A method, system, equipment, and medium for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions.
By integrating physical information neural networks and topology analysis technology, the constraint coupling problem in reactive power regulation of pumped storage units was solved, enabling refined evaluation and rapid response of the reactive power regulation capability of pumped storage units, and supporting dynamic reactive power support of the power grid.
Patent Information
- Application Number
- CN202611286186.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies fail to effectively consider the coupling of mechanical, electrical, and grid constraints in reactive power regulation of pumped storage units, resulting in large computational loads, response delays, and control malfunctions. They also lack global topological understanding of constraint boundaries and bottleneck location capabilities, making it difficult to meet the complex reactive power regulation requirements of the power grid.
A multi-condition reactive power regulation constraint boundary model is constructed by combining Physical Information Neural Network (PINN) with differentiable Morse-Smale complex analysis, Grey Wolf optimization, Convolutional Neural Network (CNN), Multi-head Cross-Attention (MHCA), and Bidirectional Long Short-Term Memory Network (BiLSTM). By fusing topological features, local constraint features, and conflict intensity coefficients, the model predicts steady-state dominant constraint switching and introduces an inertial travel time compensation mechanism to achieve a systematic characterization of the reactive power regulation capability of pumped-storage units.
It enables refined evaluation of the reactive power regulation capability of pumped storage units, automatically identifies steady-state dominant constraints, reduces response delay, improves computational efficiency and accuracy, provides reactive power regulation boundary guidance under multiple operating conditions, and supports dynamic reactive power support of the power grid.
Smart Images

Figure CN123121871A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of reactive power regulation technology for pumped storage units, specifically to a method, system, equipment, and medium for determining the reactive power regulation constraint boundaries of pumped storage units under multiple operating conditions. Background Technology
[0002] Pumped storage units, as important peak-shaving and frequency-regulating power sources in the power grid, also possess reactive power regulation capabilities, providing dynamic reactive power support to the grid under various operating conditions such as power generation, pumping, and phase regulation. With the rapid development of "dual-high" (high proportion of renewable energy and high proportion of power electronic equipment) power systems, the grid's demand for reactive power regulation is becoming increasingly complex, placing higher demands on the refined assessment of the reactive power regulation capabilities of pumped storage units.
[0003] The reactive power regulation capability of pumped-storage units under different operating conditions is limited by multiple constraints. Existing research mainly focuses on the reactive power regulation capability analysis under a single operating condition, lacking systematic modeling of the three types of constraints: mechanical constraints, electrical constraints, and grid constraints. Specifically, mechanical constraints involve the limitations of hydraulic-mechanical parameters such as rotational speed, guide vane opening, and head on reactive power regulation; electrical constraints involve the limitations of electromagnetic parameters such as stator current, excitation current, and converter capacity; and grid constraints involve the limitations of grid operation requirements such as voltage limits, voltage sensitivity, and transient stability. These three types of constraints are coupled and jointly determine the reactive power regulation boundary of the pumped-storage unit under each operating condition.
[0004] Existing technologies for solving reactive power feasible regions with multiple types of constraints generally employ ergonomic methods or traditional optimization solvers. These methods suffer from three prominent problems: First, they require repeated solving of optimization problems with nonlinear constraints at each operating point, resulting in high computational costs and difficulty meeting the timeliness requirements of online evaluation. Second, existing methods only output the final reactive power boundary value and cannot automatically identify which of the three types of constraints plays a decisive role in the current boundary. Third, existing methods treat all three types of constraints equally, failing to quantify the relative contribution of each constraint to the boundary. This leads to a lack of clear bottleneck location criteria for the dispatch center when facing insufficient reactive power capacity, and it becomes unclear from which physical link to start in expanding reactive power capacity. Furthermore, existing methods lack the ability to recognize the topological structure of the constraint boundary, making it difficult to predict the switching point of the steady-state dominant constraint and unable to perform control compensation based on the inertial characteristics of mechanical / electrical systems. This results in response delays and the risk of malfunctions in the closed loop from constraint identification to control execution.
[0005] In recent years, Physics-Informed Neural Networks (PINNs), as a new paradigm that embeds physical laws into neural network training, have demonstrated powerful capabilities in solving complex problems with physical constraints. Its core idea is to incorporate physical equations as part of the loss function, enabling the network to satisfy physical laws while fitting data. However, while there are reports of deep integration of differentiable Morse-Smale complex analysis, GWO-CNN-MHCA-BiLSTM with PINN and the introduction of boundary shrinkage rate ratio prediction and inertial travel time compensation mechanisms, no method has yet proposed using differential topological analysis to extract global topological features of constrained manifolds to guide neural network training. Therefore, there is an urgent need for a reactive power regulation constraint boundary determination method that can comprehensively consider three types of constraints, automatically predict steady-state dominant constraint switching, and provide control compensation strategies. Summary of the Invention
[0006] To address the problems in related technologies, this invention proposes a method, system, equipment, and medium for determining the reactive power regulation constraint boundary of pumped storage units under multiple operating conditions, in order to overcome the aforementioned technical problems existing in the existing related technologies.
[0007] Therefore, the specific technical solution adopted by the present invention is as follows:
[0008] According to a first aspect of the present invention, a method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions is provided, comprising the following steps:
[0009] Obtain the mechanical and electrical parameters of the pumped-storage unit and the operating parameters of the power grid where it is located, and construct a constraint condition model that includes mechanical constraints, electrical constraints and power grid constraints;
[0010] Based on the constraint model, mechanical constraints, electrical constraints, and power grid constraints are established for each operating condition.
[0011] Based on the physical information neural network, topological prior features, local constraint features, conflict intensity coefficients and reactive power boundary sequence features are integrated. After training, the integrated physical information neural network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints. The model is then solved to obtain the steady-state reactive power output boundary of the pumped storage unit under various operating conditions.
[0012] The steady-state dominant constraint at the current operating point is determined based on the conflict intensity coefficient. The contraction rate ratio of various constraint boundaries to active power changes is calculated, and the critical active power value for the switching of steady-state dominant constraint is predicted based on the contraction rate ratio. The inertial time constants of the mechanical and electrical systems are calculated, and the inertial travel time and control dead zone compensation amount for switching control modes to cope with the switching of steady-state dominant constraint are determined.
[0013] By introducing response speed constraints and duration constraints, a transient reactive power response capability boundary model is established. This model is then used to deduce the transient reactive power response capability boundary of the pumped-storage unit under different operating conditions. The reactive power output boundary diagram, transient reactive power response capability boundary, critical active power value, inertial travel time, and control dead zone compensation amount for each operating condition are output to guide the reactive power scheduling and control of the pumped-storage unit.
[0014] According to a second aspect of the present invention, a system for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions is provided, comprising:
[0015] The constraint model construction module is used to obtain the mechanical parameters, electrical parameters and operating parameters of the power grid of the pumped storage unit, and construct a constraint model that includes mechanical constraints, electrical constraints and power grid constraints.
[0016] The operating condition constraint establishment module is used to establish mechanical constraints, electrical constraints, and power grid constraints for each operating condition based on the constraint condition model.
[0017] The steady-state reactive power output boundary solution module is used to integrate topological prior features, local constraint features, conflict intensity coefficients and reactive power boundary sequence features based on physical information neural networks. After training, the integrated physical information neural network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints, and the model is solved to obtain the steady-state reactive power output boundary of the pumped storage unit under various operating conditions.
[0018] The constraint switching prediction and parameter tuning module is used to determine the steady-state dominant constraint at the current operating point based on the conflict intensity coefficient, calculate the contraction rate ratio of various constraint boundaries to active power changes, and predict the critical active power value for steady-state dominant constraint switching based on the contraction rate ratio; calculate the inertial time constant of the mechanical system and electrical system, and determine the inertial travel time and control dead zone compensation amount for switching control modes to cope with steady-state dominant constraint switching.
[0019] The pumped-storage unit transient reactive power capacity assessment module is used to introduce response speed constraints and duration constraints, establish a transient reactive power response capacity boundary model, and use the transient reactive power response capacity boundary model to deduce the transient reactive power response capacity boundary of the pumped-storage unit under different operating conditions. It outputs the reactive power output boundary diagram, transient reactive power response capacity boundary, critical active power value, inertial travel time, and control dead zone compensation amount under each operating condition, which are used to guide the reactive power scheduling and control of the pumped-storage unit.
[0020] According to a third aspect of the present invention, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.
[0021] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the above-described method.
[0022] The beneficial effects of this invention are as follows:
[0023] 1) This invention achieves a systematic characterization of the reactive power regulation capability of pumped-storage units by constructing a three-category constraint model that includes mechanical constraints, electrical constraints, and grid constraints, overcoming the limitations of existing studies that only consider a single type of constraint. The coupled analysis of the three types of constraints can more accurately reflect the reactive power regulation capability boundary of pumped-storage units in actual operation.
[0024] 2) This invention introduces the Differentiable Morse-Smale Complex Layer (DMSL) into the field of reactive power constraint boundary solution for pumped-storage units for the first time. By performing differential topological analysis on the constraint manifold on the power-reactive power plane, critical points, stable / unstable manifolds, and Betti numbers are extracted as topological prior features. These topological features not only guide PINN to avoid getting trapped in local feasible regions, but also provide early warnings when topological changes occur in the feasible region (such as splitting, merging, or the appearance of holes), thus solving the fundamental deficiency of existing neural network methods in lacking global topological awareness.
