Optimization method and system for tidal current calculation

Abstract

The application relates to the field of power system flow calculation, and relates to an optimization method and system for flow calculation, which comprises the following steps: setting a determination threshold and model parameters of scene characteristics, acquiring real-time data of a power grid system, determining a scene label, and then constructing a calculation neural network; wherein the real-time data comprises measured voltage and measured power; inputting the real-time data into the calculation neural network to obtain preliminary voltage calculation values and preliminary power calculation values; constructing a loss function based on power conservation constraints and voltage phase angle balance constraints, inputting the measured voltage and measured power obtained in the power grid system, the preliminary voltage calculation values and the preliminary power calculation values into the loss function, performing iterative calculation, and outputting the calculation value of the loss function; adjusting the model parameters until the loss function tends to a target value; optimizing the calculation neural network based on the adjusted model parameters; and finally performing flow calculation through the optimized calculation neural network. The application solves the problem of imbalance between the accuracy and the calculation efficiency of flow calculation.

Description

Technical Field

[0001] This application relates to the field of power flow in power systems, and in particular to an optimization method and system for power flow calculation. Background Technology

[0002] Power flow calculation, as a fundamental computational method for evaluating and determining the steady-state operating conditions of power systems, is crucial for ensuring the reliability, stability, and optimal performance of power systems. With the development of deep learning technology, it has been gradually introduced into power system power flow calculation.

[0003] Currently, the neural network structures used in deep learning for power flow calculation are relatively fixed. However, the operating scenarios of power systems are constantly changing, and fixed neural networks cannot adapt to these dynamic changes, leading to an imbalance between computational efficiency and accuracy. Summary of the Invention

[0004] This application provides an optimization method and system for power flow calculation, which at least solves the problem of imbalance between computational efficiency and computational accuracy in power flow calculation in related technologies.

[0005] In a first aspect, embodiments of this application provide an optimization method for power flow calculation, comprising: The steps for constructing a neural network include setting a threshold for determining scene features and model parameters, acquiring real-time data from the power grid system, determining scene labels based on the real-time data and the threshold for determining scene features, and constructing a computational neural network based on the scene labels and the model parameters; wherein, the real-time data includes measured voltage and measured power. In the initial value calculation step, the real-time data is input into the computational neural network to obtain preliminary voltage and power calculation values. The neural network optimization process involves constructing a loss function based on power conservation constraints and voltage phase angle balance constraints. The measured voltage and power, along with the preliminary voltage and power calculations obtained from the power grid system, are input into the loss function for iterative calculation, and the calculated value of the loss function is output. The model parameters are adjusted until the loss function approaches the target value. Based on the adjusted model parameters, the computational neural network is optimized. Finally, power flow calculation is performed using the optimized computational neural network.

[0006] In some embodiments, the step of optimizing the neural network further includes: A physical constraint model is constructed based on power conservation constraints and voltage phase angle balance constraints, and the loss function is constructed based on the physical constraint model. The physical constraint model includes, ; in, As the first weighting coefficient, This is the second weighting coefficient. For the power conservation constraint, This refers to the voltage phase angle balance constraint.

[0007] In some embodiments, the step of optimizing the neural network further includes: The constructed loss function is as follows: ; in, This is the third weighting coefficient. It is the fourth weighting coefficient. For the measured voltage, For the measured power, The initial voltage calculation value is... This is the initial power calculation value.

[0008] In some embodiments, the step of optimizing the neural network further includes: Based on the power conservation constraint and the voltage phase angle balance constraint, physical constraint conditions are set; The preliminary voltage calculation value is input into the physical constraint model. When the calculated value of the physical constraint model satisfies the physical constraint conditions, the preliminary voltage calculation value is directly used as the output value. When the calculated value of the physical constraint model does not meet the physical constraint conditions, the initial voltage calculation value is adjusted until the physical constraint conditions are met, and the adjusted value is output as the correction value. Adjust the model parameters based on the correction value or the output value.

[0009] In some embodiments, the model parameters include calculating the weights of the fully connected layers and the convolutional kernel parameters of the neural network, and the step of optimizing the neural network further includes: When the difference between the correction value or the output value and the measured voltage exceeds a preset value, the weights of the fully connected layer in the computational neural network are periodically adjusted along the direction in which the loss function tends to the target value using an adaptive moment estimation algorithm, while keeping the convolution kernel parameters unchanged, until the calculated value of the loss function tends to the target value.

[0010] In some embodiments, the scenario features include renewable energy volatility, load mutation rate, and number of topological changes; the scenario labels are simple, complex, and critical; and the thresholds for determining the scenario features include a first threshold for renewable energy volatility, a second threshold for renewable energy volatility, a first threshold for load mutation rate, a second threshold for load mutation rate, and a preset value for topological changes. The neural network construction step further includes: When the volatility of the new energy source is lower than the first threshold of the volatility of the new energy source, the load mutation rate is lower than the first threshold of the load mutation rate, and the number of topology changes is zero, the power grid system is judged to be operating normally, the scenario label is output as "simple", and the computational neural network is a first-level neural network. When the volatility of the new energy source is higher than the second threshold of the new energy source volatility, or the load mutation rate is higher than the second threshold of the load mutation rate, or the number of topology changes is higher than the preset value of the topology change, it is determined that a fault has occurred in the power grid system, and the scenario label is output as "complex", and the computational neural network is a second-level neural network; When the volatility of the new energy source is higher than or equal to the first threshold of the new energy source volatility and lower than or equal to the second threshold of the new energy source volatility, it is determined that a fault has occurred in the power grid system. The scenario label is output as the critical point. A computational neural network for parallel computation of the first-level neural network and the second-level neural network is constructed, and the computational weights of the first-level neural network and the second-level neural network are set. The computational load in the first-level neural network is less than that in the second-level neural network, the first threshold for new energy volatility is lower than the second threshold for new energy volatility, and the first threshold for load mutation rate is lower than the second threshold for load mutation rate.

