Ultra-short-term Photovoltaic Power Prediction Based on Improved PatchTST Considering Data Drift

doi:10.16182/j.issn1004731x.joss.25-0396

Abstract

Abstract:

Existing PV power prediction methods often suffer from limited accuracy and robustness due to three key shortcomings: relying on single-point mapping that cannot fully extract local temporal patterns; inadequate exploration of the global temporal dependencies in PV output, and failure to account for prevalent data drift phenomena. To overcome these limitations, an improved patch time series transformer (PatchTST) based approach is proposed for ultra-short-term PV power prediction. The methodology applies rough set theory for feature dimensionality reduction, effectively preserving critical decision information by analyzing both feature-label relationships and inter-feature correlations. An enhanced PatchTST model with a modified channel-independent mechanism extracts local temporal patterns and captures complex mapping relationships between meteorological variables and PV output. The global-local attention fusion framework (GLAFF) module is incorporated to identify global temporal guidance patterns. An adaptive weighting mechanism dynamically optimizes the integration of local and global features to counteract data drift effects. Experimental validation using real-world datasets from a PV power plant in Inner Mongolia, China, and the Desert Knowledge Australia Solar Centre demonstrates that the proposed model achieves superior performance in both prediction accuracy and robustness.

Key words: photovoltaic power prediction, ultra-short-term prediction, patch time series transformer (PatchTST), global-local attention fusion framework (GLAFF), data drift

CLC Number:

TP391.9

Mei Huawei, Yang Penghui, Yu Yang. Ultra-short-term Photovoltaic Power Prediction Based on Improved PatchTST Considering Data Drift[J]. Journal of System Simulation, 2026, 38(5): 1239-1254.

Figures/Tables 22

Fig. 1

Table 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Fig. 6

Fig. 7

Fig. 8

Table 2

Table 3

Fig. 9

Table 4

Fig. 10

Fig. 11

Table 5

Fig. 12

Table 6

Fig. 13

Table 7

Fig. 14

Fig. 15

References 17

[1]	国家能源局.国家能源局发布2024年全国电力工业统计数据[EB/OL].(2025-01-21)[2025-05-08]. .
[2]	Xia Yurui, Wang Jianzhou, Zhang Ziyuan, et al. Short-term PV Power Forecasting Based on Time Series Expansion and High-order Fuzzy Cognitive maps[J]. Applied Soft Computing, 2023, 135: 110037.
[3]	董存, 王铮, 白捷予, 等. 光伏发电功率超短期预测方法综述[J]. 高电压技术, 2023, 49(7): 2938-2951.
	Dong Cun, Wang Zheng, Bai Jieyu, et al. Review of Ultra-short-term Forecasting Methods for Photovoltaic Power Generation[J]. High Voltage Engineering, 2023, 49(7): 2938-2951.
[4]	王玲芝, 李晨阳, 刘婧, 等. 基于GRO-SSA-LSTM的短期光伏发电功率预测[J]. 太阳能学报, 2025, 46(2): 401-409.
	Wang Lingzhi, Li Chenyang, Liu Jing, et al. Short Term Photovoltaic Power Prediction Based on GRO-SSA-LSTM[J]. Acta Energiae Solaris Sinica, 2025, 46(2): 401-409.
[5]	汪繁荣, 梅涛, 卢璐. 基于相似日聚类和VMD-LTWDBO-BiLSTM的短期光伏功率预测[J]. 智慧电力, 2024, 52(10): 56-63, 111.
	Wang Fanrong, Mei Tao, Lu Lu. Short-term PV Power Prediction Based on Similar Day Clustering with VMD-LTWDBO-BiLSTM[J]. Smart Power, 2024, 52(10): 56-63, 111.
[6]	李庆生, 张裕, 龙家焕, 等. 基于IMFO-LSTM模型的光伏功率预测研究[J]. 电力科学与技术学报, 2024, 39(3): 199-206.
	Li Qingsheng, Zhang Yu, Long Jiahuan, et al. Photovoltaic Power Prediction Based on IMFO-LSTM Model[J]. Journal of Electric Power Science and Technology, 2024, 39(3): 199-206.
[7]	Zhao Yonghui, Peng Xunhui, Tu Teng, et al. WOA-VMD-SCINet: Hybrid Model for Accurate Prediction of Ultra-short-term Photovoltaic Generation Power Considering Seasonal Variations[J]. Energy Reports, 2024, 12: 3470-3487.
[8]	Khelifi Reski, Guermoui Mawloud, Rabehi Abdelaziz, et al. Short-term PV Power Forecasting Using a Hybrid TVF-EMD-ELM Strategy[J]. International Transactions on Electrical Energy Systems, 2023, 2023(1): 6413716.
[9]	Liu Qian, Li Yulin, Jiang Hang, et al. Short-term Photovoltaic Power Forecasting Based on Multiple Mode Decomposition and Parallel Bidirectional Long Short Term Combined with Convolutional Neural Networks[J]. Energy, 2024, 286: 129580.
[10]	吐松江⋅卡日, 雷柯松, 马小晶, 等. 基于LSTM-attention和CNN-BiGRU误差修正的光伏功率预测[J]. 太阳能学报, 2024, 45(12): 85-93.
	Kari·Tusongjiang, Lei Kesong, Ma Xiaojing, et al. Photovoltaic Power Prediction Based on LSTM-attention and CNN-BiGRU Error Correction[J]. Acta Energiae Solaris Sinica, 2024, 45(12): 85-93.
[11]	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.
	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.
[12]	董俊, 刘瑞, 束洪春, 等. 基于BIRCH聚类的L-transformer分布式光伏短期发电功率预测[J]. 高电压技术, 2024, 50(9): 3883-3893.
	Dong Jun, Liu Rui, Shu Hongchun, et al. Short-term Distributed Photovoltaic Power Generation Prediction Based on BIRCH Clustering and L-transformer[J]. High Voltage Engineering, 2024, 50(9): 3883-3893.
[13]	Pawlak Zdzisław. Rough Sets[J]. International Journal of Computer and Information Sciences, 1982, 11(5): 341-356.
[14]	Nie Yuqi, Nguyen N H, Sinthong P, et al. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers[EB/OL].(2023-03-05)[2025-05-08]..
[15]	Han Lu, Ye Hanjia, Zhan Dechuan. The Capacity and Robustness Trade-off: Revisiting the Channel Independent Strategy for Multivariate Time Series Forecasting[J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36(11): 7129-7142.
[16]	Wang Chengsen, Qi Qi, Wang Jingyu, et al. Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-local Fusion Perspective[EB/OL].(2024-11-20)[2025-05-08]..
[17]	Kim T, Kim J, Tae Y, et al. Reversible Instance Normalization for Accurate Time-series Forecasting Against Distribution Shift[C]//International Conference on Learning Representations. [S.l. : s.n.], 2021.

