基于气象协变量注意力和大语言模型的日前光伏出力预测

doi:10.11930/j.issn.1004-9649.202504060

中国电力 ›› 2025, Vol. 58 ›› Issue (12): 211-222.DOI: 10.11930/j.issn.1004-9649.202504060

• 新能源与储能 • 上一篇

基于气象协变量注意力和大语言模型的日前光伏出力预测

葛彦硕¹(), 周艳真¹(), 郭庆来¹(), 肖大军², 徐遐龄², 李鑫², 刘涛²

1. 清华大学电机工程与应用电子技术系，北京　100084
2. 国家电网有限公司华中分部，湖北武汉　430000

收稿日期:2025-04-23 修回日期:2025-09-16 发布日期:2025-12-27 出版日期:2025-12-28
作者简介:
葛彦硕（2003），男，博士研究生，从事人工智能在电力系统时序数据预测与场景生成中的应用研究， E-mail：geys21@mails.tsinghua.edu.cn
周艳真（1990），女，博士，从事电力系统分析与控制、人工智能在复杂电网调控中的应用研究，E-mail：zhouyanzhen@mail.tsinghua.edu.cn
郭庆来（1979），男，通信作者，博士，教授，从事信息物理系统和无功电压优化控制研究，E-mail：guoqinglai@tsinghua.edu.cn
基金资助:
国家电网有限公司科技项目（面向电网调度智能辅助的大模型关键技术研究，5108-202455044A-1-1-ZN）。

Day-Ahead Photovoltaic Power Forecasting Based on Large Language Model with Meteorological Covariate Attention

GE Yanshuo¹(), ZHOU Yanzhen¹(), GUO Qinglai¹(), XIAO Dajun², XU Xialing², LI Xin², LIU Tao²

1. Department of Electrical Engineering, Tsinghua University, Beijing 100084, China
2. Central China Branch of State Grid Corporation of China, Wuhan 430000, China

Received:2025-04-23 Revised:2025-09-16 Online:2025-12-27 Published:2025-12-28
Supported by:
This work is supported by Science and Technology Project of SGCC (Research on Key Technologies of Large Language Models for Intelligent Assistance in Power Grid Dispatching, No.5108-202455044A-1-1-ZN).

摘要/Abstract

摘要：

日前光伏出力预测在实际应用中精度有待提升，且存在新建光伏电站数据缺乏难以预测的问题。为此，充分利用大语言模型（large language model，LLM）的推理优势和序列各维相关性信息，提出一种气象协变量注意力增强Time-LLM模型用于日前光伏出力预测。首先，通过填充和拼接构建融合气象协变量信息和光伏出力历史信息的模型输入序列。然后，通过所提协变量注意力模块融合气象协变量序列与光伏出力序列之间的相关性信息。最后，通过Time-LLM架构实现时间序列与文本序列的模态对齐，有效利用LLM的文本分析能力进行光伏出力时间序列的准确预测。在光伏出力公开数据集上进行算例分析，结果表明：所提模型不仅在测试集上预测准确率最高，还具有最低的零样本预测误差。所提方法既提高了日前光伏出力预测任务的准确率，也为解决新建光伏电站因历史数据匮乏而难以应用传统深度模型预测的问题提供了新思路。

关键词: 光伏出力预测, 大语言模型, Time-LLM, 协变量注意力, 零样本

Abstract:

To improve the accuracy of day-ahead photovoltaic (PV) power forecasting and mitigate the challenge of data scarcity in newly built photovoltaic power stations, leveraging both the reasoning capabilities of large language models (LLMs) and the correlation information across various dimensions of sequences, this paper proposes a meteorological covariate attention-enhanced Time-LLM model for day-ahead PV power prediction. Firstly, the input sequence is constructed by padding and concatenating historical PV power series with meteorological covariate series. Then, a novel covariate attention module is introduced to capture the cross-dimensional dependencies between meteorological variables and PV power sequences. Finally, the Time-LLM architecture is employed to achieve modality alignment between time-series and text sequences, effectively leveraging the textual analysis capabilities of LLMs for accurate PV power forecasting. Experimental results on public PV datasets demonstrate that the proposed model achieves superior forecasting performance and exhibits remarkable zero-shot learning capability. The proposed method not only improves the accuracy of day-ahead photovoltaic power forecasting, but also provides a promising solution for newly built PV plants with limited historical data, where traditional deep learning models often fail due to data scarcity.

Key words: photovoltaic power forecasting, large language model, Time-LLM, covariate attention, zero-shot

葛彦硕, 周艳真, 郭庆来, 肖大军, 徐遐龄, 李鑫, 刘涛. 基于气象协变量注意力和大语言模型的日前光伏出力预测[J]. 中国电力, 2025, 58(12): 211-222.