[0011] In some embodiments, the real-time data further includes: the number of line switching operations, the number of distributed power source connections, the number of photovoltaic nodes, and the active power, reactive power, and renewable energy output of each node. The step of constructing the neural network further includes: The formula for calculating the volatility of the new energy source is as follows: ; Where N is the number of photovoltaic nodes, and i is the current node. To provide power to the new energy source at the i-th node at time t, Average power; The formula for calculating the load mutation rate is as follows: ; in, Let be the node load of the i-th node at time t. Let be the node load of the i-th node at time t; the node load is obtained by summing active power and reactive power. The number of topology changes includes the number of line switching operations and the number of times distributed power sources are connected.

[0012] Secondly, embodiments of this application provide an optimization system for power flow calculation, comprising: A neural network module is configured to set a threshold for determining scene features and model parameters, acquire real-time data of the power grid system, determine scene labels based on the real-time data and the threshold for determining scene features, and construct a computational neural network based on the scene labels and the model parameters; wherein, the real-time data includes measured voltage and measured power. The initial value calculation module is configured to input the real-time data into the computational neural network to obtain preliminary voltage calculation values ​​and preliminary power calculation values; The optimized neural network module is configured to construct a loss function based on power conservation constraints and voltage phase angle balance constraints. The measured voltage and power, the preliminary voltage calculation value and the preliminary power calculation value obtained from the power grid system are input into the loss function for iterative calculation, and the calculated value of the loss function is output. The model parameters are adjusted until the loss function tends to the target value. Based on the adjusted model parameters, the computational neural network is optimized. Finally, power flow calculation is performed through the optimized computational neural network.

[0013] In some embodiments, the optimized neural network module is further configured to, A physical constraint model is constructed based on power conservation constraints and voltage phase angle balance constraints, and the loss function is constructed based on the physical constraint model. The physical constraint model includes, ; in, As the first weighting coefficient, This is the second weighting coefficient. For the power conservation constraint, This refers to the voltage phase angle balance constraint.

[0014] In some embodiments, the optimized neural network module is further configured to, The constructed loss function is as follows: ; in, This is the third weighting coefficient. It is the fourth weighting coefficient. For the measured voltage, For the measured power, The initial voltage calculation value is... This is the initial power calculation value.

[0015] Compared to related technologies, the power flow calculation optimization method and system provided in this application solves the problem of imbalance between calculation accuracy and calculation efficiency in the prior art by optimizing the model parameters of the computational neural network, and achieves the technical effect of balancing calculation accuracy and calculation efficiency.

[0016] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of an optimization method for power flow calculation according to an embodiment of this application; Figure 2 This is a structural block diagram of an optimized power flow calculation system according to an embodiment of this application; Figure 3 This is a flowchart of an optimization method for power flow calculation according to an embodiment of this application. Attached image description: 201. Construct the neural network module; 202. Calculate the initial value module; 203. Optimize the neural network module. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0020] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0021] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0022] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0023] In the field of power system analysis, power flow calculation, as a fundamental computational method for assessing and determining the steady-state operating conditions of a power system, is crucial for ensuring the reliability, stability, and optimal performance of the power system. By calculating the voltage magnitude and phase angle of all nodes, it helps power system operators make informed decisions and mitigate potential problems such as voltage violations, overloads, and system instability.

[0024] Power flow calculation essentially involves solving nonlinear, nonconvex algebraic equations derived from the node balance of net active and reactive power injections at each node in a power system. Traditional power flow calculation methods, such as the forward-backward sweep (FBS) and the Newton-Raphson method, require numerous iterations to solve the nonlinear power flow equations. These equations involve impedance parameters, load characteristics, and generator conditions of the power system, making it difficult to obtain accurate analytical solutions. Therefore, the computational complexity of traditional power flow calculation methods is extremely high.

[0025] With the development of deep learning technology, its application in power flow calculation has gradually attracted attention, and there are increasing attempts to introduce deep neural networks to predict power flow solutions. This approach uses data-driven learning to determine the nonlinear mapping relationship between grid inputs and outputs, such as node injected power and topology parameters, and grid outputs such as node voltage and branch power, attempting to overcome the efficiency bottlenecks of traditional methods. Existing technologies are mainly divided into two categories: direct mapping models based on neural networks and hybrid models based on block-based and hierarchical optimization. While these two types of computational methods improve computational efficiency in specific scenarios, their structural composition, principles, or methods still have key shortcomings.

[0026] First, existing technologies have poor static trade-offs between efficiency and accuracy and lack dynamic adaptability. Most existing technologies employ fixed-structure neural networks, such as feedforward neural networks with a fixed number of layers and neurons, or use preset rules, such as fixed compression rates based on region divisions, to balance efficiency and accuracy. However, power grid operation scenarios are highly dynamic, with sudden changes in new energy output, abrupt load changes, and rapid topology switching. Models with fixed structures or preset rules cannot adjust their complexity according to the needs of real-time scenarios.

[0027] When the load is stable and the output of new energy sources is stable, the power grid operation scenario is relatively simple, but the model still needs to retain redundant calculations, such as multi-layer feature extraction of deep networks, which will lead to a waste of computing resources.

[0028] In areas with weak grids and high penetration of new energy sources, the grid operation scenario is more complex. In such cases, lightweight networks with fixed structures have insufficient feature extraction capabilities, resulting in significantly increased output errors, and even voltage errors exceeding 5%, which cannot meet the engineering accuracy requirements, namely, the voltage error must be ≤2%.

[0029] Secondly, the generalization ability of existing models is limited, relying on training data specific to certain scenarios. Current methods are mostly trained based on historical steady-state data, such as historical data on typical load levels and fixed renewable energy output. However, actual power grid operation involves numerous abnormal dynamic scenarios, such as sudden drops in photovoltaic output due to extreme weather or sudden load surges. Since the generalization ability of neural networks depends on the coverage of the training data, when the input data exceeds the distribution of the training set, such as sudden changes in line parameters or topology reconstruction, the model output error increases sharply. When the deviation between the test scenario and the training scenario exceeds 30%, the voltage error increases from 2% to 8%.