特征	分箱数	切割点
降雨量/mm	2	(0)
风向/(°)	4	(90,180,270)
风速/(m/s)	3	(3,6)
温度/(℃)	3	(10,25)
相对湿度/%	3	(40,70)
总水平辐照度/(W/m²)	3	(200,500)
散射水平辐照度/(W/m²)	3	(100,300)
功率/kW	3	(1,7)

模型名	RMSE/kW	MAE/kW	R²
*PatchTST-GLAFF**	0.650 7	0.248 5	0.958 4
PatchTST-GLAFF	0.668 9	0.249 4	0.956 0
PatchTST*	0.658 8	0.249 0	0.957 3
PatchTST	0.670 9	0.252 3	0.955 9

模型类别	模型名	RMSE/kW	MAE/kW	R²
循环神经网络	RNN	0.723 8	0.311 5	0.948 5
	LSTM	0.694 9	0.283 8	0.952 5
	GRU	0.677 4	0.273 4	0.954 9
卷积神经网络	CNN	0.918 8	0.486 3	0.917 0
卷积神经网络	TCN	0.849 2	0.424 7	0.929 1
集成树模型	RF	0.688 5	0.269 2	0.953 4
	XGBoost	0.706 3	0.274 8	0.951 0
	LightGBM	0.685 0	0.286 7	0.953 9
Transformer架构	Transformer	0.675 9	0.257 9	0.955 1
	Informer	0.673 0	0.311 1	0.955 6
	Autoformer	0.812 9	0.410 4	0.935 2
线性模型	Linear	0.921 7	0.538 4	0.916 7
	NLinear	0.745 0	0.328 1	0.945 6
	DLinear	0.800 8	0.382 5	0.937 1
现有研究	CNN-LSTM	0.743 7	0.339 8	0.945 6
	VMD-LSTM	0.727 8	0.357 7	0.947 9
	K-means-LSTM	0.693 2	0.285 5	0.952 8
	LSTM-PSO	0.678 8	0.270 4	0.954 7
	LSTM-BiGRU	0.675 1	0.361 3	0.955 2
	LSTM-BiGRU-PSO	0.681 6	0.288 0	0.954 3
	K-means-LSTM-BiGRU	0.679 6	0.362 4	0.954 6
PatchTST	PatchTST*-GLAFF	0.650 7	0.248 5	0.958 4

特征	KS-D
温度	0.408
功率	0.307
风速	0.155
总水平辐照度	0.130
相对湿度	0.110
散射水平辐照度	0.096

模型名	RMSE/kW	MAE/kW	R²
PatchTST*-GLAFF	0.922 8	0.266 9	0.990 4
PatchTST-GLAFF	0.953 4	0.275 6	0.989 8
PatchTST*	0.926 0	0.274 4	0.990 2
PatchTST	0.957 5	0.286 7	0.989 7