A Deep Koopman Learning-Driven Method and System for Safe Power Control of LCC-HVDC

By employing a deep Koopman learning-driven approach, we construct an upgraded state and an affine Koopman linear dynamic model, which solves the problem that the LCC-HVDC model struggles to balance voltage constraints and nonlinear response, thus achieving efficient and safe power control.

CN123121784APending Publication Date: 2026-11-17WUHAN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611608695.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The existing simplified model of LCC-HVDC cannot simultaneously characterize AC voltage constraints, DC power nonlinear response, and control mode switching, making it difficult to balance the prediction accuracy and operational safety of high-power support commands.

Method used

By employing a deep Koopman learning-driven approach, an upgraded state is constructed. A neural encoder is used to extract Koopman observable features, an affine Koopman linear dynamic model is established, and a weighted interval classifier is trained to determine the upgraded safe region, thereby achieving online safe power control.

Benefits of technology

It achieves linear characterization of voltage-power coupling and mode switching of LCC-HVDC, reduces recursive prediction bias and divergence risk, and improves the safety and accuracy of high-power support commands.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121784A_ABST
    Figure CN123121784A_ABST
Patent Text Reader

Abstract

This invention provides a deep Koopman learning-driven method and system for safe power control of LCC-HVDC, relating to the fields of high-voltage direct current transmission, data-driven modeling of nonlinear systems, and power system safety control. The method acquires time-series samples of LCC-HVDC under multiple operating conditions to construct physical states and supplementary power control inputs; it extracts observable features using a neural encoder and concatenates them with the original physical states; it identifies an affine Koopman linear dynamic model containing affine bias terms and trains it jointly with multi-step physical state prediction errors and stability regularization terms; it trains a weighted interval classifier based on safety samples and DC blocking samples to form an upgraded safety region; it performs online neural feature dimensionality upgrade, upgraded state prediction, DC active power projection, and safety screening, outputting candidate supplementary DC power commands that satisfy the upgraded safety region constraints. This invention can characterize voltage-power nonlinear coupling and mode switching online and restrict high-risk support commands.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of high-voltage direct current transmission, nonlinear system data-driven modeling, and power system safety control technology, specifically to a deep Koopman learning-driven LCC-HVDC safe power control method and system. Background Technology

[0002] A line-commutated-converter high-voltage direct current (LCC-HVDC) transmission system, or simply a DC system, can rapidly change active power transmission by adjusting DC current commands or DC power commands. This adjustment capability can be used to provide large-capacity emergency active power support when a power deficit occurs in the receiving-end grid.

[0003] The active power support capability of LCC-HVDC is closely related to the AC voltage and DC voltage of the converter bus, the commutation process, and reactive power exchange. When the AC voltage on the inverter side drops, the DC power available for transmission is limited. Under current margin control, the rectifier side and the inverter side may also switch between different control modes, causing the DC power response to control commands to exhibit obvious nonlinearity and piecewise characteristics.

[0004] Existing coordinated control typically uses first-order transfer functions or fixed-parameter linear models to describe DC power response. These models do not explicitly describe the AC voltage constraint on power support capability and struggle to cover mode switching caused by current margin control. In high-power support scenarios, if the controller issues excessively large supplementary DC power commands based on these models, it may cause a significant deviation between predicted and actual power, pushing the operating point close to the risk boundaries of commutation failure or DC blockage. Furthermore, directly embedding high-dimensional nonlinear models into online control increases computational load, hindering prediction and optimization within finite control cycles.

[0005] Therefore, there is an urgent need to provide a method that can characterize the voltage-power coupling and mode switching of LCC-HVDC, form a linear predictive structure that can be embedded in online control, and limit high-risk power commands based on operating data.

