1 State Key Laboratory of Green Building, Xi’an University of Architecture and Technology, Xi’an 710055, China
2 School of Science, Xi’an University of Architecture and Technology, Xi’an 710055, China
Show less
文章历史+
Received
Published
Revised
2025-10-27
2026-09-18
2026-02-06
Online Date
2026-10-09
PDF (7947K)
摘要
随着光伏发电占比提升,其功率的精准预测对电网稳定与消纳至关重要。然而,传统时间序列预测方法对光伏功率复杂动态的刻画能力有限,而现有基于图神经网络的时空预测模型在长时序依赖与全局空间相关性的捕捉上仍存在一定挑战,尤其在多步预测中易产生误差累积。为此,提出了一种改进的时空图神经网络模型:MTGNN-BiMamba2-NL。该模型在多元时间序列预测框架(multivariate time series forecasting framework based on graph neural networks, MTGNN)中,以双向状态空间模型(BiMamba2)替代时间卷积,以强化长时序特征提取。同时,在图卷积层后融入非局部(Non-local, NL)注意力机制,以增强站点间的全局空间关联建模。在两个真实光伏数据集上的实验表明,所提模型在多步预测任务中能有效缓解误差累积,其预测精度与稳定性均优于现有基线方法。
Abstract
With the increasing penetration of PV (photovoltaic) power generation, accurate prediction of PV output is essential for grid stability and power accommodation. However, traditional time series forecasting methods exhibit limited capability in characterizing the complex dynamics of PV power, while existing graph neural network based spatiotemporal forecasting models still face certain challenges in capturing long-term temporal dependencies and global spatial correlations, particularly in multi-step forecasting where error accumulation easily occurs. To address these limitations, an improved spatiotemporal graph neural network model, termed MTGNN-BiMamba2-NL, was proposed. Within MTGNN (multivariate time series forecasting framework based on graph neural networks), the temporal convolution module was replaced by a bidirectional state space model (BiMamba2) to enhance long-term temporal feature extraction. Meanwhile, a NL (Non-local) attention was incorporated after the graph convolution layers to strengthen the modeling of global spatial correlations among sites. Experiments were conducted on two real-world PV datasets. The results demonstrate that the proposed model effectively alleviates error accumulation in multi-step forecasting and achieves superior prediction accuracy and stability compared with existing baseline methods.
SHI Jia-rong, WANG Jiao-jiao, LI Xue-juan.
Temporal-spatial Graph Network Model with BiMamba2 and Non-local Attention for Photovoltaic Cluster Power Prediction[J].
Science Technology and Engineering, 2026, 26(26): 11318-11325 DOI:10.12404/j.issn.1671-1815.2507461
然而,现有方法仍存在较大改进空间:在处理极长序列时TCN感受野有限,且无法充分利用双向上下文信息;传统的图卷积操作往往局限于局部邻域信息聚合,难以捕捉全局空间依赖。为此,现对基于图神经网络的多元时间序列预测(multivariate time series forecasting with GNN,MTGNN)[14]架构的卷积模块和注意力机制进行改进,并将其应用于多站点光伏功率预测中,以期能有效缓解多步预测误差累积,增强复杂场景下的预测鲁棒性,为高比例光伏并网调度提供可靠支撑。
基于图神经网络的多元时间序列预测(multivariate time series forecasting with graph neural networks,MTGNN)是一种端到端的深度学习模型。它集成了自适应图学习、图卷积与时间卷积,同时捕捉变量间的空间依赖关系和时间动态特征,且不依赖预定义的图结构,其框架如图3所示。在多站点光伏功率预测任务中,MTGNN自动识别站点间由气象与地理因素驱动的潜在关联,并利用时序特征提取功率数据中的周期与波动模式,从而显著提升多站点协同预测的准确性与鲁棒性。
WangJ Z, YuY, ZengB, et al. Hybrid ultra-short-term PV power forecasting system for deterministic forecasting and uncertainty analysis[J]. Energy, 2024, 288: 129898.
[2]
YangJ, NiuM B. WaveGRU: a framework with frequency-domain spatial attention for accurate solar PV and wind power forecasting[J]. Sustainable Energy Technologies and Assessments, 2025, 83: 104572.
LiangYafeng, MaLihong, QiuJianhong, et al. Short-term photovoltaic power forecasting based on CEEMD-LSTM[J]. Science Technology and Engineering, 2024, 24(13): 5396-5405.
LiuTianhong, QiShengli, YiYang, et al. Interval forecasting of wind power based on an improved weighted GRU with quantile regression[J]. Acta Energiae Solaris Sinica, 2024, 45(12): 291-298.
LiuJunhong, FuSiyuan, WangYajun. Short-term photovoltaic power forecasting based on an optimized TCN ensemble model[J]. Science Technology and Engineering, 2025, 25(15): 6378-6388.
[9]
WangS, SunY H, ZhangW J, et al. Very short-term wind power forecasting considering static data: an improved Transformer model[J]. Energy, 2024, 312: 133577.
[10]
ZhouN, SunL F. A dual-layer state space network for multivariate time series forecasting with dimensional interdependency[J]. Applied Soft Computing, 2025, 183: 113452.
[11]
AlkilaneK, HeY H, LeeD H. Expert denoising for enhanced diffusion-based time series forecasting[J]. Expert Systems with Applications, 2026, 299: 130269.
ZengZeng, XiaoMaoran, XiaYuanyi, et al. A distributed photovoltaic aggregation uncertainty characterization method based on quantile spatiotemporal graph neural network[J]. Science Techno-logy and Engineering, 2024, 24(34): 14726-14733.
[14]
LiuJ L, OngG P, ChenX Q, et al. GraphSAGE-based traffic speed forecasting for segment network with sparse data[J]. IEEE Transactions on Intelligent Transportation Systems, 2022, 23(3): 1755-1766.
[15]
LiuJ C, LiT. Multi-step power forecasting for regional photovol-taic plants based on ITDE-GAT model[J]. Energy, 2024, 293: 130468.
[16]
CaoJ, LiuC Q, ChenC L, et al. A short-term load forecasting method for integrated community energy system based on STGCN[J]. Electric Power Systems Research, 2024, 232: 110265.
[17]
GuoS N, LinY F, FengN, et al. Attention based spatial-temporal graph convolutional networks for traffic flow forecasting[C]// Proceedings of the AAAI Conference on Artificial Intelligence. Palo Alto: AAAI Press, 2019: 922-929.
[18]
LiuR, HuangC Z. Data-driven modeling approach of heavy-duty gas turbine with physical constraint by MTGNN and Transformer[J]. Control Engineering Practice, 2024, 151: 106014.
[19]
JiangT L, LiuW Y, LinJ Q, et al. Research on an hourly heat load forecasting model for district heating systems based on heterogeneous model fusion[J]. Case Studies in Thermal Engineering, 2025, 73: 106500.
[20]
GoJ, RyuJ B. Spatial bias for attention-free non-local neural networks[J]. Expert Systems with Applications, 2024, 238: 122053.