[0030] Secondly, existing technologies lack sufficient accuracy guarantees for critical nodes, leading to global compromises. In power flow calculations, the voltage and power of some nodes are crucial to grid safety, such as balancing nodes, PV nodes, and nodes affecting the operating thresholds of protection devices, while the accuracy requirements for other nodes are relatively lower, such as PQ nodes. Existing technologies typically employ a uniform accuracy optimization strategy for all nodes, which necessitates increasing the global computational load to improve the accuracy of critical nodes, such as by increasing the number of network layers, at the expense of overall efficiency. If the global computational load is to be reduced, the accuracy of critical nodes may not meet the requirements, and the voltage error of balancing nodes may exceed the allowable error range.

[0031] Finally, existing neural network models suffer from weak adaptability to dynamic scenarios and lack online learning mechanisms. Most existing models operate in a static mode of offline training and online inference, unable to update model parameters online based on the real-time operating status of the power grid. When the power grid enters a new, untrained state, the model needs to be retrained offline, failing to meet real-time requirements.

[0032] To address the aforementioned issues, this application provides an optimization method and system for power flow calculation. By integrating power system physical information with deep learning technology and designing a dynamic balancing mechanism, it solves the problem of balancing computational efficiency and accuracy in existing technologies, improves the model's generalization ability, and enables flexible and efficient power flow calculation to meet the efficiency and accuracy requirements under different operating scenarios.

[0033] Figure 1 This is a flowchart of an optimization method for power flow calculation according to an embodiment of this application, such as... Figure 1 As shown, the process includes the following steps: In step S101 of constructing the neural network, the threshold values ​​for determining scene features and model parameters are set. Real-time data from the power grid system is acquired, and scene labels are determined based on the real-time data and the threshold values ​​for scene features. A computational neural network is then constructed based on the scene labels and model parameters. The real-time data includes measured voltage and measured power.

[0034] Specifically, by connecting to real-time power grid measurement systems, such as Supervisory Control and Data Acquisition (SCADA) and Phasor Measurement Units (PMUs), real-time data such as node injected power, node voltage, branch power, and renewable energy output, as well as power grid topology parameters, are collected. Power grid topology parameters include line impedance and node type. Node types include slack nodes, PQ nodes, and PV nodes. In the neural network construction step S101, the collected real-time data comes from the power system under different operating conditions, including real-time data from normal operation, fault conditions, and various transitional states.

[0035] In this context, a PQ node refers to a node where both active and reactive power are known quantities, while the node voltage amplitude and node voltage phase angle are quantities to be determined.

[0036] A PV node is a node whose active power and voltage amplitude are known quantities, while reactive power and node voltage phase angle are quantities to be determined.

[0037] In a power system, a node is an electrical connection point, representing a point in the power network where electrical parameters such as voltage and current have definite and uniform values.

[0038] Furthermore, the collected real-time data and power grid topology parameters are preprocessed. Data cleaning and normalization operations are performed on the collected real-time data to ensure data quality and consistency.

[0039] Specifically, according to Normalize the real-time data, whereby... Here, x represents the normalized value of the data, and x represents the original value of the data. This is the historical average. The standard deviation is used for normalization to eliminate the influence of dimensions.

[0040] according to The calculation shows that n is the number of historical data points. according to The calculation yielded the result.

[0041] In the initial value calculation step S102, real-time data is input into the computational neural network to obtain preliminary voltage and power calculation values.

[0042] In step S103 of the neural network optimization, a loss function is constructed based on power conservation constraints and voltage phase angle balance constraints. The measured voltage and power, as well as the preliminary voltage and power calculations obtained from the power grid system, are input into the loss function for iterative calculation, and the calculated value of the loss function is output. The model parameters are adjusted until the loss function converges to the target value. Based on the adjusted model parameters, the computational neural network is optimized. Finally, power flow calculation is performed using the optimized computational neural network.

[0043] By constructing computational neural networks adapted to different scenarios and combining real-time data to determine scenario labels and perform targeted modeling, the limitations of a single model are avoided. Physical constraints such as power conservation and voltage phase angle balance are incorporated into the loss function, solving the physical consistency problem of purely data-driven models. Continuous optimization of the neural network is achieved through iterative adjustment of model parameters, ultimately improving the efficiency and adaptability of power flow calculation while ensuring computational accuracy and reliability, further balancing the computational efficiency and accuracy of power flow calculation.

[0044] In some embodiments, the step S103 of optimizing the neural network further includes: A physical constraint model is constructed based on power conservation constraints and voltage phase angle balance constraints, and a loss function is constructed based on the physical constraint model.

[0045] The physical constraint model includes, .

[0046] in, As the first weighting coefficient, This is the second weighting coefficient. For the power conservation constraint, This is a voltage phase angle balance constraint. The first and second weighting coefficients can be set according to actual needs.

[0047] The possible value is 0.1. The possible value is 0.05.

[0048] By transforming power conservation and voltage phase angle balance constraints into quantifiable physical constraint models, a clear physical basis is provided for the construction of the loss function, solving the problem of integrating the physical laws of the power grid with neural network optimization. The influence ratio of different constraints can be flexibly adjusted through the first and second weight coefficients, enhancing adaptability to different power grid scenarios. The loss function can accurately capture errors in calculated values ​​that deviate from physical laws, guiding model parameters to optimize in a direction that satisfies physical constraints, further reducing the physical bias of the calculation results.

[0049] Ideally, , .

[0050] By embedding a physical constraint model at the output of the computational neural network and correcting the output value of the physical constraint model through a loss function, the output value of the computational neural network after iterative optimization conforms to the physical laws of the power grid.

[0051] In some embodiments, the step S103 of optimizing the neural network further includes: The constructed loss function is: .

