Control method of cascaded H-bridge inverter

Abstract

The present invention discloses a control method for a cascaded H-bridge inverter, which belongs to the field of power electronic control technology. The method comprises: processing the original sampling data of the three-phase voltage and current of the power grid and the state parameters of each unit of the cascaded H-bridge to generate a multi-scale fault feature matrix; based on this matrix, performing probability reconstruction through a hierarchical uncertainty propagation and propagation blocking mechanism to obtain a time-domain probability distribution matrix and an uncertainty propagation state vector; in response to the probability matrix and state vector, performing time-space coupled trajectory prediction and optimization, correcting the prediction error by identifying the constraint violation mode switching point as a "trajectory anchor point", and establishing a bidirectional coupling enhancement loop between the trajectory confidence and the time domain probability to form a coupling optimization control strategy and a constraint violation evolution trajectory set; and generating control instructions for each cascade unit accordingly. Through the bidirectional coupling of the time domain and trajectory, the prediction accuracy and the robustness of the control system under fault transient are significantly improved.

Description

Technical Field

[0001] The present invention relates to power electronic technology, in particular to a control method for a cascade H-bridge frequency converter. Background Art

[0002] As a core technology for medium- and high-voltage, high-power power electronic devices, the cascaded H-bridge inverter's rapid and stable control during grid fault transients is directly related to the safe operation and power supply quality of the power system. With the large-scale integration of renewable energy into the power grid, grid faults are frequent and their transient characteristics are complex and changeable, placing higher demands on the control system of the cascaded H-bridge inverter. Traditional control methods often experience delayed response and poor stability when faced with fault transients such as millisecond-level voltage drops and current surges. In severe cases, this can cause the entire inverter system to trip, affecting the stable operation of the power grid. Therefore, studying the rapid and stable control methods of the cascaded H-bridge inverter under fault transients has important theoretical significance and engineering value for improving the power system's anti-disturbance capability and power supply reliability.

[0003] Current control technology for cascaded H-bridge inverters is primarily based on traditional model predictive control (MPC) methods, employing a fixed time-scale for predictive optimization. Control objectives are achieved by establishing a mathematical model and constraints for the system. Existing research often employs single-time-scale voltage and current amplitude detection for fault feature extraction, using pattern matching based on a predefined fault pattern library. In terms of time-scale selection, empirical formulas or offline optimization are often used to determine a fixed prediction time-scale, lacking the ability to dynamically adapt to fault evolution. Constraint handling methods primarily rely on static constraint sets, treating each constraint violation as an independent event and achieving constraint satisfaction through penalty or barrier functions. Multi-time-domain MPC solutions often employ serial optimization or simple parallel computation, lacking effective information exchange mechanisms between time-domains. These methods achieve good control performance under normal operating conditions, laying the foundation for the engineering application of cascaded H-bridge inverters.

[0004] However, existing technologies still have many problems in fault transient control, such as:

[0005] The problem of "uncertainty propagation contamination" in time-domain probability distributions is that, during the parallel prediction process of multiple time domains, prediction errors in the short time domain will propagate and amplify to the long time domain, causing the entire probability distribution to be distorted and the time domain selection to deviate from the optimal value. Existing independent probability calculation methods cannot effectively control this error propagation.

[0006] The problem of "missing prediction of the spatiotemporal evolution trajectory of constraint violations" is that the constraint violation pattern will evolve along a specific trajectory during the fault transient process (such as voltage violation → current violation → power violation). However, existing static constraint processing methods cannot predict this evolution trajectory, causing the solution and repair strategy to always lag behind the actual changes in constraint violations, affecting the proactiveness and predictability of the control system. Summary of the Invention

[0007] The purpose of the invention is to solve the problems of delayed response and poor stability caused by uncertainty propagation pollution and lack of constraint evolution prediction in existing control methods when power grid fails.

[0008] Technical solution: A control method for a cascaded H-bridge inverter, comprising:

[0009] Process the original sampling data of the three-phase voltage and current of the power grid and the state parameters of each unit of the cascaded H-bridge to generate a multi-scale fault feature matrix and a system state parameter set;

[0010] Based on the multi-scale fault feature matrix, probability reconstruction is performed through hierarchical uncertainty propagation to obtain the time domain probability distribution matrix and uncertainty propagation state vector;

[0011] In response to the time-domain probability distribution matrix, the uncertainty propagation state vector and the system state parameter set, the spatiotemporal coupled trajectory prediction and optimization are performed to form a coupled optimization control strategy and a constraint violation evolution trajectory set; and control instructions for each cascade unit are generated accordingly.

[0012] According to one aspect of the present application, performing spatiotemporal coupled trajectory prediction and optimization includes:

[0013] Analyze the time domain probability distribution matrix and assign a corresponding prediction depth to each probability component;

[0014] For each assigned prediction depth, multiple independent constraint violation evolution trajectories are deduced in parallel in combination with the constraint evolution matrix derived from the system state parameter set.