[0006] The information disclosed in this background section is only for understanding the background of the inventive concept, and therefore may include information that does not constitute prior art. Summary of the Invention

[0007] The purpose of this invention is to provide a deep Koopman learning-driven method and system for safe power control of LCC-HVDC, which solves the technical problem that existing simplified LCC-HVDC models cannot simultaneously characterize AC voltage constraints, DC power nonlinear response and control mode switching, and lack equipment safety boundaries that can be directly embedded in online control, resulting in a difficulty in balancing the prediction accuracy and operational safety of high-power support commands.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a deep Koopman learning-driven LCC-HVDC safe power control method, comprising: The time-series samples of the LCC-HVDC high-voltage direct current transmission system with grid phase converter are obtained under various operating conditions and supplementary DC power commands, and the physical state and control input at each sampling time are constructed from them. The physical state is input into the neural encoder to extract Koopman observable features, and the physical state is concatenated with the Koopman observable features to form an upgraded state. Based on the time sequence of the upgraded state and the control input, an affine Koopman linear dynamic model is identified, and a linear projection from the upgraded state to the physical state and the change in DC active power is established. The neural encoder and the affine Koopman linear dynamic model are jointly trained with the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective. Train a weighted interval classifier and determine the upsizing safety region of LCC-HVDC based on the weighted interval classifier; The system acquires the current physical state and candidate supplementary DC power commands online, and sequentially executes neural feature dimensionality upgrade, dimensionality upgrade state prediction, DC active power projection, and safety scoring. The candidate supplementary DC power commands are constrained by the dimensionality upgrade safety region, and the candidate supplementary DC power commands that meet the dimensionality upgrade safety region constraints are output to the LCC-HVDC.

[0009] According to the present invention, a deep Koopman learning-driven LCC-HVDC safe power control method is provided, wherein the physical state includes at least the change in DC active power, DC voltage, AC voltage of the inverter-side converter bus, and reactive power flowing from the inverter-side converter bus to the adjacent AC line; the control input is a supplementary DC power command corresponding to the physical state.

[0010] According to the present invention, a deep Koopman learning-driven LCC-HVDC safe power control method is provided, wherein the neural encoder includes at least two fully connected layers and a nonlinear activation function connected in sequence; the up-dimensional state is obtained by concatenating the physical state and the Koopman observable features output by the neural encoder along the feature dimension.

[0011] According to the present invention, a deep Koopman learning-driven LCC-HVDC safe power control method is provided. The affine Koopman linear dynamic model includes a continuous-time form and a discrete-time form obtained by zero-order preserved discretization of the continuous-time form. The continuous-time form includes a continuous-time Koopman state matrix, a continuous-time control input matrix, and a continuous-time affine bias term. The discrete-time form includes the corresponding discrete-time Koopman state matrix, discrete-time control input matrix, and discrete-time affine bias term.

[0012] According to the deep Koopman learning-driven LCC-HVDC safe power control method provided by the present invention, the multi-step physical state prediction error is obtained by accumulating the squares of the L2 norms between each predicted physical state and the corresponding measured physical state in the prediction time domain; the stability regularization term is obtained by accumulating the positive values ​​of the real parts of each eigenvalue of the continuous-time Koopman state matrix.

[0013] According to the present invention, a deep Koopman learning-driven LCC-HVDC safe power control method is provided, wherein the weighted interval classifier is a linear classifier operating on the up-dimensional state, and its safety score is obtained by adding the bias to the inner product of the weight vector and the up-dimensional state. Training a weighted interval classifier includes: acquiring safe samples and DC-DC blocking samples offline, mapping safe samples and DC-DC blocking samples to a Koopman updimensional space; and assigning a higher penalty weight to DC-DC blocking samples that are misclassified as safe samples during training.

[0014] According to the deep Koopman learning-driven LCC-HVDC safe power control method provided by the present invention, the upgraded safe region is composed of upgraded states with a safety score not less than zero; during online control, a non-negative safety margin is set so that the safety score in the prediction time domain is not less than the non-negative safety margin, and the change in DC active power is also limited to a preset upper and lower limit; when the predicted upgraded state corresponding to the candidate supplementary DC power command does not meet the upgraded safe region constraint, the candidate supplementary DC power command is re-determined within the upgraded safe region.

[0015] The present invention provides a deep Koopman learning-driven LCC-HVDC safe power control method, which covers various operating conditions including AC voltage variation of the converter bus and the switching process between rectifier side constant current and inverter side constant turn-off angle modes, rectifier side current limiting and inverter side current error control modes, and rectifier side current limiting and inverter side constant current modes under current margin control.