[0025] 3) This invention proposes a method for predicting the switching of steady-state dominant constraints based on the boundary contraction rate ratio. By calculating the rate ratio of various constraint boundaries with active power change, the critical power point for the switching of steady-state dominant constraints is automatically identified, reserving reaction time for the dispatch center and realizing a leap from "current constraint identification" to "future constraint prediction".
[0026] 4) This invention introduces the mechanical / electrical inertia time constant into the control compensation mechanism for the first time, defining "inertia travel time" and "control dead zone compensation amount". By issuing mode switching commands in advance, the response delay caused by physical inertia is offset, realizing closed-loop automation from constraint recognition to control execution, and avoiding brief overshoots caused by inertial delay.
[0027] 5) This invention constructs a multi-layered nested algorithm framework of GWO-CNN-MHCA-BiLSTM-PINN, which deeply integrates gray wolf optimization, convolutional neural networks, multi-head cross-attention, bidirectional long short-term memory networks, and physical information neural networks. Among them, GWO automatically searches for the optimal hyperparameter configuration, CNN automatically extracts constraint features from multi-dimensional physical quantities, MHCA explicitly models the competitive conflict relationship between the three types of constraints, BiLSTM captures the monotonicity and turning point features of reactive boundary sequences, and PINN serves as the basic network to integrate the various modules and embed physical constraints to complete the end-to-end solution. The modules have clear division of labor and are progressively enhanced, which significantly improves the accuracy and efficiency of feasible region solution.
[0028] 6) This invention proposes for the first time an automatic identification method for steady-state dominant constraints based on MHCA. By inputting three types of constraint features into the MHCA module, the interaction attention matrix between constraints is explicitly calculated, and the bottleneck degree of each constraint on the reactive power boundary is quantified by a normalized conflict intensity coefficient. The identification results of steady-state dominant constraints under different operating conditions emerge automatically, solving the deficiency of existing technologies in quantifying bottleneck constraints.
[0029] 7) This invention introduces response speed constraints and duration constraints to construct a transient reactive power response capability boundary model, which can deduce the reactive power response capability boundary of pumped-storage units during transient processes, thus overcoming the shortcomings of existing steady-state models. The transient reactive power response capability boundary can reflect the reactive power support capability of the unit within a millisecond to second timescale after a fault occurs.
[0030] 8) This invention considers the differences in constraints under various operating conditions such as power generation, pumping, and phase regulation, and realizes unified modeling and comparative analysis of reactive power regulation capabilities under multiple operating conditions, providing a theoretical basis for the coordinated control of pumped-storage units under multiple operating conditions. By comparing the reactive power output boundaries under different operating conditions, the dispatch center can select the optimal operating condition according to the grid demand. Attached Figure Description
[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0032] Figure 1 This is a schematic diagram illustrating the principle of a method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to an embodiment of the present invention.
[0033] Figure 2 This is a schematic diagram illustrating the coupling relationship of three types of constraints in a method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to an embodiment of the present invention.
[0034] Figure 3 This is a bar chart comparing the reactive power adjustment ranges under each operating condition in a method for determining the reactive power adjustment constraint boundary of a pumped-storage unit under multiple operating conditions according to an embodiment of the present invention.
[0035] Figure 4 This is a transient reactive power response capability boundary curve diagram in a method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to an embodiment of the present invention.
[0036] Figure 5This is a schematic diagram of DMSL differential topology analysis in a method for determining the reactive power regulation constraint boundary of a pumped storage unit under multiple operating conditions according to an embodiment of the present invention.
[0037] Figure 6 This is a diagram of the DMSL-GWO-CNN-MHCA-BiLSTM-PINN framework architecture in a method for determining the reactive power regulation constraint boundary under multiple operating conditions of a pumped-storage unit according to an embodiment of the present invention.
[0038] Figure 7 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation
[0039] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.
[0040] Example 1
[0041] like Figures 1-6 As shown, a method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions includes the following steps:
[0042] S1. Obtain the mechanical and electrical parameters of the pumped-storage unit and the operating parameters of the power grid, and construct a constraint condition model that includes mechanical constraints, electrical constraints and power grid constraints;
[0043] Among them, such as Figure 2 The diagram illustrates the coupling relationships between various constraints, with mechanical parameters including moment of inertia J and synchronous speed. Permissible speed change rate and Guide vane opening range [ ], Maximum guide vane opening change rate Working head range [ , Electrical parameters include the maximum stator current. Excitation current range [ Maximum excitation current change rate Direct-axis synchronous reactor Maximum no-load electromotive force Maximum capacity of converter Power grid operating parameters include voltage operating range. Maximum voltage change rate Equivalent reactance X from unit to system, fault clearing time Low voltage ride-through factor k.
[0044] It should be noted that the impedance of the generator step-up transformer is ignored in this invention, and the generator terminal voltage is assumed to be the same as the grid connection point voltage, which is uniformly represented by the symbol. express.
[0045] Mechanical constraints include speed constraints, guide vane opening constraints, head constraints, and rotor motion equation constraints. Speed constraints limit the operating speed range of the pumped-storage unit under different operating conditions. Under power generation conditions, the speed must be within the allowable rate of decrease and increase of synchronous speed. Under pumping conditions, the speed is basically locked at synchronous speed with an allowable fluctuation range of less than one percent. Under phase adjustment conditions, the speed must be between the set minimum and maximum values. Guide vane opening constraints include amplitude limiting constraints and rate constraints. Amplitude limiting constraints restrict the relative opening of the guide vanes to between the mechanically allowable minimum and maximum values, while rate constraints restrict the rate of change of the guide vane opening to not exceed the maximum value. Head constraints limit the operating head to between the minimum and maximum heads. Rotor motion equation constraints describe the dynamic balance between mechanical power and electromagnetic power, reflecting the coupling effect of active power changes on the dynamics of the mechanical system.
[0046] Electrical constraints include stator current constraints, excitation current constraints, power circle constraints, and converter capacity constraints. Among them, stator current constraints limit the heating level of the stator windings; excitation current constraints limit the heating level of the rotor windings and the response capability of the excitation system; power circle constraints comprehensively consider the limitations imposed by the stator current limit circle, excitation current limit circle, and power angle stability boundary on the unit's operating range; and converter capacity constraints limit the apparent power output capability of the power electronic converter in the variable speed pumped-storage unit.
[0047] Grid constraints include voltage constraints, voltage rate of change constraints, voltage sensitivity constraints, and transient voltage stability constraints. Among them, voltage constraints are used to limit the voltage operating range of the grid connection point; voltage rate of change constraints are used to limit the voltage fluctuation speed caused by reactive power regulation; voltage sensitivity constraints are used to describe the degree of impact of reactive power changes on voltage; and transient voltage stability constraints are used to ensure that the units have sufficient reactive power support capacity to meet the requirements of low voltage ride-through and high voltage ride-through during grid faults.
[0048] S2. Based on the constraint model, establish the mechanical constraints, electrical constraints, and power grid constraints for each working condition;
[0049] S3. Based on the physical information neural network, topological prior features, local constraint features, conflict intensity coefficient and reactive power boundary sequence features are integrated. After training, the integrated physical information neural network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints. The model is then solved to obtain the steady-state reactive power output boundary of the pumped storage unit under various operating conditions.
[0050] This involves using a physical information neural network as a foundation, integrating topological prior features, local constraint features, conflict intensity coefficients, and reactive power boundary sequence features. Through training, the fused physical information neural network converges to obtain a feasible region model for reactive power regulation capability with multiple constraints. Solving this model yields the steady-state reactive power output boundary of the pumped-storage unit under various operating conditions, including:
[0051] S31. Differentiable Morse-Smale ComplexLayer (DMSL) is used to perform differential topological analysis on the constrained manifolds formed by the three types of constraints on the power-reactive plane, and critical points, stable / unstable manifolds and Betti numbers are extracted as topological prior features.
[0052] The topological features output by the DMSL module include:
[0053] Critical point set: Local extreme points and saddle points on the constraint boundary, corresponding to the "tightest" position in the feasible region and the potential position for constraint switching;
[0054] Stable and unstable manifolds: describe the evolution path of the constraint boundary between different active power values, and are used to determine whether the feasible region will split or merge;
[0055] Betti number: This represents the number of connected components in the feasible region. Represents the number of holes within the feasible region, when or It issues topology anomaly warnings in a timely manner, indicating that certain active power ranges are completely infeasible or there are unreachable areas under the current operating conditions.
[0056] The topological prior features are used to: 1) initialize the weights of PINN so that the initial output of the network has a topological structure similar to the real constrained manifold; 2) add a topological penalty term to the loss function, and apply additional penalties if the boundary of the network output violates the topological relationship extracted by DMSL (such as producing undesirable holes or connection errors).
[0057] Specifically, differential topological analysis is performed on the constrained manifolds formed by three types of constraints on the power-reactive plane using a differentiable Morse-Smale complex layer. Critical points, stable / unstable manifolds, and Betti numbers are extracted as topological prior features, including:
[0058] S311. The signed distance from any point on the power-reactive plane to the boundary of the multi-constraint feasible region is taken as the comprehensive constraint function.
