An online method for evaluating reactive power reserve requirement of power system
By combining graph convolutional deep learning networks and hybrid control sensitivity, the real-time and accuracy problems of reactive power reserve calculation in traditional power systems are solved, enabling online assessment of reactive power reserve requirements in new power systems. This adapts to the randomness and volatility of new energy sources and flexible loads, and improves the reactive power and voltage control capabilities of the power grid.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- STATE GRID HENAN ELECTRIC ZHOUKOU POWER SUPPLY
- Filing Date
- 2026-03-17
- Publication Date
- 2026-07-21
AI Technical Summary
Traditional methods for calculating reactive power reserves in power systems suffer from poor real-time performance and low accuracy in identifying the effective output of various types of reactive power sources. They are difficult to adapt to the characteristics of new power systems, cannot meet the randomness and volatility of resources such as new energy sources and flexible loads, and do not fully consider the active and reactive power coupling characteristics of the power grid and the impact of topology changes, resulting in significant deviations in the evaluation results.
By employing graph convolutional deep learning networks combined with data-driven methods, an accurate identification model for effective reactive power output of new energy sources and flexible loads under multi-dimensional scenarios is constructed. Combined with hybrid control sensitivity, dynamic reactive power partitioning of the power grid is performed. The power grid node branch model is described through graph theory, and the training dataset is expanded using adversarial networks to achieve online assessment of reactive power reserve requirements.
It enables online, rapid, and accurate assessment of reactive power reserve requirements in new power systems, improves the real-time performance and accuracy of the assessment, adapts to the characteristics of new power systems, provides accurate data support for transmission and distribution coordinated reactive power reserve optimization, and enhances the reactive power and voltage control level of the power grid.
Smart Images
Figure CN122437133A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of power system operation and control technology, and in particular provides an online assessment method for the reactive power reserve demand of a power system. Background Technology
[0002] With the increasing scale of DC power transmission and new energy power generation in my country, the operation mode and system characteristics of the power grid have undergone significant changes: DC power transmission itself lacks reactive power and voltage support capabilities, the significant reduction in the operating capacity of traditional thermal power units has led to insufficient dynamic reactive power support capabilities of the power grid, while the penetration rate of various types of adjustable reactive power resources such as new energy, energy storage, and flexible loads on the distribution network side is constantly increasing, and the randomness and volatility of their output have brought huge challenges to the assessment of the power grid's reactive power reserve demand.
[0003] Traditional methods for calculating reactive power reserve in power systems are offline calculations that require consideration of various fault scenarios, such as N-1, and involve multiple stages including power flow calculations and transient stability calculations. These methods are time-consuming and lack real-time performance, making them ill-suited to the linear requirements of reactive power reserve demand assessment in modern power systems. Furthermore, the theoretical calculated output of adjustable reactive power resources such as renewable energy sources and flexible loads deviates significantly from their actual effective output. Existing methods lack accurate identification of the effective reactive power output of these resources, leading to substantial biases in reactive power reserve calculation results. In addition, traditional methods do not fully consider the coupling characteristics of active and reactive power in the power grid, the aggregation characteristics of multiple types of reactive power sources, and the impact of grid topology changes. Consequently, the accuracy and adaptability of the assessment results are insufficient to meet the needs of transmission and distribution coordinated reactive power reserve optimization in modern power systems.
[0004] Currently, some research has attempted to apply artificial intelligence technology to reactive power and voltage control in power systems. However, a comprehensive online assessment method for reactive power reserve demand, integrating effective reactive power output identification, regional reactive power reserve calculation, and deep learning-based online evaluation, has yet to be developed. This method fails to address the real-time and accuracy issues in assessing reactive power reserve demand under conditions of high penetration from multiple reactive power sources. Therefore, there is an urgent need to develop an online assessment method for reactive power reserve demand adapted to the characteristics of new power systems, providing accurate and real-time data support for the joint optimization of transmission and distribution-coordinated reactive power reserves. Summary of the Invention
[0005] The purpose of this application is to provide an online assessment method for reactive power reserve requirements of power systems, which solves the technical problems of poor real-time performance, low accuracy in identifying the effective output of various types of reactive power sources, and difficulty in adapting to the characteristics of new power systems in traditional offline calculation methods, so as to achieve online, rapid and accurate assessment of reactive power reserve requirements of new power systems.
