新颖的短期太阳能辐射预测聚类方法

需积分: 9 13 浏览量更新于2024-09-10 收藏 987KB PDF 举报

本文档探讨了一种新颖的聚类方法，用于短时太阳能辐射预测。该研究发表在《太阳能》(Solar Energy)杂志2015年12月期上，DOI为10.1016/j.solener.2015.10.053。短时太阳能辐射预测是一个关键领域，因为准确的预测对于能源生产和管理、电力系统规划以及可再生能源利用至关重要。研究者们，包括来自阿米尔卡比尔理工大学(Mohadeseh Ghayekhloo)、华盛顿博塞尔大学(Mahmoud Ghofrani)和卡津伊斯兰自由大学(Rasool Azimi)的四位作者，共同提出了这一创新性方法。论文的核心是介绍了一种结合了优化技术和机器学习算法的聚类策略。这种策略旨在通过分析历史数据中的模式和趋势，对未来的短期太阳能辐射进行更精准的预估。聚类方法可能涉及到时间序列分析，如季节性分解或ARIMA模型，以捕捉日间、周内和年度周期性变化。同时，它也有可能采用了非线性降维技术，如主成分分析(PCA)，来减少数据维度，提高预测效率。作者们可能使用了监督学习或半监督学习的方法，例如支持向量机(SVM)或神经网络，将历史观测值与天气参数（如温度、湿度、云层厚度等）关联起来，形成预测模型。为了提升模型的性能，他们可能还进行了特征选择和模型调优，确保了模型能够适应各种天气条件下的预测需求。值得注意的是，这篇论文在发布后得到了5次引用，表明其在学术界引起了关注。阅读次数超过152次，显示了研究的潜在影响力。作者们在其他项目中继续深化对优化和机器学习在太阳能领域的应用，这表明他们在这个问题上的工作具有连续性和拓展性。链接到ResearchGate的蓝线引用允许读者直接访问原始文献和其他相关研究，以便深入了解和验证他们的方法。总体来说，这项工作为太阳能辐射预测提供了一个有前景的工具，有助于提高电力系统的运行效率和可再生能源的利用效率。

learn from historical data and complexity of the ANN-

based methods to specify numerous model parameters

require a hybrid of diﬀerent methods to resolve the short-

comings of the ind ividual forecasting techniques. A hybrid

of ARMA and a time delay neural network (TDNN) was

proposed in Wu and Chan (2011) which demonstrated

more accurate results than each method separately. How-

ever, the feasibility of the proposed method for short-term

solar forecasting is not justiﬁed due to its high-

computational complexity. Voyant et al. (2012) used a

combination of ARMA and ANN with numerical weather

prediction (NWP) to forecast the hourly mean global hor-

izontal irradiance. Self-organizing map (SOM) has been

also used in hybrid methods for pattern recognition and

classiﬁcation of the input data (Chen et al., 2011; Yang

et al., 2014). Meteorological predictions for relative

humidity, solar irradiance and temperature were utilized

in Chen et al. (2011) to forecast the solar power genera-

tion. However, inaccuracies of the meteorological fore-

casts degrade the quality of the solar power prediction.

Yang et al. (2014) developed a hy brid approach

combining a SOM, a learning vector quantization (LVQ)

network, the SVR and fuzzy inference methods for a

day-ahead PV power prediction. Numerical results

demonstrated that the proposed hybrid approach per-

forms better than the traditional ANN and the simple

SVR methods individually. However, SOM has some lim-

itations for non-convex or discontinuous input distribu-

tions. In addition, non-winning neurons do not

participate in the learning phase which reduces the quality

of the map and accuracy of the forecast. Game theories

have been used for SOM training to resolve these prob-

lems and globally optimize the map (Herbert and Yao,

2005; Neme et al., 2005 ). Herbert and Yao (2005) pro-

posed a training algorithm that considers the neurons as

players with strategy sets and utility functions. The algo-

rithm was used to improve the quality of the map, but

no justiﬁcation was provided by either numerical results

or experiments. In addition, the algorithm deﬁnes the

player’s utility function based on all players’ actions and

requires solving the game for all iterations which increases

its computational complexity. Neme et al. (2005) proposed

diﬀerent strategies to enable the winning neuron to select

between its neighbors for the weight adaption. This

reduces the complexity of the original SOM algorithm

where the weight vectors of all neighboring neurons are

updated. However, the proposed strategies were non-

cooperative in topographic map formation without mini -

mizing the error measure. Engelbrecht (2007) proposed a

hybrid model of self-organizing maps (SOM), support vec-

tor regression (SVR) and particle swarm optimization

(PSO) to forecast hourly global solar radiation. SOM

algorithm was applied in the ﬁrst step to divide the entire

input space into several disjointed regions or clusters. SVR

models were then used to model each cluster for detecting

characteristic correlation between the predicted and the

actual values. In order to deal with the volatility of SVR

with diﬀerent parameters, PSO algorithm was used to

improve the forecasting performance of the SVR models.

