基于KAN-MHA和非对称共形分位数回归的光伏功率短期区间预测

doi:10.19595/j.cnki.1000-6753.tces.251649

Abstract
Figure/Table
References
Related Citation (15)

Download: PDF (3511 KB) HTML (1 KB)
Export: BibTeX | EndNote (RIS)

Abstract First, this paper systematically reviews the basic principles of conformal prediction (CP), quantile regression-based prediction (QR), and conformal quantile regression (CQR). It is noted that when the original QR prediction intervals exhibit pronounced asymmetric bias, the symmetric calibration strategy used by CQR unnecessarily inflates the interval width. To address this issue, this paper proposes an asymmetric conformal quantile regression method (ACQR). In ACQR, the calibration residuals are split into two sets of nonconformity scores for the lower and upper bounds, respectively. Correction terms are then extracted by quantile and used to asymmetrically shift the bounds. In this way, the method aims to eliminate asymmetric bias and shrink the ineffective parts of the interval, while satisfying the desired coverage level.
This paper discusses the impact of the QR model’s fitting capability on the width of the final prediction interval. Improving the QR model’s fitting accuracy is a prerequisite for obtaining tight prediction intervals, thereby motivating the proposed KAN-MHA model. The structure of KAN-MHA is then described: using KAN as the basic building block, the model comprises a spatial feature-extraction module, a temporal feature-extraction module, and a fully connected layer that fuses spatial and temporal features to produce the predicted power.
Results from two real-world cases demonstrate that the proposed KAN-MHA+ACQR method achieves excellent prediction accuracy and interval tightness. In Case 1, meteorological features are strongly correlated with PV power, indicating that numerical weather prediction is relatively accurate. At the 80% and 90% confidence levels for QR-based prediction, the mean interval width (PINAW) of KAN-MHA is significantly smaller than those of the CNN-BiLSTM, LagBoostNet, ASMNet, and KAN models, achieving approximately 38%~47% interval compression. In Case 2, the correlation between meteorological features and power is weaker, representing scenarios with larger numerical weather prediction errors. Although the overall prediction accuracy of all models decreases, KAN-MHA still achieves the best performance. A further comparison between ACQR and CQR shows that in the vast majority of experimental settings, ACQR yields lower PINAW than CQR, and in the best case, PINAW can be reduced to 90.6% of that obtained by CQR. Visualization results indicate that, compared with CQR, ACQR’s upper and lower bounds of the prediction intervals are more symmetric and better aligned with the actual power distribution.
The following conclusions can be drawn. (1) Compared with CQR, the proposed ACQR method can effectively reduce the prediction interval width while ensuring the required coverage, thereby alleviating the interval inflation problem caused by asymmetric bias. (2) The proposed KAN-MHA model achieves tighter and more reliable prediction intervals across different case studies. (3) The combination of the KAN-MHA model with the ACQR method provides reliable and compact short-term PV power prediction intervals. This paper offers an effective and valuable new approach for engineering applications of short-term interval prediction of PV power.

Key words： Photovoltaic power range prediction Kolmogorov-Arnold network multi-head attention mechanism asymmetric conformal quantile regression

Received: 22 September 2025

PACS:

TM614

	Service

	E-mail this article
	Add to my bookshelf
	Add to citation manager
	E-mail Alert
	RSS
	Articles by authors
	Li Song
	Xia Chengjun
	Liu Yifu
	Wang Yukang

Cite this article:

Li Song,Xia Chengjun,Liu Yifu等. Photovoltaic Power Interval Prediction Based on KAN-MHA and Asymmetric Conformal Quantile Regression[J]. Transactions of China Electrotechnical Society, 2026, 41(18): 6336-6350.

