Load margin estimation method and system based on mechanism enhanced neural network
By introducing the power system voltage stability mechanism into the neural network model and constructing a composite loss function that includes empirical loss and mechanism loss, the problems of generalization ability and prediction accuracy of traditional load margin estimation methods in complex environments are solved, and a more reliable load margin estimation is achieved.
Patent Information
- Application Number
- CN202511030497.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-24
- Publication Date
- 2025-10-28
AI Technical Summary
Traditional load margin estimation methods are difficult to balance generalization ability, physical consistency and prediction accuracy in complex environments with changing operating scenarios and insufficient data samples.
By extracting the key relationship between load power factor and load margin from the voltage stability mechanism of the power system, a neural network model integrating physical mechanism constraints is constructed, and a composite loss function including empirical loss and mechanism loss is set. The neural network model is trained using sample data sets of continuous load growth and emergencies until the training termination conditions are met, thereby obtaining a load margin estimation model.
It improves the generalization ability, prediction accuracy and physical consistency of the load margin estimation model, thereby enhancing the reliability and practicality of power system load margin estimation.
Smart Images
Figure CN120851292A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a load margin estimation method and system based on mechanism-enhanced neural networks. Background Technology
[0002] With the large-scale application of induction motors and power electronic control equipment in smart grids, and the high proportion of renewable energy integration, the operating status of power systems is becoming increasingly complex and volatile, and voltage stability issues are becoming more prominent. Load margin, as a core indicator for measuring power system voltage stability, can provide a direct basis for power system safety assessment and control decisions through accurate estimation.
[0003] In related technologies, traditional load margin estimation methods are mainly divided into two categories. One category is based on physical models, including the PV (power-voltage) curve method obtained through continuous power flow calculations, the optimal power flow method, and the Thevenin equivalent method. These methods construct models through rigorous logical derivation, possessing clear physical interpretability and high estimation accuracy, and can analyze and predict voltage stability boundaries from the perspective of system operating mechanisms. However, their adaptability to changes in system operating states is weak, and their computational efficiency scalability is insufficient in actual online voltage security assessments. Especially in scenarios with high renewable energy penetration, they struggle to meet real-time requirements, limiting their widespread application in engineering practice.
[0004] Another type is data-driven machine learning methods. These methods collect massive amounts of system operation data, construct regression mappings between system state variables and load margins, and use models such as support vector machines, random forests, and neural networks to extract variable features and fit the dynamic evolution process of high-dimensional systems. However, because the mappings constructed by these methods rely entirely on fixed historical training data, they are only effective in specific scenarios, and their performance is constrained by the quality of the prior knowledge database. When the operating conditions of the power system change, the model's generalization ability decreases significantly, and problems such as prediction results deviating from the system's physical mechanisms and insufficient reliability are prone to occur.
[0005] In summary, traditional load margin estimation methods face technical challenges in complex environments with varying operating scenarios and insufficient data samples, making it difficult to balance generalization ability, physical consistency, and prediction accuracy. Summary of the Invention
[0006] This invention provides a load margin estimation method and system based on mechanism-enhanced neural networks, which addresses the shortcomings of traditional load margin estimation methods in complex environments with variable operating scenarios and insufficient data samples, where it is difficult to balance generalization ability, physical consistency and prediction accuracy.
[0007] On one hand, the present invention provides a load margin estimation method based on a mechanism-enhanced neural network, comprising: The key relationship between load power factor and load margin is extracted from the voltage stability mechanism of power system, and the key relationship is transformed into a quantitative constraint formula; A neural network model integrating physical mechanism constraints is constructed, and a composite loss function including empirical loss and mechanism loss is set according to the quantification constraint formula. A sample dataset containing continuous load growth and sudden events is constructed. The sample dataset is input into the neural network model, and the model parameters of the neural network model are iteratively updated according to the composite loss function until the training termination condition is met, thereby obtaining the load margin estimation model. The measured state data of the power system is obtained and input into the load margin estimation model to obtain the load margin estimate.
