Online Calculation Method and Device for Power System Operation Risk with Integrated Knowledge Transfer

Abstract

The present invention discloses an online calculation method and device for the operation risk of a power system with integrated knowledge transfer. The method includes: 1) Offline training of an initial neural network; 2) Monitoring devices return the current line fault information and topological information of the power system; 3) Determine whether there is a line break fault in the power system. If there is a line break fault, it proves that the topology of the system has changed, and it is necessary to take the measures in step 4) to make the neural network adapt to the current topological state; otherwise, execute step 5); 4) Based on integrated knowledge transfer, transform the initial neural network obtained in step 1), and apply the transformed neural network, then enter step 5); 5) Online apply the transformed neural network to calculate the risk under the current operating state. If the predicted value of the load loss of a certain node exceeds the set standard, the operation dispatcher arranges an emergency power supply vehicle or other mobile power supply devices for the corresponding location. The device includes: a processor and a memory. The present invention can continuously maintain the update and monitoring of the power system risk, and better control the operation risk level of the power system.

Description

Technical Field

[0001] The present invention relates to the field of power system risk control, and particularly to an online calculation method and device for the operation risk of a power system with integrated knowledge transfer. Background Art

[0002] In order to control the operation risk level of a power system and quickly feedback the control effect of a risk control strategy in real time, online and rapid calculation of operation risk is an indispensable key. The calculation process of operation risk can be divided into three stages: state sampling, state analysis, and index summarization. For system risk calculation and risk control research, improvements have also been proposed for these three stages, which are divided into two categories: model-driven and data-driven.

[0003] The model-driven method is the current mainstream idea. In the system state sampling stage, it usually adopts the Monte Carlo simulation method and the state enumeration method. However, both of them need to analyze and calculate a large number of system fault states to obtain risk indicators that meet the accuracy requirements, and their calculation efficiency is low. Therefore, the concept of importance sampling is combined and a series of improved methods based on groups [1] , based on information entropy [2] are generated; in the state analysis stage, methods based on Lagrange multipliers [3] , shadow prices [4] are proposed to decide the optimal load shedding; in order to converge the reliability index earlier, the impact increment method [5] , variance reduction [6] and other technologies are proposed. In-depth discussions on model-driven power system risk calculation and risk control technologies have been carried out at home and abroad, but the existing methods are difficult to quickly process large-scale system state sets and difficult to meet the real-time requirements of online applications.

[0004] The data-driven method is an emerging method following the informatization development of the power system. The data-driven method is a recognized calculation method with online application prospects. Advanced technical means such as machine learning are used to accelerate the state selection or state evaluation process and thus achieve online application. In the system state generation stage, the analytic hierarchy process [7]-[8] is used to associate environmental factors, which can improve the short-term fault modeling of components to a certain extent and enhance the accuracy of sampling. However, the analytic hierarchy process is fundamentally still a quantification of subjective judgment and cannot avoid human errors. Therefore, in the state generation stage, manual intervention is avoided, and the sampling process is mainly accelerated by machine learning state classification. Among them, the method combining support vector machine (Support Vector Machine, SVM) and Monte Carlo simulation was explored earlier. The combination of SVM and Latin hypercube sampling method can break through the limitation of system scale in calculating reliability [9] . These early results lack consideration of operating conditions such as line capacity in reliability analysis, and different solutions have been proposed for this[10-12] Professor Rocco first attempted to apply machine learning to dynamic reliability calculation using the local state space.

[10] The literature [11,12] considered the state changes of the system and approximately calculated the probabilistic reliability indices of two-state and multi-state systems using the machine learning classification method; the literature

[13] then considered the time series characteristics to establish a deep belief incorporated into the network, and then used the network to predict the associated reuse pattern of power grid equipment, with the potential to consider cascading failures. Most of the above studies focused on using power data for state classification, and the output was probability data in the range of 0-1, which was only suitable for calculating state probability indices such as Loss of Load Probability (LOLP), etc. To more comprehensively reflect reliability, it is necessary to establish an index system from multiple aspects. In addition to probability indices, the rapid calculation of expectation-type indices such as Expected Demand Not Supplied (EDNS) requires more in-depth research by researchers.

[0005] In terms of accelerating state analysis, the combined use of data-driven algorithms can not only make full use of energy data to obtain more accurate quantitative information of the power system. Traditional methods use the security boundary as a constraint for the system optimal power flow to ensure safe operation and solve the system's load-carrying capacity. With the help of data-driven in the state analysis stage, it is possible to avoid the time-consuming iterative solution of the system optimal power flow and save the time of state analysis. The literature

[14] took the component reliability parameters as inputs, trained an artificial neural network using the improved Back Propagation (BP) algorithm to obtain the actually supplied electric energy and thus determined the load shedding strategy; similarly, the literature [15,16] proposed to establish a probabilistic power flow and optimal power flow model using a deep network to accurately mine the high-order features of the non-linear power flow equation without using the relatively time-consuming iterative solution algorithm;

[17] further combined the prior knowledge of the physical model of Optimal Power Flow (OPF) to reduce the training difficulty and prediction error. From the perspective of reinforcement learning, the literature

[18] designed a real-time optimal power flow calculation method based on the Lagrangian-based deep reinforcement learning algorithm to respond to the changes of intermittent renewable energy.

