针对确定性光伏功率预测无法计算预测结果概率和波动范围的问题,采用改进时间卷积神经网络（TCNN）开展光伏功率概率预测。TCNN已用于各种时序预测任务,但其在输入序列很长情况下需增加卷积层来提升预测性能。在TCNN中引入稀疏注意力机制,构建注意力时间卷积神经网络（ATCNN）,通过分层卷积结构提取时间依赖关系,利用稀疏注意力关注重要的时间步,构建的稀疏注意力层无需更深的架构即可扩展感受野,并使预测结果更具可解释性。在两个光伏数据集上的功率概率预测结果表明,ATCNN的预测准确性优于TCNN、时间融合解码器（TFT）等先进深度学习模型,同时对于感受野的扩展,ATCNN比TCNN需要的卷积层更少、训练速度更快,并能可视化预测过程中最重要的时间步。同卷积层情况下,ATCNN比TCNN的点预测损失小15.7%,概率预测损失小15.9%。
Abstract
Aiming at the problem that deterministic photovoltaic power forecasting cannot calculate the probability and fluctuation range of forecasting results, the improved temporal convolutional neural network （ TCNN ） is used to forecast the probability of photovoltaic power. TCNN has been used in various time series prediction tasks, but it needs to add the convolution layer to improve the forecasting performance when the input sequence is long. The sparse attention mechanism is introduced into the TCNN to construct the attention temporal convolutional neural network （ATCNN）. The time-dependent is extracted by the hierarchical convolution structure, and the sparse attention is used to focus on the important time step. The constructed sparse attention layer can expand the receptive field without deeper architecture, and make the forecasting results more interpretable. The power probability forecasting results on two photovoltaic data sets show that the forecasting accuracy of ATCNN is better than that of advanced deep learning models such as TCNN and DeepAR. At the same time, for the expansion of receptive field, ATCNN requires fewer convolutional layers than TCNN, has faster training speed, and can visualize the most important time step in the forecasting process.In the case of the same convolutional layer, the point prediction loss of ATCNN is 15.7% smaller than that of TCNN, and the probability pr
关键词
光伏功率 /
预测 /
时间卷积网络 /
稀疏注意力机制 /
可解释性
Key words
photovoltaic power /
forecasting /
temporal convolutional network /
sparse attention mechanism /
interpretability
{{custom_sec.title}}
{{custom_sec.title}}
{{custom_sec.content}}
参考文献
[1] FAN H J, XU W, ZHANG J, et al.Steam temperature regulation characteristics in a flexible ultra-supercritical boiler with a double reheat cycle based on a cell model[J]. Energy, 2021, 229: 120701.
[2] 王瑞, 王英洲, 逯静. 基于ICEEMDAN-DTW和ISMA-WLSSVM的光伏发电功率预测[J]. 热能动力工程, 2023, 38(9): 131-140.
WANG R, WANG Y Z, LU J.Prediction of photovoltaic power generation based on ICEEMDAN-DTW and ISMA-WLSSVM[J]. Journal of engineering for thermal energy and power, 2023, 38(9): 131-140.
[3] 薛阳, 燕宇铖, 贾巍, 等. 基于改进灰狼算法优化长短期记忆网络的光伏功率预测[J]. 太阳能学报, 2023, 44(7): 207-213.
XUE Y, YAN Y C, JIA W, et al.Photovoltaic power prediction model based on IGWO-LSTM[J]. Acta energiae solaris sinica, 2023, 44(7): 207-213.
[4] LI B H, ZHANG J.A review on the integration of probabilistic solar forecasting in power systems[J]. Solar energy, 2020, 210: 68-86.
[5] 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 C, WANG Z, BAI J Y, et al.Review of ultra-short-term forecasting methods for photovoltaic power generation[J]. High voltage engineering, 2023, 49(7): 2938-2951.
[6] IHEANETU K J.Solar photovoltaic power forecasting: a review[J]. Sustainability, 2022, 14(24): 17005.
[7] LU T H, WANG C S, CAO Y, et al.Photovoltaic power prediction under insufficient historical data based on dendrite network and coupled information analysis[J]. Energy reports, 2023, 9: 1490-1500.
[8] BAI M L, CHEN Y X, ZHAO X Y, et al.Deep attention ConvLSTM-based adaptive fusion of clear-sky physical prior knowledge and multivariable historical information for probabilistic prediction of photovoltaic power[J]. Expert systems with applications, 2022, 202: 117335.
[9] ZENG J W, QIAO W.Short-term solar power prediction using an RBF neural network[C]//2011 IEEE Power and Energy Society General Meeting. Detroit, MI, USA, 2011: 1-8.
[10] LIU L Y, ZHAO Y, CHANG D L, et al.Prediction of short-term PV power output and uncertainty analysis[J]. Applied energy, 2018, 228: 700-711.
[11] 许彪, 徐青山, 黄煜, 等. 基于藤copula分位数回归的光伏功率日前概率预测[J]. 电网技术, 2021, 45(11): 4426-4435.
XU B, XU Q S, HUANG Y, et al.Day-ahead probabilistic forecasting of photovoltaic power based on vine copula quantile regression[J]. Power system technology, 2021, 45(11): 4426-4435.
[12] 邢晨, 张照贝. 基于改进时间卷积网络的短期光伏出力概率预测方法[J]. 太阳能学报, 2023, 44(2): 373-380.
XING C, ZHANG Z B.Short-term probabilistic forecasting method of photovoltaic output power based on improved temporal convolutional network[J]. Acta energiae solaris sinica, 2023, 44(2): 373-380.
[13] GAO Y, LIU J W.Adaptively sparse transformers hawkes process[J]. International journal of uncertainty, fuzziness and knowledge-based systems, 2023, 31(4): 669-689.
[14] LI M P, ZHOU G X, CAI W W, et al.Multi-scale Sparse Network with Cross-Attention Mechanism for image-based butterflies fine-grained classification[J]. Applied soft computing, 2022, 117: 108419.
[15] ZHANG Q, CHEN L, SHAO M W, et al.ESAMask: real-time instance segmentation fused with efficient sparse attention[J]. Sensors, 2023, 23(14): 6446.
[16] LI C, ZHANG L X, WANG S P, et al.Sparse attention double-channel FCN network for numerical analysis tracheid features in larch[J]. Frontiers in plant science, 2022, 13: 1079556.
[17] CHEN C L, YU J, LING Q.Sparse attention block: aggregating contextual information for object detection[J]. Pattern recognition, 2022, 124: 108418.
[18] SALINAS D, FLUNKERT V, GASTHAUS J, et al.DeepAR: probabilistic forecasting with autoregressive recurrent networks[J]. International journal of forecasting, 2020, 36(3): 1181-1191.
[19] LIN Y, KOPRINSKA I, RANA M.SpringNet: transformer and spring DTW for time series forecasting[M]. Cham: Springer International Publishing, 2020: 616-628.
[20] RANGAPURAM S S, SEEGER M W, GASTHAUS J, et al.Deep state space models for time series forecasting[C]//Proceedings of the 31st International Conference on Neural Information Processing Systems. Montreal, Canada, 2018: 7796-7805.
PDF(1827 KB)