[0008] The load margin estimation method based on mechanism-enhanced neural networks provided by this invention extracts the key relationship between load power factor and load margin from the voltage stability mechanism of power systems, and transforms the key relationship into a quantitative constraint formula, including: By using the continuous power flow method, based on the theoretical relationship between bus voltage and bus power of a single load bus, the relationship between bus power and load margin is determined, thus obtaining the first mechanism knowledge; By determining the relationship between load power factor and bus power, the second mechanism knowledge is obtained; For the same load bus, at a set time step, the first mechanism knowledge and the second mechanism knowledge are correlated to obtain the key relationship between load power factor and load margin; The key relationships are represented using mathematical logic to obtain a quantitative constraint formula.
[0009] According to the load margin estimation method based on mechanism-enhanced neural network provided by the present invention, the first mechanism knowledge is used to describe the decrease of load margin as the bus power increases; The second mechanism is used to describe how the load power factor increases as the bus power increases; The quantization constraint formula is used to describe that for the same load bus, at a set time step, if the first load power factor is greater than the second load power factor, then the corresponding first load margin is less than or equal to the second load margin.
[0010] According to the load margin estimation method based on mechanism-enhanced neural networks provided by the present invention, the quantization constraint formula is as follows:
[0011] in,k Represents mechanistic knowledge, The representation is defined as follows: k 1 represents the first mechanism knowledge. k 2 represents the second mechanism knowledge. λ 1[ t [] represents the first load power factor at time step t. λ 2[ t [] represents the second load power factor at time step t. K p1 [ t [] indicates the first load margin. K p2 [ t ] indicates the second load margin.
[0012] According to the load margin estimation method based on mechanism-enhanced neural networks provided by the present invention, a composite loss function including empirical loss and mechanism loss is set according to the quantization constraint formula, including: The mechanism knowledge corresponding to the quantification constraint formula is extended to any pair of continuous load power factors, and the difference in load margin is calculated at a set time step. The difference in the load margin is used as a penalty term to establish a mechanistic loss function; An empirical loss function based on mean squared error is established, and the empirical loss function, the mechanistic loss function, and a pre-defined regularization term are linearly combined to obtain a composite loss function.
[0013] According to the load margin estimation method based on mechanism-enhanced neural networks provided by the present invention, the mechanism loss function is:
[0014] Where ReLU() represents the modified linear unit function, This represents the difference in load margin between the i-th load power factor and the (i+1)-th load power factor at time step t. Indicates the predicted value. Let T represent the true value, T represent the length of the time series, and K represent the number of samples.
[0015] According to the load margin estimation method based on mechanism-enhanced neural networks provided by the present invention, the empirical loss function, the mechanism loss function, and a pre-defined regularization term are linearly combined to obtain a composite loss function, including: The empirical loss function, the mechanistic loss function, and the pre-defined regularization term are weighted and summed to obtain the composite loss function.
[0016] According to the load margin estimation method based on mechanism-enhanced neural networks provided by this invention, a sample dataset containing continuous load growth and sudden events is constructed, including: The original sample data is obtained by acquiring load voltage amplitude samples, operating point angle samples, load power factor samples, and load margin samples corresponding to the voltage system under continuous load growth and sudden events. The original sample data is preprocessed to obtain a sample dataset.
[0017] According to the load margin estimation method based on mechanism-enhanced neural networks provided by the present invention, the original sample data is preprocessed to obtain a sample dataset, including: The original sample data is cleaned to obtain intermediate sample data; The intermediate sample data is standardized to obtain the sample dataset.