[19] Using a Deep Reinforcement Learning Method for Dynamic Calculation of Electric Power Transfer Capability Considering System Uncertainties. Existing data-driven techniques have established regression models for system operating states (including power generation and demand) and reliability indices and continuously improved their accuracy. However, when structural faults or maintenance outages occur in the system, the original Load Curtailment (LC) solver cannot adapt to changes in the topological structure. After changes in system topology, line selection, etc., retraining must be carried out again. For a large number of potential operating scenarios, continuous network retraining and separate storage consume a large amount of computing time and storage space, which limit practical applications.

[0006] Equipment failures, maintenance, changes in renewable energy output, and fluctuations in load demand can all cause changes in the source-load and topology, thereby affecting the reliable operation of the power system. Such frequent state changes pose higher requirements for operation risk calculation and tracking. Achieving fast and accurate risk calculation is crucial for the safe and reliable operation of the system.

[0007] Although significant breakthroughs have been made in the speed of data-driven risk calculation models in recent years, due to their time-consuming training process, existing methods cannot quickly adapt to topological changes in the power system, which poses a potential hazard to the safe operation of the power system and reduces its security.

[0008] References

[0009] [1] M.R Gholami, S.H Hoseini, M.M Taheri. Assessment of power composite system annualized reliability indices based on improved particle swarm optimization and comparative study between the behaviour of GA and PSO[C]. 2008 IEEE 2nd International Power and Energy Conference, 2008: 1609 - 1612.

[0010] [2] Liu XR, Wang H, Sun QY, et al. Research on fault scenario prediction and resilience enhancement strategy of active distribution network under ice disaster[J]. Electrical Power and Energy Systems, 2022; 135: 107478.

[0011] [3] Liu ZY, Hou K, Jia HJ, et al. A Lagrange Multiplier Based State Enumeration Reliability Assessment for Power Systems With Multiple Types of Loads and Renewable Generations. IEEE Transactions on Power Systems 2021; 36(4): 3260 - 3270, 2021.

[0012] [4] Hou K, Tang PT, Liu ZY, Jia HJ, Zhu LW. Reliability assessment of power systems with high renewable energy penetration using shadow price and impact increment methods. Front Energy Res 2021; 9: 635071, 2021.

[0013] [5] Hou K, Jia HJ, Li X, et al. Impact - increment based decoupled reliability assessment approach for composite generation and transmission systems. IET Gener. Transmiss. Distrib., 2018; 12(3): 586 - 595.

[0014] [6] L. Pinto, M. Pereira. A variance reduction technique to the reliability analysis of a generation / transmission system. SAE 2003 World Congress & Exhibition. 1991; 2:892 - 895.

[0015] [7] Xin ZL, Chen L, Yang YJ, et al. Application of AHP - DEA model in comprehensive evaluation of distribution network planning reliability[C]. 2019 IEEE Sustainable Power and Energy Conference (iSPEC), 2019:290–294

[0016] [8] V. Naseri, B. Kashanizadeh, F. Varposhti. Online operational budgeting of electricity distribution network based on integration strategic and operational planning by Swot & AHP methods[C]. 2018 Electrical Power Distribution Conference (EPDC), 2018:63–68.

[0017] [9] Liang ZY. Grid risk assessment algorithm based on improved Monte Carlo and LSSVM[D]. Wuhan University, 2018.

[0018]

[10] C.M. Rocco S and J.A. Moreno. System reliability evaluation using Monte Carlo & support vector machine[C]. Annual Reliability and Maintainability Symposium, Tampa, FL, USA. 2003:482 - 486.

[0019]

[11] Muselli M. Empirical models based on machine learning techniques for determining approximate reliability expressions[J]. Reliability Engineering & System Safety 2004:83(3); 301-309.

[0020]

[12] Muselli M. Approximate multi-state reliability expressions using a new machine learning technique[J]. Reliability Engineering & System Safety 2005;89(3), 261-270.

[0021]

[13] Cai BP, Kong XD, Liu YH, et al. Application of Bayesian Networks in Reliability Evaluation[J]. IEEE Transactions on Industrial Informatics, 2019;15(4):2146-2157.

[0022]

[14] Xie Kaigui, Zhou Jiaqi. Risk assessment of power generation and transmission combined system based on ANN load shedding[J]. Automation of Electric Power Systems, 2002(22):31-33+44.

[0023]

[15] Yang Y, Yang ZF, Yu J, Zhang BS, Zhang YQ, and Yu HX. Fast Calculation of Probabilistic Power Flow: A Model-based Deep Learning Approach[J]. IEEE Transactions on Smart Grid 2020;11(3):2235-2244.

[0024]

[16] Z. Dong, K. Hou, X. Yu, et al. “Data-driven Power System Reliability Evaluation Based on Stacked Denoising Auto-encoders,” Energy Reports, 2022, vol. 8, no. 1, pp: 920-927.

[0025]

[17] Lei XY, Yang XF, Yu J, Zhao JB, Gao Q, Yu HX. Data-Driven Optimal Power Flow: A Physics-Informed Machine Learning Approach[J]. IEEE Transactions on Power Systems 2021; 36(1): 346-354.

[0026]

[18] Yan, Ziming, Xu, Yan. Real-Time Optimal Power Flow: A Lagrangian Based Deep Reinforcement Learning Approach. IEEE TRANSACTIONS ON POWER SYSTEMS. 2020, vol. 35, no. 4, pp. 3270-3273.

