基于K-means聚类和极限学习机组合算法的短期光伏功率预测

考虑光伏功率的预测精度强依赖于天气模态和气候条件等因素影响,提出了基于极限学习机组合算法的短期光伏功率预测方法。首先,基于K-means聚类算法进行天气分型,分为4个季节下晴天、多云天气、阴雨天气共12组不同天气类别。其次,针对天气分型结果,基于极限学习机ELM、遗传算法改进的极限学习机GA-ELM、鸟群算法改进的极限学习机BSA-ELM3种算法构建光伏功率预测模型。最后,以某光伏电站数据进行所提模型验证。预测结果表明,BSA-ELM预测精度最高,12种天气预测精度达到90%左右,各季节中预测精度最高的天气类型均为晴天,多云天气精度高于阴雨天气精度,可为含高比例光伏并网的新型电力系统安全稳定运行提供有效数据支撑。     

基于优化聚类的组合风速短期预测

准确的风功率预测对电力系统安全、稳定运行具有重要意义,而风速预测是风功率预测的关键。文章提出一种基于优化模糊C均值(Optimal Fuzzy C means,OFCM)聚类的组合风速短期预测方法。首先,采用模拟退火遗传算法优化模糊C均值聚类算法的初始聚类中心;其次,基于优化模糊C均值聚类算法将初始风速属性样本数据进行分组;再根据不同风速样本组,运用极限学习机(Extremely Learning Machine,ELM)构建组合风速预测模型;最后,通过风速实测值与预测值的对比,验证了该方法的可行性。     

基于k-means聚类的SVR短期风速预测

准确的风速预测是风力发电功率预测的重要基础。为了进一步提高风速预测精度,文章提出一种基于k-means聚类的支持向量回归机(SVR)的短期风速组合预测新方法。首先分析影响风速变化的因素,计算不同风速属性相对于风速序列的皮尔逊相关系数(PCC)值,并对其进行加权;然后采用k-means聚类方法对风速样本进行聚类;再利用SVR针对每组样本建模;最后结合实际风电场进行仿真,结果表明,该方法具有较高的准确性和可行性。     

基于主成分约简聚类的优化ELM短期风速组合预测

提出基于主成分属性约简聚类的粒子群优化极限学习机短期风速预测方法.考虑到不同的属性特征对于风速变化的影响不同,利用主成分分析法计算各成分特征值,选取方差贡献率较高的成分,然后采用k-均值聚类方法对风速样本进行聚类,再利用粒子群算法对极限学习机进行优化,进而构建风速组合预测模型.最后结合风电场实测历史数据进行实验预测对比...     

基于模糊聚类和BP神经网络的小水电短期发电能力预测方法

小水电大多为无调节能力的径流式电站,汛期与大中型水电挤占输电通道,导致水电弃水及地区窝电现象日趋严重,因此尽可能准确地掌握小水电发电能力、制定合理的大小水电协调调度计划愈显重要。对此提出了基于模糊聚类和BP神经网络相结合的小水电短期发电能力FC-BP预测方法,将训练样本根据历史运行数据分类,建立相应的BP网络,对待测样本识别归类,预测小水电装机日利用小时数,并将该方法应用于云南省盈江县和云龙县小水电短期发电能力预测中。结果表明,FC-BP预测方法较传统ANN模型预测精度有所提高,且泛化能力更强。     

基于SSA和K均值的TD-BP神经网络超短期光伏功率预测

针对光伏功率预测精度不高的特点,提出一种融合了奇异谱分析、K均值聚类法、时延特性和BP神经网络的光伏功率预测模型.利用相似日理论选取各天气类型训练样本,通过奇异谱分析的分解及重构,提取出蕴含在时间序列中的趋势及准周期成分,利用K均值聚类法将重构后的各天气样本聚类分析为K类.每类以温度、风速、天气类型和历史功率作为样本属...     

基于模糊聚类的光伏组件功率输出模型研究

摘要： 在光伏组件工程用功率输出模型的基础上提出了一种改进模型。基于模糊聚类方法将多晶硅组件户外测试数据、环境参数以天为单位聚类,聚类结果为晴天、多云、阴天、雨天、多云转阴和多云转雨。再由每一类聚类结果依次分析组件输出功率与组件温度以及太阳辐照度之间关系,最后验证了该模型的准确性。     

基于相似日聚类和PCC-VMD-SSA-KELM模型的短期光伏功率预测

由于光伏发电的随机性和不稳定性会影响功率预测的精度,提出一种基于皮尔逊相关系数（PCC）、K-均值算法（K-means）、变分模态分解（VMD）、麻雀搜索算法（SSA）、核函数极限学习机（KELM）的光伏功率短期预测模型。首先,用PCC选取主要因素作为输入;K-均值算法进行相似日聚类,将历史数据聚类为晴天、多云和雨天;其次,VMD对原始信号进行分解,充分提取集合中的输入因素信息,提高数据质量;SSA优化KELM模型的核函数参数和正则化系数解决其参数选择敏感问题;最后,将不同序列预测值叠加得到最终预测结果。仿真结果表明,所提相似日聚类下PCC-VMD-SSA-KELM模型具有较小的预测误差。     

基于Kohonen聚类和SVM组合算法的电网日最大负荷预测

电力系统负荷受到气象等多种因素的共同影响,为提高电力系统短期预测的准确度,提出了一种将Kohonen聚类和SVM结合的预测算法.该方法考虑到电力系统变化周期性的特点,通过对学习样本进行聚类,选出具有相同特征的历史相似日,构造相似日训练样本并利用SVM模型进行预测.样本特征向量中同时包含了多种气象因子及节假日等因素,经过仿真实验,验证了该方法的有效性.     

