一种结合CNN和GRU网络的超短期风电预测模型
|
摘要
在大型电网和小型微电网中,风力发电预测对电力系统安全和经济运行发挥着至关重要的作用。针对传统建模中风电功率时变特性的局限,进一步探索风电时间序列波动特征的潜在信息,文章提出一种结合卷积神经网络(Convolutional Neural Network,CNN)和门控循环单元(Gated Recurrent Unit,GRU)的超短期风电预测模型。首先,该模型利用NWP气象数据为输入并将其归一化处理;然后,采用结合CNN和GRU网络对多变量时间序列进行动态时间建模,引入CNN来压缩GRU隐藏状态以减少计算模型的输出,克服训练过程中的梯度爆炸和消失问题;最后,根据风速和风力发电功率特性实现风电预测。通过实验仿真结果可知,与目前已投入运行的基于ANN的风电预测系统和近年来新兴的LSTM深度学习算法相比,该方法具有更高的预测精度,具有一定的工程价值。
Wind power forecasting plays a crucial role in the safety and economic operation of power systems in large-scale power grids and small micro-grids. The paper aims at the limitation of the traditional modeling of the time-varying characteristics of wind power, after exploring the potential information of the fluctuation characteristics of wind power time series, an ultra-short-term wind power forecast model based on CNN(Convolutional Neural Network)and GRU(Gated Recurrent Unit)is proposed. First, the model is based on NWP(Numerical Weather Prediction)data as input and normalized. Then, the combination of CNN and GRU networks are used to model dynamic time of multi-variable time series. The model introduces CNN to compress the GRU hidden state to reduce the output of the calculation model and overcome the problem of gradient explosion and disappearance during training. Finally, wind speed and wind power characteristics are used to achieve wind power forecasting. The experimental simulation results show that, compared with the ANN-based wind power prediction system and the emerging LSTM deep learning algorithm, the proposed method has higher prediction accuracy and has certain engineering value.
Wind power forecasting plays a crucial role in the safety and economic operation of power systems in large-scale power grids and small micro-grids. The paper aims at the limitation of the traditional modeling of the time-varying characteristics of wind power, after exploring the potential information of the fluctuation characteristics of wind power time series, an ultra-short-term wind power forecast model based on CNN(Convolutional Neural Network)and GRU(Gated Recurrent Unit)is proposed. First, the model is based on NWP(Numerical Weather Prediction)data as input and normalized. Then, the combination of CNN and GRU networks are used to model dynamic time of multi-variable time series. The model introduces CNN to compress the GRU hidden state to reduce the output of the calculation model and overcome the problem of gradient explosion and disappearance during training. Finally, wind speed and wind power characteristics are used to achieve wind power forecasting. The experimental simulation results show that, compared with the ANN-based wind power prediction system and the emerging LSTM deep learning algorithm, the proposed method has higher prediction accuracy and has certain engineering value.
引文
[1]牛东晓,范磊磊.风电功率预测方法综述及发展研究[J].现代电力,2013,30(4):24-28.
[2]钱政,裴岩,曹利宵,等.风电功率预测方法综述[J].高电压技术,2016,42(4):1047-1060.
[3] Yatiyana E,Rajakaruna S,Ghosh A.Wind speed anddirection forecasting for wind power generation usingARIMA model[A].2017 Australasian Universities PowerEngineering Conference(AUPEC)[C].Melbourne:IEEE,2017.1-6.
[4]杨茂,黄宾阳,江博,等.基于卡尔曼滤波和支持向量机的风电功率实时预测研究[J].东北电力大学学报,2017,37(2):45-51.
[5]江岳春,杨旭琼,贺飞,等.基于EEMD-IGSA-LSSVM的超短期风电功率预测[J].湖南大学学报(自然科学版),2016,43(10):70-78.
[6]王一珺,贾嵘.基于Elman和实测风速功率数据的短期风功率预测[J].高压电器,2017,53(9):125-129.
[7]尹宝才,王文通,王立春.深度学习研究综述[J].北京工业大学学报,2015,41(1):48-59.
[8]朱乔木,李弘毅,王子琪,等.基于长短期记忆网络的风电场发电功率超短期预测[J].电网技术,2017,41(12):3797-3802.
[9]赵永宁,叶林.区域风电场短期风电功率预测的最大相关-最小冗余数值天气预报特征选取策略[J].中国电机工程学报,2015,35(23):5985-5994.
[10] Zhou G B,Wu J X,Zhang C L,et al.Minimal gated unitfor recurrent neural networks[J].International Journal ofAutomation and Computing,2016,13(3):226-234.
[11] Jiao R,Huang X,Ma X.,et al.A model combiningstacked auto encoder and back propagation algorithmfor short-term wind power forecasting[J].IEEE Access,2018,6:17851-17858.
[12] Chen H,Li R,Wang Y,et al.Wind power forecastingbased on refined LSTAR-GARCH model[J].CIRED-Open Access Proceedings Journal,2017,1(10):2590-2593.
[13] Phaisangittisagul E.An analysis of the regularizationbetween L2 and dropout in single hidden layer neuralnetwork[A].2016 7th International Conference onIntelligent Systems,Modelling and Simulation(ISMS)[C].Bangkok:IEEE,2016.174-179.
[2]钱政,裴岩,曹利宵,等.风电功率预测方法综述[J].高电压技术,2016,42(4):1047-1060.
[3] Yatiyana E,Rajakaruna S,Ghosh A.Wind speed anddirection forecasting for wind power generation usingARIMA model[A].2017 Australasian Universities PowerEngineering Conference(AUPEC)[C].Melbourne:IEEE,2017.1-6.
[4]杨茂,黄宾阳,江博,等.基于卡尔曼滤波和支持向量机的风电功率实时预测研究[J].东北电力大学学报,2017,37(2):45-51.
[5]江岳春,杨旭琼,贺飞,等.基于EEMD-IGSA-LSSVM的超短期风电功率预测[J].湖南大学学报(自然科学版),2016,43(10):70-78.
[6]王一珺,贾嵘.基于Elman和实测风速功率数据的短期风功率预测[J].高压电器,2017,53(9):125-129.
[7]尹宝才,王文通,王立春.深度学习研究综述[J].北京工业大学学报,2015,41(1):48-59.
[8]朱乔木,李弘毅,王子琪,等.基于长短期记忆网络的风电场发电功率超短期预测[J].电网技术,2017,41(12):3797-3802.
[9]赵永宁,叶林.区域风电场短期风电功率预测的最大相关-最小冗余数值天气预报特征选取策略[J].中国电机工程学报,2015,35(23):5985-5994.
[10] Zhou G B,Wu J X,Zhang C L,et al.Minimal gated unitfor recurrent neural networks[J].International Journal ofAutomation and Computing,2016,13(3):226-234.
[11] Jiao R,Huang X,Ma X.,et al.A model combiningstacked auto encoder and back propagation algorithmfor short-term wind power forecasting[J].IEEE Access,2018,6:17851-17858.
[12] Chen H,Li R,Wang Y,et al.Wind power forecastingbased on refined LSTAR-GARCH model[J].CIRED-Open Access Proceedings Journal,2017,1(10):2590-2593.
[13] Phaisangittisagul E.An analysis of the regularizationbetween L2 and dropout in single hidden layer neuralnetwork[A].2016 7th International Conference onIntelligent Systems,Modelling and Simulation(ISMS)[C].Bangkok:IEEE,2016.174-179.