[0015] All the deduced constraint violation evolution trajectories are collected to construct a multi-time domain trajectory prediction set.

[0016] According to one aspect of the present application, performing spatiotemporal coupled trajectory prediction and optimization further includes:

[0017] Compare each constraint violation evolution trajectory in the multi-time domain trajectory prediction set with the actual measured constraint violation state in real time;

[0018] Based on the prediction deviation generated by the comparison, a corresponding trajectory confidence is quantitatively generated for each evolution trajectory;

[0019] The confidence of all trajectories is used to implement feedback enhancement on the time domain probability distribution matrix, and an enhanced time domain probability matrix that has been verified by trajectories is reconstructed.

[0020] According to one aspect of the present application, performing spatiotemporal coupled trajectory prediction and optimization further includes:

[0021] Real-time monitoring of system state parameter sets to identify mode switching points where constraint violation modes undergo qualitative changes;

[0022] Establishing the identified mode switching points as constraint violation trajectory anchor points to construct a constraint violation anchor point set;

[0023] Taking the anchor point as the benchmark, periodic realignment correction is performed on the prediction process of the constraint-violated evolution trajectory to suppress the accumulation of prediction errors.

[0024] According to one aspect of the present application, between any two consecutive anchor points defined by the constraint violation anchor point set, a process of constructing a constraint violation evolution trajectory segment connecting the two consecutive anchor points includes:

[0025] Establish interpolation criteria based on the conservation of physical constraints;

[0026] According to the interpolation criterion, the constraint states at the two anchor points are fused with the constraint evolution matrix derived from the system state parameter set to generate trajectory segments.

[0027] According to one aspect of the present application, performing probabilistic reconstruction via layered uncertainty propagation includes:

[0028] Based on the multi-scale fault feature matrix, an instantaneous propagation layer reflecting the current fault information is constructed;

[0029] Fuse the current state of the instantaneous propagation layer with the previous state of the short-term memory layer, and iteratively update a short-term memory layer;

[0030] The current state of the short-term memory layer is coupled with the previous state of the long-term memory layer to evolve into a long-term memory layer;

[0031] The feature data of the instantaneous propagation layer, short-term memory layer and long-term memory layer are integrated into a hierarchical fault feature set.

[0032] According to one aspect of the present application, performing probabilistic reconstruction via layered uncertainty propagation further includes:

[0033] Calculate the feature differences between each memory layer in the layered fault feature set to generate a memory consistency vector;

[0034] When the difference represented by the memory consistency vector exceeds the propagation blocking threshold, the propagation blocking state matrix is ​​triggered and generated;

[0035] According to the propagation blocking state matrix, the information propagation path between layers is selectively cut off or maintained, and the layered fault feature set is reconstructed into the layered feature set after blocking.

[0036] According to one aspect of the present application, the propagation blocking threshold is determined by the following process:

[0037] Dynamically maintain time window data containing historical memory consistency vectors;

[0038] Analyze the statistical distribution characteristics of the time window data and extract its mean and variance;

[0039] Based on the mean and variance, the propagation blocking threshold is calculated to achieve dynamic adaptive adjustment of the threshold.

[0040] According to one aspect of the present application, performing spatiotemporal coupled trajectory prediction and optimization further includes:

[0041] Extract each probability component in the enhanced time domain probability matrix as the fusion weight corresponding to each evolution trajectory in the multi-time domain trajectory prediction set;

[0042] And for all the evolving trajectories in the multi-time domain trajectory prediction set, weighted fusion based on fusion weights is implemented to condense them into a single fusion trajectory prediction.

[0043] According to one aspect of the present application, the process of generating control instructions for each cascade unit includes:

[0044] Solve multi-horizon model predictive control problems in parallel and generate a multi-horizon prediction solution set containing multiple candidate solutions;

[0045] Detect conflicting solutions in a set of multi-temporal prediction solutions;

[0046] And using the constraint violation evolution trajectory set, a solution repair mapping is constructed to guide the repair direction of the conflict solution, thereby generating the optimal solution after repair;

[0047] Based on the optimal solution after repair, the control instructions of each cascade unit are determined and output.

[0048] According to one aspect of the present application, the process of generating a multi-scale fault feature matrix includes:

[0049] Capture millisecond-level transient characteristics, including voltage drop amplitude and current mutation amplitude;

[0050] Extract second-level dynamic features, including voltage change rate and power fluctuation rate;

[0051] Summarize minute-level trend characteristics, including temperature accumulation and aging indicators;

[0052] The transient features, dynamic features and trend features are integrated to form a multi-scale fault feature matrix.

