Wind power forecasting is an important means to cope with problems when large scale wind power is integrated into grid, which is of great significance for development programs and dispatching in smart grid.Probabilistic forecast is different expection forecast by the capability of forecasting the distribution of random variables, in addition it also provides a measure of the model uncertainty for wind power generation. In view of successful application in time series forecasting by using Gaussian processes(GP) methods, under the framework of time series analysis with strong nonlinear characteristics, combinated the NWP information with wind power data, on the basis of select the key input varable by feature extraction methods, the project will study short term wind power probabilistic forecasting methods for wind farm using Gaussian process regression techniques to build enhanced and more higher versatility wind power forecasting model.This main contents of the project are as follows: to attempt to solve some main problems of modelling large data sets as well as online learning for wind power probabilistic forecasting,studying several sparse Gaussian processes models with different sparse approximation algorithms, several effective sparse online Gaussian process models with different sparse online approximation algorithms,the extension of sparse robust Gaussian process models to non-Gaussian likelihoods, the extension of nonlinear state space model with online Gaussian processes method so that the super-short term and short term wind power probabilistic forecasting will be well performed. Moreover,to further improve the prediction accuracy and generalization of the forcasting model, combination forecasting approaches based on online local gaussian processes model using ensembling multi-model learning strategies are proposed, simutianeously, a kind of combined forecasting method based on complete ensemble empirical model decomposition and local Gaussian process model is also studied. According to the imperious demands of the national development strategy as well as the construction of wind power projects in Gansu region,this project possesses a significant value in the theory and engineering research for short-term wind power probabilistic forecasting.
风电功率预测是应对风电大规模并网运行问题的重要手段,对智能电网的发展规划和调度具有重要意义。概率预测有别于期望值预测,能够提供被预测量的分布信息,并对模型输出的不确定性提供度量。鉴于高斯过程(GP)在时间序列预测上的成功应用,结合数值天气预报信息,在解决输入特征信息提取等关键问题基础上,统一在非线性时间序列分析的框架下,研究基于GP的短期风电功率概率预测方法,以建立适应性更强的概率预测模型。具体包括:为解决大数据集的离线训练和在线学习,提出研究基于不同稀疏逼近算法的GP模型;基于非线性状态空间推理和学习的在线GP模型;延伸至非高斯似然性的鲁棒稀疏GP模型。为进一步提高预测模型的准确度和泛化能力,结合局部学习与集成学习策略,研究基于集成经验模态分解与局部GP结合的组合预测以及基于局部多模型GP集成的组合预测方法。本项目从国家建设与甘肃地区风电发展需求出发,具有重要的理论与工程应用价值。
风电功率预测是应对风电大规模并网运行问题的重要手段,对智能电网的发展规划和调度具有重要意义。概率预测有别于期望值预测,能够提供被预测量的分布信息,并对模型输出的不确定性提供度量。鉴于高斯过程(GP)在时间序列预测上的成功应用,结合数值天气预报信息,在解决输入特征信息提取等关键问题基础上,统一在非线性时间序列分析的框架下,研究包括GP、稀疏GP在内的一系列混合计算智能方法,将其应用于短期风电功率预测中,以建立适应性更强的点预测与概率预测模型。具体包括:针对时间序列预测,在单隐层前馈神经网络的基础上,基于进化计算的优化策略,提出了一种优化的核极限学习机(O-KELM)方法,在KELM的基础上,分别将遗传算法(GA)、模拟退火(SA)、微分演化(DE)三种进化算法用于模型的结构输入选择、正则化系数以及核参数的优化选取,以进一步提高网络的性能。为解决大数据集的离线训练和在线学习,提出一种基于泄漏积分型回声状态网络(LiESN)的建模方法,给出LiESN 的岭回归离线学习算法与递推最小二乘(RLS)在线学习算法。通过随机选取训练数据集的数据子集,给出基于数据点子集(SoD)逼近、回归量子集(SoR)逼近、投影过程(PP)逼近的一类基于稀疏高斯过程(Sparse-GP)的概率预测方法,该方法能自适应确定先验协方差函数中的“超参数”,在给出模型预测输出的同时,还能获取预测输出的方差,从而很好地解释模型的置信水平。提出一种基于核主成分分析(KPCA)与核最小最大概率回归机(KMPMR)相结合的方法。为进一步提高预测模型的准确度和泛化能力,结合局部学习与集成学习策略,提出一种具有自适应噪声的完整集成经验模态分解( CEEMDAN) - 排列熵和泄漏积分回声状态网络( LIESN) 的组合预测方法,进一步给出一种基于经验小波变换与多核学习算法结合的组合预测方法。为处理建模过程的不确定性并使不确定性的影响达到最小,研究了不同区间二型Mamdani模糊逻辑系统(FLS)方法,进一步还研究了基于A1-C1,A2-C0和A2-C1三种类型的区间二型TSK FLS方法,并应用反向传播(BP)算法调整模型前件和后件的参数,在此基础上,提出一类基于主成分分析 (PCA)和区间二型Mamdani FLS、区间TSK FLS相结合的方法,成功应用于短期风电功率预测中。
{{i.achievement_title}}
数据更新时间:2023-05-31
基于LASSO-SVMR模型城市生活需水量的预测
内点最大化与冗余点控制的小型无人机遥感图像配准
基于多模态信息特征融合的犯罪预测算法研究
基于FTA-BN模型的页岩气井口装置失效概率分析
F_q上一类周期为2p~2的四元广义分圆序列的线性复杂度
风电功率概率区间预测方法研究
基于组合预测模型的并网风电场短期风电功率预测研究
风电功率超短期多步预测的可预测性研究
基于风电功率在线概率预测的风储联合系统随机预测控制