The risk of transmission overload (TO) in power grids is increasing with the large-scale integration of intermittent renewable energy sources. An effective online preventive control schemes proves to be vital in safeguarding the security of power systems. In this paper, we formulate the online preventive control problem for TO alleviation as a constrained Markov decision process (CMDP), targeted to reduce the load rate of overloaded lines by implementing generation re-dispatch, transmission and busbar switching actions. The CMDP is solved with a state-of-the-art safe deep reinforcement learning method, based on the computationally efficient interior-point policy optimization (IPO), which facilitates desirable learning behavior towards constraint satisfaction and policy improvement simultaneously. The performance of IPO method is further improved with an enhanced perception of spatial-temporal correlations in power gird nodal and edge features, combining the strength of edge conditioned convolutional network and long short-term memory network, fostering more effective and robust preventive control policies to be devised. Case studies on a real-world system and a large-scale system validate the superior performance of the proposed method in TO alleviation, constraint handling, uncertainty adaptability and stability preservation, as well as its favorable computational performance, through benchmarking against both model-based and reinforcement learning-based baseline methods.

Active power dispatch is one of the major operation tasks for power system that keeps the power generation and consumption in a real-time balance. Real-time decisions for active power dispatch can be seen as a modification on the Day-ahead OPF (DAOPF) and are constrained by power system safety rules. Consequently, active power dispatch focuses mainly on the perspective of operation safety which is influenced by renewable generation power fluctuation, load variation and maintenance. It is impossible to incorporate these stochastic factors in a DAOPF problem. To solve this problem, active power dispatch is formulated into a CMDP problem, where the target is to optimize the dispatch policy that maximize the reward without breaching the safety constraints. Risk of blackout arises from the violation of safety constraints and potential human and property damage would be enormous, which makes the problem different from common RL task. To handle this problem, IPO algorithm, which belongs to safe RL, is adopted. In IEEE 14-bus system, the proposed method is implemented to show the advantages from full aspects.

Data-driven methods have been intensively investigated in transient stability prediction due to the advantages on speed and accuracy. However, the variability of power systems disables the well-trained model when the contingencies or operation points are not covered in original training set. To address this issue, this paper proposes a combinational transfer learning framework to update transient stability prediction model in time-varying power systems, where convolutional neural network (CNN) is selected as the classifier. An innovative sample transfer algorithm is proposed to select applicable samples from source system, which decreases the time for time-domain simulation. Meanwhile, different model transfer schemes are compared for better accuracy and training efficiency of CNN. Test results on IEEE 39-bus system and an actual power grid verifies the efficiency and scalability of the proposed method. In addition, it performs well in the imbalanced training set and data with random noise.

As the terminal of electricity consumption, the distribution network is a vital field to lower the carbon emission of the power system. With the integration of distributed energy resources, the flexibility of the distribution network has been promoted significantly where dispatch actions can be employed to lower carbon emissions without compromising the accessibility of reliable electricity. This study proposes a security constrained dispatch policy based on safe reinforcement learning for the distribution network. The researched problem is set up as a constrained Markov decision process, where continuous-discrete mixed action space and high-dimensional state space are in place. In addition, security-related rules are embedded into the problem formulation. To guarantee the generalization of the reinforcement learning agent, various scenarios are generated in the offline training stage, including randomness of renewables, scheduled maintenance, and different load profiles. A case study is performed on a modified version of the IEEE 33-bus system, and the numerical results verify the effectiveness of the proposed method in decarbonization.

从数据驱动的角度研究暂稳预测问题存在模型变迁和样本匮乏的困扰.针对该问题提出了一种基于继承思想的暂态功角稳定预测方法,该方法适用于暂稳数据增量或暂稳镜像变化的电网过渡阶段,填补了现有预测方法的空缺区间.分别针对电力系统暂稳样本增量的数据时变性和暂稳特性变化的暂稳镜像时变性特点,提出可计及现有预测模型参数的深度继承方法,以及可考虑样本与特征集可延拓性的广度继承方法.基于增量学习的深度继承利用新增暂稳样本,计算暂稳预测模型参数的变化量,降低模型更新所需训练时间.基于迁移学习的广度继承建立了源系统与目标系统的特征集与样本集迁移通道,实现目标系统在小样本条件下的暂稳预测模型构建.算例结果表明,该文所提方法能够反映时变电力系统的暂态稳定特性的动态变化,在样本匮乏的过渡阶段,具有速度和精度优势.

Dispatching system is an essential part for power system operation, where control signals are sent out based on the analysis of real-time data. Dispatch system itself is a complex-structured computer system, containing lots of devices. Consequently, potential faults of software and hardware endanger the security of power system. To address this issue, an early waring method for dispatching system faults is proposed based on knowledge graph. Firstly, dispatch system warning signals are digitalized into vectors, which is convenient for rule-based knowledge graph. Then, a discrete-continuous combined algorithm is proposed to calculate the similarity of event chains. Finally, a knowledge graph is established to save the event chains, where real-time warning vectors are analyzed based on the knowledge graph for upcoming faults. Numerical test is conducted in a simulated dataset. The test results prove the efficiency and accuracy of the proposed method.