[0059] S312. Calculate the gradient field of the synthetic constraint function on the discrete grid using numerical methods, and divide the critical points into local maxima, local minima and saddle points according to the eigenvalue signs of the Hessian matrix.
[0060] S313. Starting from each saddle point, trace the gradient streamlines forward and backward along the eigenvector direction of the Hessian matrix to obtain the stable and unstable manifolds, and determine the critical points (maximum points, minimum points, saddle points) based on the intersection of the stable and unstable manifolds.
[0061] S314. After completing the manifold partitioning, the power-reactive plane is decomposed into several cavities, and all points in each cavity have the same gradient flow towards the endpoint. The connectivity of the cavities is counted, and the number of connected components and the number of voids are determined.
[0062] S32. Use the Grey Wolf Optimization Algorithm (GWO) to automatically optimize the network hyperparameters and determine the optimal network hyperparameters;
[0063] S33. Construct a convolutional neural network (CNN) to extract features from the multidimensional physical parameters of the three types of constraints, and obtain local constraint features;
[0064] S34. Explicitly model the interaction conflict relationship between the three types of constraints using the multi-head cross attention mechanism (MHCA) to determine the conflict intensity coefficient;
[0065] Specifically, the attention mechanism layer is used to automatically identify the type of steady-state dominant constraint under each operating condition: after training convergence under the power generation condition, the attention weight corresponding to the electrical constraint is significantly higher than that of the mechanical constraint and the grid constraint, and the power angle stability boundary in the electrical constraint is the steady-state dominant constraint; after training convergence under the pumping condition, the attention weight corresponding to the mechanical constraint is significantly higher than that of the electrical constraint and the grid constraint, and the lower limit of the power factor in the mechanical constraint is the steady-state dominant constraint; after training convergence under the phase adjustment condition, the excitation current sub-term in the electrical constraint has the highest weight, and the excitation current constraint in the electrical constraint is the steady-state dominant constraint.
[0066] S35. Capture the sequence pattern of reactive power boundary changes with active power through a bidirectional long short-term memory network (BiLSTM) to determine the characteristics of the reactive power boundary sequence.
[0067] S36. Using the Physical Information Neural Network (PINN) as the base network, topological prior features, local constraint features, conflict intensity coefficients, and reactive boundary sequence features are integrated into the Physical Information Neural Network framework to obtain the fused Physical Information Neural Network.
[0068] S37. Traverse the range of active power values, take active power as input and the upper and lower boundaries of reactive power as output, combine the optimal network hyperparameters and fusion loss function, train the fused physical information neural network until the network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints.
[0069] The fusion loss function is composed of three weighted components: data fitting loss, physical constraint loss, and topology consistency loss. The data fitting loss uses mean squared error to measure the deviation between the reactive boundary of the network output and the labels of the training samples. The physical constraint loss is composed of the violation penalty terms of the three types of constraints, weighted by the conflict intensity coefficient. The topology consistency loss is composed of Betti number prediction bias and critical point position bias, and is used to penalize situations where the network output boundary is inconsistent with the topology of the real constraint manifold.
[0070] Specifically, the expression for the fusion loss function is:
[0071]
[0072]
[0073]
[0074] In the formula, For the loss of fusion, For data fitting loss, For physical constraint loss, For topology consistency loss, The weights for the physical constraint loss, The weights for the topology consistency loss are... For mechanical constraint loss, For electrical constraint loss, For grid constraint losses, The conflict intensity coefficient is the normalized mechanical constraint loss. The conflict intensity coefficient is the normalized electrical constraint loss. The conflict intensity coefficient is the normalized power grid constraint loss. , These represent the number of connected components in the feasible region predicted by the network and the actual value calculated by the differentiable Morse-Smale complex layer, respectively. , The numbers are the number of holes predicted by the network and the actual value calculated by the Morse-Smale complex layer, respectively. , These represent the critical point locations output by the network and the true critical point locations extracted by the differentiable Morse-Smale complex layer, respectively. , , These are the weighting coefficients for the deviation of the number of connected components, the deviation of the number of voids, and the deviation of the critical point location, respectively. The preferred values are as follows: , The total number of critical points. For critical point index, Let represent the square of the Euclidean distance.
[0075] During network training, attention weights (i.e. conflict intensity coefficients) are updated synchronously with other network parameters. When a certain type of constraint is frequently violated under the current operating conditions, the network will adaptively increase its attention weights to impose stronger physical constraint penalties, so that the network output conforms to the constraint.
[0076] After training convergence, the physical meaning of the attention weights is as follows: the higher the weight value, the more severely the constraint was violated during training, and the stronger its limiting effect on the reactive power boundary, which is the steady-state dominant constraint under the current operating condition. The identification results of the steady-state dominant constraints under the three operating conditions are: electrical constraints dominate under the power generation condition, mechanical constraints dominate under the pumping condition, and the excitation current constraint dominates the electrical constraints under the phase-shifting condition. This steady-state dominant constraint's pattern of switching with the operating condition is not a pre-set judgment logic, but a physical law that the network autonomously learns from the data.
[0077] The training samples are generated as follows: Upper and lower limits of reactive power corresponding to different active power values are extracted from historical operating data or simulation data as labels; for active power ranges lacking labels, supplementary training samples are generated using a traversal method; the total number of training samples covers the active power value ranges for each of the three operating conditions: power generation, pumping, and phase regulation. After training, for any given active power input, the network can directly infer and output the corresponding reactive power boundary value without needing to re-execute the traversal or iterative solution.
[0078] S38. Based on the feasible domain model of reactive power regulation capability with multi-constraint coupling, determine the steady-state reactive power output boundary of pumped storage unit under various operating conditions.
[0079] Specifically, the multiple operating conditions include power generation, pumping, power generation phase modulation, and pumping phase modulation.
[0080] Under different operating conditions, the reactive power allowable ranges defined by the three types of constraints are different, causing the steady-state dominant constraint to switch with the operating conditions, specifically manifested as follows:
[0081] Under power generation conditions, the power angle stability boundary is the key constraint, the reactive power regulation range is mainly in the late-phase operation region, and the steady-state reactive power output boundary is determined by the intersection of electrical constraints and grid constraints.
[0082] Under pumping conditions, the lower limit of the power factor is a key constraint, the reactive power adjustment range is relatively narrow, and the steady-state reactive power output boundary is determined by the intersection of mechanical and electrical constraints.
[0083] Under phase-shifting conditions, the active power is close to zero, the reactive power regulation range is widest, the excitation current constraint is the main limiting factor, the steady-state reactive power output boundary is determined by the intersection of electrical constraints and grid constraints, and the transient reactive power response boundary is dominated by electrical constraints.
[0084] Specifically, the DMSL-GWO-CNN-MHCA-BiLSTM-PINN framework constructed in step S3 consists of five nested algorithm modules:
[0085] The first layer is the DMSL differential topology analysis module: it treats the boundary curves of the three types of constraints on the PQ plane as a level set of a scalar function, and calculates the critical points (maximum points, minimum points, saddle points), stable and unstable manifolds connecting the critical points, and the Betti number of each connected region using a differentiable Morse-Smale complex. The number of connected components. (This refers to the number of holes). The above topological features are embedded as prior knowledge in the loss function and network initialization of PINN to guide the network to avoid getting trapped in local feasible regions and to maintain boundary topological consistency.
[0086] The second layer is the GWO hyperparameter optimization module, which uses the Grey Wolf optimization algorithm to automatically search for the number of network layers, the number of nodes per layer, the learning rate, and the physical loss weights.
[0087] The third layer is the CNN constraint feature extraction module, which organizes physical parameters such as guide vane opening, head, rotational speed, stator current, excitation current, power angle, grid connection point voltage and voltage change rate into a one-dimensional feature sequence, and automatically extracts the local correlation patterns inside various constraints through multi-layer one-dimensional convolution.
[0088] The fourth layer is the MHCA interaction modeling module, which explicitly calculates the interaction attention matrix between the three types of constraints, and uses the normalized average attention weight as the conflict intensity coefficient of each type of constraint.
[0089] The fifth layer is the BiLSTM sequence modeling module, which captures the monotonicity of the reactive boundary sequence and the boundary transition characteristics when the steady-state dominant constraint switches through two long short-term memory layers, forward and backward.
[0090] The outputs of the above five modules (topology prior features, hyperparameter configuration, local constraint features, conflict intensity coefficient, and sequence features) are integrated into the PINN basic network. PINN also embeds physical constraints to complete the end-to-end solution.
[0091] S4. Determine the steady-state dominant constraint at the current operating point based on the conflict intensity coefficient, calculate the contraction rate ratio of various constraint boundaries to active power changes, and predict the critical active power value for steady-state dominant constraint switching based on the contraction rate ratio; calculate the inertial time constant of the mechanical system and electrical system, and determine the inertial travel time and control dead zone compensation amount for control mode switching to cope with steady-state dominant constraint switching. Here, control mode switching refers to the control system switching from the control strategy corresponding to the current constraint to the control strategy corresponding to the new dominant constraint when the steady-state dominant constraint changes.
[0092] The specific calculation process for the shrinkage rate ratio is as follows:
[0093] After the network training is completed, for active power P, the reactive power boundary value corresponding to the i-th type of constraint is defined as follows: the reactive power boundary value determined by the i-th type of constraint is... Then the shrinkage rate of the constraint boundary Define the shrinkage rate ratio of constraint type i to constraint type j as . .