[0006] To achieve the above objectives, this application adopts the following technical solution: A method for online assessment of reactive power reserve demand in a power system includes the following steps: a. Collect historical reactive power output data from wind power, photovoltaic, energy storage, and flexible loads; analyze the reactive power uncertainty distribution characteristics of various types of adjustable reactive power resources; and obtain a set of typical scenarios by adopting scenario generation and reduction strategies. Based on the data-driven approach, construct an accurate identification model for the effective reactive power output of new energy sources and flexible loads under multi-dimensional scenarios, and efficiently train the model to obtain an accurate prediction model. b. Determine the typical operation mode of the power grid, construct the corresponding set of power grid faults, study the basic calculation method of power grid reactive power reserve, propose a reactive power partitioning method that considers the dynamic reactive power reserve of the power grid, and obtain a calculation method for power grid reactive power reserve demand under multi-dimensional scenarios based on the partitioning method. c. Analyze the input characteristics of graph convolutional deep learning networks, use graph theory to describe the node branch model of the power grid, and study the parameter characteristics of the reactive power reserve demand input model and the topology change characteristics of the power grid. d. An efficient training method for convolutional deep learning networks based on cross-sectional research of historical power grid operation data. An adversarial network is used to construct a set of simulated data of historical real scenarios, expand the training dataset, and optimize the model training strategy. e. Input the effective reactive power output prediction results of step 1) and the regional reactive power reserve demand calculation results of step 2) into the trained and optimized graph convolutional deep learning model to realize the online assessment of the reactive power reserve demand of the new power system.
[0007] Furthermore, the multi-dimensional scenarios mentioned in step 1) include one or more of the following: new energy output fluctuation scenario, flexible load response scenario, grid fault disturbance scenario, and energy storage charging and discharging characteristic scenario.
[0008] Furthermore, the training method for the effective reactive power output accurate identification model described in step 1) is an iterative training method based on gradient descent. During the training process, the model parameters are dynamically adjusted with the goal of minimizing the error between the model prediction value and the actual output value.
[0009] Furthermore, the reactive power zoning method considering the dynamic reactive power reserve of the power grid described in step 2) is implemented based on the calculation results of the hybrid control sensitivity. The hybrid control sensitivity measures the physical response characteristics of the PV / PQ nodes of the power system and uses a successive solution method to calculate the sensitivity of the controlled node voltage to the reactive power injection of the control node.
[0010] Furthermore, the calculation process for the hybrid control sensitivity includes: a) Based on the PQ decomposition method, the QV iterative equation takes into account the control characteristics of the PV node to maintain constant voltage, and sets the maximum value on the corresponding diagonal element of the PV node. b) For substation nodes equipped with switchable reactive power compensation devices, the node admittance to ground is corrected according to the capacitor / reactor capacity, and the PQ node model is not used. c) Using a successive solution method, each generator node is sequentially set as a PQ node, and the remaining generator nodes are set as PV or PQ nodes according to their actual control characteristics. The reactive voltage sensitivity of each control node is calculated to obtain the hybrid control sensitivity matrix.
[0011] Furthermore, the input parameters of the graph convolutional deep learning network in step 3) include one or more of the following: power grid topology parameters, node voltage / reactive power data, output data of multiple types of reactive power sources, power grid fault type and location data, and reactive power partitioning information.
[0012] Furthermore, the adversarial network described in step 4) includes a generator and a discriminator. The generator generates a simulated scene dataset based on historical real scene data, and the discriminator distinguishes between real data and simulated data. Through adversarial training between the generator and the discriminator, the realism of the simulated scene data is improved.
[0013] Furthermore, the output results of the online assessment described in step 5) include the reactive power reserve demand value of each reactive power zone, the reactive power support demand of key nodes of the power grid, and the predicted reactive power deficit value of areas with high penetration of new energy sources. The update frequency of the output results matches the acquisition frequency of real-time monitoring data of the power grid.
[0014] Furthermore, the method also includes an accuracy verification step for the online evaluation results: the online evaluation results of the model are compared with the offline fine calculation results, and when the error exceeds a preset threshold, the graph convolutional deep learning model is retrained and the parameters are optimized based on new power grid operation data.