However, the SOM algorithms may converge to non-

optimal clustering results depending on the initialization

and learning rate considered for the algorithm. In addi-

tion, neighborhood violations occur if the output space

topology does not match with the data shape.

The amoun t of solar radiation that reaches the ground is

extremely variable because of the apparent motion of the

sun and changing conditions of the atmosphere, which

make its forecasting complicated. The chaotic nature of

the solar data disrupts the neural network learning process

and presents the forecasting results with high errors (Di

Piazza et al., 2011). An improved version of SOM algo-

rithm is proposed in this paper, to group the solar time ser-

ies data into clusters to better characterize its irregular

features. This technique allows handling groups of data

separately, whic h by identifying anomalies and irregular

patterns as well as neglecting outliers, provides a better

understanding of the collected information. This leads to

more appropriate learning for neural networks and

improves the accuracy of the forecasts.

This paper develops a hybrid solar radiation forecasting

based on a novel clustering algorithm. Game-theoretic con-

cepts with new strategies are used in conjunction with SOM

to provide a modiﬁed GTSOM clustering. The proposed

clustering method increases the non-winning neurons’ par-

ticipation in the learning and enables their competition

with the winning neurons. This provides a better clustering

performance as compared to the original SOM where the

dead neurons have negligible chances to obtain input

patterns. Unlike the original SOM which deﬁnes the

neighborhood ba sed on the neurons’ distances in the two-

dimensional lattice, NG is combined with CHL to deﬁne

the neighborhood based on the neurons’ distances in the

input space. This speeds up the learning and enhances

the mapping quality. Wavelet decomposition and time

series analysis are used to preprocess the solar radiation

data. The time series analysis develops the structure of

the training and forecasting for BNNs. The proposed

GTSOM groups the pre-pr ocessed solar radiation datasets

into a number of clusters determined by the elbow method.

A cluster selection technique is proposed to select the clus-

ter that is used for the solar foreca sting of each individual

hour. Temperat ure, wind speed and wind direction data are

also included in the inputs to the BNN whose outputs

provide the solar radiation forecasts.

The rest of the paper is organized as follows. The orig-

inal SOM algorithm and game-theoretic concepts are

explained in Section 2. This section also describes the pro-

posed GTSOM clustering and the hybrid solar radiation

forecasting methods. Section 3 provides a case study where

the accuracy results are calculated for the hybrid forecast-

ing with diﬀerent clustering algorithms including the pro-

posed GTSOM method and the existing K-means, the

original SOM and NG clustering. Conclusions are given

in Section 4.

1372 M. Ghayekhloo et al. / Solar Energy 122 (2015) 1371–1383

剩余13页未读，继续阅读

oxygency

粉丝: 0
资源: 5

新颖的短期太阳能辐射预测聚类方法

Compress-Filtering and Transfer-Expanding of Data Set for Short-Term Load Forecasting

Short-Term-Load-Forecasting-for-Electric-Power-Systems

A Novel Three-Way Clustering Algorithm for Mixed-Type Data

Novel-Clustering-and-Outlier-Detection

Discovering-Topic-Representative-Terms-for-Short-Text-Clustering:发现短文本聚类的主题代表词

峰值密度聚类matlab代码-Clustering--based--density--peaks:通过快速搜索和查找密度峰值的聚类算法

A Semi-supervised Three-Way Clustering Framework for Multi-view Data

词向量进行聚类word-vector-clustering-master.zip

jsantarc/Dynamic-Time-Alignment-K-Means-Kernel-Clustering-For-Time-Sequence-Clustering:用于时间序列聚类的动态时间对齐 (DTA) K-Means 内核聚类-matlab开发

K均值聚类(K-Means聚类)-聚类算法-聚类可视化-MATLAB代码

最新资源