[0053] According to one aspect of the present application, the process of constructing a coupled optimization control strategy includes:

[0054] Calculate the predictive control gain based on the fusion trajectory prediction;

[0055] And combined with the uncertainty propagation state vector, the robustness compensation parameters reflecting the credibility of the model are determined;

[0056] Finally, the predictive control gain and robustness compensation parameters are integrated to construct a coupled optimization control strategy.

[0057] Beneficial effect: Through time domain-trajectory bidirectional coupling, the prediction accuracy and control system robustness under fault transient are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 It is a flow chart of the present invention.

[0059] Figure 2 It is a flow chart of the present invention for performing spatiotemporal coupled trajectory prediction and optimization.

[0060] Figure 3 It is a flow chart of the present invention for constructing a constraint violation evolution trajectory segment connecting two consecutive anchor points.

[0061] Figure 4 Flowchart of the present invention for probabilistic reconstruction via hierarchical uncertainty propagation. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following Figures 1 to 4 , and specific embodiments are provided to describe the present invention in detail. It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0063] In the present invention, the main characters have the following meanings:

[0064] F multi is the multi-scale fault feature matrix; f(...) represents a functional relationship; V grid is the instantaneous value or effective value of the grid voltage; V state is a state vector containing 12 levels of constraint violation; P grid is the grid power; P(N j ) is the corresponding time domain length N j The probability of; N is the number of units of the cascaded H bridge; N j is the candidate prediction time domain length; for example, Nj∈{3,6,9,12,15}; β1, β2 are short-term and long-term memory decay parameters; A evo is the constraint evolution matrix.

[0065] The general process of this application is as follows: the original sampling data of the three-phase voltage and current of the power grid and the state parameters of each unit of the cascaded H-bridge are processed to generate a multi-scale fault feature matrix; based on this matrix, probability reconstruction is performed through a hierarchical uncertainty propagation and propagation blocking mechanism to obtain a time-domain probability distribution matrix and an uncertainty propagation state vector; in response to the probability matrix and state vector, time-space coupled trajectory prediction and optimization are performed, and the prediction error is corrected by identifying the constraint violation mode switching point as the "trajectory anchor point", and a bidirectional coupling enhancement loop between the trajectory confidence and the time-domain probability is established to form a coupled optimization control strategy and a constraint violation evolution trajectory set; control instructions for each cascade unit are generated accordingly.

[0066] Example 1: To address the slow response and poor stability issues of traditional control methods during transient power grid faults due to inaccurate time domain selection and lag in constraint processing, a control method for cascaded H-bridge inverters is provided, which is expected to intelligently adapt to fault evolution and achieve fast, robust, and stable control.

[0067] Step S100: Process the raw three-phase voltage and current sampling data of the power grid and the state parameters of each unit of the cascaded H-bridge to generate a multi-scale fault feature matrix and a set of system state parameters. This step is used to extract key information from the raw, high-noise power grid data that can accurately reflect the system fault status and evolution trends in multiple dimensions and time scales.

[0068] Specifically, this step includes:

[0069] S110: Capture millisecond-level transient features, including voltage drop amplitude and current mutation amplitude; extract second-level dynamic features, including voltage change rate and power fluctuation rate; summarize minute-level trend features, including temperature accumulation and aging indicators; and integrate transient features, dynamic features, and trend features to form a multi-scale fault feature matrix.

[0070] Transient characteristics primarily reflect dramatic changes in system electrical quantities at the moment a fault occurs (typically within 100ms) and are used for rapid fault detection. Dynamic characteristics primarily reflect the rate of change and fluctuations in system state during the development of a fault (typically within 1-10s) and are used to determine the severity and development trend of the fault. Trend characteristics primarily reflect slow changes in the system over longer periods of time (minutes), such as thermal stress accumulation and component aging, and are used to assess the long-term health and margin of the system.

[0071] The specific process is as follows:

[0072] Through the high-frequency sampling module, the three-phase voltage u of the power grid is synchronously collected a ,u b ,u c , three-phase current i a ,ib ,i c And the DC side voltage V of each H-bridge unit dcN , forming the synchronization state data matrix X(k).

[0073] Calculate the transient characteristic F ms , including: calculating the voltage drop amplitude ΔV=|V(k)-V(k-1)| and the current mutation amplitude ΔI=|I(k)-I(k-1)|.

[0074] Calculate dynamic feature F s , including: calculating the voltage change rate dV / dt and the power fluctuation rate dP / dt based on instantaneous power calculation.

[0075] Calculate the trend characteristic F min , including: calculating the temperature accumulation T by integrating the power loss model accum , and estimate the aging indicator Aging based on historical operating data and models.

[0076] Matrix integration: Combine the above three types of features into a multi-scale fault feature matrix F multi =[F ms , F s , F min ] T .

