【作者】 林楠；

【导师】 王怀远；

【作者基本信息】 福州大学 ， 电气工程（专业学位）， 2022， 硕士

【摘要】 快速准确的暂态稳定评估(TSA)对电力系统的安全稳定运行至关重要,能够帮助电网运营商在电力系统发生故障后采取有效的紧急控制措施。传统的TSA方法难以实现在线实时评估,因此具有一定的局限性。智能电网的发展和人工智能技术使得机器学习在TSA中的应用成为可能。深度学习算法作为机器学习的重要领域,具有精度高、速度快等特点,可以实现在线实时应用。在实际电网中存在着失稳样本数量远小于稳定样本数量的情况,导致深度学习算法在运用过程中会出现评估倾向性现象。因此,本文提出了基于后验分布信息的暂态稳定评估模型样本不平衡修正方法,主要工作内容如下:(1)电力系统运行信息如功角、角速度等都是时序相关信息,其并不是相互独立而是存在着时序逻辑关系。现有的数据驱动的TSA研究工作大都是在固定时刻输出暂态稳定评估结果,往往忽略了信息之间的时间相关性。为了适应电力系统信息的时序特点,基于长短时记忆网络(LSTM)构建时间自适应TSA模型。提出的TSA模型的优化目标是最小化所有时刻的评估误差。相较于以最后一个时刻的评估误差值为优化目标,本文提出的基于LSTM的TSA模型能够根据不同的故障情况自适应地决定评估时刻,在整个评估过程中都表现出良好的评估性能。(2)在电力系统中,实际的稳定样本数量远远多于不稳定样本数量。为了解决样本不平衡带来的评估倾向性问题,本文从深度学习模型的损失函数出发,分析样本不平衡对评估模型的影响,发现训练过程中的损失函数均值比能够反映样本的不平衡程度。因此,本文提出基于样本后验分布信息的代价敏感修正方法。针对不同的不平衡样本集,该方法均能够有效改善模型的评估倾向性问题。基于样本后验分布信息的代价敏感修正方法具有模型独立性。它可以与任何具有损失函数的TSA模型相结合以修正评估倾向性问题。(3)将时间自适应方法和模型评估倾向性修正方法相结合,建立了基于后验分布信息的时间自适应TSA模型,以实现准确快速的暂态稳定评估。将多个具有不同决策时刻的基于LSTM的TSA模型聚合成一个自适应的TSA模型。利用基于后验分布信息的不平衡修正方法对每一个特定时刻的LSTM子模型进行不平衡修正。最后得到的集成模型在实现时间自适应的同时,能够很好地修正样本不平衡带来的模型评估倾向性。更多还原

【Abstract】 An accurate and fast transient stability assessment(TSA)is important for the safe and stable operation of power system.TSA can help grid operators to take effective emergency control measures after the power system suffers from large disturbances.Traditional TSA methods have limitations in real-time asseesment.The development of smart grids and artificial intelligence technology make the application of machine learning in TSA possible.As an important field of machine learning,deep learning algorithm has the characteristics of high accuracy and fast calculation speed,and it can realize online applications.In the actual power grid,the number of unstable samples is much smaller than that of stable samples,which leads to the evaluation tendency in deep learning algorithm.Therefore,an imbalance correction method of TSA model based on posterior distribution information is proposed.The main work is as follows:(1)Power system operation information such as power angle,angular velocity,etc.,are time-series related information.They have a time-series logic relationship instead of an independent relationship.However,most of the existing research work outputs the evaluation results at a fixed time,which often ignores the time correlation between time-series information.In order to adapt to the time-series features,a time-adaptive TSA model is constructed based on long short-term memory(LSTM).The optimization goal of the proposed LSTM-based TSA model is to minimize the evaluation error of all time steps.Compared with the LSTM-based TSA model,which optimization target is to minimize the evaluation error at the last time step,the proposed LSTM-based TSA model can adaptively determine the evaluation time step according to different fault conditions,and show good performance in the whole evaluation process.(2)In a power system,the number of stable samples is much larger than that of unstable samples.To solve the evaluation tendency problem caused by sample imbalance,the loss function of the deep learning model is considered and the impact of sample imbalance on the evaluation model is analyzed.It is found that the imbalance degree of samples can be reflected by the loss value in the training process.Therefore,a cost-sensitive correction method based on posterior distribution information is proposed.For different imbalanced sample sets,this method can effectively improve the evaluation tendency of the model.Moreover,the proposed imbalance correction method is model-independent.It can be combined with any TSA model which has a loss function to correct the evaluation tendency problem.(3)A time-adaptive TSA model based on posterior distribution information is constructed by combining the time-adaptive method and the sample imbalance correction method to achieve accurate and fast TSA.Firstly,multiple LSTM-based TSA models with different evaluation time steps are integrated into a time-adaptive TSA model.Then,the imbalance correction method based on posterior distribution information is applied to each sub-model at each specific time step.The final ensemble model can not only achieve time adaption but also greatly improve the evaluation tendency caused by sample imbalance.更多还原