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

Improved Deep Learning Algorithm Based on S-BGD and Gradient Pile Strategy and Its Application in PV Power Forecasting

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。

To improve accuracy of photovoltaic（PV） power forecasting, this paper proposes a new forecasting method based on improved deep learning algorithm. Firstly, aiming at the problem of low training speed of conventional deep learning algorithm often using batch gradient descent（BGD） training method, a method combing stochastic gradient descent（SGD） and BGD methods are proposed. By using SGD method, training speed can be greatly improved. Secondly, to eliminate the problem of falling into local optimal points and saddle points during parameter training process, an improved method of gradient pile（GP） is proposed, using kinematic theory for reference. GP method uses cumulative gradient as the modified value to avoid local optimal points and saddle points. Finally, based on the data from Australia＇s Alice Springs PV power station, the proposed method is applied in its PV power forecasting. Forecasting results show that the proposed method has good performances in PV power forecasting.