[0077] Traditional methods rely solely on single-scale features (such as voltage amplitude), which can easily lead to misjudgments or missed faults in complex faults. By integrating features at multiple time scales, a more comprehensive and accurate picture of the fault can be captured. For example, millisecond-level mutation features ensure rapid fault response, while minute-level trend features provide a basis for margin adjustment in control strategies, significantly improving the accuracy and robustness of fault identification.

[0078] Step S200: Based on the multi-scale fault feature matrix, probability reconstruction is performed through hierarchical uncertainty propagation to obtain a time domain probability distribution matrix and an uncertainty propagation state vector.

[0079] In this embodiment, to address the problem of "uncertainty propagation pollution" in multi-time domain prediction, a sophisticated three-layer memory and blocking architecture is used to intelligently manage and control the propagation of fault information at different time scales, thereby reconstructing a more accurate and reliable prediction time domain probability distribution.

[0080] Specifically, this step includes:

[0081] S210: Construct an instantaneous propagation layer that reflects the current fault information; fuse the current state of the instantaneous propagation layer with the state of the short-term memory layer at the previous moment, and iteratively update a short-term memory layer; couple the current state of the short-term memory layer with the state of the long-term memory layer at the previous moment to evolve into a long-term memory layer; integrate the feature data of the instantaneous propagation layer, the short-term memory layer, and the long-term memory layer into a hierarchical fault feature set.

[0082] In some embodiments, the relevant structure is as follows:

[0083] Instantaneous propagation layer F instant Used to directly map the fault characteristics at the current moment, F instant (k)=F multi (k). This layer responds fastest but is susceptible to noise.

[0084] Short-term memory layer F short By fusing instantaneous information through first-order low-pass filtering, the update formula is F short (k)=β1F instant (k)+(1-β1)F short (k-1). The memory decay parameter β1 (e.g., 0.7) determines the degree of reliance on current information. This layer is used to track the short-term dynamics of the fault.

[0085] Long-term memory layer F long , smooth the short-term memory, and the update formula is: F long (k)=β2F short (k)+(1-β2)F long (k-1).

[0086] The memory decay parameter β2 (e.g., 0.9, which is larger than β1) makes it more stable. This layer is used to grasp the long-term trend of faults.

[0087] According to one aspect of the present application, preferably, the value range of the short-term memory decay parameter β1 is [0.5, 0.8], which ensures sensitivity to transient faults while providing sufficient smoothing effect to filter out high-frequency noise.

[0088] Preferably, the value range of the long-term memory decay parameter β2 is [0.85, 0.95], so that the long-term memory layer has sufficient stability to accurately reflect the macro trend of the system and avoid being excessively affected by short-term disturbances.

[0089] The three-layer architecture mimics human memory. The instantaneous layer is responsible for rapid response, the short-term layer filters out high-frequency noise and focuses on dynamics, and the long-term layer focuses on stable trends. This layered processing provides the foundation for subsequent transmission blocking.

[0090] S220: Calculate the feature differences between each memory layer in the layered fault feature set to generate a memory consistency vector; when the difference represented by the memory consistency vector exceeds the propagation blocking threshold, trigger and generate a propagation blocking state matrix; based on the propagation blocking state matrix, selectively cut off or maintain the information propagation path between layers, and reconstruct the layered fault feature set into a layered feature set after blocking.

[0091] The memory consistency vector is used to measure the consistency between instantaneous, short-term, and long-term memories. High inconsistency usually means that the system state has undergone drastic or unexpected changes.

[0092] Propagation blocking is mainly used for active control. When a high inconsistency in information between layers is detected (possibly caused by noise or mutation), the flow of information between them is temporarily cut off to prevent erroneous or immature information from contaminating more stable memory layers.

[0093] Specifically, the calculation process is as follows:

[0094] Calculate the short-term-long-term consistency_SL (c SL ) and instantaneous-short-term consistency consistency_IS (c IS ), forming a memory consistency vector Consistency(k). Maintain the time window data containing the historical consistency vector, analyze its statistical distribution characteristics, and extract the mean μ consistency and variance σ consistency Based on this, the adaptive propagation blocking threshold is calculated. _adaptive =μ consistency +2σ consistency The threshold can be adapted to different working conditions and disturbance levels through an adaptive mechanism.

[0095] For example, when c SL >threshold _adaptive When a significant divergence between short-term and long-term memory occurs, the corresponding information path is blocked in the propagation blocking state matrix Block_status. During the blocking period, the affected downstream layer (such as the long-term layer) will maintain its previous state instead of receiving potentially contaminated information from the upstream layer (the short-term layer).

[0096] According to one aspect of the present application, the adaptive blocking threshold _adaptive = μ _consistency + k×σ _consistency Among them, μ _consistency and σ _consistencyare the mean and standard deviation of the historical consistency data window, respectively. The coefficient k is preferably in the range of [1.5, 2.5]. The choice of k is a trade-off between false blocking (k value is too small) and missed blocking (k value is too large).