[0094] when When the boundary contraction is faster for the i-th type of constraint, its constraint effect on the reactive boundary is stronger than that of the j-th type, meaning the i-th type is the dominant steady-state constraint; when When, the j-th type is the steady-state dominant constraint; when When the value changes from greater than 1 to less than 1 (or vice versa), it indicates that the restrictive effect of the i-th and j-th types of constraints has reversed, and the corresponding active power value is... This is the predicted critical point for the switching of the dominant steady-state constraint. This prediction is used to guide the dispatch center when power changes approach... Adjust the control strategy in advance to avoid voltage instability caused by constraint switching.
[0095] The specific process of inertial travel time and control dead zone compensation is as follows:
[0096] Calculate the inertial time constant of a mechanical system Where H is the unit's inertial constant, Calculate the synchronous angular velocity; calculate the inertial time constant of the electrical system. ,in For the excitation winding inductance, This represents the excitation winding resistance. Defines the inertial travel time for control mode switching. ,in To control communication delay, inertial travel time is used here. This represents the minimum lead time required from the detection of the steady-state dominant constraint to the actual execution control mode, used to compensate for mechanical / electrical inertia and communication delays.
[0097] Based on the current rate of change of active power Calculate the control dead zone compensation amount When the active power is detected to be approaching the predicted switching point. And enter the interval In such cases, the control system issues a mode switching command in advance to compensate for the response delay caused by mechanical / electrical inertia. This compensation mechanism ensures a seamless transition from constraint recognition to control execution, avoiding brief limit exceedances due to inertial delays.
[0098] S5. Introduce response speed constraints and duration constraints to establish a transient reactive power response capability boundary model. Use this model to deduce the transient reactive power response capability boundary of the pumped-storage unit under different operating conditions. The selection of the steady-state basic reactive power value in the transient reactive power response capability boundary model is based on the reactive power boundary value corresponding to the steady-state dominant constraint. Output reactive power output boundary diagrams (with the steady-state dominant constraint type marked on each boundary segment), transient reactive power response capability boundaries (with the steady-state dominant constraint switching nodes at different time stages marked in the transient reactive power response capability boundary parameters), critical active power values, inertial travel time, and control dead zone compensation for each operating condition to guide the reactive power scheduling and control of the pumped-storage unit.
[0099] This involves introducing response speed constraints and duration constraints to establish a transient reactive power response capability boundary model. This model is then used to deduce the transient reactive power response capability boundary of the pumped-storage unit under different operating conditions. The model outputs reactive power output boundary diagrams, transient reactive power response capability boundaries, critical active power values, inertial travel time, and control dead zone compensation for each operating condition. These outputs guide the reactive power scheduling and control of the pumped-storage unit, including:
[0100] S51. Based on the guide vane opening rate constraint in the mechanical constraints and the excitation current rate constraint in the electrical constraints, determine the upper limit of the response speed of reactive power regulation. The upper limit of the response speed is determined by the slowest of the excitation current change rate, the guide vane opening rate, and the voltage change rate.
[0101] S52. Based on the upper limit of the response speed of reactive power regulation, calculate the reactive power ramp-up at each moment during the transient process. The reactive power ramp-up is the integral of the upper limit of the response speed over time, representing the total amount of reactive power that the unit can generate from the moment the fault occurred to the current moment.
[0102] S53. Based on the thermal accumulation constraints of the stator winding and the excitation winding, determine the duration of reactive power regulation. Under the thermal accumulation constraint, the reactive power decreases in an S-shaped curve as the duration increases. Determine the upper limit of reactive power limited by thermal accumulation at each moment.
[0103] S54. The reactive power ramp-up is superimposed on the steady-state basic reactive power value and compared with the upper limit of reactive power capacity of thermal accumulation limit and transient stability limit value. The minimum value among the three is taken as the boundary of transient reactive power response capability at the current moment. The steady-state basic reactive power value is the reactive power output of the unit before the fault, and the transient stability limit value is affected by voltage, excitation electromotive force and power angle.
[0104] S55. Based on the transient reactive response capability boundary, simulate the reactive response process at different time scales, obtain the time-reactive power curve, and determine the transient reactive response capability boundary value at each moment in the transient process.
[0105] S56. Output the steady-state reactive power output boundary under power generation, pumping and phase adjustment conditions as a reactive power output boundary map in the form of a feasible region on the power-reactive power plane, and the feasible region is composed of the intersection of three types of constraints: mechanical constraints, electrical constraints and grid constraints.
[0106] S57. Output the transient reactive power response capability boundary in the form of a time-reactive power curve to reflect the reactive power support capability of pumped storage units in the millisecond to second timescale after a fault occurs; based on the reactive power output boundary diagram, inertial travel time, control dead zone compensation, critical active power value and the time-reactive power curve of transient reactive power response capability boundary parameters under various operating conditions, guide the generation of reactive power dispatch instructions and grid reactive power reserve configuration of pumped storage units under different operating conditions, help the dispatch center to pre-assess the reactive power support capability of pumped storage units before grid faults, and quickly call up the corresponding reactive power regulation capability when a fault occurs.
[0107] In addition, the output results can also include the distribution of the three types of constraint conflict intensity coefficients learned by MHCA under various operating conditions, presented intuitively in the form of charts. The dispatch center not only knows the range of reactive power that can be adjusted, but also clearly knows the key physical constraints that currently limit this range and their bottleneck degree. When it is necessary to expand the reactive power adjustment capability, the dispatch center determines the physical links that should be adjusted first based on the conflict intensity coefficient distribution.
[0108] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle.
[0109] I. Construction of Mathematical Expressions for Mechanical Constraints
[0110] Mechanical constraints originate from the hydraulic-mechanical system characteristics of pumped storage units and mainly include speed constraints, guide vane opening constraints, head constraints, and rotor motion equation constraints.
[0111] The mathematical expressions for the speed constraint are as shown in equations (1) to (3):
[0112] Under power generation conditions:
[0113]
[0114] in, , Where n is the actual operating speed of the pumped-storage unit under the current operating conditions. , These are the lower and upper limits of the actual operating speed of the pumped-storage unit under the current operating conditions, respectively. and These are the allowable rate of decrease and rate of increase in rotational speed, typically ranging from 3% to 5%.
[0115] Under pumping conditions:
[0116]
[0117] in, To achieve synchronous speed, This refers to the permissible speed fluctuation range under pumping conditions, typically less than 1%.
[0118] Under phase adjustment conditions:
[0119]
[0120] in, This is the minimum permissible speed under phase-shifting conditions. This is the maximum permissible speed under phase-shifting conditions.
[0121] The mathematical expressions for the guide vane opening constraint are shown in equations (4) to (5):
[0122] Amplitude limiting constraints:
[0123]
[0124] in, The relative opening of the guide vanes, The value is typically between 0.1 and 0.2. The value is typically between 0.95 and 1.0.
[0125] Rate constraints:
[0126]
[0127] in, The maximum opening change rate is typically set to 0.1 to 0.3 pu / s, and this constraint directly affects the response speed of reactive power regulation.
[0128] The mathematical expression for head constraint is shown in equation (6):
[0129]
[0130] in, For working head, and These are the minimum and maximum working heads, respectively.
[0131] The mathematical expression for the rotor motion equation constraint is shown in equation (7):
[0132]
[0133] in, For rotational inertia, For mechanical angular velocity, and These are mechanical torque and electromagnetic torque, respectively. For mechanical power, Let be the electromagnetic active power. This equation describes the coupling relationship between the change in active power and the dynamics of the mechanical system.
[0134] II. Construction of Mathematical Expressions for Electrical Constraints
[0135] Electrical constraints originate from the electromagnetic characteristics of generators / motors and the operational limitations of power electronic equipment, and mainly include stator current constraints, excitation current constraints, power circle diagram constraints, and converter capacity constraints.
[0136] The mathematical expressions for stator current constraint are shown in equations (8) to (9):
[0137]
[0138]
[0139] in, For stator current, For the maximum permissible stator current, Q is the active power, U is the reactive power, and U is the terminal voltage. Equation (9) represents the stator current limiting circle, which is a circle with the origin as the center on the PQ plane.
[0140] The mathematical expressions for the excitation current constraint are as shown in equations (10) to (12):
[0141] Steady-state limits:
[0142]
[0143] in, This is the excitation current.
[0144] Rate constraints:
[0145]
[0146] in, This represents the maximum rate of change of excitation current.
[0147] Excitation current limiting circle:
[0148]
[0149] in, It is a direct-axis synchronous reactance. The maximum no-load electromotive force is given by equation (12). Equation (12) represents the excitation current limiting circle, whose center is on the negative half-axis of the Q-axis.
[0150] The power circle diagram constraints include the stator current limiting circle, the excitation current limiting circle, and the power angle stability boundary, which together constitute the operating range of the generator in the PQ plane.
[0151] The mathematical expression for the converter capacity constraint is shown in equation (13):
[0152]
[0153] in, For the apparent power of the converter, , These represent the active and reactive power on the converter side, respectively. This is the maximum capacity of the converter. This constraint mainly applies to variable-speed pumped-storage units that use full-power or doubly-fed converters.