[0018] On the other hand, the present invention also provides a load margin estimation system based on a mechanism-enhanced neural network, comprising: The processing module is used to extract the key relationship between load power factor and load margin from the voltage stability mechanism of the power system, and to convert the key relationship into a quantitative constraint formula; The construction module is used to construct a neural network model that integrates physical mechanism constraints, and to set a composite loss function that includes empirical loss and mechanism loss according to the quantification constraint formula. The training module is used to construct a sample dataset containing continuous load growth and sudden events, input the sample dataset into the neural network model, and iteratively update the model parameters of the neural network model according to the composite loss function until the training termination condition is met, thereby obtaining the load margin estimation model. The estimation module is used to acquire measured state data of the power system and input the measured state data into the load margin estimation model to obtain the load margin estimate.
[0019] This invention provides a load margin estimation method and system based on a mechanism-enhanced neural network. The method extracts the key relationship between load power factor and load margin from the voltage stability mechanism of the power system and transforms this key relationship into a quantitative constraint formula. It constructs a neural network model that integrates physical mechanism constraints and sets a composite loss function containing empirical and mechanistic losses based on the quantitative constraint formula. It constructs a sample dataset containing continuous load growth and sudden events, inputs the sample dataset into the neural network model, and iteratively updates the model parameters according to the composite loss function until the training termination condition is met, thus obtaining a load margin estimation model. Finally, it acquires measured state data of the power system and inputs this measured state data into the load margin estimation model to obtain the estimated load margin value. Since the composite loss function used in the neural network model training process is obtained from the transformed quantitative constraint formula based on the key relationship between load power factor and load margin extracted from the voltage stability mechanism of the power system, it can effectively improve the generalization ability, prediction accuracy, and physical consistency of the load margin estimation model, thereby improving the reliability and practicality of power system load margin estimation. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0021] Figure 1 This is a flowchart illustrating the load margin estimation method based on mechanism-enhanced neural networks provided in an embodiment of the present invention. Figure 2 This is a statistical diagram illustrating the estimation accuracy indices RMSE and MI for the four models; Figure 3 and Figure 4 These are schematic diagrams showing the RMSE results obtained after setting the proportion of missing data in the training dataset and the test dataset to 10%, 20%, 30%, and 40% respectively, training the model, and applying it to the test sets with missing data and without missing data. Figure 5 This is a schematic diagram of the comparative test results after modifying and comparing the loss functions of RF and SVM models using the extracted mechanistic knowledge; Figure 6 This is a schematic diagram of the load margin estimation system based on mechanism-enhanced neural networks provided in an embodiment of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0023] This embodiment relates to the field of data processing and can be specifically applied to the load margin estimation scenario of power systems. Addressing the problems of poor generalization ability, prediction results inconsistent with system physical mechanisms, and insufficient reliability of traditional load margin estimation methods in complex environments such as variable operating scenarios and insufficient data samples, this embodiment proposes a data-driven load margin estimation method that integrates physical mechanism constraints. By introducing voltage stability-related mechanistic knowledge during the training process of the machine learning model, the generalization ability, prediction accuracy, and physical consistency of the model are improved, thereby enhancing the reliability and practicality of power system voltage stability assessment.
[0024] In practical applications, considering a steady-state power system model, a model is constructed to accommodate incremental load demand and any possible changes in power growth, as detailed below: (1) in, P l , Q l respectively load l The active power demand and reactive power demand; P l0 and Q l0 These are the initial values for active power and reactive power, respectively. λt Let be the load power factor at time t; P G0 This represents the initial active power generation. k pl , k Ql and k pG These are all directions of actual growth.
[0025] when λt As time changes, the equilibrium point will vary in the state space, causing the voltage amplitude to decrease slowly. This occurs at the upper limit of the critical load power factor. λmax At this point, the voltage amplitude drops sharply, and the system loses stability. Therefore, the load margin for voltage stability can be defined as: (2) in, λ cThis is the current load factor.
[0026] Due to the highly dynamic and nonlinear operating characteristics of power systems, the application of ML-based models in power systems is limited in the following two aspects: firstly, the quantity and quality of the training set affect the accuracy and reliability of the results; secondly, the generalization performance is weak when the stochasticity of large-scale renewable energy grid connection is enhanced.