[0052] in, This is the third weighting coefficient. It is the fourth weighting coefficient. To measure voltage, To measure power, These are preliminary voltage calculation values. These are preliminary power calculation values. The third and fourth weighting coefficients can be set according to actual needs.

[0053] The possible value is 1. The possible value is 0.5.

[0054] The loss function incorporates both the deviation between measured and calculated values ​​and physical constraints, forming a dual error monitoring system that ensures both data fitting accuracy and compliance with physical laws. This prevents the computational neural network from merely fitting measurement noise while ignoring the underlying physical properties. The third and fourth weight coefficients can be adjusted to control the weights of voltage and power deviations, adapting to different scenarios such as voltage-sensitive systems or power scheduling cores. The collaborative optimization of multi-dimensional errors enables the neural network to converge simultaneously towards directions that closely match measured data and satisfy physical constraints, reducing optimization oscillations and achieving a balance between convergence speed and computational accuracy.

[0055] In some embodiments, the step S103 of optimizing the neural network further includes: Physical constraints are set based on power conservation constraints and voltage phase angle balance constraints.

[0056] The initial voltage calculation value is input into the physical constraint model. When the calculated value of the physical constraint model meets the physical constraint conditions, the initial voltage calculation value is directly used as the output value.

[0057] When the calculated values ​​of the physical constraint model do not meet the physical constraint conditions, the initial voltage calculation values ​​are adjusted until the physical constraint conditions are met. The adjusted values ​​are then output as correction values.

[0058] Adjust the model parameters based on the correction or output values.

[0059] Before adjusting model parameters, physical constraints are verified on the initial calculated values. Calculated values ​​that meet the constraints are directly output, avoiding invalid iterations and significantly improving power flow calculation efficiency. Calculated values ​​that do not meet the constraints are specifically adjusted, ensuring the physical validity of the output values ​​while reducing the computational cost of model optimization. Model parameters are adjusted based on corrected values ​​or valid output values, enabling model parameter optimization to precisely focus on the weak links that cause constraint non-compliance, accelerating the evolution of the computational neural network towards output values ​​that naturally satisfy physical constraints.

[0060] The initial voltage calculation value is input into the physical constraint model, and the loss function is run simultaneously. The calculated value of the physical constraint model is adjusted by the loss function so that the calculated value of the physical constraint model tends to zero, ensuring that both the output value and the correction value satisfy the power conservation constraint and the voltage phase angle balance constraint.

[0061] Embedding a loss function with physical constraints at the output of a computational neural network can force correction of solutions that are not power-conserved or voltage-unbalanced from the output of the computational neural network.

[0062] By embedding power conservation constraints and voltage phase angle balance constraints, the computational neural network is forced to output power flow solutions that conform to the physical laws of the power grid. Even when the input data exceeds historical data, the voltage error can still be controlled within 3% in the face of sudden changes in line parameters and topology reconfiguration. Compared with the pure data-driven models of existing technologies, this framework can still control the voltage error within 3% in extreme scenarios, significantly improving generalization ability, reducing dependence on massive labeled data, and enhancing the model's ability to learn implicit physical laws.

[0063] In some embodiments, the model parameters include calculating the weights of the fully connected layers and the convolutional kernel parameters of the neural network, and the step S103 of optimizing the neural network further includes: When the difference between the corrected value or output value and the measured voltage exceeds the preset value, the weights of the fully connected layers in the computational neural network are periodically adjusted in the direction that the loss function tends toward the target value through the adaptive moment estimation algorithm, while keeping the convolution kernel parameters unchanged, until the calculated value of the loss function tends toward the target value.

[0064] Furthermore, the target value can be set to zero. The closer the value calculated by the loss function is to zero, the more accurate the power flow calculation of the computational neural network will be.

[0065] By adjusting only the weights of the fully connected layers while keeping the convolutional kernel parameters unchanged, the flexibility of the fully connected layers in adapting to local errors is utilized, while preserving the stability of the convolutional kernel in capturing global features of the power grid, thus avoiding model performance fluctuations caused by full parameter adjustments. An adaptive moment estimation algorithm is employed, which dynamically adjusts the update step size based on changes in the loss function gradient, enabling faster convergence to optimal parameters and reducing the number of iterations. Optimization is only initiated periodically when the deviation between the calculated and measured values ​​exceeds a preset value, avoiding unnecessary parameter adjustments, reducing the system's computational burden, and ensuring that the solution output by the computational neural network can be corrected promptly when errors exceed the limit.

[0066] Specifically, the corrected value or output value of the calculated neural network is compared with the measured voltage to verify the result. If the comparison result satisfies... If the condition is not met, then the current result will be marked as unreliable. If the result is positive, it is marked as a reliable result. This is the correction value or output value of the neural network for the i-th node. Let be the measured voltage of the i-th node.

[0067] After marking untrusted results, an online learning mode is triggered. In the online learning mode, an incremental learning framework is used to collect trustworthy and untrustworthy results within a preset time period every 300 seconds and calculate the loss function. The model parameters of the computational neural network are updated via backpropagation using an adaptive moment estimation optimizer. Specifically, only the fully connected layers are updated, while the parameters of the feature extraction layers are preserved to maintain generalization ability. The learning rate of the adaptive moment estimation optimizer is set to 0.001.

[0068] Specifically, only the weights of the fully connected layers, accounting for 10% of the total model parameters, are updated, while the parameters of the feature extraction layers, accounting for 90% of the total model parameters, are retained. This is to maintain generalization ability and avoid catastrophic forgetting, thereby controlling the accuracy loss of the computational neural network in old scenarios to be less than 1%, while enabling the computational neural network to quickly adapt to new states.

[0069] Compared to the offline retraining mode of existing technologies, this application can complete the computational neural network update in just 300 seconds, reducing the adaptation time to new scenarios by 90%, meeting the real-time requirements of the power grid, and supporting plug-and-play expansion.