[0016] Secondly, the present invention provides a deep Koopman learning-driven LCC-HVDC safe power control system, including a sample acquisition module, a state construction module, a neural feature dimensionality enhancement module, a linear dynamic identification module, an output projection module, a stability training module, a safety boundary learning module, and an online control module; The sample acquisition module is used to acquire time-series samples of the grid-connected high-voltage direct current transmission system under various operating conditions and supplementary DC power commands. The state construction module is used to construct physical states and control inputs based on time-series samples; The neural feature upscaling module is used to extract Koopman observable features and concatenate the physical state with the Koopman observable features to form an upscaled state; The linear dynamic identification module is used to identify affine Koopman linear dynamic models based on the time sequence of the upgraded state and the control input. The output projection module is used to obtain the physical state and the change in DC active power from the projection of the upgraded state; The stability training module is used to jointly train the neural encoder and the affine Koopman linear dynamic model with the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective. The safety boundary learning module is used to train a weighted interval classifier based on safety samples and DC blocking samples and to determine the upgraded safety region. The online control module is used for online physical state and candidate supplementary DC power commands to perform dimensionality-upgrading state prediction, DC active power projection and safety scoring, and outputs supplementary DC power commands that meet the dimensionality-upgrading safety region constraints.

[0017] According to the present invention, a deep Koopman learning-driven LCC-HVDC safe power control system is provided, in which the upgraded state output by the neural feature upgrade module retains the original physical state, so that the output projection module can reconstruct the physical state through the state selection matrix and obtain the predicted value of the change in DC active power through the active power selection matrix.

[0018] Compared with the prior art, the present invention has at least the following technical effects: 1. Utilize neural networks (i.e. neural encoders) to automatically learn finite-dimensional Koopman observable features, and balance nonlinear representation capabilities and physical quantity recoverability through state-preserving dimensionality enhancement; 2. The complex LCC-HVDC response is converted into a linear up-dimensional model containing control input and affine bias (i.e., an affine Koopman linear dynamic model), which is convenient for embedding in online predictive control; 3. By using the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective, the neural encoder and the affine Koopman linear dynamic model can be trained together, which can reduce the recursive prediction bias and divergence risk. 4. The higher-dimensional safety region, which can be linearly represented by safety samples and DC blocking samples, is learned, so that high-risk power support commands can be constrained before they are issued. Attached Figure Description

[0019] 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.

[0020] In the attached diagram: Figure 1 This is a flowchart of the LCC-HVDC safe power control method driven by deep Koopman learning according to the present invention; Figure 2 This is a schematic diagram of the LCC-HVDC safe power control system driven by deep Koopman learning according to the present invention. Figure 3 The following are safety scores and DC voltage curves of the method proposed in the simulation embodiment of the present invention, where (a) is the safety score and (b) is the DC voltage. Figure 4 The following are safety scores and DC voltage curves for a first-order model and static constraints in a simulation embodiment of the present invention, where (a) is the safety score and (b) is the DC voltage. Detailed Implementation 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.

[0021] The following detailed description of some embodiments of the present invention will be provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0022] Please see Figure 1 This invention provides a deep Koopman learning-driven method for safe power control of LCC-HVDC, comprising the following steps: Step 1: Obtain timing samples of the LCC-HVDC under various operating conditions and supplementary DC power commands, and construct the physical state and control input at each sampling time from them; Specifically, the physical state and control input at each sampling time are constructed from the time-series samples, including: constructing the physical state of the DC device at each sampling time from the time-series samples, and using the supplementary DC power command corresponding to the physical state as the control input.

[0023] Step 2: Input the physical state into the neural encoder to extract Koopman observable features, and then concatenate the physical state with the Koopman observable features to form an upgraded state. Step 3: Based on the time sequence of the upgraded state and the control input, identify the affine Koopman linear dynamic model and establish a linear projection from the upgraded state to the physical state and the change in DC active power. Specifically, establishing a linear projection from the upgraded state to the physical state and the change in DC active power includes: establishing a linear projection from the upgraded state to the physical state, and obtaining the predicted value of the change in DC active power by selecting the DC active power component in the physical state.