[0027] To address the aforementioned technical issues, this embodiment integrates the constraints of the strong causal relationship between the load power factor and load margin obtained from analyzing the PV curve into the ML method. This allows system knowledge to help improve the performance of machine learning methods, thereby consistently providing more accurate load margin estimates. The following section combines... Figures 1 to 6 This invention describes the detailed scheme of the load margin estimation method and system based on mechanism-enhanced neural networks provided in the embodiments of the present invention.
[0028] like Figure 1 As shown, the load margin estimation method based on mechanism-enhanced neural networks provided in this embodiment of the invention mainly includes the following steps: Step 110: Extract the key relationship between load power factor and load margin from the voltage stability mechanism of the power system, and transform the key relationship into a quantitative constraint formula.
[0029] Understandably, in order to incorporate the knowledge of voltage stability mechanisms into the training process of neural networks, this embodiment uses the key relationship between load power factor and load margin as the proposed mechanism enhancement constraint to guide the training of neural network models.
[0030] Step 120: Construct a neural network model that integrates physical mechanism constraints, and set a composite loss function that includes empirical loss and mechanism loss based on the quantification constraint formula.
[0031] In this embodiment, the neural network model can provide a model basis for subsequent load margin estimation. According to the quantification constraint formula, a penalty term can be set as part of the mechanism loss, making the composite loss function more accurate and effective.
[0032] Step 130: Construct a sample dataset containing continuous load growth and sudden events, input the sample dataset into the neural network model, and iteratively update the model parameters of the neural network model according to the composite loss function until the training termination condition is met, thus obtaining the load margin estimation model.
[0033] Understandably, load margin estimation models can output corresponding load margin estimates based on power system state data.
[0034] Step 140: Obtain the measured state data of the power system and input the measured state data into the load margin estimation model to obtain the load margin estimate.
[0035] In one embodiment, the key relationship between load power factor and load margin is extracted from the power system voltage stability mechanism, and the key relationship is transformed into a quantitative constraint formula, specifically including: First, by using the continuous power flow method, based on the theoretical relationship between the bus voltage and the bus power of a single load bus, the relationship between the bus power and the load margin is determined, thus obtaining the first mechanism knowledge.
[0036] It is understandable that the voltage and load power of a single load bus in a power system can be obtained using the continuous power flow method. For each load bus in a power system, under the premise of voltage stability, the bus voltage gradually decreases as the bus power increases, which leads to a decrease in available power increment, i.e., a decrease in load margin. At the same time step t, if the power of the first bus is greater than that of the second bus, then the corresponding first load margin is less than that of the second load margin. That is, the first mechanism knowledge is used to describe the decrease in load margin as the bus power increases. Accordingly, the first mechanism knowledge can be expressed as: (3) in, k 1 represents the first mechanism knowledge; It is a mathematical symbol, that is, defined as; P 1[ t [] represents the power of the first bus at time step t; P 2[ t [] represents the power of the second bus at time step t; K p1 [ t [] indicates the first load margin at time step t; K p2 [ t ] represents the second load margin at time step t.
[0037] Then, the relationship between load power factor and bus power is determined to obtain the second mechanism knowledge.
[0038] In practical applications, the load power factor gradually increases with the increase of bus power. That is, at the same time step t, if the power of the first bus is greater than the power of the second bus, then the corresponding first load power factor is greater than the second load power factor. This second mechanism describes how the load power factor increases with the increase of bus power. Therefore, the second mechanism can be expressed as: (4) in, k 2 represents the second mechanism knowledge;P 1[ t ] represents the power of the first bus at time step t; P2[ t [] represents the power of the second bus at time step t; λ 1[ t [] indicates the first load power factor at time step t; λ 2[ t ] represents the second load power factor at time step t.