[0070] Furthermore, incremental learning to update parameters can be replaced by a meta-learning framework, where the computational neural network is pre-trained offline to quickly adapt to new tasks, such as adding distributed power supply access, and only a small number of samples are needed for fine-tuning in the online phase.

[0071] Furthermore, each update to the computational neural network allows for the addition of new scene labels. This application can update the scene feature library, adding newly emerging scenes, such as highly volatile scenes under extreme weather conditions, to the scene labels. Through dynamic expansion of the scene feature library, the computational neural network's ability to remember power grid evolution is continuously enhanced, maintaining over 95% of its initial accuracy even after long-term operation.

[0072] In some embodiments, scenario features include renewable energy volatility, load mutation rate, and number of topological changes. Scenario labels are simple, complex, and critical. The thresholds for determining scenario features include a first threshold for renewable energy volatility, a second threshold for renewable energy volatility, a first threshold for load mutation rate, a second threshold for load mutation rate, and a preset value for topological changes. The neural network construction step S101 further includes: When the volatility of new energy sources is lower than the first threshold of new energy volatility, the load mutation rate is lower than the first threshold of load mutation rate, and the number of topology changes is zero, the power grid system is judged to be operating normally, the output scenario label is simple, and the computational neural network is a first-level neural network.

[0073] When the volatility of new energy sources exceeds the second threshold of new energy volatility, or the load mutation rate exceeds the second threshold of load mutation rate, or the number of topology changes exceeds the preset value of topology changes, it is determined that a fault has occurred in the power grid system, the output scenario label is complex, and the computational neural network is a second-level neural network.

[0074] When the volatility of new energy sources is higher than or equal to the first threshold of new energy volatility and lower than or equal to the second threshold of new energy volatility, it is determined that a fault has occurred in the power grid system. The output scenario label is critical. A computational neural network is constructed that performs parallel computation of the first-level neural network and the second-level neural network, and the computational weights of the first-level neural network and the second-level neural network are set.

[0075] Among them, the computational cost in the first-level neural network is less than that in the second-level neural network, the first threshold of new energy volatility is lower than the second threshold of new energy volatility, and the first threshold of load mutation rate is lower than the second threshold of load mutation rate.

[0076] Based on the volatility of new energy sources, load mutation rate, and the number and threshold of topology changes, this system accurately identifies three scenarios: normal grid operation, critical transition, and fault disturbance, providing an objective basis for the construction of computational neural networks. For simple scenarios, a first-level neural network with low computational cost ensures efficiency; for complex scenarios, a second-level neural network with high computational cost ensures accuracy; and for critical scenarios, parallel computing balances efficiency and accuracy, solving the problem of traditional models where efficiency and accuracy cannot be simultaneously achieved. Parallel computing in critical scenarios integrates the advantages of both network levels through weight configuration, effectively addressing the complex state of the power grid transitioning from steady state to fault, and improving the computational reliability in complex scenarios.

[0077] Specifically, the first threshold for new energy volatility can be set to 5%. The second threshold for new energy volatility can be set to 15%. The first threshold for load mutation rate can be set to 2%. The second threshold for load mutation rate can be set to 5%. The preset value for topology change can be set to 2 times.

[0078] The second-level neural network can be set as a high-precision sub-network, using a 5-layer residual neural network. The input layer has 10-dimensional features, the hidden layer has 64 neurons per layer, it has batch normalization, and the output layer is 2×N-dimensional. It has a large computational load, introduces residual connections to avoid gradient vanishing, and is suitable for complex scenarios.

[0079] The first-level neural network can be set as a lightweight sub-network, using a 3-layer multilayer perceptron. The input layer has 10-dimensional features, the hidden layer has 32 neurons per layer and is activated by modified linear units, and the output layer is 2×N-dimensional. It has low computational cost and is suitable for simple scenarios.

[0080] When the first-level neural network and the second-level neural network are computed in parallel, the computational weights of the first-level neural network and the second-level neural network are set. For example, the computational weight of the first-level neural network is 0.3 and the computational weight of the second-level neural network is 0.7. The final output is a weighted average result.

[0081] By setting up neural networks with different computational loads, the overall computational load is reduced by more than 40% while ensuring that the computational accuracy error is ≤1%.

[0082] The lightweight subnetwork computation process involves inputting scene features and preprocessed real-time data, which are then processed by a 3-layer multilayer perceptron (MLP) to output preliminary voltage and power calculation values.

[0083] MLP is an artificial neural network composed of multiple neuron layers.

[0084] The high-precision sub-network calculation process involves inputting scene features and preprocessed real-time data, which are then processed by a 5-layer residual neural network (ResNet) to output preliminary voltage and power calculation values.

[0085] The parallel computation process of the first-level neural network and the second-level neural network is as follows: Input scene features and preprocessed real-time data are processed according to... Calculate the initial voltage value, where w is the calculation weight of the first-level neural network. The initial voltage calculation value is obtained from the first-level neural network. This is the initial voltage calculation value obtained from the second-level neural network. The initial power calculation value can be obtained similarly.

[0086] In simple scenarios, calling the lightweight subnetwork reduces computational load by more than 90% compared to existing technologies. In practical applications, traditional feedforward neural networks require 100ms per operation, while the lightweight subnetwork of this invention only requires 10ms per operation, meeting real-time scheduling requirements.

[0087] In complex scenarios, by using a high-precision sub-network, the voltage error is strictly controlled within 2%, which is 8% compared to the error of the prior art in complex scenarios. The calculation accuracy of this application is significantly improved.

[0088] By dividing the scenarios into simple, critical, and complex levels, the applicable scenario range is improved by more than 50% compared to existing technologies that only cover steady-state or single dynamic scenarios. It can cope with more than 90% of actual operation scenarios, such as sudden changes in new energy sources and load jumps.

[0089] In practical applications, based on voltage sensitivity and protection settings, key nodes are identified, such as balancing nodes and important load nodes. These key nodes are then calculated individually using high-precision sub-networks, with voltage errors strictly controlled within 1% to meet the protection device's operating threshold. The device's operating threshold typically requires an error ≤1%.