[0027]

[19] Wang, Tianjing, Tang, Yong. Transfer-Reinforcement-Learning-Based rescheduling of differential power grids considering security constraints, APPLIED ENERGY. Vol. 306 Part B, 2022, Article 118121.

[0028]

[20] Canada AESO power system yearly load curve. (Accessed on: Dec. 15, 2021). URL http: / / www.aeso.ca / assets / Uploads / PlanningRegions-Nov26-PRINT.pdf Summary of the Invention

[0029] The present invention provides an online calculation method and device for the operation risk of a power system with integrated knowledge transfer. The present invention performs self-adaptive update according to the topological changes in the actual operation of the power system, avoids the time-consuming process of training a new network, realizes the rapid transformation of the data-driven model with the power grid topology, and thus more timely puts it into the work of online risk calculation; therefore, the present invention can continuously maintain the update and monitoring of the power system risk, so as to better control the operation risk level of the power system, as described in detail below:

[0030] In a first aspect, an online calculation method for the operation risk of a power system with integrated knowledge transfer, the method includes:

[0031] 1) Offline training of the initial neural network;

[0032] 2) The monitoring device returns the current line fault information and topological information of the power system;

[0033] 3) Determine whether there is a line break fault in the power system. If there is a line break fault, it proves that the system has undergone topological changes, and it is necessary to take the measures in step 4) to make the neural network adapt to the current topological state;; otherwise, execute step 5);;

[0034] 4) Based on integrated knowledge transfer, transform the initial neural network obtained in step 1), and apply the transformed neural network, and enter step 5);

[0035] 5) Online apply the transformed neural network to calculate the risk in the current operating state. If the predicted value of the load loss of a certain node exceeds the set standard, the operation dispatcher arranges an emergency power supply vehicle or other movable power supply devices for the corresponding location.

[0036] Among them, the method divides the operation scenario structure of the power system into two parts: source-load fluctuation and topological structure change, and the solution method of the operation risk index is:

[0037]

[0038]

[0039] Among them, s A represents the power system state involving topological changes, s B represents the power system state involving source-load fluctuations; s (A,B) represents the power system state involving both topological changes and source-load fluctuations; P(s B |s A ) represents the conditional probability of source-load fluctuations occurring under the condition of topological changes; EDNS represents the expected load loss index; represents the node load reduction vector of state s (A,B) ; P(sA ) represents the probability of state s A ; represents the set of possible topological changes at time t; represents the set of states of possible topological changes and source-load fluctuations at time t; LOLP represents the loss of load probability index; is the node load loss flag bit of state s (A,B) .

[0040] Furthermore, when the operating scenario is source-load fluctuation, on any migration domain, models with different hyperparameters will jointly participate in the prediction of the final result. The output result of each stacked denoising autoencoder network is regarded as the input of the stacking layer and used for the training of the stacking layer with the goal of minimizing the comprehensive prediction error.

[0041] Among them, when the operating scenario is topological structure change, according to the mapping change characteristics before and after the topological change of the power system, knowledge transfer training is designed;

[0042] The knowledge transfer is used to establish a mapping function from the original space to the new domain space:

[0043] f t : Y s →Y t

[0044] Among them, Y t , Y s represent the label quantity spaces of the migration domain and the source domain respectively.

[0045] Furthermore, the transformation of the initial neural network is specifically as follows:

[0046] Change the hyperparameter r for fine-tuning training to obtain neural networks with different performance for the same topology k;

[0047] Set an additional layer on the obtained multiple solvers and train the additional layer. Use the outputs of different neural networks as the inputs of the additional layer, and train the weights and other parameters of the additional layer through supervised learning to achieve the joint participation of multiple neural networks in the optimal load shedding decision.

[0048] Among them, the training of the additional layer is:

[0049]

[0050]

[0051] Among them, ω j is the weight corresponding to the output of the jth neural network; a is the learning rate; ε is a very small value close to 0; m j τ , v jτ is the first moment and the second moment of the historical gradient value ω j ; τ is the number of iterations; β1 and β2 are the parameters for calculating the first moment and the second moment respectively; is the final predicted value calculated from the weight ω τ-1

[0052] In a second aspect, an on-line calculation device for the operation risk of a power system with integrated knowledge transfer is characterized in that the device comprises: a processor and a memory, wherein program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method steps described in any one of the first aspect.

[0053] In a third aspect, a computer-readable storage medium stores a computer program, the computer program includes program instructions, and when the program instructions are executed by a processor, the processor is enabled to execute the method steps described in any one of the first aspect.

