针对风电功率预测精度低且模型不稳定的问题,提出基于双阶段注意力机制生成对抗网络（SAM-WGAN-GP）的短期风电功率预测模型。首先,在生成对抗网络的生成模型中引入自注意力机制和时间注意力机制,通过自注意力机制自适应的选择输入特征,并通过时间注意力机制捕获风电数据时间序列的长时间依赖性;判别模型采用卷积神经网络,提高模型的预测精度。其次,将SAM-WGAN-GP网络的生成器损失函数和均方根误差结合作为目标函数,以提高模型的稳定性,同时为解决判别器缓慢学习的问题,引入双时间尺度更新规则（TTUR）以平衡网络的训练过程。最后,以甘肃省酒泉市某风电场的实际运行数据为例,验证SAM-WGAN-GP模型不仅能自适应选择输入特征,而且可捕捉风电数据的长时间依赖性,并提高预测精度。

Aiming at the problem of low prediction accuracy and unstable model of wind power prediction, a short-term wind power prediction model based on dual-stage self-attention mechanism wasserstein generative adversarial networks with gradient penalty （SAM-WGAN-GP） was proposed in this paper. Firstly, self-attention mechanism and time attention mechanism are introduced in the generative model of generative adversarial network to construct a dual-stage attention-based wasserstein generative adversarial networks with Gradient Penalty, the self-attention mechanism is brought into the generator to adaptively select the input features, a time attention mechanism is brought into the generator to capture the long-term dependence of wind power data time series; convolutional neural network is selected as the discriminator to improve the prediction accuracy of the model. Secondly, the generator loss function and root mean square error of SAM-WGAN-GP network are combined as the objective function to improve the stability of the model. Meanwhile, in order to solve the problem of slow learning of discriminators, TTUR is introduced to balance the training process of the network. Finally, the actual operation data of a wind farm in Jiuquan, Gansu Province are used as an example to verify that the SAM-WGAN-GP model. The results indicate that the model can not only adaptively select input features, but also capture the long time dependence of wind power date and improve the prediction accuracy.

针对传统电能质量扰动分类方法中人工选取特征困难、步骤繁琐和分类准确率低等问题，提出了一种基于粒子群优化(particle swarm optimization，PSO)算法与卷积神经网络(convolutional neural network，CNN)的扰动分类方法。首先，利用reshape函数将各电能质量扰动信号的一维时间序列分别转成行列相等的二维矩阵，并对这些二维矩阵进行适当划分，形成训练数据集和测试数据集；其次，基于CNN构建电能质量扰动的分类模型；再次，采用PSO算法对该分类模型的参数进行优化，使用训练数据集对优化后的电能质量扰动分类模型进行训练；最后，使用测试数据集对经过训练的电能质量扰动分类模型进行测试，根据输出标签得到各类电能质量扰动的分类结果。仿真结果表明：该分类模型可以自行提取电能质量扰动数据的特征，相较于其他电能质量扰动分类模型，其对电能质量扰动信号的分类准确率更高。

Aiming at the problems of difficult manual selection of features, cumbersome classification steps and low accuracy in traditional power quality disturbance classification methods, a disturbance classification method based on particle swarm optimization (PSO) and convolutional neural network (CNN) was proposed. Firstly, the one-dimensional time series of power quality disturbance signals were converted into two-dimensional matrices with equal rows and columns by using the reshaping function, and these two-dimensional matrices were properly divided into training data set and test data set. Secondly, the classification model of power quality disturbance was built based on CNN. Thirdly, the PSO algorithm was used to optimize the parameters of the classification model, and the trained data set was used to train the optimized power quality disturbance classification model. Finally, the trained power quality disturbance classification model was tested by using the test data set, and the class results of various power quality disturbances were obtained according to the output labels. Simulation results show that the classification model can extract the characteristics of power quality disturbance data by itself. Compared with other power quality disturbance classification models, this method has higher classification accuracy for power quality disturbance signals.

针对传统电力负荷预测方法存在精度较低、负荷数据噪声大等问题, 提出一种基于改进变分模态分解(VMD)、卷积神经网络(CNN)和形变长短期记忆(Mogrifier LSTM)网络的多因素电力负荷预测方法。首先, 运用麻雀搜索算法(SSA)对变分模态分解进行优化, 得到最佳效果的分解子序列, 有效减轻负荷数据噪声对预测精度的影响; 其次, 分析各因素对负荷预测的影响机理, 利用皮尔逊相关系数推导各影响因素与负荷之间的相关性, 剔除冗余特征, 大幅降低模型失准的发生概率; 最后, 采用CNN提取特征向量, 将分解后的负荷数据及温度、湿度等特征数据输入到CNN-Mogrifier LSTM深度网络模型中, 通过CNN-Mogrifier LSTM深度网络模型对特征数据进行多维分析, 提高短期负荷预测精度。算例分析结果表明, 所提出的多因素电力负荷预测模型具有较好的适配性和预测效果, 与次优VMD-CNN-Mogrifier LSTM模型相比, 在两份所用真实数据集上的预测精度分别提升0.5和2.4百分点, 为短期电力负荷预测提供一种可行的解决思路。

