基于高斯混合聚类和改进条件变分自编码的多风电场功率日场景生成方法

doi:10.11930/j.issn.1004-9649.202311050

摘要/Abstract

摘要：

大量出力不确定的风电场并入电网会带来运行隐患和不可控风险，基于变分自编码器的场景生成模型方法能生成确定性场景集合以描述风电出力的不确定性。针对多风电场出力复杂的时空相关性以及在传统变分自编码器模型训练过程中可能存在的“KL坍缩”等问题，提出一种基于高斯混合聚类和改进条件变分自编码器的多风电场时空功率日场景生成方法。通过引入二维卷积技术提取时空相关性进行降维，并采用最大化最小夹角独立正则化技术，强化隐特征的独立性；采用超球面分布替代高斯分布，避免模型出现“KL坍缩”，提高模型场景生成训练的稳定性和准确性；另外，进一步考虑多风电场功率日场景的多样性和灵活性，引入高斯混合聚类技术，使模型可根据特定的条件标签生成具有差异化特征的确定性场景集。实际算例的结果表明，相较于常见方法，所提方法累积概率分布误差下降了17%~71%，时空相关性平均误差分别下降了85%~97%和55%~91%，且能精准生成不同风况类别占比的多风电场功率日场景集，提高了场景生成的多样性和灵活性。

关键词: 风电场景生成, 高斯混合模型, 特征提取, 条件变分自编码器, 超球面分布, 正则化技术

Abstract:

The integration of a large number of wind farms with uncertain output into the power grid will bring potential hazards in operation and uncontrollable risks. The uncertainty of wind power output is described by uncertain scenario sets generated from the variational autoencoder-based scenario generation method. Aimed at the complex spatiotemporal correlation of multi-wind farm output and the possible "KL collapse" during the traditional variational autoencoder model training, this paper proposes a daily scenario generation method of spatiotemporal power based on the Gaussian mixture model and improved conditional variational autoencoding. The two-dimensional convolution technique is introduced to extract the spatiotemporal correlation for dimension reduction, and the maximizing min-angle regularization technique is used to strengthen the independence of latent features. Hyperspherical distribution, instead of Gaussian distribution, is used to avoid "KL collapse" and improve the stability and accuracy of scene generation training. In addition, considering the diversity and flexibility of daily power scenarios of multi-wind farm, Gaussian mixture clustering technology is introduced to generate uncertain scenario sets with differentiated and changing characteristics, enabling the generation of certain scenario sets with varied characteristics according to specific condition labels. The results of real examples show that compared with conventional methods, the proposed method reduces the accumulated error distribution of probability by 17% to 71%, and the average error of temporal and spatial correlation by 85% to 97% and 55% to 91%, respectively. Besides, the proposed method can accurately generate daily power scenario sets of multi-wind farm in different wind conditions, improving scene generation's diversity and flexibility.

Key words: wind power scenario generation, Gaussian mixture model, feature extraction, conditional variational autoencoder, hyperspherical distribution, regularization technique

李丹, 梁云嫣, 缪书唯, 方泽仁, 胡越, 贺帅. 基于高斯混合聚类和改进条件变分自编码的多风电场功率日场景生成方法[J]. 中国电力, 2024, 57(12): 17-29.

Dan LI, Yunyan LIANG, Shuwei MIAO, Zeren FANG, Yue HU, Shuai HE. Daily Power Scenario Generation Method for Multiple Wind Farms Based on Gaussian Mixture Clustering and Improved Conditional Variational Autoencoder[J]. Electric Power, 2024, 57(12): 17-29.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.electricpower.com.cn/CN/10.11930/j.issn.1004-9649.202311050

https://www.electricpower.com.cn/CN/Y2024/V57/I12/17

图/表 16

图 1 多风电场时空功率日场景生成总流程

Fig.1 Daily power scenario generation in spatiotemporal multi-wind farms

图 2 GMM流程

Fig.2 Flow of GMM

图 3 二维卷积的CVAE场景生成模型框架

Fig.3 Two-dimensional convolved CVAE scenario generation

图 4 高斯分布与超球面分布隐特征向量采样对比

Fig.4 Comparison of latent feature vectors sampling between Gaussian and hyperspherical distributions

图 5 多风电场地理位置

Fig.5 Distribution of multiple wind farms

表 1 各方法的超参数设置

Table 1 Hyperparameter settings of methods

模型	参数设置明细
VAE	N=433； d_m=24×18； d_z=100； E=10³	p(Z)~N(0,1)；
GAN
CVAE		K=3；p(Z)~N(0,1)
CGAN		K=3
vMF-CVAE		K=3；p(Z)~vMF(κ)；κ=10⁵
1D-GMM-ICVAE		K=3；p(Z)~vMF(κ)； κ=10⁵；λ=100
2D-GMM-ICVAE		K=3；p(Z)~vMF(κ)； κ=10⁵； H=150(3×2)；λ=100