[0054] The beneficial effects of the technical solution provided by the present invention are:

[0055] 1. The present invention can monitor the operation risk of a power system online, extract the value information of energy data based on a data mining method, and without establishing a fine model of the power system, only by means of simulation or historical operation data, the rapid calculation of the operation risk of the power system can be realized;

[0056] 2. The present invention can consider the situation of power system topology changes, and can adapt to more diverse power system operation scenarios compared with other devices for quickly calculating operation risks;

[0057] 3. The present invention can provide risk information accurate to the internal node level of the power system, can visually feedback the risk distribution of the power system in a visual form, timely repair and detect the occasions where risks exist, reduce the failure rate of the power system, improve the operation safety of the power system, and meet the needs in practical applications;

[0058] 4. The present invention can visually display the risks of each energy-consuming node of the power system, and the risk calculation time of the present invention is less than 1 second, which meets the needs of real-time update and intuitive display of node risks, and is convenient for operation personnel to timely discover and locate the risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a schematic structural diagram of an autoencoder network for quickly solving load shedding;

[0060] Figure 2 is a schematic diagram of the source domain and migration domain tasks of the present invention;

[0061] ​Figure 3 Schematic diagram of the operation risk indicators of the system under different operation scenarios;

[0062] Figure 4 Schematic diagram of node risk visualization;

[0063] Figure 5 Flow chart of the online calculation method for the operation risk of a power system with integrated knowledge transfer;

[0064] Figure 6 Schematic diagram of the structure of the online calculation device for the operation risk of a power system with integrated knowledge transfer. Specific implementation manners

[0065] To make the objectives, technical solutions and advantages of the present invention clearer, the following further describes in detail the implementation manners of the present invention.

[0066] Embodiment 1

[0067] To solve the problems existing in the background art, the embodiment of the present invention proposes a way to hierarchically process power system faults, and based on the technology of integrated knowledge transfer, speeds up the calculation process to achieve real-time monitoring of the operation risk of the power system. The embodiment of the present invention simultaneously takes the source-load fluctuation and topological structure characteristics of the power system as risks, quickly calculates to establish a load curtailment (LC) fast solver, proposes an integrated deep knowledge transfer training method to update the neural network to adapt to the new power grid topological structure, and on this basis, establishes an online calculation device for operation risk.

[0068] Among them, the initial neural network is established based on the power flow characteristics of the complete power system and the encoder network in the literature

[17] to calculate the operation risk associated with the fluctuations in both power generation and power consumption of the power system. The main task undertaken by the knowledge transfer training is to partially change the parameters in the encoder network to adapt to the implicit mapping relationship between the source-load fluctuation and the risk indicators under the new power system topological structure. In this way, most of the knowledge in the initial neural network is retained and can play a role in the power system after the topological structure changes. In addition, the embodiment of the present invention also combines the integrated learning mechanism and the knowledge transfer training, thereby weakening the dependence of the transfer effect on the hyperparameters and further improving the accuracy. After the offline training is completed, the established model will use the key measurement information to calculate the risk for the current power system state. Finally, the risk information and geographical information of the weak points of the power system energy supply can be intuitively obtained to take targeted improvement measures.

[0069] For the fluctuations of power sources and loads, a deep neural network based on the autoencoder structure undertakes the tasks of feature mining and mapping construction. For the new topological features after the topological structure update, a new neural network is established based on the mapping established by the original stacked denoising autoencoder, combined with the techniques of knowledge transfer training and ensemble learning.

[0070] Risk calculation mainly calculates the continuous power supply capacity of the power system under various possible operating scenarios. Operating risk focuses on the operating scenarios that may occur in the short term. Among them, the influencing factors include: fluctuations in load levels, component failures, changes in generator output, etc. These changes may cause the power system operation to cross the safety domain, so necessary load shedding measures need to be taken.

[0071] In order to quickly determine the optimal load shedding amount, the embodiment of the present invention is based on the established LC fast solver. Without establishing an accurate model of the power system for solving the optimal power flow, the implicit security constraints affecting the operation results will be included in the historical data and simulation information, and a fitting relationship can be established for the internal correlation between power distribution, security constraints, upper and lower limits of equipment operation constraints, and the resulting load shedding. Different from the traditional model parsing method, the embodiment of the present invention divides the operating scenario structure into two parts: source-load fluctuations and topological structure changes, and the solution method of its operating risk index correspondingly changes to:

[0072]

[0073]

[0074] Among them, s A represents the power system state involving topological changes, and s B represents the power system state involving source-load fluctuations; s (A,B) represents the power system state involving both topological changes and source-load fluctuations; P(s B |s A ) represents the conditional probability of source-load fluctuations occurring under the condition of topological changes; EDNS represents the expected demand not served index; represents the node load shedding vector of state s (A,B) ; P(s A ) represents the probability of state s A ; represents the set of possible topological changes at time t; represents the set of states of possible topological changes and source-load fluctuations at time t; LOLP represents the loss of load probability index; is the node load loss flag of state s (A,B) .

[0075] In a scenario involving only source-load fluctuations, the constraints included in the optimal power flow model are the same as the power flow equations, and there is an obvious non-linear and non-convex relationship between its power injection and the minimum load curtailment. Therefore, an autoencoder is used to establish a mapping model for this implicit function and calculate the load curtailment. Its deep network structure is as Figure 1 shown.

[0076] Among them, the neural network based on the autoencoder neural network can independently mine and characterize high-order features, and has a progressive data encoding / decoding process, which can be expressed as:

[0077] Z l = h l (h l-1 (h l-2 (···h 1 (X)))) (3)

[0078] Z j = h j (X j ) = s(W j X j + b j ) (4)

[0079]

[0080] Among them, s(~) represents the activation function; h j (~) represents the mapping function of the input-output relationship of the j-th layer, that is, the encoding function, W j and b j represent the weights and biases inside the network layer respectively; l is the number of layers; X is the input vector, Z is the implicit feature vector output by the encoding layer, b' j represents the bias of the decoding layer, X j represents the input feature vector of the j-th layer, Z j represents the output vector of the j-th layer, g j represents the decoding function of the j-th layer, T represents the transpose operation, Y j represents the output vector of the decoding layer of the j-th layer, represents the reconstructed feature vector of the j-th layer, h l represents the encoding function of the l-th layer.