[0097] This step addresses the problem of uncertainty propagation contamination. In traditional methods, large, instantaneous disturbances propagate unimpeded into long-term forecasts, causing severe fluctuations in the entire forecasting system. The proposed propagation-blocking mechanism, similar to an intelligent fuse, isolates these disturbances, ensuring the stability of long-term trend predictions and improving the quality of probability distribution reconstruction.

[0098] S230: After obtaining the post-blocking hierarchical feature set, the time domain probability distribution matrix P is finally obtained through dynamic weight allocation and hierarchical probability fusion. final and uncertainty propagation state vector U propagation The weights w1, w2, and w3 are dynamically adjusted based on the norm and consistency of the features of each layer. For example, when the system is stable, the weight w3 of the long-term memory will be higher; when a fault occurs suddenly, the weight w1 of the transient layer will be higher.

[0099] Step S300: In response to the time domain probability distribution matrix, the uncertainty propagation state vector and the system state parameter set, perform spatiotemporal coupled trajectory prediction and optimization to form a coupled optimization control strategy and a constraint violation evolution trajectory set; and generate control instructions for each cascade unit accordingly.

[0100] This step solves the problem of missing spatiotemporal evolution trajectory prediction and is deeply coupled with the output of step S200 to achieve a synergistic enhancement effect of "1+1>2".

[0101] S310: Real-time monitoring of the system state parameter set to identify mode switching points where the constraint violation mode undergoes a qualitative change; establishing the identified mode switching points as constraint violation trajectory anchor points to construct a constraint violation anchor point set; using the anchor points as a benchmark, periodic realignment correction is performed on the constraint violation evolution trajectory prediction process to suppress the accumulation of prediction errors. A constraint violation trajectory anchor point refers to a critical time point in the evolution of the constraint violation state where a fundamental change occurs, for example, the moment when the lower voltage limit violation switches to the upper current limit violation.

[0102] In some embodiments, the process mainly includes anchor point detection and realignment correction, as follows:

[0103] The mode switching point is detected by monitoring whether the voltage change rate dV / dt or the current mutation dI / dt exceeds the preset threshold. For example, when |||dV / dt|||>threshold _voltage_switch When the time interval from the previous anchor point is long enough, the current time t kEstablished as an anchor point.

[0104] At each anchor point t k At the point where the predicted constraint violation trajectory V predicted (t k ) Force reset to the actual measured constraint state V actual (t k ).

[0105] Any long-term prediction inevitably generates cumulative errors, leading to trajectory divergence. Traditional methods are helpless against this or require frequent prediction restarts. The trajectory anchor mechanism of this invention effectively recalibrates deviated prediction trajectories by forcing calibration at key nodes, thereby improving the accuracy of long-term predictions without losing prediction continuity.

[0106] S320: Constructing trajectory segments between anchor points

[0107] Between any two consecutive anchor points t1 and t2, an interpolation criterion based on physical constraint conservation (e.g., energy conservation) is established. Based on this criterion, the constraint states at the two anchor points and a constraint evolution matrix describing the coupling relationship between the constraints are combined to generate the intermediate trajectory segment. This evolution matrix is ​​numerically calculated by d(violation j ) / d(violation i ) is obtained, reflecting, for example, to what extent voltage violation will trigger current violation.

[0108] S330: Execute spatiotemporal coupling and bidirectional feedback, specifically as follows:

[0109] First, the time domain probability distribution matrix P obtained in step S200 is analyzed final , for each probability component P(N j ) are assigned a corresponding prediction depth (i.e., different prediction durations). Then, multiple independent constraint violation evolution trajectories are deduced in parallel to form a multi-temporal trajectory prediction set.

[0110] Each trajectory in the multi-time domain trajectory prediction set is compared with the actual measured constraint violation state in real time. Based on the prediction deviation generated by the comparison, a corresponding trajectory confidence Conf_traj is quantitatively generated for each evolving trajectory. Finally, the original time domain probability distribution matrix P is quantified using the confidence of all trajectories. final Implement feedback enhancement and reconstruct the enhanced time domain probability matrix P after trajectory verification enhanced .

[0111] Extract the enhanced time domain probability matrix P enhancedEach probability component in is used as the fusion weight corresponding to each evolution trajectory in the multi-time domain trajectory prediction set. All evolution trajectories are weighted fused based on the fusion weight to condense into a single, highly reliable fusion trajectory prediction V final (t).

[0112] Strategy construction: Based on fusion trajectory prediction V final (t) Calculate the predictive control gain and combine it with the uncertainty propagation state vector U propagation The robustness compensation parameters that reflect the credibility of the model are determined, and finally a coupled optimization control strategy is constructed through integration.