GE Yanshuo, ZHOU Yanzhen, GUO Qinglai, XIAO Dajun, XU Xialing, LI Xin, LIU Tao. Day-Ahead Photovoltaic Power Forecasting Based on Large Language Model with Meteorological Covariate Attention[J]. Electric Power, 2025, 58(12): 211-222.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.electricpower.com.cn/CN/10.11930/j.issn.1004-9649.202504060

https://www.electricpower.com.cn/CN/Y2025/V58/I12/211

图/表 12

图 1 模型输入重构示意

Fig.1 Schematic of model input reconstruction

图 2 Time-LLM模型结构

Fig.2 Structure of Time-LLM

图 3 气象协变量注意力增强Time-LLM

Fig.3 Improved Time-LLM with meteorological covariate attention

表 1 不同基线模型的最优超参数组合

Table 1 Optimal hyperparameter configuration of different baseline models

超参数名称	Informer	iTransformer	PatchTST	CNN- LSTM	Bi- LSTM	TFT
d_model	512	512	256			128
d_ff	2048	2048	1024
hidden_size				32	16	1280
n_layers	1	1	2	3	1	2
learning_rate	3×10^–5	1×10^–4	1×10^–4	3×10^–5	1×10^–3	1×10^–5

表 2 模型整体预测误差及准确率分析

Table 2 Analysis of the overall prediction error and accuracy of different models

模型	E_MA/MW	E_RMS/MW	$\eta $/%	p
Bi-LSTM	36.98	80.15	94.00	0.0000
CNN-LSTM	31.70	75.03	94.40	0.0000
Informer	24.69	59.65	95.55	0.0002
iTransformer	29.13	61.99	95.37	0.0450
PatchTST	36.20	74.81	94.37	0.0000
TFT	29.10	71.56	94.68	0.0000
Time-LLM	43.60	83.41	93.74	0.0000
改进Time-LLM	22.72	56.33	95.84

图 4 模型一周内日前滚动预测结果展示

Fig.4 Display of one-week day-ahead rolling prediction results of models

图 5 不同模型预测均方根误差分布

Fig.5 RMSE distribution of the prediction results of different models

图 6 不同场景下模型预测结果展示

Fig.6 Display of the prediction results of models under different scenarios

表 3 改进Time-LLM在不同窗口长度和滑动距离下的RMSE测试结果

Table 3 ERMS of improved Time-LLM with different patch length and stride length on test set 单位：MW

L_S	不同P下的E_RMS
L_S	12	24	48	96	192
1/6P	61.82	61.73	57.50	59.82	56.33
1/4P	59.77	57.24	57.41	59.02	61.82
1/2P	61.16	56.75	60.88	58.90	62.09
1P	56.85	58.88	61.87	58.02	59.62
2P	61.51	63.47	58.67	58.57	67.33

图 7 模型输入及气象协变量注意力权重的可视化结果

Fig.7 Visualization of model inputs and meteorological covariate attention weights

表 4 零样本预测任务中模型整体预测误差及准确率分析

Table 4 Overall prediction error and accuracy of different models in zero-shot tasks

模型	E_MA/MW	E_RMS/MW	$\eta $/%	$ \Delta E_{{\text{RMS}}}^{{\text{zero}}} $/%	p
Bi-LSTM	31.64	68.23	93.84	2.76	0.0000
CNN-LSTM	28.07	67.01	93.97	7.62	0.0000
Informer	21.72	51.67	95.38	3.80	0.0000
iTransformer	25.77	55.17	95.04	7.12	0.0048
PatchTST	32.12	67.21	93.90	8.39	0.0000
TFT	24.79	60.79	94.55	2.45	0.0000
Time-LLM	38.30	73.92	93.30	7.06	0.0000
改进Time-LLM	19.33	47.59	95.77	1.80

图 8 零样本任务中模型一周内日前滚动预测结果展示

Fig.8 Display of one-week day-ahead rolling prediction results of models in zero-shot tasks

