基于改进堆栈降噪自动编码器的预想事故频率指标评估方法研究

Research on Frequency Indicators Evaluation of Disturbance Events Based on Improved Stacked Denoising Autoencoders

可再生能源大规模并网导致电力系统转动惯量降低,在扰动事件下的频率稳定问题突出。时域仿真存在计算量大、运算耗时长等缺陷,难以满足复杂多变运行方式和海量预想事故下的频率指标快速评估需求。为了实现功率扰动事件下系统惯性中心多维频率指标(极值频率、最大频率变化率、准稳态频率)的快速评估,该文将深度学习引入到频率稳定研究中,提出一种基于改进堆栈降噪自动编码器(improved stacked denoising autoencoders,ISDAE)的智能化评估方法。首先,利用随机森林算法筛选出重要特征变量作为输入数据,实现输入数据降维;然后,将多个降噪自动编码器堆叠,构建深度学习网络结构;采用"预训练-参数微调"方法训练网络参数,引入Dropout技术提高算法泛化能力、防止过拟合,基于均方根反向传播(root mean square back propagation,RMSprop)优化方法对网络参数进行微调,减小陷入局部最优的概率;最后,根据离线训练得到的ISDAE网络结构实现扰动事件后系统惯性中心的多维频率指标在线评估。在修改后的IEEE RTS-79系统进行测试,与时域仿真、浅层神经网络以及未改进的SDAE方法所得结果进行比较,验证所提方法的快速性、准确性以及良好的泛化能力。

Since the massive integration of renewable generation reduced the system inertia level, frequency stability issues occurred in the events of sudden power imbalances. Due to the excessive computation burden of the time-domain simulation method to simulate credible disturbances, it is difficult to meet the rapid assessment requirement of frequency stability under uncertain operational scenarios and massive credible contingencies. In order to realize rapid assessment of multiple frequency metrics of the center of inertia(i.e.,frequency nadir, maximum rate-of-change of frequency,quasi-steady state frequency), a deep learning method based on improved stacked denoising autoencoders(ISDAE) was applied to frequency stability assessment. First, the random forest algorithm was used to screen the data. The importance feature variables were used as the inputs to achieve data dimensionality reduction and reduce model complexity.Second, multiple denoising autoencoders were stacked, the"pre-training, fine-tuning" method was used to train network parameters, and the nonlinear complex mapping between input data and output data was established. In the pre-training process, the Dropout method was used to improve algorithm generalization ability and prevent over-fitting. Then, the rootmean square back propagation(RMSprop) optimization was deployed to fine-tune network parameters so as to reduce the possibility of falling into local optimums. Finally, according to the established ISDAE network, online assessment of multi-frequency indicators was realized. Case studies on the modified IEEE RTS-79 system demonstrate the rapidity, high accuracy and well generalization ability of the proposed method.