[0090] The matching of scene labels with neural networks can be replaced with a lightweight NAS algorithm to dynamically search for the optimal subnetwork structure, determine the number of layers and neurons, and adapt to more complex scene changes. In some embodiments, the real-time data further includes: the number of line switching operations in the power grid system, the number of distributed power source connections, the number of photovoltaic nodes, and the active power, reactive power, and renewable energy output of each node. The neural network construction step further includes: The formula for calculating the volatility of new energy sources is: .

[0091] Where N is the number of photovoltaic nodes, and i is the current node. To provide power to the new energy source at the i-th node at time t, This represents the average power.

[0092] The formula for calculating the mutation rate of the load is: .

[0093] in, Let be the node load of the i-th node at time t. Let be the node load of the i-th node at time t. The node load is obtained by summing the active power and reactive power.

[0094] The number of topology changes includes the number of line switching operations and the number of times distributed power sources are connected.

[0095] Specific calculation formulas for renewable energy volatility, load mutation rate, and topology change frequency are provided, solving the problem of extracting scenario features from real-time data. This ensures that scenario labeling is operable and consistent, avoiding subjective judgment errors. Renewable energy volatility combines the number of photovoltaic nodes and the output of each node; load mutation rate is based on the sum of active and reactive power of nodes; and the number of topology changes covers line switching and distributed power generation access, ensuring that scenario labels accurately reflect the dynamic changes of the power grid. Quantified scenario features allow relevant thresholds to be calibrated based on historical operating data, improving the scientific rigor of scenario classification.

[0096] Specifically, when calculating scene features, real-time data and a historical scene database are used, and scene features are statistically analyzed through a sliding window. The window length is set to 10 seconds.

[0097] In practical design of computational neural networks, a computational neural network includes an input layer, multiple hidden layers, and an output layer. The input layer receives real-time operating data of the power system, including information such as active power injection, reactive power injection, node voltage magnitude, and initial phase angles at each node. It also receives encoded information reflecting the power system topology.

[0098] In the hidden layers, a combination of an improved Convolutional Neural Network (CNN) structure and a Long Short-Term Memory Network (LSTM) is employed. The CNN structure is used to extract local features and spatial correlations from the power system data, such as scanning the power system node data through convolutional kernels to capture features like electrical connections between nodes. The LSTM network is used to process the time-series features of the data, considering the changing trends of the power system's operating state over time, such as tracking dynamic changes in load.

[0099] Incorporating physical constraints into computational neural networks involves integrating the physical laws and constraints of power systems, such as Kirchhoff's current law and voltage law, into the neural network as soft constraints. By adding a penalty term related to physical constraints to the loss function, the computational neural network, during training, not only minimizes the error between the predicted results and the actual power flow values ​​but also strives to satisfy the physical constraints of the power system, thereby enhancing the model's understanding and learning ability of the physical nature of the power system.

[0100] When training the computational neural network, a large amount of actual operating data of the power system under different operating conditions is first collected, including data on normal operation, fault conditions, and various transitional states. This data is then preprocessed, including data cleaning and normalization, to ensure data quality and consistency.

[0101] The preprocessed data was divided into training, validation, and test sets. The stochastic gradient descent algorithm and its adaptive moment estimation algorithm were used to train the computational neural network. During training, the learning rate was dynamically adjusted; as the number of training epochs increased, the learning rate was gradually decreased to improve the convergence speed and stability of the computational neural network.

[0102] The computational neural network (CNN) is evaluated during training using a validation set to monitor its computational accuracy and loss function value. When the CNN's performance on the validation set no longer improves or overfitting occurs, the network's structure or model parameters are adjusted promptly, such as increasing or decreasing the number of neurons in hidden layers or adjusting regularization parameters, to prevent overfitting and improve the model's generalization ability.

[0103] When using computational neural networks, it is necessary to determine the operating scenario of the power system and monitor its operating parameters in real time, including the system's load change rate, node voltage fluctuation range, and changes in line transmission power. Based on these parameters, fuzzy logic algorithms or decision tree algorithms are used to classify and evaluate the current power system operating scenario, determining whether it belongs to different types such as normal operation, minor fault, or severe fault.

[0104] To effectively balance computational accuracy and efficiency, priority weights for computational efficiency and accuracy are set according to different operating scenarios. For example, under normal operating conditions, the priority weight for computational efficiency is set to 0.7, and the priority weight for computational accuracy is set to 0.3. Under minor fault conditions, the priority weights for both are set to 0.5. Under severe fault conditions, the priority weight for computational accuracy is set to 0.7, and the priority weight for computational efficiency is set to 0.3. These weights can be flexibly adjusted based on the actual needs and experience of power system operators.

[0105] Based on the identified operating scenario, the model parameters of the computational neural network can be adjusted accordingly. When computational efficiency is prioritized, the computational load of the computational neural network can be reduced. For example, this can be achieved by decreasing the number of convolutional kernels in a CNN layer, reducing the hidden layer dimension of an LSTM network, or using model pruning techniques to remove some unimportant connections and neurons, thereby accelerating the computational speed of the computational neural network.

[0106] In scenarios where computational accuracy is paramount, increasing the complexity and computational cost of the computational neural network (CNN) can improve its accuracy. For example, increasing the number of neurons in the hidden layers, deepening the network layers, and adjusting the parameters of the activation function can enable the CNN to learn more complex power system operation modes and characteristics. Simultaneously, adjusting the coefficients of the penalty terms in the physical constraint layer strengthens the satisfaction of physical constraints, further improving the accuracy of the CNN's computational results.

[0107] Based on the determined operating scenario, priorities, and dynamic adjustments to model parameters, computing resources are allocated rationally. If computational efficiency is prioritized in the current scenario and the computational load of the neural network has been reduced, the number of CPU cores or GPU memory allocated to power flow computing tasks will be reduced, and the freed-up resources will be used for other tasks with high real-time requirements, such as real-time monitoring and early warning tasks for power systems.