[0154] III. Construction of Mathematical Expressions for Power Grid Constraints
[0155] Grid constraints originate from grid operation requirements, voltage stability, and equipment protection limitations, and mainly include voltage constraints, voltage change rate constraints, voltage sensitivity constraints, and transient voltage stability constraints.
[0156] The mathematical expression for voltage constraint is shown in equation (14):
[0157]
[0158] Where U is the grid connection point voltage, under normal operation =0.95pu, =1.05pu, which can be relaxed to 0.85pu to 1.15pu in case of failure.
[0159] The mathematical expression for the voltage change rate constraint is shown in equation (15):
[0160]
[0161] in, According to power grid specifications, the value is usually between 0.05 and 0.1 pu / s.
[0162] The mathematical expression for the voltage sensitivity constraint is shown in equation (16):
[0163]
[0164] Where X represents the equivalent reactance from the generator unit to the system. This constraint reflects the limitation of the power grid structure on the reactive power regulation effect.
[0165] The mathematical expressions for transient voltage stability constraints are shown in equations (17) to (19):
[0166] Voltage sag allowable range:
[0167]
[0168] in, This is the critical voltage. The actual voltage at the grid connection point at time t after the fault occurs is a monitored quantity that changes over time. This is the time required to clear the fault.
[0169] Low voltage ride-through requirements:
[0170]
[0171] High voltage ride-through requirements:
[0172]
[0173] in, The reactive current injected into the power grid by the generator unit, where k is the low voltage ride-through factor.
[0174] IV. DMSL Differential Topology Analysis
[0175] To overcome the lack of global topological awareness in existing neural network methods, this invention introduces for the first time a differentiable Morse-Smale complex layer (DMSL) to perform differential topological analysis on the constrained manifolds formed by three types of constraints on the PQ plane. The core idea of DMSL is to treat the boundary of the multi-constraint coupled feasible region as a synthetic constraint function. By analyzing the global structure of the gradient field of the function, the set of contour lines is used to extract topological invariants that are instructive for solving the constraint boundary.
[0176] Step 1: Construction of Constrained Manifold Functions
[0177] Define synthesis constraint functions For: any point on the power-reactive plane Place, This represents the signed distance from the point to the nearest constraint boundary. Specifically:
[0178]
[0179] in, Let Euclidean distance be the distance from the point to the boundary. These are the boundary curves of the feasible region enclosed by the three types of constraints—mechanical, electrical, and power grid—on the PQ plane. Let be a signed function, with the following sign convention: it takes a positive value inside the feasible region (where all constraints are satisfied), and a negative value outside the feasible region (where at least one type of constraint is violated). Therefore, level set That is, the total boundary after multiple constraints are coupled, and gradient field The direction points to the "most lenient" direction within the feasible region.
[0180] Step 2: Gradient field calculation and critical point identification
[0181] Calculate using numerical methods The gradient on a discrete grid is calculated by first dividing the PQ plane into an N×N uniform grid, then calculating the signed distance of each of the three types of constraints at each grid point, and taking the minimum value. The value is then used to calculate the gradient using the central difference method.
[0182]
[0183]
[0184] in, Active power Discretization step size of the direction, reactive power Discretization step size of the direction;
[0185] The critical point in a gradient field is defined as The point. According to the Hessian matrix. The eigenvalue sign is used to classify critical points as follows:
[0186] Local maxima (Hessian negative definite): correspond to the "least lenient" point in the feasible region, that is, the center of the region where all constraints are far from the boundary;
[0187] Local minimum (Hessian positive definite): corresponds to the "tightest" point in the feasible region, usually the intersection of multiple constraint boundaries;
[0188] Saddle point (Hessian eigenvalues are positive and negative): This corresponds to the position where the steady-state dominant constraint switches and is the focus of subsequent steps.
[0189] Step 3: Tracing Stable and Unstable Manifolds
[0190] Starting from each saddle point, the gradient streamlines are traced forward and backward along the eigenvectors of the Hessian matrix, respectively. The specific algorithm is as follows:
[0191] Let the two eigenvectors at the saddle point be... (Stable direction, corresponding to negative eigenvalues) and (Unstable direction, corresponding to positive eigenvalues). From the saddle point along... A small perturbation in direction, then proceeding in the gradient descent direction. Integrating, we obtain the stable manifold—the set of all points that eventually flow to the saddle point. From the saddle point along… Directional perturbation, according to the gradient ascent direction Integrating, we obtain an unstable manifold—the set of all points that radiate outward from the saddle point.
[0192] The intersections of stable and unstable manifolds connect different critical points, forming a "skeleton" on the PQ plane. This skeleton precisely characterizes the evolution path of the constraint boundary as power changes: when the active power P changes, the boundary of the feasible region moves along these manifolds, and the saddle point location indicates the critical power value at which the steady-state dominant constraint switches.
[0193] Step 4: Calculation of Betti number and topology warning
[0194] After completing the manifold partitioning, the PQ plane is decomposed into several cells, where all points within each cell have the same gradient flowing towards the endpoint. The connectivity of the cells is statistically analyzed, and the following Betti numbers are calculated:
[0195] (0th Betti number): Number of connected components. If This indicates that the feasible region is divided into multiple unconnected areas, meaning that there are no feasible reactive power values at all in certain active power ranges. For example, in pumping operations, when the active power is close to the rated value, mechanical constraints may split the feasible region into two isolated areas, in which case the dispatch center must skip that power range.
[0196] (First Betty Number): Hole Number. If This indicates the existence of "holes" within the feasible region, meaning that although each individual constraint is satisfied, the joint constraints render some interior points infeasible. For example, when the power circle diagram of electrical constraints intersects with the voltage limit curve of grid constraints to form a circular forbidden zone, At this point, simply using convex optimization methods would incorrectly treat the interior of the hole as feasible, while DMSL can identify and warn of this in advance.
[0197] when or At this time, the DMSL module issues a topology anomaly warning, indicating that the scheduling center has complex feasible regions that are non-convex, disconnected, or have holes under the current operating conditions. Conventional optimization algorithms may fail, and a global search or hierarchical solution strategy needs to be adopted.
[0198] Step 5: Fusion of topological prior features and PINN
[0199] The topological prior features output by the DMSL module include: a set of critical point coordinates. ,in, For the first Active power coordinates at critical points For the first Reactive power coordinates at critical points; discrete trajectory points of stable and unstable manifolds; number Topology anomaly flag.
[0200] These features are incorporated into the PINN framework in two ways:
[0201] (1) Weight initialization: The critical point coordinates extracted by DMSL are used as the initial bias values of some neurons in the hidden layer of PINN. Specifically, for neurons that output the upper limit of reactive power boundary, their bias is initialized to the reciprocal of the reactive power value closest to the critical point (i.e., the closer to the critical point, the steeper the initial boundary). This initialization method makes the network sensitive to the "corner points" and "turning points" of the constraint boundary at the beginning of training, avoiding training instability and slow convergence caused by random initialization.
[0202] (2) Topology consistency loss function: Add a topology penalty term to the total loss function. The penalty consists of two parts:
[0203]
[0204] in, and These represent the number of feasible region connected components predicted by the network and the actual value calculated by DMSL, respectively. , These represent the number of connected components in the feasible region predicted by the network and the actual value calculated by the differentiable Morse-Smale complex layer, respectively. , The numbers are the number of holes predicted by the network and the actual value calculated by the Morse-Smale complex layer, respectively. , These represent the critical point locations output by the network and the true critical point locations extracted by the differentiable Morse-Smale complex layer, respectively. , , These are the weighting coefficients for the deviation of the number of connected components, the deviation of the number of voids, and the deviation of the critical point location, respectively. The preferred values are as follows: , The total number of critical points. For critical point index, Let represent the square of the Euclidean distance.
[0205] This loss term forces the output of PINN to maintain the same topology as the real constrained manifold, for example, preventing the network from generating additional holes in the originally simply connected feasible region, or preventing the network from incorrectly connecting two regions that should be separated.
[0206] Step 6: DMSL Analysis Results under Typical Operating Conditions
[0207] Taking the power generation condition as an example, DMSL analysis shows that on the PQ plane, the feasible region boundary consists of three parts: the left side is the power angle stability boundary (sloping downwards), the right side is the stator current limiting circle (curving upwards), and the bottom is the excitation current limiting circle (arc-shaped). DMSL identifies two saddle points: saddle point S1 is located at the intersection of the power angle boundary and the stator current boundary. Saddle point S2 is located at the intersection of the power angle boundary and the excitation current boundary. The unstable manifold extends laterally from S1, forming the upper boundary of the feasible region; the stable manifold extends downwards from S2, determining the shape of the lower boundary. The Betti number is calculated as follows: This indicates that the feasible region is a simply connected region with no holes.
[0208] Under pumping conditions, DMSL detected when When the pumping power is relatively high, the feasible region is split into two independent regions. DMSL issued a topology anomaly warning, indicating that the region was unschedulable. The results of the DMSL differential topology analysis are illustrated below. Figure 5 As shown.
[0209] All outputs of the DMSL analysis above (critical points, manifolds, Betti numbers, warning signs) are integrated into the subsequent PINN training as prior knowledge, which significantly improves the network's fitting accuracy and topological correctness for complex constraint boundaries.
