基于改进麻雀算法的深度信念网络短期光伏功率预测  

Short-term PV power prediction of deep belief network based on improved sparrow search algorithm

作　　者：常东峰 南新元[1] CHANG Dongfeng;NAN Xinyuan(School of Electrical Engineering,Xinjiang University,Urumqi 830047,China)

机构地区：[1]新疆大学电气工程学院,新疆乌鲁木齐830047

出　　处：《现代电子技术》2022年第17期135-140,共6页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(52065064)。

摘　　要：为进一步提高光伏发电预测精度,提出了一种改进麻雀算法(ISSA)优化深度信念网络(DBN)的预测模型。首先,为提高麻雀算法的收敛速度和避免陷入局部最优,将精英反向学习策略和Metropolis准则引入麻雀算法用于初始化SSA种群和改进SSA更新策略。其次为提高DBN模型的性能,运用ISSA对DBN模型的权值进行优化选择,同时为避免冗余的气象因子影响光伏输出,采用最大相关信息系数的特征选择法(NFCBF),选择与光伏输出相关性最好的气象特征向量作为模型输入;基于NFCBF法选好的特征向量,采用一种结合欧氏距离和灰色关联度的综合相似指数法,选择相似日作为训练集。最后建立ISSA-DBN预测模型,并利用新疆某光伏电站的实际数据进行实验分析。结果表明在训练集的选择方法相同的情况下,与DBN模型、PSO-DBN模型、SSA-DBN模型相比,ISSA-DBN的平均绝对百分比误差指标在晴天仅为3.69%,晴转多云天为5.23%,雨天为9.53%,预测精度均高于其他三种模型。由此验证了ISSA-DBN模型良好的预测精度。In order to further improve the prediction accuracy of PV power generation,a prediction model of deep belief network(DBN)optimized by improved sparrow search algorithm(ISSA)is proposed.The elite opposition-based learning(EOBL)strategy and Metropolis criterion are introduced into the sparrow search algorithm(SSA)to initialize the SSA population and improve the SSA update strategy,so as to improve its convergence speed and avoid falling into local optimum.In order to improve the performance of DBN model,the weight of DBN model is optimized by ISSA.Meanwhile,in order to avoid redundant meteorological factors affecting PV output,the feature selection method with maximum correlation information coefficient,named NFCBF(new fast correlation-based filter)is used to select the meteorological feature vector with the best correlation with PV output as the model input.On the basis of the feature vectors selected by NFCBF method,a comprehensive similarity index method combining Euclidean distance and grey relational degree was used to select similar days as the training set.The ISSA-DBN prediction model was established,and the actual data of a PV power station in Xinjiang were used for experimental analysis.The results show that,in comparison with the DBN model,PSO-DBN model and SSA-DBN model,the average absolute percentage error index of ISSA-DBN is only 3.69%on sunny days,5.23%on sunny to cloudy days at most,and 9.53%on rainy days under the condition that the training set selection methods are the same.Therefore,the prediction accuracy of ISSA-DBN model has been verified.

关 键 词：DBN模型 NFCBF法 改进麻雀算法 综合相似指数 光伏发电 功率预测 精英反向学习策略 METROPOLIS准则 

分 类 号：TN99-34[电子电信—信号与信息处理] TN711[电子电信—信息与通信工程] TM615[电气工程—电力系统及自动化]