针对光伏发电功率波动性强,传统功率预测方法难以满足高精度需求,提出一种基于多时序特征融合的TBiLSTM混合网络模型。该模型首先对输入的多维时间序列特征归一化处理,并引入可学习的位置编码以保留时序信息。其次,将预处理后的特征矩阵输入Transformer编码器,利用多头自注意力机制建模跨时间步的全局依赖关系。编码器输出经残差连接与层归一化后,送入双向长短期记忆网络（BiLSTM）,通过前向与后向传播提取局部时序动态特征。最后,BiLSTM的隐藏状态通过全连接层映射为预测功率序列。实验采用两组公开数据集,涵盖辐照度、温度、压强、湿度等多维时序特征。实验结果表明,与BiLSTM、CNN、CNN-BiLSTM模型相比,所提模型的平均绝对误差平均降低了37.16%、47.63%、12.43%,平均绝对百分比误差平均降低了55.59%、68.80%、8.79%,均方误差平均降低了52.11%、68.47%、20.98%,均方根误差平均降低了33.75%、45.49%、11.13%,拟合系数平均提高了4.03%、4.03%、2.13%,验证了该模型的可靠性与优越性。
Abstract
Aiming at the strong power volatility of photovoltaic power generation and the fact that traditional power prediction methods are difficult to meet the high-precision requirements, a TBiLSTM hybrid network model based on multi-temporal sequence feature fusion is proposed. Firstly,this model normalizes the input multi-dimensional temporal sequence features and introduces learnable positional encoding to retain the sequential information of the time series. Secondly,the preprocessed feature matrix is input into the Transformer encoder,and the multi-head self-attention mechanism is used to model the global dependencies among time steps. After the output of the encoder undergoes residual connection and layer normalization,it is fed into the Bidirectional Long Short-Term Memory Network （BiLSTM）,and the BiLSTM extracts the local temporal dynamic characteristics through forward and backward propagation. Finally,the hidden states of the BiLSTM are mapped to the predicted power sequence through the fully connected layer. Two groups of public datasets are used in the experiment,covering multi-dimensional temporal sequence features such as irradiance,temperature,pressure,and humidity. The experimental results show that compared with the BiLSTM,CNN,and CNN-BiLSTM prediction models,the proposed model reduces the mean absolute error by an average of 37.16%,47.63%,and 12.43%,the mean absolute percentage error by an average of 55.59%,68.80%,and 8.79%, the mean squared error by an average of 52.11%,68.47%,and 20.98%,the root mean squared error by an average of 33.75%,45.49%,and 11.13%, and increases the coefficient of determination by an average of 4.03%,4.03%,and 2.13%. This verifies the reliability and superiority of the proposed model.
关键词
光伏发电 /
功率预测 /
时间序列 /
多头自注意力机制 /
Transformer /
BiLSTM
Key words
photovoltaic power generation /
power forecasting /
time series /
multi-head self-attention mechanism /
transformer /
BiLSTM
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] 尤海侠. 光伏发电效率影响因素分析[J]. 能源技术与管理, 2022, 47(6): 147-149.
You H X.Analysis on influencing factors of PV power generation efficiency[J]. Energy Technology and Management, 2022, 47(6): 147-149.
[2] 韩方杰. 基于SVMD-KPCA-BiGRU-Transformer的光伏发电功率预测研究[J]. 东莞理工学院学报, 2025, 32(1): 57-67, 134.
Han F J.Research on photovoltaic power generation prediction based on SVMD-KPCA-BiGRU-Transformer[J]. Journal of Dongguan University of Technology, 2025, 32(1): 57-67, 134.
[3] 陈大为, 张玮, 慕龙. 基于物联网的光伏发电功率预测系统设计[J]. 智能物联技术, 2024, 56(5): 88-91.
Chen D W, Zhang W, Mu L.Design of photovoltaic power generation prediction system based on Internet of Things[J]. Technology of IoT & AI, 2024, 56(5): 88-91.
[4] 黄博阳, 何肖蒙, 肖小兵, 等. 基于FCM-WS-BP的光伏日前出力预测研究[J]. 控制工程, 2023, 30(12): 2254-2260.
Huang B Y, He X M, Xiao X B, et al.Research of photovoltaic day-ahead output power prediction based on FCM-WS-BP[J]. Control engineering of China, 2023, 30(12): 2254-2260.
[5] 常青松, 杨昭, 杨熠辉, 等. 基于相似日聚类的超短期光伏功率组合预测模型[J]. 热力发电, 2023, 52(11): 123-131.
Chang Q S, Yang Z, Yang Y H, et al.Ultrashort term photovoltaic power combinatorial forecasting model based on similar day clustering[J]. Thermal Power Generation, 2023, 52(11): 123-131.
[6] 李国庆, 李欣彤, 边竞, 等. 基于每时晴空指数的大规模光伏电站出力多维时间序列模拟[J]. 电网技术, 2020, 44(9): 3254-3262.