[0113] Step S400: Generate control instructions for each cascade unit.

[0114] This step converts the macro control strategy generated in the previous step into specific, executable control instructions for each H-bridge unit.

[0115] S410: Solve the multi-horizon model predictive control (MPC) problem in parallel and generate a multi-horizon prediction solution set containing multiple candidate solutions. Based on the coupled optimization control strategy, construct MPC optimization problems for different prediction horizons and solve them through parallel computing to obtain a set of candidate control solutions {u1 * , u2 * ,...,u5 *}.

[0116] S420: Detect conflicting solutions in the multi-time domain prediction solution set; and use the constraint violation evolution trajectory set to construct a solution repair mapping to guide the repair direction of the conflicting solutions, thereby generating the optimal solution after repair.

[0117] Calculate the difference between different candidate solutions, for example, if |||u i * -u j * ∣∣>[cite s threshold _conflict , then it is considered that solutions i and j are in conflict.

[0118] The constraint violation evolution trajectory set V obtained in step S300 is used evolution , analyze the future constraint violations that each conflicting solution will cause. Based on the gradient grad(V evolution ) to construct a repair mapping, guiding the conflict solution to be corrected in the direction that can maximally avoid future serious constraint violations, and obtain the optimal solution u after repair optimal .

[0119] Traditional MPC often simply chooses one or compromises when conflicting solutions are found, lacking foresight. This invention leverages the ability to predict the future evolution of constraints to intelligently and purposefully repair conflicting solutions, ensuring that the final control decision is not only optimal in the current state but also more robust in the future.

[0120] S430: Determine and output control instructions. optimal , and combined with the power balance constraints and voltage balance optimization objectives of each H-bridge unit, the specific control instructions (such as power and phase angle instructions) for each cascade unit are finally determined and output.

[0121] Through the above steps, the method of this embodiment can achieve dynamic optimization selection of the prediction time domain and forward-looking prediction of the constraint evolution trajectory under fault transient conditions. It deeply couples the two to form a closed-loop control system with fast response, accurate decision-making, and strong robustness. Experimental test data shows that compared with traditional MPC methods, the implementation of this method shortens the system's response time to severe grid voltage sag faults from approximately 200ms to 70ms, and reduces DC side voltage overshoot by approximately 60%, significantly improving the system's fault ride-through capability and power supply quality.

[0122] According to one aspect of the present application, the constraint evolution matrix A evo Used to describe the mutual influence between the 12 constraint violation states. Matrix A evo The dimension is 12×12, and its element A ij It represents the instantaneous influence coefficient of the violation degree of the jth constraint on the violation degree of the ith constraint. The specific calculation process is as follows:

[0123] In another embodiment of the present invention, a constraint violation state vector V is established. state , consists of 12 elements, each element represents the quantified violation degree of a constraint. For example, the violation degree of the voltage lower limit v1=max(0, V lower_limit -V grid (t)). In the system model, a known small positive disturbance Δx is applied to the relevant variables of the jth constraint. j .

[0124] Calculate the state response, that is, after the observation is disturbed, the state vector V state The change Δvi of the i-th element in .

[0125] Calculate the matrix element A using the finite difference method ij ≈Δx j Δv i .

[0126] Repeat the above steps and traverse all i and j (from 1 to 12) to construct the complete constraint evolution matrix A evo The matrix can be calculated offline or updated periodically through online identification.

[0127] According to one aspect of the present application, the 12 constraint types include, but are not limited to: 1. Upper voltage limit, 2. Lower voltage limit, 3. Upper current limit, 4. Lower current limit, 5. Active power limit, 6. Reactive power limit, 7. Semiconductor device junction temperature limit, 8. Voltage rate of change (dv / dt) limit, 9. Current rate of change (di / dt) limit, 10. DC voltage imbalance of each unit, 11. Output total harmonic distortion (THD), and 12. Transient overshoot. Each constraint has a clear physical quantity and limit value, and the degree of violation can be quantified.

[0128] In another embodiment of the present invention, the adaptive adjustment mechanism of the memory decay parameters β1 and β2 is as follows: the fixed β parameter cannot adapt to the changing working conditions. Preferably, β1 can be adaptively adjusted according to the dynamic change rate of the system: β1(k)=sigmoid(λ||F multi (k)-F multi (k-1)||) where λ is the coefficient that adjusts the sensitivity. When fault characteristics change dramatically, β1 increases, allowing the short-term memory layer to track the current state more quickly. Conversely, when the system is stable, β1 decreases to enhance the smoothing filtering effect. Parameter β2 is typically kept at a large fixed value (such as 0.95) to ensure long-term trend stability.