[0210] V. Feasible Region Model of Reactive Power Regulation Capability with Multi-Constraint Coupling
[0211] Taking into account the three types of constraints, the reactive power output boundary of the pumped-storage unit in steady state is shown in equation (24):
[0212]
[0213] in, The steady-state reactive power output boundary after multi-constraint coupling. The reactive power boundary is determined by mechanical constraints. The reactive power boundary is determined by electrical constraints. The reactive power boundary is determined by grid constraints.
[0214] Reactive boundary determined by mechanical constraints The expression is as shown in equation (25):
[0215]
[0216] in, This is a nonlinear mapping function between the reactive boundary determined by mechanical constraints and various mechanical parameters. Mechanical power;
[0217] Reactive power boundary determined by electrical constraints The expression is as shown in equation (26):
[0218]
[0219] in, The reactive power boundary is determined by the excitation current constraint. The reactive power boundary is determined by the stator current constraint. The reactive power boundary is determined by the power circle diagram constraints. The reactive power boundary is determined by the converter capacity constraint.
[0220] Reactive power boundary determined by grid constraints The expression is as shown in equation (27):
[0221]
[0222] in, The reactive power boundary is determined by voltage constraints. The reactive power boundary is determined by the grid capacity constraint. The reactive power boundary is determined by the low voltage ride-through requirement. The reactive power boundary is determined by the high voltage ride-through requirements.
[0223] By traversing the range of active power P, and solving for the maximum and minimum values of reactive power Q that satisfy all constraints for each value of P, the steady-state reactive power boundary on the PQ plane can be obtained.
[0224] VI. Multi-constraint Coupled Solution and Steady-State Dominant Constraint Recognition Based on GWO-CNN-MHCA-BiLSTM-PINN
[0225] The core innovation of this invention lies in constructing a multi-layered nested algorithm framework, such as GWO-CNN-MHCA-BiLSTM-PINN. Figure 6 As shown, this framework uses PINN as the base network and deeply integrates four enhancement modules: GWO, CNN, MHCA, and BiLSTM. Each module has a clear division of labor and is progressively enhanced.
[0226] 1. Network architecture design
[0227] The overall network uses PINN as its backbone, employs a feedforward fully connected structure, and embeds a CNN constrained feature extraction layer, an MHCA layer, and a BiLSTM sequence modeling layer. The network consists of an input layer, a CNN constrained feature extraction layer, three fully connected hidden layers, one MHCA layer, a BiLSTM sequence modeling layer, and an output layer. The specific design of each layer is as follows:
[0228] Input layer: 1 node, receiving the input variable active power P.
[0229] CNN constraint feature extraction layer: Organizes the multidimensional physical parameters involved in the three types of constraints into a one-dimensional feature sequence, including: mechanical side parameters—guide vane opening α, head H, rotational speed n, and rotational speed change rate dn / dt; electrical side parameters—stator current. Excitation current , and angle Converter power Grid-side parameters—grid-connected point voltage U, voltage change rate dU / dt, and equivalent reactance X. Local correlation patterns and pre-limit features within various constraints are automatically extracted using three layers of one-dimensional convolution (kernel size 3, stride 1, output channel numbers 64, 128, and 256 respectively, followed by batch normalization and ReLU activation after each layer). After global average pooling, high-dimensional feature representations of each of the three types of constraints are output. , which serves as the input for subsequent hidden layers.
[0230] Hidden Layer 1: A fully connected layer containing 128 nodes, using the ReLU activation function. This layer is used for preliminary feature extraction and dimensionality enhancement of the input variables.
[0231] Hidden Layer 2: A fully connected layer containing 128 nodes, using the ReLU activation function. This layer, preceding the attention mechanism layer, further abstracts the features, enabling subsequent attention layers to allocate weights based on richer feature representations.
[0232] MHCA Layer: This layer is the core innovative structure of this network. It maps the features output from the second hidden layer to representations of three types of constraints, and explicitly calculates the interaction attention matrix between these three types of constraints through multiple attention heads. Specifically, using mechanical constraint features as queries, and electrical and power grid constraints as keys and values respectively, the dependencies of mechanical constraints on the other two types of constraints are calculated; similarly, using electrical and power grid constraints as queries, cross-attention is calculated. The outputs of each attention head are concatenated and fused through linear projection to form joint features. .
[0233] While calculating cross-attention, the average attention weights corresponding to the three types of constraints in each attention head are extracted, and after softmax normalization, they are used as the conflict intensity coefficients for each type of constraint. The constraint type that satisfies the condition that the sum of the three factors is 1 is the steady-state dominant constraint at the current running point.
[0234] The output of this layer has a dual purpose: firstly, it provides a conflict intensity coefficient. As a weighting coefficient, the penalty terms of the three types of physical constraints are weighted and summed in the loss function; secondly, after training convergence, the three coefficient values are directly read, and the constraint type with the largest coefficient is the steady-state dominant constraint under the current working condition.
[0235] BiLSTM sequence modeling layer: Combines the features output by MHCA Input a BiLSTM network. The forward LSTM scans the sequence according to the active power P from smallest to largest, while the backward LSTM scans the sequence according to P from largest to smallest. The hidden states of the two directions are concatenated and output at each time step. The core role of BiLSTM in this framework is to capture the monotonicity of the reactive power boundary sequence—as P increases, the stator current limit circle gradually tightens, and the upper limit of reactive power increases. Monotonically decreasing; identifying boundary transition features—when the steady-state dominant constraint switches at a certain P value, the slope of the Q boundary curve changes, and the gating mechanism of BiLSTM can effectively capture such transitions.
[0236] Hidden Layer 3: A fully connected layer containing 128 nodes, using the ReLU activation function. This layer receives features processed by BiLSTM and further performs nonlinear mapping.
[0237] Output layer: 2 nodes, each outputting the upper limit of reactive power. and the lower limit of reactive power No activation function is set for the output layer.
[0238] During training, the trainable parameters of the MHCA layer, along with the weights and biases of all fully connected layers in the network, are updated synchronously via backpropagation and the Adam optimizer. The GWO hyperparameters are evaluated for fitness and the gray wolf positions are updated after each complete training iteration. While reducing data fitting error, the network autonomously learns the relative constraints on the reactive boundary through gradient feedback from the physical loss term, ultimately ensuring that the steady-state dominant constraint achieves the highest conflict intensity coefficient.
[0239] 2. Loss Function Design
[0240] Network loss function Loss due to data fitting Physical constraint loss and topology consistency loss Three-part weighted composition:
[0241]
[0242] in, The weights for the physical constraint loss, The weights for the topology consistency loss are typically taken as... Data fitting loss Mean squared error is used to measure the deviation between the Q boundary value of the network output and the training sample labels.
[0243] Physical constraint loss It consists of a weighted sum of the penalty terms for violating the three types of constraints, calculated using attention weights:
[0244]
[0245] in, , , The normalized conflict intensity coefficient is automatically calculated by the Multi-Head Cross-Attention Mechanism (MHCA) based on the current input data, satisfying the following conditions: No fixed values need to be preset. The losses for each constraint are defined as follows:
[0246] Mechanical constraint losses – penalties for exceeding guide vane opening limits and exceeding head limits:
[0247]
[0248] in, Let i be the number of samples, and i be the sample index. Let be the relative opening of the guide vane for the i-th sample. These are the minimum and maximum allowable values for the guide vane opening. To take the maximum value between zero and the value in parentheses, ensure that positive loss only occurs when the value exceeds the allowed range (loss is zero within the allowed range).
[0249] Electrical constraint losses – penalties for stator current exceeding limits, excitation current exceeding limits, and power angle exceeding limits:
[0250]
[0251] Grid constraint losses – penalties for exceeding voltage limits and exceeding voltage change rate limits:
[0252]
[0253] When the physical quantity corresponding to the network output value is within the allowable range, the penalty term is zero; when it exceeds the range, the penalty is calculated as the square of the excess amount.
[0254] 3. Training process and steady-state dominant constraint identification
[0255] Historical operating data or simulation data were used as training samples. The input of each sample was active power P, and the labels were the corresponding upper and lower limits of reactive power. Independent network models were trained for three operating conditions: power generation, pumping, and phase regulation. The GWO algorithm was used to automatically optimize the network hyperparameters (number of layers, number of nodes, learning rate, and weight hyperparameters), and the Adam optimizer was used to update the network parameters and attention weights synchronously. During training, when a constraint was continuously violated in a large number of samples, MHCA adaptively increased the conflict intensity coefficient corresponding to that constraint, thus making the optimizer more focused on satisfying that constraint. After training convergence, the constraint type with the largest conflict intensity coefficient was the steady-state dominant constraint under that operating condition.
[0256] 4. Steady-state dominant constraint identification results for three operating conditions
[0257] Under power generation conditions, the conflict intensity coefficient corresponding to electrical constraints is significantly higher than that of mechanical and grid constraints, indicating that electrical constraints are the dominant steady-state constraint under the current conditions. Under pumping conditions, the conflict intensity coefficient corresponding to mechanical constraints is much higher than that of the other two types of constraints, indicating that the dominant steady-state constraint has switched from electrical constraints to mechanical constraints. Under phase-shifting conditions, the excitation current sub-term in the electrical constraints accounts for the highest proportion of the conflict intensity coefficient, indicating that the excitation current constraint is the dominant steady-state constraint.