[0108] When computational accuracy is prioritized and the computational load of the computational neural network increases, the computational resources allocated to the power flow calculation task should be increased to ensure that the computational neural network has sufficient resources to perform complex operations and guarantee high accuracy of the calculation results. For example, more CPU cores can be allocated to the power flow calculation task, or more GPU memory can be allocated to the computational neural network to accelerate the training and inference process of the model.

[0109] like Figure 3 As shown, in practical applications, the first step is to acquire real-time data and grid topology parameters from the grid real-time measurement system SCADA or PMU. The data is then normalized to eliminate the influence of dimensions. The real-time data and grid topology parameters are input into the calculation formulas for renewable energy volatility and load mutation rate. Based on a sliding window, the values ​​of renewable energy volatility, load mutation rate, and the number of topology changes are calculated.

[0110] Based on the values ​​of renewable energy volatility, load mutation rate, and the number of topology changes, a scenario label is determined to identify whether the current power grid is in a simple, complex, or critical scenario. After determination, a computational neural network is constructed according to the neural network used for each scenario. The computational neural network outputs preliminary voltage and power calculation values.

[0111] A physical constraint model is set up, and the initial voltage and power calculation values ​​are corrected by physical constraints to obtain the corrected values ​​or output values. The results are verified to confirm that the corrected values ​​or output values ​​are reliable results, and a power flow solution that conforms to the physical constraints is obtained.

[0112] Meanwhile, throughout the entire operation of the computational neural network, unreliable results are marked through result verification, triggering online learning. Based on physical constraints and loss functions, the model parameters are optimized in reverse, and the model and scene features are updated to continuously optimize the computational neural network with more accurate calculation results.

[0113] like Figure 2 As shown in the embodiments of this application, an optimization system for power flow calculation is also provided, including: A neural network module 201 is configured to set threshold values ​​and model parameters for scene features, acquire real-time data from the power grid system, and determine scene labels based on the real-time data and the threshold values ​​for scene features. A computational neural network is then constructed based on the scene labels and model parameters. The real-time data includes measured voltage and measured power.

[0114] The initial value calculation module 202 is configured to input real-time data into the computational neural network to obtain preliminary voltage calculation values ​​and preliminary power calculation values.

[0115] The optimized neural network module 203 is configured to construct a loss function based on power conservation constraints and voltage phase angle balance constraints. It inputs measured voltage and power data, along with preliminary voltage and power calculations from the power grid system, into the loss function for iterative calculation, outputting the calculated value of the loss function. Model parameters are adjusted until the loss function converges to the target value. Based on the adjusted model parameters, the computational neural network is optimized. Finally, power flow calculation is performed using the optimized computational neural network.

[0116] The three core steps of constructing the neural network, calculating initial values, and optimizing the neural network are broken down into independent modules. Each module has a clearly defined function and interface, facilitating development, debugging, and maintenance. For example, modules can be upgraded or optimized independently without affecting the initial value calculation module. Each module has clearly defined input, processing logic, and output, making it valuable for engineering applications. The modular architecture allows for flexible expansion of functionality according to the needs of power grid development, such as adding energy storage output characteristics or introducing new constraints, thereby enhancing the system's lifecycle value.

[0117] In some embodiments, the optimized neural network module 203 is further configured to, A physical constraint model is constructed based on power conservation constraints and voltage phase angle balance constraints, and a loss function is constructed based on the physical constraint model.

[0118] The physical constraint model includes, .

[0119] in, As the first weighting coefficient, This is the second weighting coefficient. For the power conservation constraint, This is a voltage phase angle balance constraint.

[0120] This system explicitly optimizes the configuration logic of the neural network module for constructing the physical constraint model, ensuring that the calculated values ​​output by the system are always constrained by the physical laws of the power grid. Through module configuration, it supports the setting and adjustment of the first and second weight coefficients, enabling the system to dynamically optimize the influence weights of physical constraints based on external needs such as power grid dispatch instructions and measurement system upgrades, thus enhancing flexibility. The integration of the physical constraint model within the module provides the system optimization process with underlying physical verification logic, preventing optimization from deviating from physical laws due to independent module operation and ensuring the reliability of the system output.

[0121] In some embodiments, the optimized neural network module 203 is further configured to, The constructed loss function is: .

[0122] in, This is the third weighting coefficient. It is the fourth weighting coefficient. To measure voltage, To measure power, These are preliminary voltage calculation values. These are preliminary power calculation values.

[0123] The system explicitly optimizes the construction and configuration of the neural network module for multi-dimensional loss functions, ensuring that the system can simultaneously monitor data fitting errors and physical constraint errors. Through module configuration, it supports the adjustment of third and fourth weight coefficients, enabling operation and maintenance personnel to flexibly configure error weights according to the power grid operating status, thereby improving the system's responsiveness to actual operational needs. The integration of multi-dimensional loss functions allows the system to optimize calculation results from both data matching and physical compliance perspectives. Compared to systems with a single error target, its calculation accuracy is more capable of meeting the needs of core scenarios such as power grid dispatching and fault analysis.

[0124] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0125] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. An optimization method for power flow calculation, characterized in that, include: The steps for constructing a neural network include setting a threshold for determining scene features and model parameters, acquiring real-time data from the power grid system, determining scene labels based on the real-time data and the threshold for determining scene features, and constructing a computational neural network based on the scene labels and the model parameters; wherein, the real-time data includes measured voltage and measured power. In the initial value calculation step, the real-time data is input into the computational neural network to obtain preliminary voltage and power calculation values. The neural network optimization process involves constructing a loss function based on power conservation constraints and voltage phase angle balance constraints. The measured voltage and power, along with the preliminary voltage and power calculations obtained from the power grid system, are input into the loss function for iterative calculation, and the calculated value of the loss function is output. The model parameters are adjusted until the loss function approaches the target value. Based on the adjusted model parameters, the computational neural network is optimized. Finally, power flow calculation is performed using the optimized computational neural network.