[0039] Next, for the same load bus, at a set time step, the first mechanism knowledge and the second mechanism knowledge are correlated to obtain the key relationship between load power factor and load margin.
[0040] Finally, the key relationships are represented through mathematical logic to obtain the quantitative constraint formula.
[0041] In this embodiment, the quantization constraint formula is used to describe that for the same load bus, at a set time step, if the first load power factor is greater than the second load power factor, then the corresponding first load margin is less than or equal to the second load margin.
[0042] Specifically, the quantization constraint formula can be expressed as: (5) in, k This represents the mechanistic knowledge after association; It is defined as; k 1 represents the first mechanism knowledge; k 2 represents the second mechanism knowledge; λ 1[ t [] indicates the first load power factor at time step t; λ 2[ t [] indicates the second load power factor at time step t; K p1 [ t [] indicates the first load margin; K p2 [ t ] indicates the second load margin.
[0043] In one embodiment, based on the quantization constraint formula, a composite loss function comprising empirical loss and mechanistic loss is defined, specifically including: First, the mechanistic knowledge corresponding to the quantification constraint formula is extended to any pair of continuous load power factors, and the difference in load margin is calculated at a set time step.
[0044] Then, the difference in load margin is used as a penalty term to establish a mechanistic loss function.
[0045] Finally, an empirical loss function based on mean squared error is established, and the empirical loss function, the mechanistic loss function, and the pre-defined regularization term are linearly combined to obtain a composite loss function.
[0046] In this embodiment, the initial and boundary training samples can be represented as: , where i and N are the index and number of training data samples, respectively; and These are the i-th input data and the output target data, respectively. To approximate an optimal function, we will... Input the neural network model and define an empirical loss function based on mean squared error, as follows: (6) in, and These are the predicted value and the actual value, respectively.
[0047] The empirical loss function calculates the deviation between the predicted value and the true value based on the mean square error, ensuring the model's basic fitting ability to the training data and ensuring that the output results are numerically close to the actual situation.
[0048] The underlying mechanism of the quantification constraint formula is extended to any pair of continuous load power factors. and The difference in load margin is calculated at time step t, as follows: (7) in, This represents the difference in load margin between the i-th load power factor and the (i+1)-th load power factor at time step t. This indicates that at time step t, the load power factor is... The load margin of a specific load bus; This indicates that at time step t, the load power factor is... The load margin of a specific load bus.
[0049] In this embodiment, the difference in load margin It can be used as a penalty term in the mechanistic loss function to improve the performance of neural network models in load margin estimation.
[0050] Furthermore, the mechanistic loss function can be expressed as: (8) Where ReLU() represents the modified linear unit function; This represents the difference in load margin between the i-th load power factor and the (i+1)-th load power factor at time step t. Indicates the predicted value; T represents the true value; T represents the length of the time series; K represents the number of samples.
[0051] The mechanism loss function penalizes predictions that violate the mechanism of the negative correlation between load power factor and load margin by correcting the linear unit function, thus forcing the model to follow the physical rules of power system voltage stability and avoiding erroneous predictions that do not conform to the mechanism.
[0052] In this embodiment, the empirical loss function, the mechanistic loss function, and the pre-defined regularization term are linearly combined to obtain a composite loss function, which specifically includes: The composite loss function is obtained by weighting and summing the empirical loss function, the mechanistic loss function, and the pre-defined regularization term.
[0053] Specifically, the composite loss function can be expressed as: (9) in, Represents the composite loss function; Represents the regularization term; Represents the network parameters of a neural network model; and Represents the regularization parameter; and These represent the L1 and L2 norms of the network weights, respectively; parameters , and All of these represent weighting coefficients.
[0054] By introducing a regularization term into the composite loss function, we can prevent the model from overfitting to noise or local features in the training data, thus enhancing its adaptability to unseen data. Simultaneously, by incorporating mechanistic constraints, the mapping relationships learned by the model can more closely approximate the fundamental laws of the system, rather than solely relying on the statistical correlation of the training data, further improving generalization performance.