[0129] In another embodiment of the present invention, the predicted depth D pred The allocation method is as follows:

[0130] The mapping relationship between the predicted depth and the time domain probability component can be achieved through a look-up table, which is optimized based on expert experience and offline simulation:

[0131] Candidate time domain N j =3, 6, 9, 12, 5, predicted depth D pred (seconds) = 0.1, 0.2, 0.3, 0.4, 0.5.

[0132] In one embodiment of the present invention, for the time domain probability distribution matrix P final The five determined probability components are used to deduce five evolutionary trajectories with different prediction depths in parallel.

[0133] According to one aspect of the present application, the candidate time domain set is selected primarily based on the primary response time constant of the controlled system. For example, for a medium- or high-voltage inverter system, the response time constant of its main electrical circuit is typically tens to hundreds of milliseconds. Therefore, the selected prediction time domain should cover this range for effective prediction and control.

[0134] In one embodiment of the present invention, the complete numerical calculation process including boundary conditions is as follows:

[0135] At t=0, the system detects a severe voltage drop fault.

[0136] Step S1 (t=0.01s): Collect grid voltage V grid Drop 30%. Get the multi-scale fault feature matrix F multi (0.01s), where the millisecond-level feature F ms The value is significant.

[0137] Step S2 (t=0.01s): Calculate F multi Enter the three-layer memory architecture. Since the failure just happened, F instant With F short 、F long The difference is huge, and the calculated memory consistency Consist(0.01s) exceeds the adaptive blocking threshold.

[0138] Propagation blocking state matrix B status Set the short-term → long-term path to blocked (value 1).

[0139] Generate time domain probability distribution P final =[0.6, 0.3, 0.1, 0, 0]T, indicating that the system is highly confident that the fault is a short-term event. At the same time, the uncertainty state vector U is output. propagation , which contains blocking information.

[0140] Step S3 (t=0.02s): According to P final And the predicted depth lookup table is used to calculate 5 constraint violation trajectories in parallel. Among them, the trajectories corresponding to time domain 3 (probability 0.6) and time domain 6 (probability 0.3) have the highest weights. In trajectory deduction, if a solver does not converge within the specified time (such as 5ms), the confidence C of the trajectory is onftraj Set it to 0 directly.

[0141] At t=0.02s, the actual constraint violation state V is measured state_actual Calculate the deviation between the 5 predicted trajectories and the actual values ​​to obtain the trajectory confidence vector, such as C onftraj =[0.9, 0.7, 0.4, 0.2, 0.1] T .

[0142] Using C onftraj P final Enhance and get P enhanced =V final_traj (Ni)*(1+β C onftraj );

[0143] Use P enhanced As weights, the five trajectories are weighted and fused into a single V final_traj (t).

[0144] Step S4 (t=0.02s): Based on V final_traj (t) and P enhanced , solve multiple time domain MPCs in parallel and obtain a set of candidate solutions, such as u1 * and u2 * .

[0145] u1 detected * and u2 * There is a conflict. Using V final_traj (t) predicts u2 * In the next 0.05s, the current will be seriously exceeded. Therefore, the solution repair map will be u2 * To u1 * direction of the device.

[0146] Generate the optimal solution u after repair optimal , and issue the final control instructions.

[0147] According to one aspect of the present application, based on the aforementioned embodiments, a further, preferred or more refined description is provided for several steps in the control method proposed in the present invention.

[0148] In step S100 described in the first embodiment, in addition to generating a multi-scale fault feature matrix, a preferred implementation also includes online identification of system dynamic parameters.

[0149] Specifically, after obtaining the synchronous state data matrix X(k) and the multi-scale fault feature matrix F multi After that, the recursive least squares method (RLS) can be used to estimate and update the key parameters of the system in real time. grid (k), online identification of load characteristics and update of system equivalent inertia parameters. These real-time updated parameters are integrated into the system state parameter set {A(k), B(k), Z grid (k)} provides a more accurate and timely model for the subsequent model predictive control (MPC), thereby improving the accuracy of control.

[0150] To further enhance the intelligence and robustness of layered uncertainty propagation, the present invention preferably includes a gradual mechanism for propagation recovery and state evaluation of the short-term and long-term memory layers:

[0151] After the propagation blocking in step S220 is triggered, the system needs a clear recovery mechanism. The present invention proposes a propagation recovery condition monitoring method. When the memory consistency indicator is monitored to be restored to below the safety threshold (for example, consistency _total <0.8threshold _adaptive ) and the blocking has lasted for the minimum time, the recovery process is triggered.

[0152] Preferably, the recovery process is not instantaneous, but rather a gradual recovery weight recovery _weight To smooth the transition, the weight can be a sigmoid function with respect to time. This gradual recovery mechanism avoids secondary oscillations of the system caused by sudden state changes.