[0081] For scenarios involving topological changes, embodiments of the present invention design knowledge transfer training based on the mapping change characteristics before and after the topological change of the power system, so as to gain the upper hand in adapting to topological changes by reusing prior knowledge. The fundamental reason for the mapping change characteristics before and after the topological change of the power system lies in the significant difference in its power distribution characteristics from those of the original system, resulting in the possibility that the original solver established based on the stacked denoising autoencoder may not be able to establish an accurate mapping relationship for power supply and demand in the new scenario. The mapping change characteristics before and after the topological change are manifested as follows: for an actual power system, when the topological structure of the power system changes due to the outage of a branch for fault or maintenance, the corresponding feature space of the input of the data-driven model remains unchanged, while the output feature space of the neural network undergoes a shift away from zero, that is, the input features still fall within the interval formed by the power supply capacity and the peak load, and the value range of the feature space of the output load reduction amount is higher than that in the case of the complete topology. The constraint conditions in the optimal power flow model before and after the topological change of the power system highly coincide, providing a similarity premise for knowledge transfer. Therefore, to address this issue, embodiments of the present invention adopt knowledge transfer by adjusting the original function mapping, so that the original fast calculation method can adapt to the new topological structure of the power system.

[0082] Component failures and maintenance will lead to changes in the topological structure of the power system, thus changing the implicit relationship between load shedding and power injection. Therefore, the source domain and the target domain have the same feature data distribution but different label data spaces. The main task of knowledge transfer is to establish a mapping function from the original space to the new domain space:

[0083] f t :Y s →Y t (6)

[0084] where Y t ,, Y s represent the label quantity spaces of the transfer domain and the source domain respectively. For the power system, the complete and intact power system structure is regarded as the source domain, and the power system after topological change is regarded as the transfer domain:

[0085]

[0086]

[0087] where X s , X t represent the spaces of the input features of the source domain and the transfer domain respectively,,Y s Y trespectively represent the spaces of the output features of the migration domain and the source domain; P(Y|X) represents the conditional probability of X occurring given that Y has occurred. Establishing an initial load shedding solver based on an autoencoder network can be used as the learning task of the source domain, and the learning task of the migration domain is to establish a mapping function for the corresponding output space, so as to adjust the original solver, such as Figure 2 as shown.

[0088] The transfer and application of existing knowledge from the source domain to the target domain realizes the reuse of high-order features in other similar domains. Using the updated encoder by knowledge transfer to predict the optimal load rejection can satisfy the optimal power flow constraints under the new topological structure and minimize the prediction error of the load loss under the new topological structure.

[0089] However, selecting appropriate hyperparameters for each migration domain is still a difficult problem. To improve the comprehensive performance of the migration model on systems with numerous topological changes, the embodiments of the present invention further design an ensemble learning strategy. On any migration domain, models with different hyperparameters will jointly participate in the prediction of the final result. The output result of each stacked denoising autoencoder (SDAE) network is regarded as the input of the stacking layer and is used for the training of the stacking layer with the goal of minimizing the comprehensive prediction error.

[0090] Figure 5 is the operation flowchart of the method of the present invention, which is divided into three stages: data preparation, model training, and online application. The first two links are both offline. The offline work of link 1 mainly includes system simulation to accumulate states and data processing, and the main work of link 2 is divided into the training of the source domain and the migration domain.

[0091] Step 1: Data preparation;

[0092] Sub-step 1: Monte Carlo simulation of the operating state of the power system and data accumulation.

[0093] Through the Monte Carlo method, the output of generators on the source side of the power system and the transmission line faults on the network side are simulated. Considering the fluctuations of renewable energy output and load demand in time simulation, a simulation database is established to supplement the deficiencies of historical data.

[0094] Sub-step 2: Establish a data warehouse based on the simulation database obtained in sub-step 1. The data warehouse is separated by a dimension table and a fact data table, where the topological data is used as the classification basis (dimension table), and the supply and demand information of the power system is used to fill the data in the fact data table.

[0095] The fact data table is stored in numerical type, and the number of its data items increases as the status accumulates. For example, "the power generation of each node at 0:00 of the system with no wireless line fault is recorded in array G, and the load demand of each node is recorded in array L". The dimension information that can be extracted is "no broken line fault", and the fact information is matrices G and L corresponding to power generation and power consumption respectively.

[0096] Sub-step 3: Data screening. For the data used for training in the original fact data table, select an equal number of states with and without load loss to form a class-balanced training data set.

[0097] Step 2: Offline training of the initial neural network;

[0098] Through a two-stage process of "pre-training - fine-tuning", train the SDAE network to obtain the initial neural network, and apply the initial neural network in the device of the present invention. The parameter update during the process of training the initial neural network is expressed as follows taking the weight matrix as an example.