[0055] Meanwhile, the composite loss function can adjust the proportions of empirical loss, mechanistic loss, and regularization term through weight coefficients, guiding iterative optimization of network parameters during training. When the proportion of empirical loss is too high, the model may ignore physical mechanisms; when the proportion of mechanistic loss is too high, it may sacrifice data fitting accuracy. By balancing the weights of these three factors, the model minimizes prediction error while strictly adhering to the intrinsic relationship between load power factor and load margin, ultimately outputting an accurate load margin estimate that conforms to physical logic.
[0056] In one embodiment, a sample dataset containing continuous load growth and sudden events is constructed, specifically including: First, the original sample data is obtained by acquiring the load voltage amplitude samples, operating point angle samples, load power factor samples, and load margin samples corresponding to the voltage system under continuous load growth and sudden events.
[0057] In practical applications, raw sample data such as voltage amplitude, operating point angle, load power factor, and load margin of the current power system can be obtained through sensors or SCADA systems.
[0058] Then, the original sample data is preprocessed to obtain the sample dataset.
[0059] In one specific implementation, the original sample data is preprocessed to obtain a sample dataset, including: First, the original sample data is cleaned to obtain intermediate sample data.
[0060] Understandably, data cleaning can remove outliers from the original sample data, such as removing potentially erroneous, duplicate, and noisy data, and fill in missing values, thereby effectively improving the data quality of the sample data.
[0061] Then, the intermediate sample data is standardized to obtain the sample dataset.
[0062] In the standardization process, data can be scaled to a specific range to make data with different characteristics comparable. This helps with subsequent data analysis and model training, allowing the model to better learn the features and patterns in the data.
[0063] The feasibility of the load margin estimation method provided in this embodiment is verified below using the IEEE 39 bus test system.
[0064] This embodiment generates raw sample data such as load voltage amplitude, operating point angles, load power factor, and load margin in PSAT. The constructed database includes two scenarios: continuous load increase and sudden events. The root mean square error (RMSE) is used as the evaluation index to assess the estimation accuracy of the load margin. The specific formula is as follows: (10) In addition to RMSE guaranteeing the model's generalizability, this embodiment also employs mechanism inconsistency (MI) to ensure the learning of predictions from mechanism-consistent models. The specific formula is as follows: (11) To verify the accuracy of the load margin estimation method provided in this embodiment, random forest (RF), support vector machine (SVM), conventional neural network (NN), and the mechanism-enhanced neural network (The proposed) provided in this embodiment were used for verification. Offline training was performed using training data generated by the IEEE 39 bus system. The four models obtained from the above training were labeled LMEM1, LMEM2, LMEM3, and LMEM4, respectively, and tested using a test set. The estimation accuracy metrics RMSE and MI of the four models are as follows: Figure 2 As shown.
[0065] The training dataset of the IEEE 39 bus system was randomly selected, and sub-databases with 1000, 5000, and 10000 data samples were constructed together with the initial whole database. Random forest, support vector machine, conventional neural network and mechanism-enhanced neural network provided in this embodiment were trained. The constructed model was used for the test set. The results of RMSE and MI are listed in Table 1 below.
[0066] Table 1. Results of RMSE and MI obtained from the test set test.
[0067] The proportions of missing data in the training and test datasets were set to 10%, 20%, 30%, and 40%, respectively. Models were trained using random forest, support vector machine, regular neural network, and the mechanism-enhanced neural network provided in this embodiment. These models were then applied to test sets with and without missing data, respectively. The RMSE results are shown below. Figure 3 and Figure 4 As shown.
[0068] Without considering changes in operating state, training and test sets were collected under different system operating states. The performance of random forest, support vector machine, conventional neural network and mechanism-enhanced neural network provided in this embodiment on the test set was observed in combination with the scenario. The results are shown in Table 2.