[0015] Compared with the prior art, the present invention has the following advantages: 1. Improved real-time performance of assessment: This invention combines data-driven and deep learning technologies to construct an online assessment model based on graph convolutional deep learning. This avoids the complex power flow and transient stability calculation steps in traditional offline calculations, significantly shortens the calculation time, and enables online assessment of reactive power reserve requirements. The output results are updated synchronously with real-time power grid monitoring data, adapting to the real-time requirements of reactive power reserve assessment in new power systems.
[0016] 2. Improved assessment accuracy: This invention constructs an accurate identification model for the effective reactive power output of various types of adjustable reactive power resources, solving the problem of large deviations between theoretical and actual output of resources such as new energy and flexible loads; at the same time, it realizes dynamic reactive power zoning of the power grid based on hybrid control sensitivity, fully considering the active and reactive power coupling characteristics of the power grid and the physical response characteristics of nodes, making the reactive power reserve calculation of the zoning more in line with the actual operating state of the power grid, and significantly improving the accuracy of reactive power reserve demand assessment.
[0017] 3. Adaptation to the characteristics of new power systems: This invention fully considers the characteristics of power grids with high penetration of new energy sources, energy storage, and flexible loads. Through scenario generation and reduction, and adversarial network to expand the dataset, it adapts to the randomness and volatility of reactive power output of multiple types of sources. At the same time, it establishes a mapping relationship between power grid topology changes and model input parameters, ensuring the model's adaptability to power grid topology changes and making it suitable for new power systems with a high proportion of new energy sources.
[0018] 4. Provide support for the optimization of reactive power reserve in transmission and distribution coordination: The online evaluation results of this invention include multi-dimensional information such as the reactive power reserve requirements of each reactive power zone and the reactive power support requirements of key nodes. This can provide accurate and real-time data support for the joint optimization of reactive power reserve requirements in subsequent transmission and distribution coordination, which helps to fully tap the potential of adjustable reactive power resources on the distribution network side and improve the economy and security of reactive power reserve optimization in the power grid. Attached Figure Description
[0019] Exemplary embodiments of this application will now be described in detail with reference to the accompanying drawings. It should be understood that the embodiments described below are for illustrative purposes only and are not intended to limit the scope of this application. In the accompanying drawings: Figure 1 A flowchart illustrating the overall process of the online assessment method for reactive power reserve requirements of the power system provided in this application. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.
[0021] This embodiment uses the online assessment of reactive power reserve demand of the new power system in Zhoukou City, Henan Province, as an example to illustrate the technical solution of the present invention in detail, but the scope of protection of the present invention is not limited to this embodiment.
[0022] The Zhoukou power grid in this embodiment contains a high proportion of wind power and photovoltaic new energy, and a large number of energy storage devices and flexible loads are connected to the distribution network. The main grid has DC feeds in, and traditional offline calculation methods can no longer meet the real-time requirements of reactive power reserve demand assessment. Figure 1 As shown, the online assessment method for reactive power reserve demand in power systems according to the present invention comprises the following specific steps: Step 1: Construct an effective reactive power output accurate identification model Historical reactive power output data from wind power, photovoltaic (PV), energy storage, and flexible loads in the Zhoukou power grid over the past three years were collected. Combined with real-time monitoring data from the power grid's SCADA system, the analysis revealed a fluctuation range of 0-100MW for wind power output, 0-80MW for PV output, a reactive power response time of 0.52s for flexible loads, and a charging / discharging power range of 0-50MW for energy storage. Monte Carlo simulations were used to generate 1000 multi-dimensional scenarios, which were then reduced using K-means clustering to obtain a set of 20 typical scenarios.
[0023] A precise identification model for effective reactive power output was constructed based on a BP neural network. A set of typical scenarios was used as training samples, and the model was trained using the stochastic gradient descent method. The goal was to minimize the mean square error between the model's predicted value and the actual output value. The weights and thresholds of the model were dynamically adjusted. After training, the model's identification error for effective reactive power output of new energy and flexible loads was controlled within 3%, resulting in a precise prediction model.
[0024] Step 2: Determine the method for calculating reactive power reserve requirements in a multi-dimensional partitioned scenario. Based on the Zhoukou City power grid plan, three typical operating modes—peak load, flat load, and valley load—were identified, and a power grid fault set was constructed, including N-1 line faults, DC blocking faults, and renewable energy disconnection faults. The reactive power reserve of the power grid was calculated using the PQ decomposition method, yielding the basic reactive power deficit of the power grid under different operating modes and fault scenarios.
