基于DBN深度学习残差修正的AUKF超短期光伏功率预测模型

AUKF ultra-short-term photovoltaic power prediction model based on DBN deep learning residual correction

针对线性模型和测量噪声时变因素降低光伏功率预测精度的局限,提出一种分层的光伏发电功率预测模型。通过自适应算法改进无迹卡尔曼滤波实时噪声估计,降低光伏功率预测系统状态空间模型中辐照度、光伏功率测量噪声时变对预测精度的影响,实现光伏发电功率的初步预测。二层预测中,基于DBN深度学习网络修正初步预测残差,降低非线性气象因素对预测精度的影响。利用现场实测数据,通过仿真验证了改进模型具有较好的预测精度、良好的泛化能力和工程应用价值。

Aiming at the limitation of linear model and time-varying factors of measurement noise reducing the accuracy of photovoltaic power prediction,a hierarchical photovoltaic power prediction model is put forward.The adaptive algorithm is designed to amend the actual-time noise appraisal of unscented Kalman filter,which reduces the influence of irradiance in the state space model of photovoltaic power prediction system and time-varying noise of photovoltaic power measurement on the prediction accuracy,and achieves the preliminary prediction of photovoltaic power generation.In the two-level prediction,the elemental prediction residual is corrected based on the DBN deep learning network to reduce the influence of nonlinear meteorological factors on the prediction accuracy.The simulation results show that the modified model has better prediction accuracy,better generalization ability and engineering application value.