图 6 真实功率场景与生成功率场景

Fig.6 Real power scenarios and generated power scenarios

图 7 2014年风电时空功率日场景的聚类结果

Fig.7 Clustering results of daily spatiotemporal scenarios of wind power in 2014

图 8 不同类别条件下原始场景和生成场景对比

Fig.8 Comparison of original and generated scenarios in different label conditions

表 2 不同聚类方式下生成场景的评价指标对比

Table 2 Evaluation indexes of generated scenarios in different clustering ways

聚类方式	E_MAAP		隐特征平均相关系数		时/空相关系数的 E_RMS		E_RMS-CDF
聚类方式	数值/ %	Ip/ 百分比	数值	Ip/%	数值	Ip/%	数值/ MW	Ip/ %
GMM 聚类	6.10	0	0.17	0	0.02/0.09	0/0	0.13	0
月份	13.69	7.59	0.37	53.38	0.13/0.23	86.40/60.79	0.25	46.80
四季	11.92	5.82	0.35	50.62	0.11/0.19	84.64/52.91	0.23	40.89
多/少风季	11.09	4.99	0.28	39.15	0.07/0.34	80.23/75.82	0.15	11.92
不分类	14.97	8.87	0.43	59.95	0.24/0.98	93.03/90.87	0.30	55.81

图 9 MMA对隐特征的影响对比

Fig.9 Comparison of MMA's influence on latent features

表 3 不同方法下生成场景的评价指标对比

Table 3 Evaluation index comparisons of the scenes generated by different methods

模型	E_MAAP					隐特征平均相关系数					时空相关系数的E_RMS					E_RMS-CDF
	全部		小风日	中风日	大风日	全部		小风日	中风日	大风日	全部		小风日	中风日	大风日	全部		小风日	中风日	大风日
	值/ %	Ip/ 百分比	小风日	中风日	大风日	值	Ip/ %	小风日	中风日	大风日	值	Ip/ %	小风日	中风日	大风日	值/ MW	Ip/ %	小风日	中风日	大风日
2D-GMM- ICVAE	6.10	0	6.07	5.85	6.42	0.17	0	0.15	0.15	0.21	0.02/0.09	0	0.02/0.09	0.01/0.07	0.03/0.12	0.13	0	0.13	0.12	0.15
1D-GMM- ICVAE	6.90	0.80	6.87	6.53	7.12	0.23	25	0.22	0.19	0.24	0.11/0.20	85/55	0.11/0.19	0.10/0.17	0.14/0.22	0.16	17	0.15	0.14	0.20
vMF-CVAE	7.39	1.29	7.33	6.79	8.06	0.33	48	0.31	0.28	0.36	0.13/0.31	87/71	0.13/0.25	0.13/0.22	0.14/0.33	0.20	32	0.15	0.19	0.26
CVAE	12.82	6.72	12.63	11.49	13.67	0.51	66	0.50	0.48	0.54	0.22/0.43	92/79	0.22/0.42	0.21/0.39	0.23/0.43	0.28	52	0.24	0.24	0.34
CGAN											0.32/0.92	95/90				0.33	59	\	\	\
GAN											0.50/0.93	97/90				0.46	71	\	\	\
VAE	15.02	8.92				0.54	68				0.32/0.98	95/91				0.39	65	\	\	\

图 10 时空相关性系数矩阵对比热图

Fig.10 Heatmap comparison of spatiotemporal correlation coefficient matrixes

图 11 CDF曲线对比

Fig.11 CDF curve comparisons

图 12 消融模型的误差指标对比

Fig.12 Error index comparisons of ablation models

图 13 本文模型不同风况误差指标对比

Fig.13 Error index comparisons of different wind conditions in the proposed model

