其他干扰项,使得SE值较大。随着$K$值的增大,SE值逐渐变小,当取得适当的$K$值时,SE值骤减,此时再增大分解次数$K$,SE值变化较小,并且逐渐趋于稳定。因此,将SE骤减趋于稳定的转折点作为VMD分解的次数,以避免过度分解。在本文中,取K=9,即利用VMD将原始风向信号分解为9个子序列,如图6所示。

      Figure 6.  Original wind direction decomposition by VMD

    • 为验证VMD分解对预测结果的影响,进行VMD分解前后对比算法建模预测结果的比较。由表2图7图8可知,VMD-LSTM在4个季度的24步风向预测的平均MAE、RMSE、MAPE为8.430°、16.870°、9.155,比未分解的LSTM模型平均减少77.91%、69.30%、69.42%,因此,VMD将原始风向序列分解为相对稳定的模态信号,可以有效地降低风向的非线性、非平稳性和随机性,提高风向序列预测的准确性。

      预测步长第1季度第2季度第3季度第4季度
      MAE/(°)RMSE/(°)MAPEMAE/(°)RMSE/(°)MAPEMAE/(°)RMSE/(°)MAPEMAE/(°)RMSE/(°)MAPE
      17.78215.4458.1376.69311.6324.7045.3588.7728.7724.59215.0974.983
      27.71215.3258.0456.54911.4044.6015.3168.6828.6824.55314.9264.934
      37.71115.3388.0126.44911.3154.5055.3068.5718.5714.53214.7304.872
      47.64415.2517.9126.74411.8584.6885.3318.5988.5984.47814.4434.788
      57.67415.3237.9397.35312.9694.9975.4608.7668.7664.42114.2294.755
      67.61815.3127.9697.57913.2065.2085.6418.9918.9914.50014.4724.872
      77.74715.7308.1527.82913.5025.4135.9119.3619.3614.77515.1885.213
      88.14016.6578.5018.08513.9505.6585.8559.4259.4255.12716.1675.637
      98.81017.9579.2528.23114.1185.7856.0039.7509.7505.62017.8486.211
      109.28018.5909.5918.22914.2335.7936.19410.15610.1566.22019.6916.790
      119.84219.33210.3098.35514.4645.8536.49410.68510.6856.73621.4557.257
      1210.59120.64711.1528.31714.4325.7966.89111.28611.2867.03122.3527.520
      1311.96823.26712.4498.42714.8145.8647.26311.83611.8367.29222.5687.639
      1412.89025.43313.3838.61015.1686.0137.97712.84812.8487.67422.9687.936
      1513.05925.33113.5848.70415.3886.1417.88312.82712.8278.14623.4528.423
      1614.07226.94214.8869.03415.9016.4228.04513.03613.0368.62724.4618.829
      1713.60725.73114.6479.42816.4646.7628.46213.61313.6138.44023.3328.432
      1813.04224.30614.04510.18617.5847.3128.60113.71013.7108.53422.8898.355
      1913.09824.20514.08910.05517.2857.2588.57813.82313.8238.91823.0678.585
      2013.29224.44914.19810.28517.6267.4538.75614.30014.3008.78621.9858.495
      2113.74025.28914.88310.78318.2537.7729.18215.18115.1818.95422.0068.740
      2213.75725.23714.86910.62818.0607.7229.66216.07116.0719.08721.8908.988
      2313.97725.42315.13510.41217.6677.64410.01116.60016.6009.45722.3509.440
      2413.75525.00314.83410.34217.6577.65110.42117.37117.37110.02923.1669.975

      Table 2.  Wind direction multistep forecast results based on VMD-LSTM

      Figure 7.  Comparison of LSTM and VMD-LSTM forecast results

      Figure 8.  Wind direction multistep forecast results based on VMD-LSTM

      预测误差随着预测时间的增加而累加,进而导致误差逐步增大,LSTM模型平均每步误差增长约为1.26°,VMD-LSTM模型误差稳步增长,增长幅度较小,平均每步误差增长约为0.29°,说明VMD-LSEM模型可以有效地降低误差的增长速度,稳定误差增长幅度。

    • 不同的预测方法对超短期风向多步预测有一定的影响。本文基于不同季度的风向分别构建ARMA、RF、VMD-RF、LSTM、VMD-LSTM模型。

      表3图9可知,VMD-LSTM在每个季度的各个误差评价指标均优于其他算法。VMD-LSTM模型较表现次好的VMD-RF模型,其平均MAE、RMSE、MAPE分别减少37.19%、23.80%、26.85%。通过不同建模算法的试验,结果表明,VMD-LSTM在不同季度下的风向具有更高的准确性和较好的预测能力。此外,VMD-LSTM和VMD-RF比未分解的模型LSTM和RF的MSE分别降低了77.91%和57.86%,说明经过分解后的建模精度有较大幅度的提高,VMD在提取风向趋势信息方面具有较好的能力。

      预测模型季度MAE/(°)RMSE/(°)MAPE
      ARMA118.17136.76221.131
      221.63637.29315.900
      314.18325.07215.230
      48.77427.9759.853
      RF133.89752.22430.148
      228.18945.54419.395
      318.73428.46219.032
      446.55456.74956.749
      VMD-RF116.13027.31017.054
      212.90920.4219.514
      310.78616.02111.003
      413.85724.80212.495
      LSTM141.67761.16234.512
      238.96859.59828.449
      323.14033.85122.536
      448.87765.17134.266
      VMD-LSTM110.86720.89711.499
      28.63814.9566.126
      37.27511.84411.844
      46.93919.7807.153

      Table 3.  Wind direction forecast results of different methods

      Figure 9.  Wind direction forecast results of different methods

    • 为检验该方法对不同风机的风向预测效果,本文对比分析了同一风场1#、2#、3#号风机,绝对平均误差MAE见表4

      风机第1季度第2季度第3季度第4季度合计
      1#14.5958.3498.4929.86510.325
      2#13.3458.2319.7508.92710.063
      3#12.3908.95410.8088.93410.271

      Table 4.  MAE of wind direction forecast for different wind turbines (°)

      虽然1#、2#、3#号风机同处一风场,但受风机排布位置及尾流的影响,导致不同风机在每个季度的预测结果也有所差异,但不同风机的预测总体误差差距较小,说明VMD-LSTM可适用于同一风场不同风机的预测,具有良好的泛化能力。

    • 由于风向具有强波动性和随机性等特点,目前的超短期风向预测步长较短,难以满足风电场的实际生产需求。为此,本文提出1种基于VMD-LSTM的风向多步预测算法,通过算例分析,得到以下主要结论:

      1)VMD将原始风向序列分解为相对稳定的模态信号,可以有效地降低风向的非线性、非平稳性和随机性,提高风向序列预测的准确性。利用最小样本熵的值确定VMD分解的子模态数,可优化VMD的风向信号分解性能,提高VMD的风向趋势信息提取能力。

      2)针对不同的预测方法对超短期风向多步预测的影响,分别构建ARMA、RF、VMD-RF、LSTM、VMD-LSTM预测模型。比较发现,VMD-LSTM在4个季度的24步风向预测的平均MAE、RMSE、MAPE为8.430°、16.870°、9.155,在每个季度不同时间尺度的各个误差评价指标均优于其他算法,所提算法可满足风电场的实际生产中优化控制偏航角的要求。

      3)由于风向数据具有较大的随机性和不稳定性等特点,且受风机排布位置及尾流的影响,进一步增大了风向预测的难度。通过对比不同风机的预测结果,证明了模型的鲁棒性,具有良好的泛化能力。

Reference (17)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return