[0099]

[0100] W i ,W i ’ are the weight matrices before and after the parameter update of the neurons in the i-th layer respectively, T represents the matrix transpose operation, Y l,T is the target value of the i-th layer of the network, Y l,P is the actual output of the l-th layer, W(i) is the weight of the i-th layer, Y i is the output of the i-th layer, ||Y l,T -Y l ||2 represents the square of the Euclidean distance between the actual output and the target value of the l-th layer, max(0,Y i ) represents the result of setting the negative elements in Y i to zero, l is the number of layers of the solver network, and r is the learning rate.

[0101] Step 3: The monitoring device returns the current line fault information and topology information of the power system;

[0102] Step 4: Determine whether there is a broken line fault in the power system. If there is a broken line fault, then a topology change occurs, and execute Step 5; otherwise, the system topology is complete, continue to apply the initial neural network in the device of the present invention and execute Step 6;

[0103] Step 5: Based on integrated knowledge transfer, transform the initial neural network obtained in Step 2, and apply the transformed neural network, then enter Step 6.

[0104] Among them, Step 5 is divided into the following sub-steps:

[0105] Sub-step 1: Knowledge transfer training of the transfer domain. The transfer domain inherits the encoder layer of the SDAE model from the source domain and only needs to be fine-tuned. The migration process of the neural network can be expressed as formula (10), and formula (9) is used for fine-tuning training after the neural network migration.

[0106] W i k =W i ,b i k =b i ,i=1,2,…l-1 (10)

[0107] Among them, W i , W i k are the weight parameters of the i-th layer neurons in the initial neural network and the k-th neural network generated for topology k, respectively, and b i , b i k They are the bias parameters of the i-th layer neurons in the initial neural network and the k-th neural network generated for topology k, respectively.

[0108] Sub-step 2: Change the hyperparameter r used for fine-tuning training and repeat sub-step 2 to obtain neural networks with different performance for the same topology k.

[0109] Sub-step 3: Set up additional layers on top of the multiple solvers obtained in sub-step 2 and train the additional layers using formulas (11)-(12). Use different neural network outputs as inputs to the additional layers, and train the weights and other parameters of the additional layers through supervised learning, so that multiple neural networks can jointly participate in the optimal shelving decision.

[0110]

[0111]

[0112]

[0113]

[0114] ω j is the weight corresponding to the output of the jth neural network; a is the learning rate; ε is close to the minimum value of 0; m j τ ,v j τ is the historical gradient value ω j The first and second order moments of ; τ is the number of iterations; β1 and β2 are the parameters for calculating the first and second order moments respectively; is the weight ω τ-1 The final predicted value calculated; YT is the target output under the current topological structure; is the square of the two-norm of the residual vector.

[0115] Step 6: Online apply the neural network to calculate the risk under the current operating state.

[0116] According to the topological state of the current power system and the fault probability based on the current state, use the Monte Carlo method to generate a set of potential scenarios, use the neural network returned in the previous step to batch calculate the load curtailment (LC) of the scenario set, and then take the expected value of the obtained load curtailment to get the risk value.

[0117] Step 7: Risk visualization

[0118] According to the risk value in Step 6 and combined with the geographical location information of the power system, draw a risk distribution map. As Figure 4 shown, the corresponding load loss prediction will be displayed at the node positions where the power system has risks. If the system risk is within the allowable range, no alarm signal will be generated; if the predicted load loss of a certain node exceeds the set standard, the operation dispatcher can arrange an emergency power supply vehicle or other mobile power supply devices for this location.

[0119] Embodiment 2

[0120] The following further introduces the solution in Embodiment 1 in combination with specific experiments, as detailed in the following description:

[0121] This method and device are tested on the IEEE-RTS79 system. This power system includes 24 nodes, 32 generator sets, 38 branches, and the peak load is 2850MW respectively. Among them, the 38 branches include: 5 transformer branches, 1 cable branch and 32 transmission branches.

[0122] The expected expected demand not served (EDNS) and loss of load probability (LOLP) are used to calculate the indicators of the power system risk. For the traditional method, one state is added each time until the coefficient of variation of the EDNS index reaches 5%, and for the other methods, 10,000 states are added each time for convenient batch processing, and the operating state of the power system is determined whether it belongs to the state with load curtailment based on 10% of the EDNS, and then the LOLP index in the above table is determined. The five methods compared in the table are all calculated using the same set of states.

[0123] Table 1 Comparison of methods for calculating system risk

[0124]

[0125]

[0126] Table 1 compares the present invention with other technical methods in terms of computing performance. The values given in the table indicate that the time taken to retrain the SDAE model or the training using one-hot encoding is too long or the accuracy is not ideal enough. The embodiments of the present invention balance the speed and accuracy of the operation risk calculation of the power system, and the comprehensive performance is better than other methods.

[0127] Table 2 (original RTS 79 system) Multi-type fault operation status table

[0128]

[0129] Table 2 verifies the effectiveness of the embodiments of the present invention in dealing with various types of faults in the power system. Seven different types of fault scenarios are used as embodiments to show the application effects of the present invention:

[0130] Scenario 1: N-1 fault scenario, a branch line 4 has a disconnection fault, taking into account the operation status of conventional generators in the original rts79 power system and the uncertainty of load fluctuations.

[0131] Scenario 2: N-1 fault scenario, considering the generator fault on node 23, taking into account the operation status of conventional generators in the system and the uncertainty of load fluctuations. When the operation status of components is determined, the randomness of wind power output and the fluctuation of load level are taken into account.