参考文献 24

1	徐询, 谢丽蓉, 梁武星, 等. 考虑风电预测误差时序性及风电可信度的双层优化模型[J]. 电工技术学报, 2023, 38 (6): 1620- 1632, 1661.
	XU Xun, XIE Lirong, LIANG Wuxing, et al. Bi-level optimization model considering time series characteristic of wind power forecast error and wind power reliability[J]. Transactions of China Electrotechnical Society, 2023, 38 (6): 1620- 1632, 1661.
2	国家能源局. 国家能源局发布2023年全国电力工业统计数据. (2023-12-20)[2024-08-20]. https://www.nea.gov.cn/2023-12/20/c_1310756286.htm.
3	彭小圣, 熊磊, 文劲宇, 等. 风电集群短期及超短期功率预测精度改进方法综述[J]. 中国电机工程学报, 2016, 36 (23): 6315- 6326, 6596.
	PENG Xiaosheng, XIONG Lei, WEN Jinyu, et al. A summary of the state of the art for short-term and ultra-short-term wind power prediction of regions[J]. Proceedings of the CSEE, 2016, 36 (23): 6315- 6326, 6596.
4	黄大为, 张伟, 韩学山. 基于自适应风电功率场景选取的有功调度模型[J]. 电力系统自动化, 2013, 37 (19): 68- 73, 92. DOI
	HUANG Dawei, ZHANG Wei, HAN Xueshan. Active power dispatch based on self-adaptive wind power scenario selection[J]. Automation of Electric Power Systems, 2013, 37 (19): 68- 73, 92. DOI
5	张宁, 康重庆. 风电出力分析中的相依概率性序列运算[J]. 清华大学学报(自然科学版), 2012, 52 (5): 704- 709.
	ZHANG Ning, KANG Chongqing. Dependent probabilistic sequence operations for wind power output analyses[J]. Journal of Tsinghua University (Science and Technology), 2012, 52 (5): 704- 709.
6	邹斌, 李冬. 基于有效容量分布的含风电场电力系统随机生产模拟[J]. 中国电机工程学报, 2012, 32 (7): 23- 31, 187.
	ZOU Bin, LI Dong. Power system probabilistic production simulation with wind generation based on available capacity distribution[J]. Proceedings of the CSEE, 2012, 32 (7): 23- 31, 187.
7	顾洁, 刘书琪, 胡玉, 等. 基于深度卷积生成对抗网络场景生成的间歇式分布式电源优化配置[J]. 电网技术, 2021, 45 (5): 1742- 1751.
	GU Jie, LIU Shuqi, HU Yu, et al. Optimal allocation of intermittent distributed generation based on deep convolutions generative adversarial network in scenario generation[J]. Power System Technology, 2021, 45 (5): 1742- 1751.
8	王新迎, 李烨, 董骁翀, 等. 基于变分自编码器的主动配电网多源-荷场景生成方法[J]. 电网技术, 2021, 45 (8): 2962- 2969.
	WANG Xinying, LI Ye, DONG Xiaochong, et al. Multi-source-load scenario generation of active distribution network based on variational autoencoder[J]. Power System Technology, 2021, 45 (8): 2962- 2969.
9	丁明, 宋晓皖, 孙磊, 等. 考虑时空相关性的多风电场出力场景生成与评价方法[J]. 电力自动化设备, 2019, 39 (10): 39- 47.
	DING Ming, SONG Xiaowan, SUN Lei, et al. Scenario generation and evaluation method of multiple wind farms output considering spatial-temporal correlation[J]. Electric Power Automation Equipment, 2019, 39 (10): 39- 47.
10	BILLINTON R, KARKI B, KARKI R, et al. Unit commitment risk analysis of wind integrated power systems[J]. IEEE Transactions on Power Systems, 2009, 24 (2): 930- 939. DOI
11	丁明, 鲍玉莹, 毕锐. 应用改进马尔科夫链的光伏出力时间序列模拟[J]. 电网技术, 2016, 40 (2): 459- 464.
	DING Ming, BAO Yuying, BI Rui. Simulation of PV output time series used improved Markov chain[J]. Power System Technology, 2016, 40 (2): 459- 464.
12	董骁翀. 基于数据驱动的可再生能源场景生成与约简方法研究[D]. 北京: 华北电力大学, 2020.
	DONG Xiaochong. Research on data-driven renewable energy scene generation and reduction method [D]. Beijing: North China Electric Power University, 2020.
13	李丹, 王奇, 缪书唯, 等. 基于张量SOM和VAE的多风电时空功率日场景生成[J]. 可再生能源, 2022, 40 (12): 1658- 1665. DOI
	LI Dan, WANG Qi, MIAO Shuwei, et al. A daily scenario generation for spatio-temporal power of multi-wind power based on tensor SOM and VAE[J]. Renewable Energy Resources, 2022, 40 (12): 1658- 1665. DOI
14	彭雨筝, 李晓露, 李聪利, 等. 基于残差卷积自编码的风光荷场景生成方法[J]. 电力建设, 2021, 42 (8): 10- 17. DOI
	PENG Yuzheng, LI Xiaolu, LI Congli, et al. Typical wind-PV-load scenario generation based on residual convolutional auto-encoders[J]. Electric Power Construction, 2021, 42 (8): 10- 17. DOI
15	刘德宝. 基于电动汽车充电负荷场景区间预测的孤岛型微电网鲁棒优化调度[D]. 吉林: 东北电力大学, 2021.
	LIU Debao. Robust optimal dispatch of islanded microgrid based on interval prediction of electric vehicle charging load scene[D]. Jilin: Northeast Electric Power University, 2021.
16	李旭霞, 张琳娜, 郑晓明, 等. 基于KL散度的储能电站分布鲁棒规划方法[J]. 太阳能学报, 2022, 43 (4): 46- 55.
	LI Xuxia, ZHANG Linna, ZHENG Xiaoming, et al. KL divergence-based distributionally robust planning method for energy storage plants[J]. Acta Energiae Solaris Sinica, 2022, 43 (4): 46- 55.
17	DAVIDSON T R, FALORSI L, NICOLA D C. Hyperspherical variational auto-encoders[C]//34th Conference on Uncertainty in Artificial Intelligence, United States, 2018: 856–865.
18	XU J C, DURRETT G. Spherical latent spaces for stable variational autoencoders[C]//Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Brussels, Belgium. Stroudsburg, PA, USA: Association for Computational Linguistics, 2018: 4503–4513.
19	李丹, 王奇, 杨保华, 等. 基于独立稀疏SAE的多风电场超短期功率预测[J]. 电力系统及其自动化学报, 2022, 34 (2): 23- 30.
	LI Dan, WANG Qi, YANG Baohua, et al. Ultra-short-term power prediction of multiple wind farms based on independent sparse SAE[J]. Proceedings of the CSU-EPSA, 2022, 34 (2): 23- 30.
20	WANG Z N, XIANG C Q, ZOU W B, et al. MMA regularization: decorrelating weights of neural networks by maximizing the minimal angles[EB/OL]. (2006-06-05)[2024-08-20]. http://arxiv.org/abs/2006.06527v2.
21	王守相, 陈海文, 李小平, 等. 风电和光伏随机场景生成的条件变分自动编码器方法[J]. 电网技术, 2018, 42 (6): 1860- 1869.
	WANG Shouxiang, CHEN Haiwen, LI Xiaoping, et al. Conditional variational automatic encoder method for stochastic scenario generation of wind power and photovoltaic system[J]. Power System Technology, 2018, 42 (6): 1860- 1869.
22	王磊, 张志勇, 曾维贵, 等. 基于数据场联合决策图改进的GMM聚类[J]. 系统工程与电子技术, 2022, 44 (9): 2743- 2751. DOI
	WANG Lei, ZHANG Zhiyong, ZENG Weigui, et al. An improved GMM clustering based on data field and decision graph[J]. Systems Engineering and Electronics, 2022, 44 (9): 2743- 2751. DOI
23	KIM S, KIM H. A new metric of absolute percentage error for intermittent demand forecasts[J]. International Journal of Forecasting, 2016, 32 (3): 669- 679. DOI
24	黄越辉, 孙亚南, 李驰, 等. 基于条件生成对抗网络的多区域风电短期出力场景生成方法[J]. 电网技术, 2023, 47 (1): 63- 77.
	HUANG Yuehui, SUN Yanan, LI Chi, et al. Constructing method of short-term output scenarios for multi-regional wind power based on conditional generative adversarial network[J]. Power System Technology, 2023, 47 (1): 63- 77.