[0025] Dynamic reactive power zoning of the power grid is realized based on the calculation results of hybrid control sensitivity: First, according to the QV iterative equation of the PQ decomposition method, taking into account the constant voltage control characteristics of 12 PV nodes in the Zhoukou power grid, the maximum value is set on the corresponding diagonal element; for the 8 substation nodes equipped with switchable capacitors, the node admittance to ground is corrected according to the capacitor capacity; the hybrid control sensitivity matrix is calculated by the successive solution method, and the Zhoukou power grid is divided into 3 reactive power control zones according to the matrix. The reactive power reserve demand of each zone under 20 typical scenarios is calculated, and the reactive power reserve demand calculation method under multi-dimensional scenarios of the zones is obtained.
[0026] Step 3: Analyze the input characteristics of the graph convolutional deep learning network and the description of the power grid topology. Graph Convolutional Neural Network (GCN) was selected as the base network for the online evaluation model. Using graph theory, 38 nodes of the Zhoukou power grid were used as vertices and 56 branches as edges. The input parameters of the model were determined to include: power grid topology parameters (node admittance matrix), node voltage / reactive power data, output data of wind power / photovoltaic / energy storage / flexible loads, power grid fault type and location data, and information of three reactive power zones.
[0027] Analysis revealed that the impact weights of node voltage and renewable energy output data on the evaluation results were 35% and 28%, respectively. A real-time mapping relationship was established between power grid topology changes (such as branch switching) and model input parameters (node admittance matrix) to ensure that the model can quickly adapt to changes in power grid topology.
[0028] Step 4: Train and optimize the graph convolutional deep learning model for online evaluation. One thousand historical operation data sections of the Zhoukou power grid over the past year were selected as training samples, covering peak, flat, and valley operation modes and various fault scenarios. A generative adversarial network (GAN) was constructed to generate a simulated scenario dataset: the generator generates 500 simulated scenario data based on historical real data, and the discriminator distinguishes between real and simulated data through binary classification training. After 1 thrice adversarial training, the realism of the simulated data reached over 95%.
[0029] By fusing real and simulated datasets to obtain 1500 training samples, a graph convolutional neural network was trained. Early stopping was used to prevent overfitting. After training, the offline test error of the model was 4.2%, which meets the engineering requirements.
[0030] Step 5: Implement online assessment of reactive power reserve demand for the Zhoukou power grid. The effective reactive power output prediction results from step 1 and the reactive power reserve demand calculation results from step 2 are input into the trained and optimized graph convolutional deep learning model in real time. The model is connected to the Zhoukou power grid SCADA system, with a data acquisition frequency of 5 seconds and a synchronous update frequency of 5 seconds for the model output results. The model outputs the reactive power reserve demand values for the three reactive power zones, the reactive power support demand for key DC feed-in nodes of the power grid, and the reactive power deficit prediction values for areas with high penetration of new energy sources in real time, thereby realizing the online assessment of the reactive power reserve demand of the Zhoukou power grid.
[0031] Step 6: Accuracy Verification and Model Optimization The online evaluation results of the model were compared with the offline fine-grained calculation results based on PSASP software, and an error threshold of 5% was set. After 30 days of continuous monitoring, the average error of the online evaluation results was 3.8%, and the maximum error was 4.7%, neither of which exceeded the threshold. For a case where the evaluation deviation of a sudden grid disconnection failure of a new energy source was close to the threshold, the model was retrained based on the operational data of this failure. After optimization, the evaluation deviation of the model for this type of failure was reduced to 2.9%, further improving the model accuracy.
[0032] This embodiment achieves online, rapid, and accurate assessment of the reactive power reserve requirements of the new power system in Zhoukou City through the method of the present invention. The assessment results are highly real-time and meet engineering requirements, providing accurate data support for the joint optimization of reactive power reserve in the transmission and distribution of the Zhoukou power grid, and effectively improving the reactive power and voltage control level of the power grid.