[0024] Step 4: Using the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective, jointly train the neural encoder and the affine Koopman linear dynamic model. Step 5: Train the weighted interval classifier and determine the safe region for dimensionality upgrade of LCC-HVDC based on the weighted interval classifier; Step 6: Obtain the current physical state and candidate supplementary DC power commands online, and sequentially execute neural feature dimensionality upgrade, dimensionality upgrade state prediction, DC active power projection, and safety scoring. Constrain the candidate supplementary DC power commands with the dimensionality upgrade safety region, and output the candidate supplementary DC power commands that meet the dimensionality upgrade safety region constraints to LCC-HVDC.

[0025] Furthermore, the physical state includes at least the change in DC active power, DC voltage, AC voltage of the inverter-side converter bus, and reactive power flowing from the inverter-side converter bus to adjacent AC lines; depending on the converter station wiring and measurement configuration, the reactive power may include the reactive power of one or more adjacent AC lines.

[0026] Furthermore, the neural encoder comprises at least two fully connected layers connected sequentially and a nonlinear activation function; the upgraded state is obtained by concatenating the physical state with the Koopman observable features output by the neural encoder along the feature dimension. Therefore, the state-preserving upgraded state incorporates the original physical state as part of the upgraded state, thus allowing the physical quantity to be directly recovered through a constant coefficient selection matrix.

[0027] Furthermore, the affine Koopman linear dynamic model includes a continuous-time form and a discrete-time form obtained by zero-order preserved discretization of the continuous-time form; the continuous-time form includes a continuous-time Koopman state matrix, a continuous-time control input matrix, and a continuous-time affine bias term, while the discrete-time form includes the corresponding discrete-time Koopman state matrix, discrete-time control input matrix, and discrete-time affine bias term.

[0028] Furthermore, the multi-step physical state prediction error is obtained by summing the squared L2 norms between each predicted physical state and its corresponding measured physical state in the prediction time domain; the stability regularization term is obtained by summing the positive values ​​of the real parts of the eigenvalues ​​of the continuous-time Koopman state matrix. It should be noted that the stability regularization term in the training objective penalizes the positive values ​​of the real parts of the eigenvalues ​​of the continuous-time Koopman state matrix to suppress recursive prediction divergence.

[0029] Furthermore, the weighted interval classifier is a linear classifier operating on the upgraded state, and its safety score is obtained by adding a bias to the inner product of the weight vector and the upgraded state. Training the weighted interval classifier includes: offline acquisition of safe samples and DC-DC blocking samples, mapping the safe samples and DC-DC blocking samples to the Koopman upgraded space; during training, DC-DC blocking samples that are misclassified as safe are assigned a higher penalty weight than safe samples. Setting a higher penalty weight for DC-DC blocking samples reduces the probability of misclassifying high-risk operating points as safe operating points.

[0030] Furthermore, the upgraded safety region is composed of upgraded states with a safety score of not less than zero; during online control, a non-negative safety margin can be set on the basis of a zero safety score threshold (the right side of equation (12) below is 0, and the right side can be a positive value, i.e., non-negative safety margin, during actual control), so that the safety score in the prediction time domain is not less than the non-negative safety margin, and the change in DC active power is also limited to the preset upper and lower limits; when the predicted upgraded state corresponding to the candidate supplementary DC power command does not meet the upgraded safety region constraint, the candidate supplementary DC power command is re-determined within the upgraded safety region.

[0031] Furthermore, multiple operating conditions cover the switching process between AC voltage changes on the converter bus and the rectifier-side constant current and inverter-side constant turn-off angle modes under current margin control, rectifier-side current limiting and inverter-side current error control modes, and rectifier-side current limiting and inverter-side constant current modes.

[0032] Based on the same inventive concept, such as Figure 2As shown, another embodiment of the present invention provides a deep Koopman learning-driven LCC-HVDC safe power control system, including a sample acquisition module, a state construction module, a neural feature dimensionality enhancement module, a linear dynamic identification module, an output projection module, a stability training module, a safety boundary learning module, and an online control module. Each module collaboratively executes the above control method. The sample acquisition module is used to acquire time-series samples of the grid-connected high-voltage direct current transmission system under various operating conditions and supplementary DC power commands. The state construction module is used to construct physical states and control inputs based on time-series samples; The neural feature upscaling module is used to extract Koopman observable features and concatenate the physical state with the Koopman observable features to form an upscaled state; The linear dynamic identification module is used to identify affine Koopman linear dynamic models based on the time sequence of the upgraded state and the control input. The output projection module is used to obtain the physical state and the change in DC active power from the projection of the upgraded state; The stability training module is used to jointly train the neural encoder and the affine Koopman linear dynamic model with the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective. The safety boundary learning module is used to train a weighted interval classifier based on safety samples and DC blocking samples and to determine the upgraded safety region. The online control module is used for online physical state and candidate supplementary DC power commands to perform dimensionality-upgrading state prediction, DC active power projection and safety scoring, and outputs supplementary DC power commands that meet the dimensionality-upgrading safety region constraints.