[1]	郑鹏, 韩鹏程, 王国栋, 娄颖. 中压配电线路断线高阻接地故障精细化诊断方法[J]. 中国电力, 2024, 57(4): 220-228.
[2]	张博智, 张茹, 焦东翔, 王龙宇, 周一凡, 周丽霞. 基于VMD-SAST的电能质量扰动分类识别方法[J]. 中国电力, 2024, 57(2): 34-40.
[3]	马海飞, 滕伟, 彭迪康, 柳亦兵, 靳涛. 基于DRS与改进Autogram的风电齿轮箱复合故障特征提取[J]. 中国电力, 2023, 56(10): 71-79.
[4]	胡京, 邓颖, 蒋兴良, 曾蕴睿. 输电线路覆冰垭口微地形的特征提取与识别方法[J]. 中国电力, 2022, 55(8): 135-142.
[5]	樊江川, 于昊正, 刘慧婷, 杨丽君, 安佳坤. 基于多分支门控残差卷积神经网络的短期电力负荷预测[J]. 中国电力, 2022, 55(11): 155-162,174.
[6]	黄冬梅, 王玥琦, 胡安铎, 孙锦中, 时帅, 孙园, 房岭锋. 融合多维度特征的绝缘子状态边缘识别方法[J]. 中国电力, 2022, 55(1): 133-141.
[7]	刘文慧, 徐遵义, 张旭冉, 张海燕. 基于互信息和PCA理论的湿法烟气脱硫工况特征提取方法[J]. 中国电力, 2020, 53(8): 158-163.
[8]	林女贵, 洪兰秀, 黄道姗, 易扬, 刘智煖, 徐启峰. 基于改进深度自编码网络的异常用电行为辨识[J]. 中国电力, 2020, 53(6): 18-26.
[9]	耿江海, 王平, 田姗. SF₆气体泄漏光学检测图像特征提取方法研究[J]. 中国电力, 2014, 47(1): 8-12.