URL:

https://dgjsxb.ces-transaction.com/EN/10.19595/j.cnki.1000-6753.tces.251649 OR https://dgjsxb.ces-transaction.com/EN/Y2026/V41/I18/6336

[1] 杨明, 李梦林, 王勃, 等. 面向新型电力系统的数值天气预报技术及应用综述[J]. 电力系统自动化, 2025, 49(17): 1-20.
Yang Ming, Li Menglin, Wang Bo, et al.Review on technologies and applications of numerical weather prediction for new power system[J]. Automation of Electric Power Systems, 2025, 49(17): 1-20.
[2] 彭春华, 张浩旗, 孙惠娟, 等. 基于延迟感知多智能体深度强化学习的多光储直柔系统优化调度[J]. 电工技术学报, 2026, 41(11): 3742-3754.
Peng Chunhua, Zhang Haoqi, Sun Huijuan, et al.Optimal scheduling of multi PV-storage-DC-flexible systems based on delay-aware multi-agent deep reinforcement learning[J]. Transactions of China Electrotechnical Society, 2026, 41(11): 3742-3754.
[3] 杨茂, 张书天, 王勃, 等. 基于混合分位数回归长短期记忆神经网络的风电功率短期区间预测[J]. 太阳能学报, 2025, 46(2): 582-590.
Yang Mao, Zhang Shutian, Wang Bo, et al.Short- term wind power interval prediction based on mixed quantile regression long and short-term memory neural network[J]. Acta Energiae Solaris Sinica, 2025, 46(2): 582-590.
[4] 柳天虹, 齐胜利, 裔扬, 等. 基于分位数回归的改进权重GRU风电功率区间预测[J]. 太阳能学报, 2024, 45(12): 291-298.
Liu Tianhong, Qi Shengli, Yi Yang, et al.Improved weighted GRU wind power interval prediction based on quantile regression[J]. Acta Energiae Solaris Sinica, 2024, 45(12): 291-298.
[5] 赵耀, 高少炜, 李东东, 等. 基于天气相似聚类与QRNN的短期光伏功率区间概率预测[J]. 电力系统自动化, 2023, 47(23): 152-161.
Zhao Yao, Gao Shaowei, Li Dongdong, et al.Short-term interval probability prediction of photo- voltaic power based on weather similarity clustering and quantile regression neural network[J]. Auto- mation of Electric Power Systems, 2023, 47(23): 152-161.
[6] 张昆明, 马龙义, 章天晗, 等. 基于iDGA-LSTM的分布式光伏出力时空协同概率预测方法[J]. 电力系统自动化, 2025, 49(1): 128-138.
Zhang Kunming, Ma Longyi, Zhang Tianhan, et al.Spatio-temporal cooperative probability forecasting method for distributed photovoltaic output based on iDGA-LSTM[J]. Automation of Electric Power Systems, 2025, 49(1): 128-138.
[7] 杨茂, 张书天, 王勃, 等. 基于门控循环加权共形分位数回归的风电功率短期区间预测[J]. 中国电机工程学报, 2025, 45(19): 7565-7574.
Yang Mao, Zhang Shutian, Wang Bo, et al.Short- term wind power interval prediction based on weighted gated recurrent conformalized quantile regression[J]. Proceedings of the CSEE, 2025, 45(19): 7565-7574.
[8] 彭依明, 王佳, 周长城, 等. 基于BiLSTM-SA的分布式光伏功率超短期概率预测[J]. 南方电网技术, 2025, 19(11): 172-181.
Peng Yiming, Wang Jia, Zhou Changcheng, et al.Ultra-short-term probability prediction of distributed photovoltaic power based on BiLSTM-SA[J]. Southern Power System Technology, 2025, 19(11): 172-181.
[9] 彭昱, 符琛, 郭昕, 等. 基于FCM-BOA-TCN-GRU的分布式光伏出力异常检测方法[J]. 电工技术学报, 2025, 40(17): 5389-5401.
Peng Yu, Fu Chen, Guo Xin, et al.Distributed photovoltaic power output anomaly detection method based on FCM-BOA-TCN-GRU[J]. Transactions of China Electrotechnical Society, 2025, 40(17): 5389-5401.