参考文献 31

1	SUN H B, GUO Q L, SHEN X W, et al. Energy Internet: redefinition and categories[J]. Energy Internet, 2024, 1 (1): 3- 8. DOI
2	FAN S X, GUO J B, MA S C, et al. Enabling artificial intelligence-based scenario application in new type power systems[J]. Energy Internet, 2024, 1 (1): 9- 13. DOI
3	刘大贵, 王维庆, 张慧娥, 等. 基于隐马尔科夫修正的光伏中长期电量预测及调度计划应用[J]. 高电压技术, 2023, 49 (2): 840- 848. DOI
	LIU Dagui, WANG Weiqing, ZHANG Huie, et al. Mid-long term available quantity of electricity forecasting with error calibration by hidden Markov model in photovoltaic and application of dispatching plan[J]. High Voltage Engineering, 2023, 49 (2): 840- 848. DOI
4	解振学, 林帆, 王若谷, 等. 基于时序动态回归的超短期光伏发电功率预测方法[J]. 智慧电力, 2022, 50 (7): 45- 51.
	XIE Zhenxue, LIN Fan, WANG Ruogu, et al. Very short-term photovoltaic power forecasting method based on time series dynamic regression[J]. Smart Power, 2022, 50 (7): 45- 51.
5	彭曙蓉, 陈慧霞, 孙万通, 等. 基于改进LSTM的光伏发电功率预测方法研究[J]. 太阳能学报, 2024, 45 (11): 296- 302. DOI
	PENG Shurong, CHEN Huixia, SUN Wantong, et al. Research on photovoitaic power prediction method based on improved LSTM[J]. Acta Energiae Solaris Sinica, 2024, 45 (11): 296- 302. DOI
6	李丰君, 王磊, 赵健, 等. 基于天气融合和LSTM网络的分布式光伏短期功率预测方法[J]. 中国电力, 2022, 55 (11): 149- 154. DOI
	LI Fengjun, WANG Lei, ZHAO Jian, et al. Research on distributed photovoltaic short-term power prediction method based on weather fusion and LSTM-Net[J]. Electric Power, 2022, 55 (11): 149- 154. DOI
7	龙小慧, 秦际赟, 张青雷, 等. 基于相似日聚类及模态分解的短期光伏发电功率组合预测研究[J]. 电网技术, 2024, 48 (7): 2948- 2957. DOI
	LONG Xiaohui, QIN Jiyun, ZHANG Qinglei, et al. Short-term photovoltaic power prediction study based on similar day clustering and modal decomposition[J]. Power System Technology, 2024, 48 (7): 2948- 2957. DOI
8	何威, 苏中元, 史金林, 等. 基于双重注意力GRU与相似修正的光伏功率预测[J]. 太阳能学报, 2024, 45 (3): 480- 487. DOI
	HE Wei, SU Zhongyuan, SHI Jinlin, et al. Photovoltaic power forecasting based on dual-attention-GRU and similarity modification[J]. Acta Energiae Solaris Sinica, 2024, 45 (3): 480- 487. DOI
9	姜建国, 杨效岩, 毕洪波. 基于VMD-FE-CNN-BiLSTM的短期光伏发电功率预测[J]. 太阳能学报, 2024, 45 (7): 462- 473.
	JIANG Jianguo, YANG Xiaoyan, BI Hongbo. Photovoltaic power forecasting method based on VMD-FE-CNN-BiLSTM[J]. Acta Energiae Solaris Sinica, 2024, 45 (7): 462- 473.
10	冯裕祺, 李辉, 李利娟, 等. 基于CNN-GRU的光伏电站电压轨迹预测[J]. 中国电力, 2022, 55 (7): 163- 171. DOI
	FENG Yuqi, LI Hui, LI Lijuan, et al. Voltage trajectory prediction of photovoltaic power station based on CNN-GRU[J]. Electric Power, 2022, 55 (7): 163- 171. DOI
11	VASWANI A, SHAZEER N, PARMAR N, et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30.
12	ZHOU H Y, ZHANG S H, PENG J Q, et al. Informer: beyond efficient transformer for long sequence time-series forecasting[J]. Proceedings of the AAAI Conference on Artificial Intelligence, 2021, 35 (12): 11106- 11115. DOI
13	NIE Y Q, NGUYEN N H. , SINTHONG P, et al. A time series is worth 64 words: Long-term forecasting with transformers[C]//The Eleventh International Conference on Learning Representations. 2022.
14	ZHANG Y H, YAN J C. Crossformer: transformer utilizing cross-dimension dependency for multivariate time series forecasting[C]//International Conference on Learning Representations, 2023.
15	LIU Y, WU H X, WANG J M, et al. Non-stationary transformers: exploring the stationarity in time series forecasting[J]. Advances in Neural Information Processing Systems, 2022, 35, 9881- 9893.
16	ZHOU T, MA Z Q, WEN Q S, et al. FEDformer: frequency enhanced decomposed transformer for long-term series forecasting[C]//International Conference on Machine Learning, 2022 : 27268–27286.
17	梁睿, 金沫含, 朱慧君, 等. 基于多场景聚类和可解释时间融合Transformer网络的乡村地区净负荷区间预测[J/OL]. 电网技术, 1–13 [2025-03-28]. https://kns.cnki.net/kcms/detail/11.2410.TM.20250327.1802.012.html.