[0033] Specifically, the upgraded state output by the neural feature upgrade module retains the original physical state, so that the output projection module can reconstruct the physical state through the state selection matrix and obtain the predicted value of the change in DC active power through the active power selection matrix.

[0034] The following is a detailed description.

[0035] I. LCC-HVDC Voltage-Power Coupling and Sample Construction For the inverter-side converter of LCC-HVDC, its quasi-steady-state active power injection P i and reactive power consumption Q i They can be represented as: (1) (2) In the formula,N i The number of series converter bridges, U aci This represents the effective value of the AC bus line voltage on the inverter side. X ci For commutation reactance, T i The transformer ratio is the ratio between the grid side and the valve side of the converter transformer. For the shut-off angle, The commutation overlap angle is given by equations (1) and (2). It can be seen from equations (1) and (2) that changes in AC bus voltage will affect both active power transmission and reactive power consumption. Therefore, AC voltage, DC voltage and reactive power exchange are retained in the emergency power support model.

[0036] Therefore, it is evident that the power of an LCC-HVDC is affected by voltage, and this process is nonlinear and varies with control. The focus of this invention is how to characterize the power-voltage coupling characteristics of an LCC-HVDC in a high-dimensional space using data-driven methods, and how to linearize these characteristics for easier control.

[0037] This invention applies different supplementary DC power commands to an electromagnetic transient simulation platform or actual operation records, and collects time-series data under various load levels, new energy output, AC voltage disturbances, DC power levels, and fault conditions. Multiple operating conditions cover the switching of current margin control modes, including rectifier-side constant current and inverter-side constant turn-off angle modes, rectifier-side current limiting and inverter-side current error control modes, and rectifier-side current limiting and inverter-side constant current modes.

[0038] In one alternative embodiment, the physical state vector at sampling time k can be represented as: ,in This represents the change in DC active power. DC voltage This refers to the AC voltage of the inverter-side converter bus. and To provide reactive power flowing from the inverter-side converter bus to two adjacent AC lines; supplementary DC power commands are used as control inputs. u LCC The sampled data are normalized according to the maximum absolute value of each variable. When the converter station wiring is different, the reactive power measurement component can be increased or decreased accordingly.

[0039] The physical state vector at the next sampling time k+1 can then be expressed as: (3) Equation (3) represents the discrete-time controlled nonlinear response that is unknown or difficult to solve explicitly in the original physical space. F .

[0040] II. State-Preserving Deep Koopman Learning Dimensional Upgrading Model like Figure 1 As shown, the neural encoder maps the sampled physical states to high-dimensional observable features, namely Koopman observables. Taking a feedforward network with two hidden layers as an example, the Koopman observables are written as: (4) In the formula, It is a non-linear activation function. These are the network parameters. The original physical state is concatenated with the neural encoder output to form a state-preserving, upgraded state: (5) In the higher-dimensional space, the controlled nonlinear response is represented by a continuous-time ray-like system as follows: (6) Zero-order preserved discretization of equation (6) yields: (7) In the formula, A LCC,c and B LCC,c These are the continuous-time Koopman linear matrix and the control input matrix, respectively. d LCC,c For affine bias term; A LCC,d , B LCC,d and d LCC,d For the corresponding discrete-time parameters.

[0041] Joint state-input observables can also be utilized. h ( x LCC ( k ), u LCC ( k ))= g ( x LCC ( k ))+ Lu LCC ( k Describe the control input, which satisfies Where L represents the input gain matrix, and A and B represent the input gain matrix respectively. A LCC,c or A LCC,d , B LCC,c or BLCC,d Regardless of whether continuous-time identification followed by discretization is used, or discrete-time parameters are directly identified, the linear or affine form of the upgraded dynamics with respect to the upgraded state and control input is maintained.