[0258] 5. The physical meaning of MHCA
[0259] The allocation of conflict intensity coefficients is entirely autonomously learned by the MHCA during training. The network automatically identifies which type of constraint is triggered most frequently and has the most severe exceedance under the current physical conditions, and assigns the highest conflict intensity coefficient to that type of constraint. This automatic emergence pattern closely matches actual physical analysis—electrical constraints are tightest during power generation, power factor lock-in occurs during pumping, and the excitation current determines the depth of failure during phase adjustment. The output includes a conflict intensity coefficient distribution map for each operating condition, which the dispatch center can use to quickly locate key physical bottlenecks limiting non-functional capacity, providing a basis for targeted adjustments.
[0260] VII. Boundary Model of Transient Reactive Power Response Capability
[0261] During transient processes, reactive power regulation capability is also constrained by response speed. The boundary model of transient reactive power response capability is shown in equation (33):
[0262]
[0263] in, For the transient reactive response limit, This represents the steady-state basic reactive power value. For thermal accumulation constraint, For transient stability constraints, reactive power ramping As shown in equation (34):
[0264]
[0265] in, For the maximum rate of change of reactive power, This represents the maximum rate of change of excitation current. The rate of change of the maximum guide vane opening. This represents the maximum rate of voltage change.
[0266] Thermal accumulation constraint As shown in equation (35):
[0267]
[0268] in, The slope of the heat accumulation curve. is the heat accumulation time constant.
[0269] Transient stability constraints As shown in equation (36):
[0270]
[0271] in, For the angle of attack.
[0272] By simulating the reactive power response process at different time scales, time-reactive power curves are obtained, and the reactive power response limit in the transient process is determined. A typical transient reactive power response capability boundary curve is shown below. Figure 4 As shown.
[0273] 8. Constraints on Response Speed and Durability
[0274] The response speed of reactive power regulation is determined by the rate constraint in both mechanical and electrical constraints, and the upper limit of the response speed is shown in equation (37):
[0275]
[0276] in, , , These are the sensitivity coefficients of reactive power to excitation current, guide vane opening, and voltage change rate, respectively.
[0277] The duration of reactive power regulation is determined by the heat accumulation constraint. Satisfying equation (38):
[0278]
[0279] And it satisfies the stator winding temperature rise constraint as shown in equation (39):
[0280]
[0281] in, Stator winding temperature The initial temperature. For stator resistance, For heat dissipation coefficient, For heat capacity, This is the maximum permissible temperature.
[0282] IX. Constraint Characteristics under Different Working Conditions
[0283] The various operating conditions include power generation, pumping, power generation phase regulation, and pumping phase regulation. The reactive power regulation range under different operating conditions is compared as follows: Figure 3 As shown. The boundary parameters of the constraints under different working conditions are set according to the characteristics of the working conditions.
[0284] Under power generation conditions, the power angle stability boundary is the key constraint, and the reactive power regulation range is mainly in the late-phase operation region. At this time, the unit operates as a generator, with positive active power, and the reactive power regulation range is subject to the combined effects of stator current limitation, excitation current limitation, and power angle stability limitation.
[0285] Under pumping conditions, the lower limit of the power factor is a critical constraint, and the reactive power adjustment range is relatively narrow. At this time, the unit operates as a motor, with negative active power, and the reactive power adjustment capability is limited by the rated power factor lower limit set by the equipment manufacturer, rather than by temperature rise limitations.
[0286] Under phase-shifting conditions, active power is close to zero, reactive power regulation range is widest, and excitation current constraint is the main limiting factor. At this time, the unit is disconnected from the grid or operates under no-load conditions, specifically providing reactive power support, and the excitation current can be adjusted within a large range.
[0287] Example 2
[0288] A system for determining the reactive power regulation constraint boundaries of a pumped-storage unit under multiple operating conditions includes:
[0289] The constraint model construction module is used to obtain the mechanical parameters, electrical parameters and operating parameters of the power grid of the pumped storage unit, and construct a constraint model that includes mechanical constraints, electrical constraints and power grid constraints.
[0290] The operating condition constraint establishment module is used to establish mechanical constraints, electrical constraints, and power grid constraints for each operating condition based on the constraint condition model.
[0291] The steady-state reactive power output boundary solution module is used to integrate topological prior features, local constraint features, conflict intensity coefficients and reactive power boundary sequence features based on physical information neural networks. After training, the integrated physical information neural network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints, and the model is solved to obtain the steady-state reactive power output boundary of the pumped storage unit under various operating conditions.
[0292] The constraint switching prediction and parameter tuning module is used to determine the steady-state dominant constraint at the current operating point based on the conflict intensity coefficient, calculate the contraction rate ratio of various constraint boundaries to active power changes, and predict the critical active power value for steady-state dominant constraint switching based on the contraction rate ratio; calculate the inertial time constant of the mechanical system and electrical system, and determine the inertial travel time and control dead zone compensation amount for switching control modes to cope with steady-state dominant constraint switching.
[0293] The pumped-storage unit transient reactive power capacity assessment module is used to introduce response speed constraints and duration constraints, establish a transient reactive power response capacity boundary model, and use the transient reactive power response capacity boundary model to deduce the transient reactive power response capacity boundary of the pumped-storage unit under different operating conditions. It outputs the reactive power output boundary diagram, transient reactive power response capacity boundary, critical active power value, inertial travel time, and control dead zone compensation amount under each operating condition, which are used to guide the reactive power scheduling and control of the pumped-storage unit.
[0294] Example 3
[0295] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.
[0296] This computer device can be a server, and its internal structure diagram can be as follows: Figure 7 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.
[0297] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0298] Example 4
[0299] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0300] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0301] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions, characterized in that, Includes the following steps: Obtain the mechanical and electrical parameters of the pumped-storage unit and the operating parameters of the power grid where it is located, and construct a constraint condition model that includes mechanical constraints, electrical constraints and power grid constraints; Based on the constraint model, mechanical constraints, electrical constraints, and power grid constraints are established for each operating condition. Based on the physical information neural network, topological prior features, local constraint features, conflict intensity coefficients and reactive power boundary sequence features are integrated. After training, the integrated physical information neural network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints. The model is then solved to obtain the steady-state reactive power output boundary of the pumped storage unit under various operating conditions. The steady-state dominant constraint at the current operating point is determined based on the conflict intensity coefficient. The contraction rate ratio of various constraint boundaries to active power changes is calculated, and the critical active power value for the switching of steady-state dominant constraint is predicted based on the contraction rate ratio. The inertial time constants of the mechanical and electrical systems are calculated, and the inertial travel time and control dead zone compensation amount for switching control modes to cope with the switching of steady-state dominant constraint are determined. By introducing response speed constraints and duration constraints, a transient reactive power response capability boundary model is established. This model is then used to deduce the transient reactive power response capability boundary of the pumped-storage unit under different operating conditions. The reactive power output boundary diagram, transient reactive power response capability boundary, critical active power value, inertial travel time, and control dead zone compensation amount for each operating condition are output to guide the reactive power scheduling and control of the pumped-storage unit.
2. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 1, characterized in that, The mechanical constraints include speed constraints, guide vane opening constraints, head constraints, and rotor motion equation constraints. The electrical constraints include stator current constraints, excitation current constraints, power circle diagram constraints, and converter capacity constraints. The grid constraints include voltage constraints, voltage change rate constraints, voltage sensitivity constraints, and transient voltage stability constraints.
3. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 1, characterized in that, The method is based on a physical information neural network, which integrates topological prior features, local constraint features, conflict intensity coefficients, and reactive power boundary sequence features. Through training, the fused physical information neural network converges to obtain a feasible region model of reactive power regulation capability with multiple constraints. Solving this model yields the steady-state reactive power output boundary of the pumped-storage unit under various operating conditions, including: Differential topological analysis of the constrained manifolds formed by three types of constraints on the power-reactive plane is performed using a differentiable Morse-Smale complex layer, and critical points, stable / unstable manifolds and Betti numbers are extracted as topological prior features. The Grey Wolf optimization algorithm is used to automatically optimize network hyperparameters and determine the optimal network hyperparameters. A convolutional neural network is constructed to extract features from the multidimensional physical parameters of three types of constraints, thereby obtaining local constraint features. The interaction conflict relationship among the three types of constraints is explicitly modeled using a multi-head cross-attention mechanism to determine the conflict intensity coefficient; By using a bidirectional long short-term memory network to capture the sequential pattern of reactive power boundary changes with active power, the characteristics of the reactive power boundary sequence are determined. Using the physical information neural network as the base network, topological prior features, local constraint features, conflict intensity coefficients, and reactive boundary sequence features are integrated into the physical information neural network framework to obtain the fused physical information neural network. Traverse the range of active power values, take active power as input and the upper and lower boundaries of reactive power as output, combine the optimal network hyperparameters and fusion loss function, train the fused physical information neural network until the network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints. Based on the feasible domain model of reactive power regulation capability with multi-constraint coupling, the steady-state reactive power output boundary of pumped storage units under various operating conditions is determined.
4. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 3, characterized in that, The method involves performing differential topological analysis on the constrained manifolds formed by three types of constraints on the power-reactive plane using a differentiable Morse-Smale complex layer, extracting critical points, stable / unstable manifolds, and Betti numbers as topological prior features, including: The signed distance from any point on the power-reactive plane to the boundary of the multi-constraint feasible region is taken as the comprehensive constraint function; The gradient field of the synthetic constraint function on the discrete grid is calculated using a numerical method, and the critical points are divided into local maxima, local minima and saddle points according to the eigenvalue signs of the Hessian matrix. Starting from each saddle point, the gradient streamlines are traced forward and backward along the eigenvector direction of the Hessian matrix to obtain the stable and unstable manifolds, and the critical point is determined based on the intersection of the stable and unstable manifolds. After completing the manifold partitioning, the power-reactive plane is decomposed into several cavities, and all points in each cavity have the same gradient flow towards the endpoint. The connectivity of the cavities is counted to determine the number of connected components and the number of voids.
5. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 3, characterized in that, The fusion loss function is composed of three weighted components: data fitting loss, physical constraint loss, and topology consistency loss. The data fitting loss uses mean square error to measure the deviation between the reactive boundary of the network output and the labels of the training samples. The physical constraint loss is composed of the violation penalty terms of the three types of constraints weighted by the conflict intensity coefficient. The topology consistency loss is composed of the Betti number prediction bias and the critical point position bias.
6. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 5, characterized in that, The expression for the fusion loss function is: ; ; ; In the formula, For the loss of fusion, For data fitting loss, For physical constraint loss, For topology consistency loss, The weights for the physical constraint loss, The weights for the topology consistency loss are... For mechanical constraint loss, For electrical constraint loss, For grid constraint losses, The conflict intensity coefficient is the normalized mechanical constraint loss. The conflict intensity coefficient is the normalized electrical constraint loss. The conflict intensity coefficient is the normalized power grid constraint loss. , These represent the number of connected components in the feasible region predicted by the network and the actual value calculated by the differentiable Morse-Smale complex layer, respectively. , The numbers are the number of holes predicted by the network and the actual value calculated by the Morse-Smale complex layer, respectively. , These represent the critical point locations output by the network and the true critical point locations extracted by the differentiable Morse-Smale complex layer, respectively. , , These are the weighting coefficients for the deviation of the number of connected components, the deviation of the number of holes, and the deviation of the critical point location, respectively. The total number of critical points. For critical point index, Let represent the square of the Euclidean distance.
7. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 1, characterized in that, The multiple operating conditions include power generation, pumping, power generation phase modulation, and pumping phase modulation. Under different operating conditions, the reactive power allowable ranges defined by the three types of constraints are different, causing the dominant steady-state constraint to switch with the operating conditions, specifically manifested as follows: Under power generation conditions, the conflict intensity coefficient of electrical constraints is greater than that of mechanical constraints and grid constraints. The power angle stability boundary in electrical constraints is the dominant steady-state constraint, and the steady-state reactive power output boundary is determined by the intersection of electrical constraints and grid constraints. Under pumping conditions, the conflict intensity coefficient of mechanical constraints is greater than that of electrical constraints and grid constraints. The lower limit of the power factor in mechanical constraints is the dominant constraint in steady state, and the boundary of steady-state reactive power output is determined by the intersection of mechanical constraints and electrical constraints. Under phase-shifting conditions, the excitation current sub-item in the electrical constraints has the highest weight. The excitation current constraint in the electrical constraints is the steady-state dominant constraint, and the steady-state reactive power output boundary is determined by the intersection of the electrical constraints and the grid constraints. The transient reactive power response boundary is dominated by the electrical constraints.
8. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 1, characterized in that, The calculation of the contraction rate ratio of various constraint boundaries to changes in active power, and the prediction of the critical active power value for steady-state dominant constraint switching based on the contraction rate ratio, includes: Based on the reactive boundary value corresponding to each constraint, determine the contraction rate of the constraint boundary and calculate the ratio of the contraction rates of any two types of constraints. When the shrinkage rate ratio changes from greater than 1 to less than 1, it indicates that the limiting effect of the two types of constraints on the reactive power boundary has reversed. The corresponding critical active power value at this time is the predicted critical point for the switching of the steady-state dominant constraint.
9. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 8, characterized in that, The inertial time constants of the computational mechanical and electrical systems are determined to be the inertial travel time and control dead zone compensation amount for switching control modes in response to the switching of steady-state dominant constraints, including: The inertial time constant of the mechanical system is calculated based on the unit's inertial constant and synchronous angular velocity, and the inertial time constant of the electrical system is calculated based on the excitation winding inductance and excitation winding resistance. Based on the inertial time constants of mechanical and electrical systems, and combined with communication delay, the inertial travel time for control mode switching is determined. The control dead zone compensation amount is calculated based on the real-time monitored rate of change of active power, and the control dead zone interval is determined by the control dead zone compensation amount. When the active power is detected to be close to the predicted critical point for the switching of the steady-state dominant constraint and enter the control dead zone, a control mode switching command is issued to offset the response delay caused by mechanical / electrical inertia.
10. The method for determining the reactive power regulation constraint boundary of a pumped-storage unit under multiple operating conditions according to claim 1, characterized in that, The introduction of response speed constraints and duration constraints establishes a transient reactive power response capability boundary model, which is then used to deduce the transient reactive power response capability boundary of the pumped-storage unit under different operating conditions. The model outputs reactive power output boundary diagrams, transient reactive power response capability boundaries, critical active power values, inertial travel time, and control dead zone compensation for each operating condition, guiding the reactive power scheduling and control of the pumped-storage unit, including: Based on the guide vane opening rate constraint in the mechanical constraints and the excitation current rate constraint in the electrical constraints, the upper limit of the response speed of reactive power regulation is determined. The upper limit of the response speed is determined by the slowest of the excitation current change rate, the guide vane opening rate, and the voltage change rate. Based on the upper limit of the response speed of reactive power regulation, the reactive power ramp-up amount at each moment during the transient process is calculated, and the reactive power ramp-up amount is the integral of the upper limit of the response speed over time, representing the total amount of reactive power that the unit can generate from the moment the fault occurred to the current moment. Based on the thermal accumulation constraints of the stator winding and the excitation winding, the duration of reactive power regulation is determined. Under the thermal accumulation constraint, the reactive power decreases in an S-shaped curve as the duration increases, and the upper limit of reactive power limited by thermal accumulation at each moment is determined. The reactive ramp rate is superimposed on the steady-state basic reactive value and compared with the upper limit of reactive capacity of thermal accumulation limit and the transient stability limit value. The minimum value among the three is taken as the boundary of transient reactive response capability at the current moment. Based on the transient reactive response capability boundary, the reactive response process at different time scales is simulated to obtain the time-reactive power curve and determine the transient reactive response capability boundary value at each moment in the transient process. The steady-state reactive power output boundary under power generation, pumping and phase adjustment conditions is output as a reactive power output boundary map in the form of a feasible region on the power-reactive power plane, and the feasible region is composed of the intersection of three types of constraints: mechanical constraints, electrical constraints and grid constraints. The transient reactive power response capability boundary is output in the form of a time-reactive power curve. Based on the reactive power output boundary diagram, inertial travel time, control dead zone compensation, critical active power value, and time-reactive power curve of the transient reactive power response capability boundary parameters under various operating conditions, the reactive power scheduling and control of pumped storage units are guided.
11. A system for determining the constraint boundary of reactive power regulation under multiple operating conditions of a pumped-storage unit, used to implement the steps of the method for determining the constraint boundary of reactive power regulation under multiple operating conditions of a pumped-storage unit as described in any one of claims 1-10, characterized in that, include: The constraint model construction module is used to obtain the mechanical parameters, electrical parameters and operating parameters of the power grid of the pumped storage unit, and construct a constraint model that includes mechanical constraints, electrical constraints and power grid constraints. The operating condition constraint establishment module is used to establish mechanical constraints, electrical constraints, and power grid constraints for each operating condition based on the constraint condition model. The steady-state reactive power output boundary solution module is used to integrate topological prior features, local constraint features, conflict intensity coefficients and reactive power boundary sequence features based on physical information neural networks. After training, the integrated physical information neural network converges to obtain a feasible domain model of reactive power regulation capability with multiple constraints, and the model is solved to obtain the steady-state reactive power output boundary of the pumped storage unit under various operating conditions. The constraint switching prediction and parameter tuning module is used to determine the steady-state dominant constraint at the current operating point based on the conflict intensity coefficient, calculate the contraction rate ratio of various constraint boundaries to active power changes, and predict the critical active power value for steady-state dominant constraint switching based on the contraction rate ratio; calculate the inertial time constant of the mechanical system and electrical system, and determine the inertial travel time and control dead zone compensation amount for switching control modes to cope with steady-state dominant constraint switching. The pumped-storage unit transient reactive power capacity assessment module is used to introduce response speed constraints and duration constraints, establish a transient reactive power response capacity boundary model, and use the transient reactive power response capacity boundary model to deduce the transient reactive power response capacity boundary of the pumped-storage unit under different operating conditions. It outputs the reactive power output boundary diagram, transient reactive power response capacity boundary, critical active power value, inertial travel time, and control dead zone compensation amount under each operating condition, which are used to guide the reactive power scheduling and control of the pumped-storage unit.
12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-10.
13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-10.