Li G Q, Li X T, Bian J, et al.Multi-dimensional time series simulation of large-scale photovoltaic power plant output based on hourly clear sky index[J]. Power System Technology, 2020, 44(9): 3254-3262.
[7] 路志英, 任一墨, 葛路琨. 基于样条估计分位数回归的光伏功率回归模型[J]. 湖南大学学报(自然科学版), 2017, 44(10): 91-98.
Lu Z Y, Ren Y M, Ge L K.Photovoltaic power regression model based on spline estimation and quantile regression[J]. Journal of Hunan University (Natural Sciences), 2017, 44(10): 91-98.
[8] 魏嫽嫽, 崔承刚, 杨宁, 等. 光伏发电功率预测斜面辐射组合模型的评估研究[J]. 可再生能源, 2018, 36(6): 842-849.
Wei L L, Cui C G, Yang N, et al.Evaluation study on combination models of solar irradiance on inclined surfaces in the PV power forecasting[J]. Renewable Energy Resources, 2018, 36(6): 842-849.
[9] Scolari E, Reyes-Chamorro L, Sossan F, et al.A comprehensive assessment of the short-term uncertainty of grid-connected PV systems[J]. IEEE Transactions on Sustainable Energy, 2018, 9(3): 1458-1467.
[10] 贾凌云, 云斯宁, 赵泽妮, 等. 神经网络短期光伏发电预测的应用研究进展[J]. 太阳能学报, 2022, 43(12): 88-97.
Jia L Y, Yun S N, Zhao Z N, et al.Recent progress of short-term forecasting of photovoltaic generation based on artificial neural networks[J]. Acta Energiae Solaris Sinica, 2022, 43(12): 88-97.
[11] 姚宏民, 杜欣慧, 秦文萍. 基于密度峰值聚类及GRNN神经网络的光伏发电功率预测方法[J]. 太阳能学报, 2020, 41(9): 184-190.
Yao H M, Du X H, Qin W P.PV power forecasting approach based on density peaks clustering and general regression neural network[J]. Acta Energiae Solaris Sinica, 2020, 41(9): 184-190.
[12] 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.
Peng S R, Chen H X, Sun W T, et al.Research on photovoitaic power prediction method based on improved lstm[J]. Acta energiae Solaris Sinica, 2024, 45(11): 296-302.
[13] Wang Y, Shen Y X, Mao S W, et al.LASSO and LSTM integrated temporal model for short-term solar intensity forecasting[J]. IEEE Internet of Things Journal, 2019, 6(2): 2933-2944.
[14] 马磊, 黄伟, 李克成, 等. 基于Attention-LSTM的光伏超短期功率预测模型[J]. 电测与仪表, 2021, 58(2): 146-152.
Ma L, Huang W, Li K C, et al.Photovoltaic ultra-short-term power prediction model based on Attention-LSTM[J]. Electrical Measurement & Instrumentation, 2021, 58(2): 146-152.
[15] 李超然, 潘鹏程, 杨伟荣, 等. 基于改进相似日优化HBA-BiLSTM-KELM的光伏发电功率预测[J]. 太阳能学报, 2024, 45(5): 508-516.
Li C R, Pan P C, Yang W R, et al.Research on pv system power prediction based on improved similar day and HBA-BiLSTM-Kelm neural network[J]. Acta Energiae Solaris Sinica, 2024, 45(5): 508-516.
[16] 李生, 于淏. 基于POA-GRU模型的光伏发电功率预测研究[J]. 电力电子技术, 2025, 59(2): 74-79, 87.
Li S, Yu H.Research on photovoltaic power generation forecasting based on the POA-GRU model[J]. Power Electronics, 2025, 59(2): 74-79, 87.
[17] 张静, 熊国江. 考虑季节特性与数据窗口的短期光伏功率预测组合模型[J]. 电力工程技术, 2025, 44(1): 183-192.
Zhang J, Xiong G J.Short-term photovoltaic power prediction combination model considering seasonal characteristic and data window[J]. Electric Power Engineering Technology, 2025, 44(1): 183-192.
[18] Vaswani A, Shazeer N, Parmar N, et al.Attention is all you need[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. 2017: 6000-6010.
[19] 焦丕华, 蔡旭, 王乐乐, 等. 考虑数据分解和进化捕食策略的BiLSTM短期光伏发电功率预测[J]. 太阳能学报, 2024, 45(2): 435-442.
Jiao P H, Cai X, Wang L L, et al.BiLSTM short-term photovoltaic power prediction considering data decomposition and evolutionary predation strategies[J]. Acta Energiae Solaris Sinica, 2024, 45(2): 435-442.
基金
国家自然科学基金（62241603）; 装备智能运用教育部重点实验室开放基金（AAIE-2023-0403）
PDF(1508 KB)