[0042] Since the original physical state is preserved in the upgraded state, the predicted physical state value and the predicted DC active power value are obtained by selecting matrices respectively: (8) (9) In the formula, C x Choose a matrix for the physical state. C P A matrix is ​​selected for the DC active power component. This structure avoids the need for a separate, highly complex decoder and allows training errors to be calculated directly in the physical quantity space.

[0043] III. Stability Constraint Training With length as N K The prediction time domain is jointly trained on the neural encoder parameters, the Koopman state matrix, the control input matrix, and the affine bias term. The training objective consists of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model. (10) In the formula, From time k The predicted physical state obtained by recursion, where i represents the time sequence, i.e., training N K The predicted step size, instead of training only one step; The Koopman state matrix in continuous time is the first... j 1 eigenvalue, The first term is a stability regularization weight. The second term constrains multi-step physical state prediction, and the third term penalizes eigenvalues ​​with positive real parts to suppress recursive prediction divergence.

[0044] In one non-limiting embodiment, the sampling interval is 0.02 seconds, the neural encoder has two hidden layers, each containing 128 neurons, and the dimensionality of the upscaling state is set to 32. These values ​​can be adjusted according to the control cycle, measurement dimension, and online computing resources, and do not constitute a limitation on the scope of protection of this invention.

[0045] IV. Learning about the safety boundaries of higher-dimensional spaces During offline sample generation, trajectories that do not experience commutation failure or DC lockout and whose equipment quantity exceeds the limit while meeting preset conditions are marked as safe samples, while trajectories that experience DC lockout or meet preset high-risk conditions are marked as unsafe samples. The physical state of each sample is mapped using equation (5) in a higher dimension, and a linear safety classifier is trained: (11) In the formula, w s and b s These are the weight vector and bias of the safety hyperplane, respectively. The upgraded safety region is defined as: (12) To reduce the probability of misclassifying high-risk operating points as safe operating points, the classifier uses a weighted interval loss: (13) In the formula, ε saf For classification interval, λ saf For unsafe samples, the penalty weight is... i =+1 and i = 1 represents the first i Each sample is divided into safe and unsafe samples. By making... A unit weight greater than that of a safe sample can increase the cost of incorrectly accepting a locked sample. After training, equation (12) can be used as a linear state constraint in online optimization.

[0046] V. Online Safety Constraint Power Control Figure 1 This illustrates the main flow of the LCC-HVDC safe power control method of the present invention. During online control, the measurement interface obtains the current physical state. The upper-level frequency controller provides candidate supplementary DC power commands. The neural feature dimensionality upgrade module obtains the current dimensionality upgrade state according to equation (5), the linear dynamic identification module predicts the subsequent dimensionality upgrade state according to equation (7), and the output projection module obtains the predicted value of the change in DC active power according to equation (9).

[0047] The safety boundary learning module calculates the safety score for each predicted dimensionality upgrade state according to equation (11). The controller requires that the safety score in the prediction time domain be no less than the non-negative safety margin, while ensuring that the change in DC active power is between the upper and lower limits allowed by the equipment. Candidate supplementary DC power commands can be output when they meet the above constraints; if the candidate supplementary DC power commands do not meet the above constraints, the candidate supplementary DC power commands are not directly issued to the DC system, but the upper-level controller redetermines the candidate supplementary DC power commands within the feasible domain formed by the constraints.

[0048] Thus, the complex LCC-HVDC nonlinear response is transformed into a recursive linear up-dimensional dynamic through a deep Koopman learning model, and the safety boundary is also expressed as a linear inequality of the up-dimensional state. Both can be directly used as predictive models and constraints for model predictive control, rolling optimization control or other online optimization control.

[0049] In the LCC-HVDC safe power control system driven by deep Koopman learning in this embodiment of the invention, the offline training part completes time-series sample acquisition, state construction, neural dimension expansion, Koopman matrix identification, and safety boundary learning; the online control part reuses the neural encoder, Koopman matrix, projection matrix, and safety classifier completed by training.

[0050] The sample acquisition module can connect to an electromagnetic transient real-time digital simulator, an offline simulation program, or a converter station historical database; the online control module can be deployed in an auxiliary control device of the DC control and protection system, a station control computing device, or a computing platform that communicates with the frequency controller. The physical deployment of the modules does not change the data relationships and functional relationships between the modules.