[0069] Table 2. Model performance test results in the test set.
[0070] Using the dataset obtained from the basic test scenario, the average CPU time estimated in a single run for the improved method provided in this embodiment and the traditional CPF method is shown in Table 3.
[0071] Table 3 Average CPU time of the improved method and the traditional CPF method
[0072] To further verify the applicability of the improved method proposed in this embodiment, the loss functions of the RF and SVM models were modified using the extracted mechanistic knowledge, and comparative tests were conducted. The results of the comparative tests are as follows: Figure 5 As shown.
[0073] Furthermore, the IEEE 39 test system can be modified by connecting four 100 MW wind farms to each system, replacing the traditional synchronous generator. Tests were conducted using the improved method proposed in this embodiment, and the results were compared with RF, SVM, and traditional NN. The results are shown in Table 4.
[0074] Table 4. Results of comparative tests between the improved method and RF, SVM, and NN.
[0075] In summary, the load margin estimation method based on mechanism-enhanced neural networks provided in this embodiment aims to achieve machine learning-based load margin estimation under variable operating conditions by mining key mechanistic knowledge of voltage stability from P-V correlation. A case study was conducted on the IEEE 39-bus test system to verify the accuracy, robustness, and scalability of the improved method in load margin estimation. Compared with traditional neural network methods, this embodiment demonstrates better generalization performance in load margin estimation, effectively improving the reliability of machine learning methods in load margin estimation.
[0076] Based on the same general inventive concept, this invention also protects a load margin estimation system based on a mechanism-enhanced neural network. The load margin estimation system based on a mechanism-enhanced neural network provided by this invention will be described below. The load margin estimation system based on a mechanism-enhanced neural network described below can be referred to in correspondence with the load margin estimation method based on a mechanism-enhanced neural network described above.
[0077] like Figure 6 As shown, the load margin estimation system based on mechanism-enhanced neural networks provided in this embodiment of the invention specifically includes: Processing module 210 is used to extract the key relationship between load power factor and load margin from the voltage stability mechanism of power system, and to convert the key relationship into a quantitative constraint formula.
[0078] Module 220 is used to construct a neural network model that integrates physical mechanism constraints, and to set a composite loss function that includes empirical loss and mechanism loss based on the quantization constraint formula.
[0079] The training module 230 is used to construct a sample dataset containing continuous load growth and sudden events. The sample dataset is input into the neural network model, and the model parameters of the neural network model are iteratively updated according to the composite loss function until the training termination condition is met, thus obtaining the load margin estimation model.
[0080] The estimation module 240 is used to acquire measured state data of the power system and input the measured state data into the load margin estimation model to obtain the load margin estimate.
[0081] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments of the relevant methods, and will not be elaborated further here.
[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A load margin estimation method based on mechanism-enhanced neural networks, characterized in that, include: The key relationship between load power factor and load margin is extracted from the voltage stability mechanism of power system, and the key relationship is transformed into a quantitative constraint formula; A neural network model integrating physical mechanism constraints is constructed, and a composite loss function including empirical loss and mechanism loss is set according to the quantification constraint formula. A sample dataset containing continuous load growth and sudden events is constructed. The sample dataset is input into the neural network model, and the model parameters of the neural network model are iteratively updated according to the composite loss function until the training termination condition is met, thereby obtaining the load margin estimation model. The measured state data of the power system is obtained and input into the load margin estimation model to obtain the load margin estimate.
2. The load margin estimation method based on mechanism-enhanced neural networks according to claim 1, characterized in that, The key relationship between load power factor and load margin is extracted from the voltage stability mechanism of power systems, and the key relationship is transformed into a quantitative constraint formula, including: By using the continuous power flow method, based on the theoretical relationship between bus voltage and bus power of a single load bus, the relationship between bus power and load margin is determined, thus obtaining the first mechanism knowledge; By determining the relationship between load power factor and bus power, the second mechanism knowledge is obtained; For the same load bus, at a set time step, the first mechanism knowledge and the second mechanism knowledge are correlated to obtain the key relationship between load power factor and load margin; The key relationships are represented using mathematical logic to obtain a quantitative constraint formula.