基于NWP相似性分析的超短期光伏发电功率预测

针对相似气象情况下,光伏电站发电功率接近的实际情况,提出一种基于数值天气预报(NWP)相似性分析的超短期光伏发电功率预测方法.该方法根据皮尔逊相关系数找到与被预测时刻相似的天气预报数据,根据相似时刻的实际功率估计出被预测时刻的功率.该方法能高效的根据天气预报数据预测发电功率,运算速度快,与神经网络相比预测有着更优的效果...     

基于多尺度卷积双向长短期记忆网络与注意力机制的滚动轴承剩余寿命预测

通过卷积神经网络(Convolutional Neural Network, CNN)处理轴承一维时域或频域信号,难以提取具有代表性的非线性特征信息,且易忽略低层次信息。针对这一问题,基于多尺度特征提取,引入一种特征注意力机制,提出一种基于卷积双向长短期记忆网络(MSAM CNN BiLSTM)的轴承剩余寿命预测方法。基于西安交通大学（Xi′an Jiao Tong University,XJTU）轴承数据集中的3组数据对MSAM CNN BiLSTM、LSTM、CNN LSTM和MSAM CNN LSTM 4种方法的预测误差进行对比分析。结果表明：MSAM CNN BiLSTM方法在3组数据集中的预测误差均小于其他3种方法,说明该模型能同时学习数据中的低层次与高层次信息,可有效提高轴承的剩余寿命预测精度。     

基于混合深度学习的风电功率预测及一次调频应用

为了解决高比例不确定性风电接入电力系统带来强烈调频需求的问题,提出了基于混合深度学习模型的风电功率预测及其一次调频应用方法。首先,采用孤立森林(Isolated Forest, IF)对历史数据进行异常值处理,提高数据质量,其次,构建卷积神经网络(Convolutional Neural Network, CNN)、双向长短期记忆(Bidirectional Long Short Term Memory, BiLSTM)和注意力机制(Attention Mechanism, AM)的混合深度学习模型对风电功率进行预测。最后,依据功率预测精度配置超级电容器储能,设计储能调频控制原则,弥补风电机组自身预测误差,并协同风电机组参与电力系统一次调频。基于预测结果为4台风电发电机组2个负荷区域仿真系统配置超级电容器储能系统,利用digsilent平台进行了风预测误差和负荷波动下的一次调频仿真。结果表明：所提IF-CNN-BiLSTM-AM模型比BP和LSTM基准模型预测误差(MSE)降低了81.53%和51.44%,具有最优的预测性能;设计的风储一次调频模型与原则可有效应对风电预测误差和负荷波动...     

基于深度学习的翼型气动弹片流场参数预测

弹片是解决翼型流动分离的重要技术手段,合理的弹片参数对翼型表面压力分布尤为重要。基于数据驱动的深度学习方法与计算流体力学(Computational Fluid Dynamics, CFD)相结合,可快速有效地完成对复杂流场特征的识别与提取。本文提出一种基于卷积神经网络(Convolutional Neural Network, CNN)的翼型表面压力分布预测方法,通过提取流场的尾流速度、压力等流动特征构建翼型表面压力分布的预测模型。首先,通过数值模拟计算了8种不同抬起角度的NACA 0012弹片翼型的流场;其次,采用提取的流场数据建立CNN预测模型;最后,将预测值和CFD计算值进行对比。结果表明：基于CNN的预测模型对翼型表面压力系数分布有较高的预测精度,其中尾流速度模型在弹片抬起角度为15°时的预测均方根误差仅为0.1,说明尾流速度中包含丰富的流场信息。     

基于MPGA-BP模型的降雨预报研究

降雨预报是水文预报的重要环节,提高其准确性是进行洪水、径流等预报的前提。针对目前预测方法中存在的易落入局部极小值、收敛速度慢和收敛对初值敏感等问题,将多种群遗传算法(MPGA)与反向传播(BP)神经网络模型相结合,提出了一种适用于降雨预报的多种群遗传神经网络模型(MPGA-BP)。实例计算结果表明,该模型具有良好的预报性能和泛化能力,为降雨准确预测提供了有力的技术支持。     

An accurate method for the PV model identification based on a genetic algorithm and the interior-point method