[0132] Scenarios 3-4: N-2 / N-k fault scenarios, more generator and branch line faults, taking into account the uncertainty of the operation status of components in the power generation system and the volatility of load levels.

[0133] Figure 3 and Figure 4 It shows that the operation risk calculation method and device of the present invention enable power practitioners to not only obtain the risk indicators of the power system, but also master the risk levels of each load node. Figure 5 It is the visualization effect of node risk information and geographical information, indicating that through the present invention, the risk information and geographical information of the weak points in the system energy supply can be intuitively obtained, which is beneficial for relevant operation and maintenance personnel to take targeted improvement measures.

[0134] The computer hardware configuration of the embodiments of the present invention includes an Intel Core i5-6500 CPU, 8G of memory, the operating system is windows10, the simulation software is MATLAB2021a, and the matpower toolbox is used for calculation when the traditional OPF calculates the optimal load shedding.

[0135] The settings of the data sets used in the numerical example are as follows: the scale of the source domain data set is 40,000, and the scale of the migration domain data set is 2,000. The hyperparameters for training the neural network based on SDAE are as follows: the number of hidden layers is 3, and each layer has 250, 200, and 250 neurons respectively. The settings of 5 different sets of hyperparameters used in the knowledge transfer training stage are shown in Table 3, where the selection of the learning rate is adapted to the size of the training batch and the number of update generations.

[0136] Table 3 Hyperparameter settings in the knowledge transfer training stage

[0137]

[0138] In order to obtain the training sample set X S,train , X T,train that represents the operating state of the power system and the corresponding operating risk information, and the label data set Y S , Y T that represents the optimal load shedding of the target output, it is necessary to first simulate the possible state set of the power system operation within a period of time through Monte Carlo simulation, and use the traditional optimal power flow calculation model to solve the load loss of each fault state as the target output of the sample;

[0139] Subsequently, the established data set is deconstructed according to the source, network, and load data sources and a database is established. Among them, the source load data is used as the fact table data, and the network side data is used as the dimension table header and classified and filled into the fact table data of the corresponding topological structure;

[0140] Secondly, select the data table corresponding to the complete topology from the dimension table information as the source domain training sample of the power system, and the rest of the data tables are migration domain data, and use the formulas (3)-(5) to train the initial neural network of the source domain;

[0141] After that, perform knowledge transfer training on the original neural network in each migration domain. The specific process is to first perform fine-tuning training on the basis of the original solver with different hyperparameters, and then perform supervised training on the stacked layer parameters to achieve the group decision-making of multiple updated solvers;

[0142] Finally, store the offline-trained solver in the model library, and according to the topological structure information of the power system monitored in real time, select the corresponding solver for the dynamic calculation update of the power system operation risk, and monitor its operation risk at all times.

[0143] Example 3

[0144] An on-line calculation device for the operation risk of a power system with integrated knowledge transfer, see Figure 6, the device includes: a processor 1 and a memory 2. Program instructions are stored in the memory 2, and the processor 1 calls the program instructions stored in the memory 2 to enable the device to execute the following method steps in the embodiment:

[0145] 1) Offline train the initial neural network;

[0146] 2) Monitor the device to return the current line fault information and topology information of the power system;

[0147] 3) Determine whether there is a line break fault in the power system. If there is a line break fault, it proves that the topology of the system has changed, and it is necessary to take the measures in step 4) to make the neural network adapt to the current topology state; otherwise, execute step 5);

[0148] 4) Based on integrated knowledge transfer, transform the initial neural network obtained in step 1), and apply the transformed neural network, then enter step 5);

[0149] 5) Online apply the transformed neural network to calculate the risk under the current operating state. If the predicted value of the load loss of a certain node exceeds the set standard, the operation dispatcher arranges an emergency power supply vehicle or other mobile power supply devices for the corresponding location.

[0150] Among them, the operating scenario structure of the power system is divided into two parts: source-load fluctuation and topology structure change. The solution method of the operating risk index is:

[0151]

[0152]

[0153] Among them, s A represents the power system state involving topology change, s B represents the power system state involving source-load fluctuation; s (A,B) represents the power system state involving both topology change and source-load fluctuation; P(s B | s A ) represents the conditional probability of source-load fluctuation occurring under the condition of topology change; EDNS represents the expected load shedding index; represents the node load curtailment vector of state s (A,B) ; P(s A ) represents the probability of state s A ; represents the set of possible topology changes at time t; represents the set of states of possible topology changes and source-load fluctuations at time t; LOLP represents the load loss probability index; is the node load loss flag bit of state s (A,B) .

[0154] Furthermore, when the operation scenario is source-load fluctuation, models with different hyperparameters will jointly participate in the prediction of the final result on any migration domain. The output result of each stack denoising autoencoder network is regarded as the input of the stacking layer, and the stacking layer is trained with the goal of minimizing the comprehensive prediction error.

[0155] Among them, when the operation scenario is topological structure change, knowledge transfer training is designed according to the mapping change characteristics before and after the topological change of the power system.

[0156] Knowledge transfer is used to establish a mapping function from the original space to the new domain space:

[0157] f t :Y s →Y t

[0158] Among them, Y t , Y s represent the label quantity spaces of the migration domain and the source domain respectively.

