【作者】 董添；

【导师】 刘富；

【作者基本信息】 吉林大学 ， 模式识别与智能系统， 2022， 博士

【摘要】 电力是现代社会发展的命脉,是经济发展、人民生活和社会稳定的基本保障。电能具有不易存储的特点,要保证电网的稳定运行,必须使电网发电侧和负荷侧保持实时功率平衡,这就需要精确预测未来电量需求,为制定发电计划提供数据支持。然而,随着清洁能源大量接入电网,以及电动汽车数量的逐年增加,电网的复杂性和不确定性程度日益加深,对传统电力负荷预测模型提出了挑战。因此,提高电力负荷预测精度,可以减少不必要的发电,提高能源利用效率,进而降低电网运行成本,具有巨大的经济效益。深度学习作为一种端到端的特征学习模型,能够通过多层非线性变换实现对复杂数据的建模,已在诸多领域中得到了广泛的应用。基于深度学习方法开展电力负荷数据的特征学习,有望提高电力负荷预测精度,满足当前电力系统需求,是当前的一个研究热点。本文针对电力负荷模式识别和电力负荷预测需求,考虑电力负荷数据不平衡、周期性的特点,结合聚类和时间序列特征学习方法,主要开展的研究工作有:(1)提出一种基于不平衡数据聚类的电力负荷模式识别算法。在传统模糊c均值算法的基础上,利用模糊度矩阵量化聚类尺寸,并用其归一化传统算法的目标函数,进而推导出聚类中心和隶属度矩阵的更新公式。实验结果表明,所提算法能够从历史负荷数据集中识别出节假日的负荷模式,与同类算法相比,在多个聚类有效性评价指标上更优。利用多种回归模型进行负荷预测,与传统聚类算法相比,提高了对节假日的负荷预测效果。(2)建立一种基于聚类和周期增强LSTM的电力负荷预测模型。首先,利用不平衡数据聚类算法对历史数据集进行聚类分析,得到若干种典型负荷模式。其次,分别为每种负荷模式建立预测模型,考虑负荷序列的周期性特点,通过将输入序列中与待测时刻同周期的负荷值和LSTM网络的输出串联,构建出全连接层,输出负荷预测结果。实验结果表明,利用待测时刻前24小时的负荷和温度序列进行负荷预测,取得了比传统LSTM 网络和其他负荷预测模型更好的预测效果,在常用的北美电力数据集上的 MAPE 值为 1.51%。(3)建立一种基于周期增强Informer的电力负荷预测模型。将输入序列中的周期负荷值与传统Informer模型的输出串联构建一个全连接层,结合卷积神经网络,同时学习长序列中的时序特征、局部特征和周期特征,从而克服了传统Informer模型中概率稀疏自注意力机制对周期特征的丢失问题。实验结果表明,以待测时刻前1周的负荷和温度序列作为输入,进一步提高了负荷预测精度,在北美电力数据集上的MAPE值降低至1.15%。(4)开发一种联合LSTM和注意力机制的电力负荷概率预测模型。首先,分别以不同随机初值运行周期增强Informer模型,得到一组负荷预测结果,以其方差作为模型的不确定性。其次,计算负荷均值与真实值之间的差的平方,利用基于注意力机制的LSTM网络拟合与输入序列之间的关系,建立数据不确定性估计模型。最后,以模型不确定性和数据不确定的和作为负荷不确定性的估计结果,利用标准正态分布临界值确定不同置信度下的负荷上界和下界。实验结果表明,所提模型在多种指标上均优于传统Informer模型和其他同类模型。综上,本文结合电力数据特点和深度学习方法,开展了电力负荷模式识别算法和预测模型研究,取得了良好的效果。更多还原

【Abstract】 Electric power is the lifeblood of modern social development and the basic guarantee for economic development,people’s life and social stability.Electric energy is not easy to store.In order to ensure the stable operation of the power grid,it is necessary to keep the real-time power balance between the power generation side and the load side of the power grid,which requires accurate prediction of future power demand and data support for the formulation of power generation plans.However,with a large number of clean energy connected to the power grid and the number of electric vehicles increasing year by year,the complexity and uncertainty of the power grid are deepening,which challenges the traditional power load forecasting model.Therefore,improving the accuracy of power load forecasting can reduce unnecessary power generation,improve energy utilization efficiency,and then reduce the operation cost of power grid,which has great economic benefits.As an end-to-end feature learning model,deep learning can model complex data through multi-layer nonlinear transformation,and has been widely used in many fields.The feature learning of power load data based on deep learning method is expected to improve the accuracy of power load forecasting and meet the demand of current power system.It is a research hotspot at present.Aiming at the demand of load pattern recognition and load forecasting and considering the unbalanced and periodic characteristics of power load data,this paper combines clustering and time series feature learning methods and carries out the following research contents:(1)A power load pattern recognition algorithm based on unbalanced data clustering is presented.On the basis of the traditional fuzzy c-mean algorithm,the fuzzy degree matrix is used to quantify the cluster size and normalize the objective function of the traditional algorithm,and then the renewal formulas of the cluster center and membership degree matrix are derived.The experimental results show that the proposed algorithm can identify the load patterns of holidays from the historical load data set,and is better than the similar algorithms in multiple clustering effectiveness evaluation indicators.Compared with the traditional clustering algorithm,the load forecasting effect on holidays is improved by using multiple regression models.(2)A power load forecasting model based on clustering and periodic enhancement LSTM is proposed.Firstly,the unbalanced data clustering algorithm is used to cluster the historical data sets,and several typical load patterns are obtained.Secondly,a prediction model is established for each load pattern.Considering the periodicity of the load sequence,the ful1 connection layer is constructed by connecting the load value in the input sequence and the output of the LSTM network in the same period with the time to be predicted in series,and the load prediction results are output.The experimental results show that the load forecasting using the load and temperature series of 24 hours before the time to be predicted achieves better forecasting results than the traditional LSTM network and other load forecasting models.The MAPE value on the commonly used North American power data set is 1.51%.(3)A power load forecasting model based on periodic enhancement Informer is proposed.The periodic load values in the input sequence and the output of the traditional Informer model are connected in series to form a full connected layer.Combined with the convolution neural network,the time-series characteristics,local characteristics and periodic characteristics in the long sequence are learned at the same time,thus overcoming the loss of periodic characteristics caused by the probability sparse self attention mechanism in the traditional Informer model.The experimental results show that the accuracy of load forecasting is further improved by taking the load and temperature series of 1 week before the time to be predicted as input,and the MAPE value on the North American power data set is reduced to 1.15%.(4)A probabilistic load forecasting model based on the combination of LSTM and attention mechanism is proposed.First,the Informer model is enhanced with different random initial values and operation periods respectively,and a group of load forecasting results are obtained,and the variance is taken as the model uncertainty.Secondly,the square of the differences between the load mean and the real value is calculated,and the data uncertainty estimation model is established by using the relationship between the LSTM network fitting based on the attention mechanism and the input sequence.Finally,taking the sum of model uncertainty and data uncertainty as the estimation result of load uncertainty,the upper and lower bounds of load under different confidence levels are determined by using the critical value of standard normal distribution.The experimental results show that the proposed model is superior to the traditional Informer model and other similar models in many indexes.To sum up,combining the characteristics of power data and deep learning method,this paper has carried out the research on power load pattern recognition algorithm and load forecasting model,and achieved good results.更多还原