To address the problems of low accuracy and large noise of load data in traditional power load forecasting methods, this study proposes a multi-factor power load forecasting method based on improved Variational Modal Decomposition(VMD), Convolutional Neural Network(CNN), and deformed length short-term memory network(Mogrifier LSTM). First, the Sparrow Search Algorithm(SSA) is used to optimize the VMD and obtain the decomposition subsequence with the best effect, which effectively reduces the influence of load data noise on the prediction accuracy. Second, the influence mechanism of each factor on load prediction is analyzed, the correlation between each influencing factor and load is derived using the Pearson's correlation coefficient, and redundant features are removed, which greatly reduces the probability of model inaccuracy. Finally, a CNN is employed to extract feature vectors. The decomposed load data and feature data such as temperature and humidity are fed into the CNN-Mogrifier LSTM deep network model. The feature data are analyzed in multiple dimensions in this model to improve the short-term load prediction accuracy. The results show that the multi-factor power load prediction model proposed in this study has good adaptability and prediction effects. Compared with the suboptimal VMD-CNN-Mogrifier LSTM model, the prediction accuracy of the proposed model on two real datasets is improved by 0.5 and 2.4 percentage points, respectively, which provides a feasible solution for short-term power load forecasting.

非侵入负荷监测是全面感知负荷数据及能效优化的有效途径。当前非侵入式负荷监测算法的主要观测对象是具有调控潜力的负荷,但对于其中功率较小、负荷曲线相似的电器辨识准确率还不够理想,算法对先验数据的依赖程度较高。基于此,提出一种基于多特征联合稀疏表达的SOM-K-means非侵入式负荷辨识算法,该算法利用负荷特征训练得出最优字典,结合最优字典与多特征联合稀疏表示构建目标函数,求解多特征联合稀疏矩阵,克服了单类负荷特征限制识别负荷种类的问题;将多特征联合稀疏矩阵作为输入,结合自组织(self-organizing map, SOM)神经网络优化的K-means算法与平均绝对误差值进行快速辨识。最后,利用PLAID数据集进行了实验验证,结果表明,所提算法仅需迭代120次辨识准确率即可达到90%,提高了算法收敛速度,证明了该方法能够准确高效地实现负荷辨识。

Nonintrusive load monitoring is an effective method for comprehensively perceiving load data and optimizing energy efficiency. At present, the main observation object of nonintrusive load monitoring algorithms is the load with a regulation potential; however, the identification accuracy is poor for electrical appliances with small power and similar load curves. Moreover, the algorithm is highly dependent on prior data. Therefore, an SOM-K-means non-intrusive load identification algorithm based on multi-feature joint sparse expression is proposed in this study. The algorithm uses load features to train the optimal dictionary. The objective function is constructed by combining the optimal dictionary and multi-feature joint sparse representation, and the multi-feature joint sparse matrix is solved, which overcomes the problem of identifying load types limited by single-type load characteristics. Considering the multi-feature joint sparse matrix as the input, combined with the K-means algorithm optimized by a self-organizing map (SOM) neural network and the mean absolute error, the load was quickly identified. Finally, experimental verification using the PLAID dataset shows that the identification accuracy of the proposed algorithm can reach 90% with only 120 iterations, improving the convergence speed of the algorithm and proving that the method can realize load identification accurately and efficiently.

在Kubernetes中, HPA具备自动扩展Pod的能力, 它可以根据流量的波动情况, 在高峰时增加Pod数量以应对需求, 而在低谷时减少数量以节省资源。然而, 由于HPA是根据当前Pod的性能指标来进行扩展的, 当流量激增时, 可能会对应用服务的可用性产生不利影响, 并且当压力较小时, 算力资源的空载会导致电子资源的浪费。针对上述问题, 研究并验证一种基于时序预测的集群资源自动缩放与智能休眠唤醒策略, 使用GC-TimesNet模型对集群资源的使用情况进行预测。当资源利用率较低时, 计算出需要关闭的算力节点数量, 将这些节点设置为不可调度状态, 并驱逐节点现有的Pod, 然后将这些机器置于睡眠状态。相反, 当资源需求增加时, 会唤醒足够数量的机器, 并通过HPA控制器增加所需数量的Pod副本。实验结果表明, 该策略能够较为准确地预测集群负载的变化趋势, 结合实施智能的休眠与唤醒策略, 提升优化集群的运维管理能力, 最大程度地提高计算资源的利用率, 为降低集群能源开销提供数据支撑, 实现节能减排。

Within Kubernetes, the Horizontal Pod Autoscaler(HPA) possesses automatic Pod-scaling capability, adjusting the number of Pods based on fluctuations in traffic, increasing the Pod count during peak periods to meet demand, and reducing it during off-peak times to conserve resources. However, because HPA scales are based on the current performance metrics of Pods, sudden traffic surges can potentially have detrimental effects on the availability of application services. In addition, during periods of low demand, idle computing resources lead to a waste of resources. To address these challenges, this study investigates and validates cluster resource autoscaling and intelligent sleep-wake strategy based on time-series forecasting. This strategy utilizes the GC-TimesNet model to predict cluster resource usage. When resource utilization is low, the strategy calculates the number of compute nodes that need to be shut down, marks these nodes as unschedulable, evicts existing Pods, and places these machines in a sleep state. Conversely, when the resource demand increases, a sufficient number of machines are awakened, and the HPA controller is used to increase the required number of Pod replicas. The experimental results demonstrate that this strategy can reasonably and accurately predict trends in cluster load changes, enhance the operational management capabilities for optimizing clusters, maximize the utilization of computing resources, provide data support for reducing cluster energy expenses, and achieve energy savings and emission reduction when combined with the implementation of intelligent sleep and wake strategies.