[0159] Furthermore, the specific transformation of the initial neural network is as follows:

[0160] Change the hyperparameter r for fine-tuning training to obtain neural networks with different performance for the same topology k.

[0161] Set an additional layer on the obtained multiple solvers and train the additional layer. Use the outputs of different neural networks as the input of the additional layer, and train the weights and other parameters of the additional layer through supervised learning to achieve the joint participation of multiple neural networks in the optimal load shedding decision.

[0162] Among them, the training of the additional layer is as follows:

[0163]

[0164]

[0165] Among them, ω j is the weight corresponding to the output of the jth neural network; a is the learning rate; ε is a very small value close to 0; m j τ , v j τ are the first moment and the second moment of the historical gradient value ω j ; τ is the number of iterations; β1 and β2 are the parameters for calculating the first moment and the second moment respectively; is the final prediction value calculated by the weight ω τ-1 .

[0166] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium. The storage medium includes a stored program, which controls the device where the storage medium is located to execute the method steps in the above embodiment when the program runs.

[0167] The computer-readable storage medium includes, but is not limited to, flash memory, hard disk, solid-state drive, etc.

[0168] It should be noted here that the description of the readable storage medium in the above embodiments corresponds to the description of the method in the embodiments, and the embodiments of the present invention will not be elaborated herein.

[0169] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present invention are generated in whole or in part.

[0170] The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through a computer-readable storage medium. The computer-readable storage medium can be any available medium that the computer can access or a data storage device such as a server or a data center that includes one or more integrated available media. The available medium can be a magnetic medium or a semiconductor medium, etc.

[0171] In the embodiments of the present invention, except for those with special descriptions of the models of each device, the models of other devices are not limited, as long as the devices can perform the above functions.

[0172] Those skilled in the art can understand that the drawings are only schematic diagrams of a preferred embodiment, and the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0173] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. An online calculation method for the operation risk of a power system with integrated knowledge transfer, characterized in that, The method includes: 1) Offline training of an initial neural network; 2) The monitoring device returns the current line fault information and topology information of the power system; 3) Determine whether there is a line break fault in the power system. If there is a line break fault, it proves that the topology of the system has changed, and it is necessary to take the measures in step 4) to make the neural network adapt to the current topology state; otherwise, execute step 5); 4) Based on integrated knowledge transfer, transform the initial neural network obtained in step 1), and apply the transformed neural network, then enter step 5); 5) Online apply the transformed neural network to calculate the risk in the current operating state. If the predicted value of the load loss of a certain node exceeds the set standard, the operation dispatcher arranges an emergency power supply vehicle or other movable power supply devices for the corresponding location; The method divides the operating scenario structure of the power system into two parts: source-load fluctuation and topology structure change. The solution method of the operation risk index is: Among them, s A represents the power system state involving topological changes, and s B represents the power system state involving source-load fluctuations; s (A,B) represents the power system state involving both topological changes and source-load fluctuations; P(s B |s A ) represents the conditional probability of source-load fluctuations occurring under the condition of topological changes; EDNS represents the expected demand not served index; represents the node load shedding vector of state s (A,B) ; P(s A ) represents the probability of state s A ; represents the set of possible topological changes at time t; represents the set of states of possible topological changes and source-load fluctuations at time t; LOLP represents the loss of load probability index; is the node load loss flag of state s (A,B) ; Among them, the specific transformation of the initial neural network is: Change the hyperparameter r used for fine-tuning training to obtain neural networks with different performance on the same topology k; Set an additional layer on the obtained multiple solvers and train the additional layer. Use the outputs of different neural networks as the inputs of the additional layer, and train the weights and other parameters of the additional layer through supervised learning to achieve the joint participation of multiple neural networks in the optimal load shedding decision; The training of the additional layer is: Among them, ω j is the weight corresponding to the output of the jth neural network; Y j represents the decoding layer output vector of the jth layer; a is the learning rate; ε is close to the minimum value of 0; m j τ ,v j τ is the historical gradient value ω j The first-order moment and the second-order moment of ; τ is the number of iterations; β1 and β2 are the parameters for calculating the first-order moment and the second-order moment respectively.

2. The online calculation method for the operation risk of a power system with integrated knowledge transfer according to claim 1, characterized in that, When the operating scenario is source-load fluctuation, on any transfer domain, models with different hyperparameters will jointly participate in the prediction of the final result. The output result of each stacked denoising autoencoder network is regarded as the input of the stacking layer, and the stacking layer is trained with the goal of minimizing the comprehensive prediction error.

3. The online calculation method for the operation risk of a power system with integrated knowledge transfer according to claim 1, characterized in that, When the operating scenario is topology structure change, design knowledge transfer training according to the mapping change characteristics before and after the topology change of the power system; The knowledge transfer is used to establish a mapping function from the original space to the new domain space: f t :Y s →Y t Among them, Y t , Y s respectively represent the label quantity spaces of the migration domain and the source domain.

4. An online calculation device for the operation risk of a power system with integrated knowledge transfer, characterized in that, The device includes: a processor and a memory. Program instructions are stored in the memory, and the processor calls the program instructions stored in the memory to enable the device to execute the method steps described in any one of claims 1-3.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program includes program instructions. When the program instructions are executed by the processor, the processor executes the method steps described in any one of claims 1-3.

Images