首页|基于注意力机制的CNN-BiGRU模型预测风力发电功率

基于注意力机制的CNN-BiGRU模型预测风力发电功率

扫码查看
风能作为清洁能源,在新能源领域深受重视,但由于风力发电不稳定的出力情况制约着其应用和发展,预测其下一阶段工作功率可以有效提高能源的利用,降低系统维护难度.基于风电场的历史时序数据、风速和叶片桨距角数据,针对风电场的短期功率预测提出一种基于注意力机制的卷积神经网络(CNN)和双向门控制循环单元(BiGRU)的融合模型(CNN-BiGRU-AM).首先通过CNN提取数据的空间特征,再将其所获取的特征传递到BiGRU,通过BiGRU学习时间特征,最后通过注意力机制捕获两个模型输出的关键时空特征,通过训练后的参数建立模型,将数值气象预报(NWP)的风速等数据输入模型中,输出所需求的风力发电功率预测结果.输出的风力发电功率预测结果由平均绝对误差(MAE)和均方根误差(RMSE)来评估模型预测的精度,并将多个经典模型、不加注意力机制的融合模型、不同位置引入注意力机制的融合模型和此模型预测效果进行对比,结果显示本文所提出的模型精度更高更稳定.
Forecast of Wind Power Generation Based on AM of CNN-BiGRU
Wind energy is highly valued as clean energy in the field of new energy.However,the unstable performance of wind energy limits its application and development.Predicting its work performance in the next stage can effectively improve the energy utilization rate and reduce the difficulty of system ma-intenance.Based on historical time series data,wind speeds and blade angle data of wind farms,a fusion model CNN BiGRU AM for short-term performance prediction of wind farms is proposed.This model adds an AM that combines convolutional neural network and bidirectionally gated recursive unit.First, the spatial characteristics of the data are extracted by CNN,then the acquired characteristics are trans-ferred to BiGRU to extract the temporal characteristics,and the model is established by training parame-ters.Finally,the most important spatiotemporal properties of the time series are captured by the AM mechanism,the wind speed data of the numerical weather forecast is entered and the final forecast results are output via the model.The final prediction results are evaluated by using two parameters:mean abso-lute error (MAE)and mean square error (RMSE),which are further compared with the prediction re-sults of several classical models,fusion models without AM mechanism and fusion models with AM mechanism at different positions.The results show that the model has a higher accuracy and stability.

attentionmechanismconvolutional neural networksbidirectional gated recurrent unitynumerical weather prediction

巩冠华、娄柯、尹杰、李冬玉

展开 >

安徽工程大学 电气工程学院,安徽 芜湖 241000

注意力机制 卷积神经网络 双向门控循环单元 数值气象预报

国家自然科学基金联合基金资助项目

U21A20146

2024

安徽工程大学学报
安徽工程大学

安徽工程大学学报

影响因子:0.289
ISSN:2095-0977
年,卷(期):2024.39(2)
  • 7