Due to the PV module simulation requirements as well as recent applications of model-based controllers, the accurate photovoltaic (PV) model identification method is becoming essential to reduce the PV power losses effectively. The classical PV model identification methods use the manufacturers provided maximum power point (MPP) at the standard test condition (STC). However, the nominal operating cell temperature (NOCT) is the more practical condition and it is shown that the extracted model is not well suited to it. The proposed method in this paper estimates an accurate equivalent electrical circuit for the PV modules using both the STC and NOCT information provided by manufacturers. A multi-objective global optimization problem is formulated using only the main equation of the PV module at these two conditions that restrains the errors due to employing the experimental temperature coefficients. A novel combination of a genetic algorithm (GA) and the interior-point method (IPM) allows the proposed method to be fast and accurate regardless the PV technology. It is shown that the overall error, which is defined by the sum of the MPP errors of both the STC and the NOCT conditions, is improved by a factor between 5.1% and 31% depending on the PV technology.     

基于深度学习与支持向量机的滚动轴承故障诊断研究

针对滚动轴承运行环境复杂,传统故障诊断方法难以从强非线性信号中提取有效故障特征,且无法充分利用信号自身特征的问题,提出CNN-LSTM-SVM故障诊断方法。以滚动轴承加速度寿命实验数据为研究对象,基于卷积神经网络(Convolutional Neural Network, CNN)与长短期记忆网络(Long Short Term Memory, LSTM)技术提取信号特征并结合支持向量机(Support Vector Machine, SVM)完成故障分类。结果显示：该方法具有良好外推性能,在变演变阶段下的平均准确率达到95.92%,与现有方法相比,至少高出11.34%,且在噪声环境下的诊断准确率均高于现有方法,稳定性更佳,体现良好的鲁棒性与泛化性。     

Proton exchange membrane fuel cell-powered bidirectional DC motor control based on adaptive sliding-mode technique with neural network estimation

Proton exchange membrane fuel cell (PEMFC), according to its merits of high energy density, zero emission, and low noise, has been widely applied in industrial appliances. A full bridge converter is used to implement PEMFC-powered DC motor bidirectional rotation in this paper. For the sake of the regulations of DC motor angular velocity as well as bus voltage, an adaptive backstepping sliding-mode control (ABSMC) technique integrated with Chebyshev neural network (CNN) is proposed. Based on the equivalent-circuit method, the control-oriented model of the PEMFC-powered motor system is structured. By constructing Lyapunov function, the adaptive laws and control laws can be obtained to achieve bus voltage and angular velocity regulations simultaneously. Moreover, the proposed neural network is applied to estimate the uncertainties of the system through orthogonal basis Chebyshev polynomials. To highlight the advantages of proposed technique, a proportional-integral (PI) control was introduced subsequently and two controllers were compared via numerical simulations. The simulation results demonstrate that CNN estimation method in conjunction with backstepping sliding-mode shows fast and accurate response even though the existence of system uncertainties and external disturbances.     

State-of-health diagnosis based on hamming neural network using output voltage pattern recognition for a PEM fuel cell

This work investigates a pattern recognition-based diagnosis approach as an application of the Hamming neural network to the identification of suitable fuel cell model parameters, which aim to diagnose state-of-health (SOH) for a polymer electrolyte membrane (PEM) fuel cell. The fuel cell output voltage (FCOV) patterns of the 20 PEM fuel cells were measured, together with the model parameters, as representative patterns. Through statistical analysis of the FCOV patterns for 20 single cells, the Hamming neural network is applied for identification of the representative FCOV pattern that matches most closely of the pattern of the arbitrary cell to be measured. Considering the equivalent circuit fuel cell model, the purpose is to select a representative loss ΔR_d, defined as the sum of two losses (activation and concentration losses). Consequently, the selected cell’s ΔR_d is properly applied to diagnose SOH of an arbitrary cell through the comparison with those of fully fresh and aged cells with the minimum and maximum of the ΔR_d in experimental cell group, respectively. This avoids the need for repeated parameter measurement. Therefore, these results could lead to interesting perspectives for diagnostic fuel cell SOH.     

A novel forecasting method based on the economic and demand response for FC/WT/PV unit and a 3 in 1 TES energy storage

Due to the lack of distribution resources and increasing demand in the daily market, the use of renewable resources is increasing. But renewable sources and market prices are uncertain behavior and cause economic problems. This paper introduces a novel market participation model include wind turbine, photovoltaic, fuel cell integrated with a novel hybrid TES energy storage system (3 in 1 concept) to minimize cost and improve load demand reliability. Also, to solve he mentioned problem a novel forecasting method are proposed. This model is a new multi artificial neural network based on the complete ensemble empirical mode decomposition which is coupled with Tanh function and using RMSE, MAPE and NMAE method the error rate of the proposed method is calculated. By using this method, the forecasting accuracy is improved and also with a novel energy storage the economic issue and market reliability are improved. Also, using the stochastic model the uncertainty system's behavior are modeled to obtain an accurate results of market participation and increase demand supply. Finally, a testing system includes wind turbine/photovoltaic/fuel cell/storage system and demand response are used to prove the superiority of the proposed model in comparison to other models.