[0051] It should be noted that the number of network layers, neurons, sampling interval, dimensionality increase, classification interval, and penalty weights in the embodiments of the present invention can all be adjusted according to the specific DC system. Any method that employs neural networks to learn Koopman observable features, preserves the original physical state in the dimensional increase state, utilizes ray-like dynamics for prediction, and learns the dimensional increase safety region based on safety samples and DC lockout samples to constrain and supplement DC power commands falls within the technical concept of this invention.

[0052] The following are simulation embodiments of the present invention.

[0053] An improved IEEE 39-node receiving-end system was established on the RTDS–MATLAB closed-loop simulation platform. The LCC-HVDC system adopted the CIGRE baseline model, and the inverter-side AC bus was configured with 337.5 Mvar reactive power support. 1500 scenarios were generated using 13-Vialatin hypercube sampling, covering ±20% variations in renewable energy output and load, load steps from 0 to 300 MW, and generator N... 1 and N 2. Events such as shutdown and pumping of pumped storage systems.

[0054] The training results of the affine Koopman linear dynamic model are as follows: The system collects data on changes in DC active power, DC voltage, AC voltage at the converter bus, and reactive power on two adjacent AC lines. The sampling interval is 0.02 s, and each variable is normalized to its maximum absolute value. The neural encoder uses two hidden layers, each containing 128 neurons, with an upgraded state dimension of 32. The prediction errors on the test set are shown in Table 1.

[0055] Table 1. Training error of the affine Koopman linear dynamic model

[0056] The training results for the safety control boundary are as follows: Trajectories where LCC-HVDC locking occurred were marked as unsafe trajectories, and all other trajectories were marked as safe trajectories. A weighted interval classifier was then trained. In the test samples used, the correct classification rate for safe samples was 99.34%, and the correct classification rate for unsafe samples was 94.02%. The proportion of safe samples misclassified as unsafe was 0.66%, and the proportion of unsafe samples misclassified as safe was 5.98%.

[0057] The closed-loop simulation verification results are as follows: In a continuous disturbance scenario where a 330MW load increase occurs first, followed by a 560MW renewable energy output loss after a 13s interval, the model and learned safety constraints proposed in this invention are compared with a first-order model and static constraints. Figure 3 As shown, the safety score of the method proposed in this invention ( Figure 3 (a) remains positive during the simulation period shown, DC voltage ( Figure 3 (b) did not cross the corresponding learning boundary; such as Figure 4 As shown, when using a first-order model and static constraints, a negative security score appears near the second perturbation. Figure 4 (a) and DC voltage ( Figure 4 (b) Out of bounds. This result shows that, under the conditions of this embodiment, the model and security constraints proposed in this invention can improve dynamic prediction and reduce the operational risks caused by high-risk power commands.

[0058] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. It should be understood that the invention is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A deep Koopman learning-driven method for safe power control of LCC-HVDC, characterized in that, include: The time-series samples of the LCC-HVDC high-voltage direct current transmission system with grid phase converter are obtained under various operating conditions and supplementary DC power commands, and the physical state and control input at each sampling time are constructed from them. The physical state is input into the neural encoder to extract Koopman observable features, and the physical state is concatenated with the Koopman observable features to form an upgraded state. Based on the time sequence of the upgraded state and the control input, an affine Koopman linear dynamic model is identified, and a linear projection from the upgraded state to the physical state and the change in DC active power is established. The neural encoder and the affine Koopman linear dynamic model are jointly trained with the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective. Train a weighted interval classifier and determine the upsizing safety region of LCC-HVDC based on the weighted interval classifier; The system acquires the current physical state and candidate supplementary DC power commands online, and sequentially executes neural feature dimensionality upgrade, dimensionality upgrade state prediction, DC active power projection, and safety scoring. The candidate supplementary DC power commands are constrained by the dimensionality upgrade safety region, and the candidate supplementary DC power commands that meet the dimensionality upgrade safety region constraints are output to the LCC-HVDC.

2. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 1, characterized in that, The physical state includes at least the change in DC active power, DC voltage, AC voltage of the inverter-side converter bus, and reactive power flowing from the inverter-side converter bus to the adjacent AC line; the control input is a supplementary DC power command corresponding to the physical state.

3. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 1, characterized in that, The neural encoder includes at least two fully connected layers and a nonlinear activation function connected in sequence; the upgraded state is obtained by concatenating the physical state with the Koopman observable features output by the neural encoder along the feature dimension.

4. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 1, characterized in that, The affine Koopman linear dynamic model includes a continuous-time form and a discrete-time form obtained by zero-order preserved discretization of the continuous-time form; the continuous-time form includes a continuous-time Koopman state matrix, a continuous-time control input matrix, and a continuous-time affine bias term, and the discrete-time form includes a corresponding discrete-time Koopman state matrix, a discrete-time control input matrix, and a discrete-time affine bias term.

5. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 4, characterized in that, The multi-step physical state prediction error is obtained by summing the squares of the L2 norms between each predicted physical state and the corresponding measured physical state in the prediction time domain; the stability regularization term is obtained by summing the positive values ​​of the real parts of each eigenvalue of the continuous-time Koopman state matrix.

6. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 2, characterized in that, The weighted interval classifier is a linear classifier that operates on the up-dimensional state. Its safety score is obtained by adding a bias to the inner product of the weight vector and the up-dimensional state. Training a weighted interval classifier includes: offline acquisition of safe samples and DC-DC blocking samples; mapping the safe samples and DC-DC blocking samples to a Koopman updimensional space to train the weighted interval classifier; during training, DC-DC blocking samples that are misclassified as safe are assigned a higher penalty weight than safe samples.

7. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 6, characterized in that, The upgraded safety region consists of upgraded states with a safety score of not less than zero. During online control, a non-negative safety margin is set so that the safety score in the prediction time domain is not less than the non-negative safety margin, and the change in DC active power is also limited to a preset upper and lower limit. When the predicted upgraded state corresponding to the candidate supplementary DC power command does not meet the constraints of the upgraded safety region, the candidate supplementary DC power command is re-determined within the upgraded safety region.

8. The deep Koopman learning-driven LCC-HVDC safe power control method according to claim 1, characterized in that, The various operating conditions cover the switching process between the rectifier side constant current and inverter side constant turn-off angle modes under the current margin control of the converter bus AC voltage variation, rectifier side current limiting and inverter side current error control modes, and rectifier side current limiting and inverter side constant current modes.

9. A deep Koopman learning-driven LCC-HVDC safe power control system, characterized in that, It includes a sample acquisition module, a state construction module, a neural feature dimensionality enhancement module, a linear dynamic identification module, an output projection module, a stability training module, a safety boundary learning module, and an online control module; The sample acquisition module is used to acquire time-series samples of the grid-commutated converter high-voltage direct current transmission system LCC-HVDC under various operating conditions and supplementary DC power commands. The state construction module is used to construct the physical state and control input at each sampling time based on the time series samples; The neural feature upscaling module is used to input the physical state into the neural encoder to extract the Koopman observable features, and to concatenate the physical state with the Koopman observable features to form an upscaled state. The linear dynamic identification module is used to identify the affine Koopman linear dynamic model based on the time sequence of the upgraded state and the control input. The output projection module is used to establish a linear projection from the upgraded state to the physical state and the change in DC active power. The stability training module is used to jointly train the neural encoder and the affine Koopman linear dynamic model with the sum of the multi-step physical state prediction error and the stability regularization term of the affine Koopman linear dynamic model as the training objective. The security boundary learning module is used to train a weighted interval classifier and determine the upgraded security region of LCC-HVDC based on the weighted interval classifier. The online control module is used to acquire the current physical state and candidate supplementary DC power commands online, and sequentially execute neural feature dimensionality upgrade, dimensionality upgrade state prediction, DC active power projection and safety scoring, constrain the candidate supplementary DC power commands with the dimensionality upgrade safety region, and output the candidate supplementary DC power commands that meet the dimensionality upgrade safety region constraints to the LCC-HVDC.

10. The deep Koopman learning-driven LCC-HVDC safe power control system according to claim 9, characterized in that, The upgraded state output by the neural feature upgrade module retains the original physical state, so that the output projection module can reconstruct the physical state through the state selection matrix and obtain the predicted value of the change in DC active power through the active power selection matrix.