	LIANG Rui, JIN Mohan, ZHU Huijun, et al. Net load interval forecasting in rural areas based on multi-scenario clustering and explainable temporal fusion transformers network[J/OL]. Power System Technology, 1–13 [2025-03-28]. https://kns.cnki.net/kcms/detail/11.2410.TM.20250327.1802.012.html.
18	张倩, 蒙飞, 李涛, 等. 基于周期信息增强的Informer光伏发电功率预测[J]. 中国电力, 2023, 56 (10): 186- 193.
	ZHANG Qian, MENG Fei, LI Tao, et al. Informer photovoltaic power generation forecasting based on cycle information enhancement[J]. Electric Power, 2023, 56 (10): 186- 193.
19	邓芳明, 刘涛, 王锦波, 等. 基于地基云图与气象因素多模态融合的光伏功率预测方法[J]. 中国电机工程学报, 2025, 45 (18): 7193- 7206. DOI
	DENG Fangming, LIU Tao, WANG Jinbo, et al. Research on photovoltaic power prediction based on multimodal fusion of ground cloud map and meteorological factors[J]. Proceedings of the CSEE, 2025, 45 (18): 7193- 7206. DOI
20	LIU Y, HU T G, ZHANG H R, et al. iTransformer: Inverted transformers are effective for time series forecasting[C]//The Twelfth International Conference on Learning Representations.
21	LIANG Y X, WEN H M, NIE Y Q, et al. Foundation models for time series analysis: a tutorial and survey[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Barcelona Spain. ACM, 2024: 6555–6565.
22	XUE H, SALIM F D. PromptCast: a new prompt-based learning paradigm for time series forecasting[J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36 (11): 6851- 6864. DOI
23	WANG X L, FENG M K, QIU J, et al. From news to forecast: integrating event analysis in LLM-based time series forecasting with reflection[J]. Advances in Neural Information Processing Systems, 2024, 37, 58118- 58153.
24	TANG H, ZHANG C, JIN M Y, et al. Time series forecasting with LLMs: understanding and enhancing model capabilities[J]. ACM SIGKDD Explorations Newsletter, 2025, 26 (2): 109- 118. DOI
25	ZHENG L N, DONG C, ZHANG W E, et al. Understanding why large language models can be ineffective in time series analysis: the impact of modality alignment[C]//Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. Toronto ON, Canada. ACM, 2025: 4026–4037.
26	JIN M, WANG S Y, MA L T, et al. Time-LLM: Time series forecasting by reprogramming large language models[C]//The Twelfth International Conference on Learning Representations.
27	ELIA. Photovoltaic power production estimation and forecast on Belgian grid (Historical)[EB/OL]. (2025-02-15)[2025-05-27]. https://opendata.elia.be/explore/dataset/ods032/information.
28	OPEN-METEO. Weather forecast API[EB/OL]. (2025-02-15)[2025-05-27]. https://open-meteo.com/en/docs.
29	刘洋, 于海东, 刘文彬, 等. 基于DTW-两阶四分位的分布式光伏发电异常数据辨识[J]. 热力发电, 2024, 53(7): 34–44.
	LIU Yang, YU Haidong, LIU Wenbin, et al. Abnormal data identification for distributed photovoltaic generation based on DTW and two-stage quartile[J]. Thermal Power Generation, 2024, 53(7): 163–171.
30	DIEBOLD F X, MARIANO R S. Comparing predictive accuracy[J]. Journal of Business & Economic Statistics, 2002, 20 (1): 134- 144.
31	LIM B, ARIK S Ö, LOEFF N, et al. Temporal fusion transformers for interpretable multi-horizon time series forecasting[J]. International Journal of Forecasting, 2021, 37 (4): 1748- 1764. DOI

[1]	张金营, 王哲峰, 谢华, 么长英, 闵艳丽, 王新颖. 基于知识图谱与大语言模型的电力行业知识检索分析系统研发与应用[J]. 中国电力, 2024, 57(12): 198-205.
[2]	刘昳娟, 陈云龙, 刘继彦, 张雪梅, 吴潇雨, 孔维政. 基于集成学习的分布式光伏发电功率日前预测[J]. 中国电力, 2022, 55(9): 38-45.

基于气象协变量注意力和大语言模型的日前光伏出力预测

Day-Ahead Photovoltaic Power Forecasting Based on Large Language Model with Meteorological Covariate Attention

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 31

相关文章 2

编辑推荐

Metrics

模态框（Modal）标题

基于气象协变量注意力和大语言模型的日前光伏出力预测

Day-Ahead Photovoltaic Power Forecasting Based on Large Language Model with Meteorological Covariate Attention

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 31

相关文章 2

编辑推荐

Metrics