[10] Liu Ziming, Wang Yixuan, Vaidya S, et al. KAN: Kolmogorov-Arnold networks[J]. arXiv:2404.19756, 2024.
[11] 刘栋, 郭国栋, 辛蜀骏, 等. 基于KAN的可解释净负荷概率预测方法[J]. 电力系统自动化, 2025, 49(15): 123-132.
Liu Dong, Guo Guodong, Xin Shujun, et al.Probabilistic prediction method for interpretable net load based on Kolmogorov-Arnold network[J]. Automation of Electric Power Systems, 2025, 49(15): 123-132.
[12] 成润坤, 王辉, 刘达, 等. 融合RSDE框架与KAN算法的现货电价异质模型集成预测[J]. 中国电机工程学报, 2024, 44(24): 9645-9657.
Cheng Runkun, Wang Hui, Liu Da, et al.Ensemble prediction of spot electricity prices using hetero- geneous models by integrating the RSDE framework and KAN algorithm[J]. Proceedings of the CSEE, 2024, 44(24): 9645-9657.
[13] 桂前进, 徐文法, 李晓阳, 等. 基于时序分解与共形分位数回归的超短期光伏功率区间预测[J]. 中国电力, 2025, 58(5): 21-32.
Gui Qianjin, Xu Wenfa, Li Xiaoyang, et al.Ultra- short-term photovoltaic power interval forecasting based on time-series decomposition and conformal quantile regression[J]. Electric Power, 2025, 58(5): 21-32.
[14] 焦治杰, 王小君, 刘曌, 等. 考虑分布式新能源出力不确定性的虚拟电厂概率可行域构建方法[J]. 电力系统自动化, 2024, 48(18): 129-138.
Jiao Zhijie, Wang Xiaojun, Liu Zhao, et al.Con- struction method for probability-based feasible region of virtual power plant considering uncertainty of distributed renewable energy output[J]. Automation of Electric Power Systems, 2024, 48(18): 129-138.
[15] Desert Knowledge Australia Solar Centre. DKA solar Centre[EB/OL]. (2025-07-09)[2025-08-29]. https://dkasolarcentre.com.au/.
[16] 国能日新第二届光伏功率预测赛[EB/OL]. (2024- 09-06)[2025-08-29]. https://gitcode.com/open-source-toolkit/6e0d3.
[17] 李楠, 王鹤潼, 张光耀, 等. 基于双重分解策略与MSTCN-BiLSTM的超短期负荷预测[J/OL]. 电工技术学报, 1-19[2025-12-15].
Li Nan, Wang Hetong, Zhang Guangyao, et al.Short- term load forecasting based on dual decomposition strategy and MSTCN-BiLSTM[J/OL]. Transactions of China Electrotechnical Society, 1-19[2025-12-15].
[18] 杨朵, 吕浩然, 李民策, 等. 基于双向长短期记忆网络和注意力机制的质子交换膜燃料电池衰退行为预测[J]. 电工技术学报, 2026, 41(2): 701-713, 724.
Yang Duo, Lü Haoran, Li Mince, et al.Prediction of degradation behavior in proton exchange membrane fuel cells based on bi-long short-term memory and attention mechanism[J]. Transactions of China Elec- trotechnical Society, 2026, 41(2): 701-713, 724.
[19] 高爽, 李辰昊, 李泽宇, 等. 考虑波浪作用和时序特性的海上漂浮式光伏阵列故障诊断方法[J]. 电工技术学报, 2026,41(3): 999-1011.
Gao Shuang, Li Chenhao, Li Zeyu, et al.Fault diagnosis method for offshore floating photovoltaic array considering wave effects and temporal chara- cteristics[J]. Transactions of China Electrotechnical Society, 2026,41(3): 999-1011.
[20] Ma Qiaoyu, Fu Xueqian, Yang Qiang, et al.Adaptive masked network for ultra-short-term photovoltaic forecast[J]. Engineering Applications of Artificial Intelligence, 2025, 139: 109555.
[21] Yang Feifei, Fu Xueqian, Li Zhengshuo, et al.A novel clustering-ensemble learning model for day- ahead photovoltaic power forecasting[J]. Engineering Applications of Artificial Intelligence, 2025, 161: 112281.