为提高风电出力的预测精度,提出一种基于Bayes优化的长短期记忆人工神经网络(long-short term memory,LSTM)的预测模型.首先,利用经验模态分解对风电历史出力序列进行分解,并对各分量及原始数据分别提取8个统计特征量,与预测前6个时刻出力值共同组成预测特征集.然后,采用绳索算法(least absolute shrinkage and selection ...operator,LASSO)从预测特征集中提取具有统计意义的特征子集,作为预测模型的输入.最后,提出基于Bayes超参数寻优的LSTM网络优化方法,以提高预测精度.选取湖北某市风电出力历史数据进行预测实验,结果表明:相较于BP神经网络、SVM、RBF网络、GRNN网络等预测模型,所提模型预测精度较高,特征提取方法较为合理.
In reliability analysis of power system, component failures are usually supposed to be independent from each other. Consequently, the number of component failure in given time interval should have ...exponential distribution and there should not be periodicity in the time series of component failure. In our previous work, it has been observed that component failure in distribution system follow power law distribution and have long term autocorrelation, i.e., it behave self organized criticality and there are phenomena like periodicity in time series of fault. The impact of extreme climate on power law distribution of fault has been identified, while the influence on periodicity of fault has not been reported. In the article, Hurst exponent, which denotes long term autocorrelation of a time series of data, is utilized to represent periodicity of power system fault. Fault data of Changsha distribution system in China is analyzed. It is found that the investigated system has long term autocorrelation. In order to investigate influence of extreme climate on periodicity in fault, the days with largest number of fault are removed and the Hurst exponent is calculated. Based on in-depth analysis of cause of removed fault data and variation of Hurst exponent, the factors contribute to periodicity of fault are identified. (6 pages)