In this paper, a prediction model of cascading commutation failure (CF) based on integration of data-driven method and model-driven method is proposed. The impact of AC system fault is quantified by a data-driven model, which is trained by multiple simulation cases with different fault locations and durations. Extreme learning machine (ELM) is applied in this data model for its fast calculation. Adjusted multi-infeed interaction factor (MIIF) is firstly proposed that considers nonlinearity of voltage interactions between converter stations. Then, voltage sensitivity matrix among converter stations is calculated based on adjusted MIIF, which is a simplified physics model established to compute the voltage drop in target station due to the voltage drop at neighbor stations. Afterwards, the total voltage drop on target station is obtained by adding the output of data model and physics model. Finally, the prediction of CF is given by comparing the predicted voltage with critical voltage. The method is verified in a two-infeed HVDC test system and East-China power grid. Extensive simulations are conducted in PSD-BPA and the result shows the advantage in speed and accuracy of the proposed model.

现代智能电网出现了电力电子化、信息物理融合和大电网复杂互联等新特征,从而对电力系统暂态问题的分析与控制方法产生了极大影响。人工智能(artificialintelligence,AI)在解决数据问题中的优势与暂态问题特点匹配程度高。该文从信息、机理、仿真、分析和控制等角度分析了电力系统暂态问题出现的新特点,归纳总结了现有将AI用于分析电力系统暂态问题时的研究成果,指出了研究中仍需解决的问题,探讨了在数据获取、特征提取和算法应用等方面的若干改进思路,并对AI应用于暂态稳定问题的研究现状进行总结。

With the development of smart grid, operation and control methods of power grid are more abundant and flexible, which makes contribution to solve the uncertain risks introduced by photovoltaic (PV) power plants, and also puts forward higher requirements for the decision-making effects of power grid. This paper first analyzes the role of three measures in the regulation and control of photovoltaic-integrated power system from the perspective of security and economy, namely inverter phase modulation, grid dynamic segmentation, and distributed power flow controller (DPFC). Then, taking DPFC as an example, a quantitative risk assessment and control method for photovoltaic-integrated power system is proposed. This method takes into account the uncertainty of photovoltaic output and the regulatory costs to quantify the operational control risk and control effect of the PV-integrated power system. Finally, IEEE RTS79 test system is established to verify the rationality and effectiveness of the proposed method.

Transient stability prediction using machine learning algorithms has been highly concerned. Existing researches have made great progress in simplified scenarios, while these simplifications cannot be made in a practical problem. Considering two typical characteristics of power systems: time-varying and data-increasing, two methods of prediction model updating are proposed in this paper. Transfer learning is applied to adding samples to a time-varying system with insufficient training samples. Incremental learning is used to update the prediction model with constantly increasing training samples. Both two methods make the prediction model more robust in a dynamic power system. In IEEE 39-bus system, the proposed methods are tested in angle and frequency stability problem respectively. Results show that transfer learning makes more accurate prediction in a time-varying power system than traditional method and incremental learning enables model updating to be fast enough to be applied online. More importantly, the proposed updating methods can be further investigated for practical application.

Grounding fault and injecting AC into DC control system in substation may lead to many protector problems, such as maloperation and miss trip, Based on the traditional unbalanced bridge and blocking condenser principle, the IUB(Improved Unbalanced Bridge) is put forward to detect ground resistance and alternating voltage occurred in substation DC system. The circuit of IUB models in MATLAB/Simulink is presented, and a DC system fault detection circuit board is constructed to achieve the detection function of grounding fault and AC intrusion. The results showed that the proposed method is correct and effective.

Frequency prediction after a disturbance has received increasing research attention given its substantial value in providing a decision-making foundation in power system emergency control. With the advancing development of machine learning, analysis power systems with machine-learning methods has become completely different from traditional approaches. In this paper, an ensemble algorithm using cross-entropy as a combination strategy is presented to address the trade-off between prediction accuracy and calculation speed. The prediction difficulty caused by inadequate numbers of severe disturbance samples is also overcome by the ensemble model. In the proposed ensemble algorithm, base learners are selected following the principle of diversity, which guarantees the ensemble algorithm's accuracy. Cross-entropy is applied to evaluate the fitting performance of the base learners and to set the weight coefficient in the ensemble algorithm. Subsequently, an online prediction model based on the algorithm is established that integrates training, prediction and updating. In the Western System Coordinating Council 9-bus (WSCC 9) system and the Institute of Electrical and Electronics Engineers 39-bus (IEEE 39) system, the algorithm is shown to significantly improve the prediction accuracy in both sample-rich and sample-poor situations, verifying the effectiveness and superiority of the proposed ensemble algorithm.