3. The load margin estimation method based on mechanism-enhanced neural networks according to claim 2, characterized in that, The first mechanism is used to describe how the load margin decreases as the bus power increases; The second mechanism is used to describe how the load power factor increases as the bus power increases; The quantization constraint formula is used to describe that for the same load bus, at a set time step, if the first load power factor is greater than the second load power factor, then the corresponding first load margin is less than or equal to the second load margin.
4. The load margin estimation method based on mechanism-enhanced neural networks according to claim 3, characterized in that, The quantization constraint formula is: in, k Represents mechanistic knowledge, The representation is defined as follows: k 1 represents the first mechanism knowledge. k 2 represents the second mechanism knowledge. λ 1[ t [] represents the first load power factor at time step t. λ 2[ t [] represents the second load power factor at time step t. K p1 [ t [Indicates the first load margin] K p2 [ t ] indicates the second load margin.
5. The load margin estimation method based on mechanism-enhanced neural networks according to claim 1, characterized in that, Based on the aforementioned quantification constraint formula, a composite loss function comprising empirical loss and mechanistic loss is defined, including: The mechanism knowledge corresponding to the quantification constraint formula is extended to any pair of continuous load power factors, and the difference in load margin is calculated at a set time step. The difference in the load margin is used as a penalty term to establish a mechanistic loss function; An empirical loss function based on mean squared error is established, and the empirical loss function, the mechanistic loss function, and a pre-defined regularization term are linearly combined to obtain a composite loss function.
6. The load margin estimation method based on mechanism-enhanced neural networks according to claim 5, characterized in that, The mechanism loss function is: Where ReLU() represents the modified linear unit function, This represents the difference in load margin between the i-th load power factor and the (i+1)-th load power factor at time step t. Indicates the predicted value. Let T represent the true value, T represent the length of the time series, and K represent the number of samples.
7. The load margin estimation method based on mechanism-enhanced neural networks according to claim 5, characterized in that, The empirical loss function, the mechanistic loss function, and the pre-defined regularization term are linearly combined to obtain a composite loss function, including: The empirical loss function, the mechanistic loss function, and the pre-defined regularization term are weighted and summed to obtain the composite loss function.
8. The load margin estimation method based on mechanism-enhanced neural networks according to claim 1, characterized in that, Construct a sample dataset containing continuous load growth and sudden events, including: The original sample data is obtained by acquiring load voltage amplitude samples, operating point angle samples, load power factor samples, and load margin samples corresponding to the voltage system under continuous load growth and sudden events. The original sample data is preprocessed to obtain a sample dataset.
9. The load margin estimation method based on mechanism-enhanced neural networks according to claim 8, characterized in that, The original sample data is preprocessed to obtain a sample dataset, including: The original sample data is cleaned to obtain intermediate sample data; The intermediate sample data is standardized to obtain the sample dataset.
10. A load margin estimation system based on a mechanism-enhanced neural network, characterized in that, include: The processing module is used to extract the key relationship between load power factor and load margin from the voltage stability mechanism of the power system, and to convert the key relationship into a quantitative constraint formula; The construction module is used to construct a neural network model that integrates physical mechanism constraints, and to set a composite loss function that includes empirical loss and mechanism loss according to the quantification constraint formula. The training module is used to construct a sample dataset containing continuous load growth and sudden events, input the sample dataset into the neural network model, and iteratively update the model parameters of the neural network model according to the composite loss function until the training termination condition is met, thereby obtaining the load margin estimation model. The estimation module is used to acquire measured state data of the power system and input the measured state data into the load margin estimation model to obtain the load margin estimate.