基于新型注意力机制的LSTM光伏发电预测方法  

LSTM Photovoltaic Power Generation Forecasting Method Based on Attention Mechanism

作　　者：陈晨 王小杨[1] 梁建盈 马武兴 CHEN Chen;WANG Xiaoyang;LIANG Jianying;MA Wuxing(Zhongshan Institute,University of Electronic Science and Technology of China,Zhongshan 528402;Hunan Anhuayuan Technology Co.,Ltd.,Xiangtan 411100;China Shuifa Singyes Energy Holdings Ltd(Zhuhai),Zhuhai 519000)

机构地区：[1]电子科技大学中山学院,中山528402 [2]湖南安华源电力科技有限公司,湘潭411100 [3]水发兴业能源(珠海)有限公司,珠海519000

出　　处：《现代计算机》2021年第11期28-32,38,共6页Modern Computer

基　　金：广东省重点领域研发计划项目(No.2019B010148001);广东省教研教改项目(No.SJY201803)。

摘　　要：结合注意力机制与长短时记忆(LSTM)循环神经网络对常规气象数据下的光伏发电量进行预测,建立一个多气象因素加权预测模型(MVWM)。首先获取光伏电站近几年气象数据及其发电量数据,对其做归一化处理,提高模型收敛速度;建立基于不同气象因素的加权注意力机制;搭建LSTM循环神经网络框架;最后使用历史数据训练该网络,并将MVWM和SVM、LSTM、线性回归等模型在同一测试数据集上进行预测,结果表明MVWM具有较好的预测准确度。Using a combined attention mechanism and Long-Short-Term Memory(LSTM)cyclic neural network to predict photovoltaic power genera⁃tion,a multi-weather factor weighted prediction model(MVWM)was established.First obtain the meteorological data and power generation data of photovoltaic power stations in recent years,and normalize them to improve the convergence speed of the model;establish a weighted attention mechanism based on different meteorological factors;build the LSTM recurrent neural network framework;finally use historical da⁃ta Train the network and predict MVWM and SVM,LSTM,linear regression and other models on the same test data set.The results show that MVWM has better prediction accuracy.

关 键 词：注意力机制 LSTM 多气象因素加权 光伏预测 

分 类 号：TM615[电气工程—电力系统及自动化]