2. The optimization method for power flow calculation according to claim 1, characterized in that, The step of optimizing the neural network further includes: A physical constraint model is constructed based on power conservation constraints and voltage phase angle balance constraints, and the loss function is constructed based on the physical constraint model. The physical constraint model includes, ； in, As the first weighting coefficient, This is the second weighting coefficient. For the power conservation constraint, This refers to the voltage phase angle balance constraint.

3. The optimization method for power flow calculation according to claim 2, characterized in that, The step of optimizing the neural network further includes: The constructed loss function is as follows: ; in, This is the third weighting coefficient. It is the fourth weighting coefficient. For the measured voltage, For the measured power, The initial voltage calculation value is... This is the initial power calculation value.

4. The optimization method for power flow calculation according to claim 2, characterized in that, The step of optimizing the neural network further includes: Based on the power conservation constraint and the voltage phase angle balance constraint, physical constraint conditions are set; The preliminary voltage calculation value is input into the physical constraint model. When the calculated value of the physical constraint model satisfies the physical constraint conditions, the preliminary voltage calculation value is directly used as the output value. When the calculated value of the physical constraint model does not meet the physical constraint conditions, the initial voltage calculation value is adjusted until the physical constraint conditions are met, and the adjusted value is output as the correction value. Adjust the model parameters based on the correction value or the output value.

5. The optimization method for power flow calculation according to claim 4, characterized in that, The model parameters include calculating the weights of the fully connected layers and the convolutional kernel parameters of the neural network. The step of optimizing the neural network further includes: When the difference between the correction value or the output value and the measured voltage exceeds a preset value, the weights of the fully connected layer in the computational neural network are periodically adjusted along the direction in which the loss function tends to the target value using an adaptive moment estimation algorithm, while keeping the convolution kernel parameters unchanged, until the calculated value of the loss function tends to the target value.

6. The optimization method for power flow calculation according to claim 1, characterized in that, The scenario features include renewable energy volatility, load mutation rate, and number of topological changes. The scenario labels are simple, complex, and critical. The threshold values ​​for determining the scenario features include a first threshold for renewable energy volatility, a second threshold for renewable energy volatility, a first threshold for load mutation rate, a second threshold for load mutation rate, and a preset value for topological changes. The step of constructing the neural network further includes: When the volatility of the new energy source is lower than the first threshold of the volatility of the new energy source, the load mutation rate is lower than the first threshold of the load mutation rate, and the number of topology changes is zero, the power grid system is judged to be operating normally, the scenario label is output as "simple", and the computational neural network is a first-level neural network. When the volatility of the new energy source is higher than the second threshold of the new energy source volatility, or the load mutation rate is higher than the second threshold of the load mutation rate, or the number of topology changes is higher than the preset value of the topology change, it is determined that a fault has occurred in the power grid system, and the scenario label is output as "complex", and the computational neural network is a second-level neural network; When the volatility of the new energy source is higher than or equal to the first threshold of the new energy source volatility and lower than or equal to the second threshold of the new energy source volatility, it is determined that a fault has occurred in the power grid system. The scenario label is output as the critical point. A computational neural network for parallel computation of the first-level neural network and the second-level neural network is constructed, and the computational weights of the first-level neural network and the second-level neural network are set. The computational load in the first-level neural network is less than that in the second-level neural network, the first threshold for new energy volatility is lower than the second threshold for new energy volatility, and the first threshold for load mutation rate is lower than the second threshold for load mutation rate.

7. The optimization method for power flow calculation according to claim 6, characterized in that, The real-time data also includes: the number of line switching operations in the power grid system, the number of distributed power source connections, the number of photovoltaic nodes, and the active power, reactive power, and renewable energy output of each node. The step of constructing the neural network further includes: The formula for calculating the volatility of the new energy source is as follows: ; Where N is the number of photovoltaic nodes, and i is the current node. To provide power to the new energy source at the i-th node at time t, Average power; The formula for calculating the load mutation rate is as follows: ; in, Let be the node load of the i-th node at time t. Let be the node load of the i-th node at time t; the node load is obtained by summing active power and reactive power. The number of topology changes includes the number of line switching operations and the number of times distributed power sources are connected.

8. An optimization system for power flow calculation, used to implement the optimization method for power flow calculation according to any one of claims 1-7, characterized in that, include: A neural network module is configured to set a threshold for determining scene features and model parameters, acquire real-time data of the power grid system, determine scene labels based on the real-time data and the threshold for determining scene features, and construct a computational neural network based on the scene labels and the model parameters; wherein, the real-time data includes measured voltage and measured power. The initial value calculation module is configured to input the real-time data into the computational neural network to obtain preliminary voltage calculation values ​​and preliminary power calculation values; The optimized neural network module is configured to construct a loss function based on power conservation constraints and voltage phase angle balance constraints. The measured voltage and power, the preliminary voltage calculation value and the preliminary power calculation value obtained from the power grid system are input into the loss function for iterative calculation, and the calculated value of the loss function is output. The model parameters are adjusted until the loss function tends to the target value. Based on the adjusted model parameters, the computational neural network is optimized. Finally, power flow calculation is performed through the optimized computational neural network.

9. The power flow calculation optimization system according to claim 8, characterized in that, The optimized neural network module is further configured to, A physical constraint model is constructed based on power conservation constraints and voltage phase angle balance constraints, and the loss function is constructed based on the physical constraint model. The physical constraint model includes, ； in, As the first weighting coefficient, This is the second weighting coefficient. For the power conservation constraint, This refers to the voltage phase angle balance constraint.

10. The power flow calculation optimization system according to claim 8, characterized in that, The optimized neural network module is further configured to, The constructed loss function is as follows: ; in, This is the third weighting coefficient. It is the fourth weighting coefficient. For the measured voltage, For the measured power, The initial voltage calculation value is... This is the initial power calculation value.

Images