[0033] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for online assessment of reactive power reserve demand in a power system, characterized in that, Includes the following steps: a. Collect historical reactive power output data from wind power, photovoltaic, energy storage, and flexible loads; analyze the reactive power uncertainty distribution characteristics of various types of adjustable reactive power resources; and obtain a set of typical scenarios by adopting scenario generation and reduction strategies. Based on the data-driven approach, construct an accurate identification model for the effective reactive power output of new energy sources and flexible loads under multi-dimensional scenarios, and efficiently train the model to obtain an accurate prediction model. b. Determine the typical operation mode of the power grid, construct the corresponding set of power grid faults, study the basic calculation method of power grid reactive power reserve, propose a reactive power partitioning method that considers the dynamic reactive power reserve of the power grid, and obtain a calculation method for power grid reactive power reserve demand under multi-dimensional scenarios based on the partitioning method. c. Analyze the input characteristics of graph convolutional deep learning networks, use graph theory to describe the node branch model of the power grid, and study the parameter characteristics of the reactive power reserve demand input model and the topology change characteristics of the power grid. d. An efficient training method for convolutional deep learning networks based on cross-sectional research of historical power grid operation data. An adversarial network is used to construct a set of simulated data of historical real scenarios, expand the training dataset, and optimize the model training strategy. e. Input the effective reactive power output prediction results of step 1) and the regional reactive power reserve demand calculation results of step 2) into the trained and optimized graph convolutional deep learning model to realize the online assessment of the reactive power reserve demand of the new power system.
2. The online assessment method for reactive power reserve demand in a power system according to claim 1, characterized in that, The multi-dimensional scenarios mentioned in step 1) include one or more of the following: new energy output fluctuation scenario, flexible load response scenario, grid fault disturbance scenario, and energy storage charging and discharging characteristic scenario.
3. The online assessment method for reactive power reserve demand in a power system according to claim 1, characterized in that, The training method for the effective reactive power output accurate identification model described in step 1) is an iterative training method based on gradient descent. During the training process, the model parameters are dynamically adjusted with the goal of minimizing the error between the model prediction value and the actual output value.
4. The online assessment method for reactive power reserve demand in a power system according to claim 1, characterized in that, The reactive power zoning method considering the dynamic reactive power reserve of the power grid in step 2) is implemented based on the calculation results of the hybrid control sensitivity. The hybrid control sensitivity measures the physical response characteristics of the PV / PQ nodes of the power system and uses a successive solution method to calculate the sensitivity of the controlled node voltage to the reactive power injection of the control node.
5. The online assessment method for reactive power reserve demand in a power system according to claim 4, characterized in that, The calculation process for the hybrid control sensitivity includes: a) Based on the PQ decomposition method, the QV iterative equation takes into account the control characteristics of the PV node to maintain constant voltage, and sets the maximum value on the corresponding diagonal element of the PV node. b) For substation nodes equipped with switchable reactive power compensation devices, the node admittance to ground is corrected according to the capacitor / reactor capacity, and the PQ node model is not used. c) Using a successive solution method, each generator node is sequentially set as a PQ node, and the remaining generator nodes are set as PV or PQ nodes according to their actual control characteristics. The reactive voltage sensitivity of each control node is calculated to obtain the hybrid control sensitivity matrix.
6. The online assessment method for reactive power reserve demand in a power system according to claim 1, characterized in that, The input parameters of the graph convolutional deep learning network mentioned in step 3) include one or more of the following: power grid topology parameters, node voltage / reactive power data, output data of multiple types of reactive power sources, power grid fault type and location data, and reactive power partitioning information.
7. The online assessment method for reactive power reserve demand in a power system according to claim 1, characterized in that, The adversarial network described in step 4) includes a generator and a discriminator. The generator generates a simulated scene dataset based on historical real scene data, and the discriminator distinguishes between real data and simulated data. Through adversarial training between the generator and the discriminator, the realism of the simulated scene data is improved.
8. The online assessment method for reactive power reserve demand in a power system according to claim 1, characterized in that, The output results of the online assessment described in step 5) include the reactive power reserve demand value of each reactive power zone, the reactive power support demand of key nodes of the power grid, and the predicted reactive power deficit value of areas with high penetration of new energy sources. The update frequency of the output results matches the acquisition frequency of real-time monitoring data of the power grid.
9. The online assessment method for reactive power reserve demand of a power system according to any one of claims 1-8, characterized in that, The method also includes an accuracy verification step for the online evaluation results: the online evaluation results of the model are compared with the offline fine calculation results, and when the error exceeds a preset threshold, the graph convolutional deep learning model is retrained and the parameters are optimized based on new power grid operation data.