基于TCN和DLinear的光伏发电功率多步预测模型  

Multistep Prediction Model for Photovoltaic Power Generation Based on Time Convolution and DLinear

作　　者：王舒雨 李豪[1] 马刚[1] 袁宇波 卜强生 叶志刚 WANG Shuyu;LI Hao;MA Gang;YUAN Yubo;BU Qiangsheng;YE Zhigang(School of Electricity and Automation,Nanjing Normal University,Nanjing 210046,China;Jiangsu Electric Power Testing and Research Institute Co.,Ltd.,Nanjing 211100,China)

机构地区：[1]南京师范大学电气与自动化学院,南京市210046 [2]江苏省电力试验研究院有限公司,南京市211100

出　　处：《电力建设》2025年第4期173-184,共12页Electric Power Construction

基　　金：江苏省科技项目(BE2022003-5)。

摘　　要：【目的】光伏发电功率预测是提高太阳能利用效率和降低运营成本的关键技术。然而,传统模型在多步光伏发电功率预测中存在时间趋势学习能力不足和误差累积的问题,限制了预测精度的提升。【方法】文章提出了一种基于时间卷积神经网络(temporal convolutional networks,TCN)和DLinear的光伏发电功率多步预测模型。首先,通过改进的自适应白噪声完全集合经验模态分解(improved complete ensemble empirical mode decomposition with adaptive noise,ICEEMDAN)对多元气象序列进行分解,揭示其潜在特征,得到更易学习多尺度特征的多维子序列。其次,利用TCN对局部时序信息进行建模,挖掘短期内的时序特征。最后,利用DLinear将序列分解为趋势分量和残差分量,通过线性网络学习多尺度特征,并直接输出多步(每步15 min)光伏发电功率预测结果。【结果】实验结果表明,所提方法的每个模块均能显著提升模型的预测性能。与集合经验模态分解-卷积神经网络(convolutional neural networks,CNN)-双向长短期记忆网络(bidirectional long short-term memory,BiLSTM)(ICEEMDAN-CNN-BiLSTM)、Informer、Autoformer三个基准模型相比,归一化均方根误差(normalized root mean square error,NRMSE)平均分别下降了22.455%、6.139%、8.504%,具有明显优势。【结论】文章通过ICEEMDAN和TCN-DLinear组合模型,有效解决了传统方法在多步预测中的不足,显著提高了光伏发电功率预测的准确性和可靠性。研究结果为光伏发电功率的精准预测提供了新的技术路径,为太阳能发电的高效管理和运营提供了理论支持,为新型电力系统的安全稳定运行提供了数据支持。未来可进一步探索该模型在不同气象条件和地理环境下的泛化能力,以推动太阳能发电技术的发展。[Objective]Photovoltaic power forecasting is a key technology to improve the efficiency of solar energy utilization and reduce operating costs.However,the traditional model has the problem of insufficient time-trend learning ability and error accumulation in multistep photovoltaic power prediction,which limits the improvement of prediction accuracy.[Methods]This paper presents a multistep prediction model for photovoltaic power generation based on a temporal convolutional network(TCN)and DLinear combined model.First,it improves the complete ensemble empirical mode decomposition with adaptive noise and ICEEMDAN decomposes multivariate meteorological sequences to reveal their potential features and obtain multidimensional subsequences that make it easier to learn multiscale features.Second,the TCN is used to model the local time sequence information and mine short-time sequence features.Finally,DLinear decomposes the sequence into trend and residual components,learns multiscale features through linear networks,and directly outputs multistep(four-step prediction,15 min per step)photovoltaic power prediction results.[Results]Experimental results show that each module of the proposed method can significantly improve the prediction performance of the model.Compared with ICEEMDAN-CNN-BiLSTM,Informer,and Autoformer,the normalized root mean square error(NRMSE)decreased by 22.455%,6.139%,and 8.504%on average,respectively,with obvious advantages.[Conclusions]Through the improved ICEEMDAN decomposition and TCN-DLinear combined model,this study effectively solves the shortcomings of traditional methods in multistep prediction and significantly improves the accuracy and reliability of photovoltaic power prediction.The research results provide a new technical path for the accurate prediction of photovoltaic power generation,theoretical support for the efficient management and operation of solar power generation,and data support for the safe and stable operation of new power systems.In the future,the generalizability of this mod

关 键 词：改进的自适应白噪声完全集合经验模态分解 时间卷积神经网络 DLinear 光伏预测 

分 类 号：TM615[电气工程—电力系统及自动化]