基于S-BGD和梯度累积策略的改进深度学习方法及其在光伏出力预测中的应用
发布时间：2018-11-27 09:32
【摘要】：为提高光伏出力的预测精度,提出了一种改进深度学习算法的光伏出力预测方法。首先,针对传统的深度学习算法采用批量梯度下降(batch gradient descent,BGD)法训练模型参数速度慢的问题,利用随机梯度下降(stochastic gradient descent,SGD)法训练快的优点,提出了一种改进的随机-批量梯度下降(stochastic-batch gradient descent,S-BGD)搜索方法,该方法兼具SGD和BGD的优点,提高了参数训练的速度。然后,针对参数训练过程中容易陷入局部最优点和鞍点的问题,借鉴运动学理论,提出了一种基于梯度累积(gradient pile,GP)的训练方法。该方法以累积梯度作为参数的修正量,可以有效地避免训练陷入局部点和鞍点,进而提高预测精度。最后,以澳大利亚艾丽斯斯普林光伏电站的数据为样本,将所提方法应用于光伏出力预测中,验证所提方法的有效性。
[Abstract]:In order to improve the accuracy of photovoltaic force prediction, an improved depth learning algorithm is proposed for photovoltaic force prediction. First of all, aiming at the problem that the traditional depth learning algorithm uses batch gradient descent (batch gradient descent,BGD) method to train the model parameters slowly, the advantage of the stochastic gradient descent (stochastic gradient descent,SGD) method is presented. An improved random-batch gradient descent (stochastic-batch gradient descent,S-BGD) search method is proposed, which combines the advantages of SGD and BGD, and improves the speed of parameter training. Then, aiming at the problem that parameter training is easy to fall into local optimum and saddle point, a training method based on gradient cumulation (gradient pile,GP) is proposed based on kinematics theory. In this method, the cumulative gradient is used as the parameter modifier, which can effectively avoid the training falling into local points and saddle points, and then improve the prediction accuracy. Finally, based on the data of Alice Spring photovoltaic power station in Australia, the proposed method is applied to photovoltaic force prediction to verify the effectiveness of the proposed method.
【作者单位】： 广西电力系统最优化与节能技术重点实验室(广西大学);
【基金】：国家重点研发计划支持项目(2016YFB0900100) 国家自然科学基金项目资助(51377027)~~
【分类号】：TM615
，
本文编号：2360232
[Abstract]:In order to improve the accuracy of photovoltaic force prediction, an improved depth learning algorithm is proposed for photovoltaic force prediction. First of all, aiming at the problem that the traditional depth learning algorithm uses batch gradient descent (batch gradient descent,BGD) method to train the model parameters slowly, the advantage of the stochastic gradient descent (stochastic gradient descent,SGD) method is presented. An improved random-batch gradient descent (stochastic-batch gradient descent,S-BGD) search method is proposed, which combines the advantages of SGD and BGD, and improves the speed of parameter training. Then, aiming at the problem that parameter training is easy to fall into local optimum and saddle point, a training method based on gradient cumulation (gradient pile,GP) is proposed based on kinematics theory. In this method, the cumulative gradient is used as the parameter modifier, which can effectively avoid the training falling into local points and saddle points, and then improve the prediction accuracy. Finally, based on the data of Alice Spring photovoltaic power station in Australia, the proposed method is applied to photovoltaic force prediction to verify the effectiveness of the proposed method.
【作者单位】： 广西电力系统最优化与节能技术重点实验室(广西大学);
【基金】：国家重点研发计划支持项目(2016YFB0900100) 国家自然科学基金项目资助(51377027)~~
【分类号】：TM615
，
本文编号：2360232
本文链接：https://www.wllwen.com/kejilunwen/dianlidianqilunwen/2360232.html
