为解决光伏序列的强噪音干扰以及单一模型在光伏功率预测方面精度偏低和泛化性较差的问题,提出了一种基于特征优化和混合改进灰狼算法优化双向长短时记忆网络(bi-directional long short-term memory,BiLSTM)的短期光伏功率预测方法。首...为解决光伏序列的强噪音干扰以及单一模型在光伏功率预测方面精度偏低和泛化性较差的问题,提出了一种基于特征优化和混合改进灰狼算法优化双向长短时记忆网络(bi-directional long short-term memory,BiLSTM)的短期光伏功率预测方法。首先,运用互信息算法进行输入数据的变量选择,以消除冗余变量。其次,通过互补集合经验模态分解和改进的小波阈值算法对筛选后的数据进行特征重构,旨在降低数据中的噪声干扰并完成输入变量的特征优化。随后,结合改进的Tent混沌映射、非线性递减因子、动态权重策略和差分进化算法对标准灰狼优化算法进行混合优化,以确定双向长短期记忆神经网络的最优超参数组合,并引入注意力机制以挖掘数据中的关键时序信息,最终构建出一种新型的短期光伏功率预测模型。仿真实验表明,相较于最小二乘支持向量机、长短期记忆网络和双向长短期记忆网络,所提模型在晴天、多云、阴天和降雨等不同工况下的均方根误差平均分别降低了12.45%、7.95%和5.37%,显示出优秀的预测性能、良好的泛化能力和潜在的工程应用价值。展开更多
为解决一些决策树受到数据噪声等因素的影响,导致它们对随机森林聚类产生有限甚至负面贡献这一问题,提出一种基于聚类集成选择的随机森林聚类方法(random forest clustering method based on cluster ensemble selection,RFCCES)。将每...为解决一些决策树受到数据噪声等因素的影响,导致它们对随机森林聚类产生有限甚至负面贡献这一问题,提出一种基于聚类集成选择的随机森林聚类方法(random forest clustering method based on cluster ensemble selection,RFCCES)。将每一棵决策树视为一个基聚类器,根据基聚类器集合的稳定和不稳定性设计两种不同的聚类集成选择方法,将评估单个决策树对随机森林的增益问题,转化为基聚类器对最终的聚类集成结果的增益问题。该算法与5种对比方法在10个数据集上进行比较,实验结果验证了RFCCES的独特优势和整体有效性。展开更多
To accelerate the selection process of feature subsets in the rough set theory （RST）, an ensemble elitist roles based quantum game （EERQG） algorithm is proposed for feature selec- tion. Firstly, the multilevel eli...To accelerate the selection process of feature subsets in the rough set theory （RST）, an ensemble elitist roles based quantum game （EERQG） algorithm is proposed for feature selec- tion. Firstly, the multilevel elitist roles based dynamics equilibrium strategy is established, and both immigration and emigration of elitists are able to be self-adaptive to balance between exploration and exploitation for feature selection. Secondly, the utility matrix of trust margins is introduced to the model of multilevel elitist roles to enhance various elitist roles＇ performance of searching the optimal feature subsets, and the win-win utility solutions for feature selec- tion can be attained. Meanwhile, a novel ensemble quantum game strategy is designed as an intriguing exhibiting structure to perfect the dynamics equilibrium of multilevel elitist roles. Finally, the en- semble manner of multilevel elitist roles is employed to achieve the global minimal feature subset, which will greatly improve the fea- sibility and effectiveness. Experiment results show the proposed EERQG algorithm has superiority compared to the existing feature selection algorithms.展开更多
文摘为解决光伏序列的强噪音干扰以及单一模型在光伏功率预测方面精度偏低和泛化性较差的问题,提出了一种基于特征优化和混合改进灰狼算法优化双向长短时记忆网络(bi-directional long short-term memory,BiLSTM)的短期光伏功率预测方法。首先,运用互信息算法进行输入数据的变量选择,以消除冗余变量。其次,通过互补集合经验模态分解和改进的小波阈值算法对筛选后的数据进行特征重构,旨在降低数据中的噪声干扰并完成输入变量的特征优化。随后,结合改进的Tent混沌映射、非线性递减因子、动态权重策略和差分进化算法对标准灰狼优化算法进行混合优化,以确定双向长短期记忆神经网络的最优超参数组合,并引入注意力机制以挖掘数据中的关键时序信息,最终构建出一种新型的短期光伏功率预测模型。仿真实验表明,相较于最小二乘支持向量机、长短期记忆网络和双向长短期记忆网络,所提模型在晴天、多云、阴天和降雨等不同工况下的均方根误差平均分别降低了12.45%、7.95%和5.37%,显示出优秀的预测性能、良好的泛化能力和潜在的工程应用价值。
文摘为解决一些决策树受到数据噪声等因素的影响,导致它们对随机森林聚类产生有限甚至负面贡献这一问题,提出一种基于聚类集成选择的随机森林聚类方法(random forest clustering method based on cluster ensemble selection,RFCCES)。将每一棵决策树视为一个基聚类器,根据基聚类器集合的稳定和不稳定性设计两种不同的聚类集成选择方法,将评估单个决策树对随机森林的增益问题,转化为基聚类器对最终的聚类集成结果的增益问题。该算法与5种对比方法在10个数据集上进行比较,实验结果验证了RFCCES的独特优势和整体有效性。
基金supported by the National Natural Science Foundation of China(6113900261171132+4 种基金61300167)the Natural Science Foundation of Jiangsu Education Department(12KJB520013)the Open Project Program of Jiangsu Provincial Key Laboratory of Computer Information Processing Technologythe Qing Lan Project of Jiangsu Provincethe Starting Foundation for Doctoral Scientific Research,Nantong University(14B20)
文摘To accelerate the selection process of feature subsets in the rough set theory （RST）, an ensemble elitist roles based quantum game （EERQG） algorithm is proposed for feature selec- tion. Firstly, the multilevel elitist roles based dynamics equilibrium strategy is established, and both immigration and emigration of elitists are able to be self-adaptive to balance between exploration and exploitation for feature selection. Secondly, the utility matrix of trust margins is introduced to the model of multilevel elitist roles to enhance various elitist roles＇ performance of searching the optimal feature subsets, and the win-win utility solutions for feature selec- tion can be attained. Meanwhile, a novel ensemble quantum game strategy is designed as an intriguing exhibiting structure to perfect the dynamics equilibrium of multilevel elitist roles. Finally, the en- semble manner of multilevel elitist roles is employed to achieve the global minimal feature subset, which will greatly improve the fea- sibility and effectiveness. Experiment results show the proposed EERQG algorithm has superiority compared to the existing feature selection algorithms.