[0153] In order to more comprehensively evaluate the status of each memory layer, in the data structure of the short-term memory layer, in addition to the updated feature F short (k), can also include short-term memory variance Var short (k) and short-term memory confidence Conf short (k). Similarly, the data structure of the long-term memory layer can include the long-term trend slope Slope long (k) and long-term memory stability index S tability-long (k) These additional state evaluation parameters can be used to refine the calculation of the adaptive weight in step S230. For example, when the long-term memory stability is high, the corresponding weight w3 will be improved.

[0154] To improve the accuracy and adaptability of trajectory prediction, the present invention preferably includes verification and optimization of trajectory anchor points, as well as weighted trajectory fusion based on energy conservation:

[0155] After identifying the initial constraint-violating trajectory anchor points in step S310, an anchor point construction verification and optimization step is preferably added. This step evaluates the validity of the anchor points by verifying interpolation accuracy (comparing the trajectory interpolated between anchor points with the actual measured values). Simultaneously, the density of anchor points is optimized according to the preset accuracy target. For example, the density of anchor points can be increased in areas with drastic changes and reduced in stable areas to reduce the computational burden while maintaining accuracy.

[0156] In the trajectory fusion stage of step S330, in addition to using the enhanced probability Penhanced as the weight, a preferred implementation method is to introduce constraint violation energy conservation modeling. Specifically, for each predicted trajectory, the implicit violation energy Energyviolation , and establish an energy transfer matrix Energy according to the energy conservation constraint transfer When performing weighted fusion, the probability weight and the energy weight calculated based on the violation energy are comprehensively considered to form a comprehensive weight W combined , and then perform weighted fusion of the trajectories. This approach not only takes probability into account in the fusion process, but also incorporates energy constraints at the physical level, making the final fusion trajectory more consistent with physical laws.

[0157] To ensure that the final output control instructions are optimal and feasible, the present invention preferably includes conflict analysis and repair based on evolution trajectory, as well as power allocation and coordinated control of cascaded units:

[0158] In the conflict resolution and repair phase of step S420, in order to improve the accuracy of the repair, it is preferred to add a conflict analysis link based on the evolution trajectory. Specifically, the constraint violation evolution trajectory set V is used. evolution , evaluate the severity of the identified conflict solutions conflict The severity not only considers the difference in solution vectors, but also the impact of these differences on the future constraint violation trajectory. Based on the severity ranking, priority is given to repairing those key conflict pairs that may lead to serious consequences in the future. The direction of repair is determined by the gradient of the evolution trajectory (V evolution )Precise guidance.

[0159] In step S420, the optimal solution u after repair is obtained optimal After that, it needs to be converted into specific control instructions for n cascaded H-bridge units. This process is a coordinated optimization solution process that needs to meet the following constraints at the same time:

[0160] Inter-unit power balance constraint: ∑Pi=P total .

[0161] Capacity limit of each unit: P i ≤P i_max .

[0162] Inter-cell voltage balance optimization target: min∑(V dcN -V dc_ref ) 2 By solving this constrained optimization problem, we can finally obtain the control instruction set Control that contains the power and phase angle instructions of each unit. commands , ensuring that each unit operates in a coordinated, safe and stable manner while meeting the overall control objectives.

[0163] Through the above technical details supplemented by this embodiment, its adaptability, robustness and control performance in practical applications are further enhanced.

[0164] According to one aspect of the present application, after identifying and establishing two consecutive anchor points t1 and t2, it is necessary to construct a constraint violation evolution trajectory segment connecting the two anchor points. Preferably, an inter-anchor constraint conservation interpolation algorithm is used, specifically including:

[0165] Establish an interpolation criterion based on physical constraint conservation. In this embodiment, the criterion is specifically the conservation of constrained energy. Its physical meaning is that between two key anchor points caused by external disturbances or mode switching, the energy conversion of each constraint violation state within the system should follow the conservation law. The conservation equation can be expressed as: Energy constraint (t)=∫[V state (τ)dV state / dτ]dτ=constant.

[0166] Construct a conservative interpolation function. According to the above criteria, and fuse the constraint states V at the two anchor points t1 and t2 state (t1),V state (t2) and the previously constructed constraint evolution matrix A evo , generate trajectory segments. Specifically, the interpolation function V interpolated (t) Construction method: V interpolated (t)=V state (t1)+∫t1 t [A evo V state (τ)]dτ{(t-t1) / (t2-t1)}. Through this algorithm, the generated inter-anchor trajectory is not only aligned with the actual state at the endpoints, but its evolution process also follows the inherent physical laws, thus ensuring the accuracy of the prediction.

[0167] According to one aspect of the present application, based on the fusion trajectory prediction V final_traj (t) and uncertainty propagation state vector U propagation , and finally integrate to construct a coupled optimization control strategy. In a preferred embodiment, the strategy consists of three core parameter parts:

[0168] Predictive control gain matrix K predictive :According to the fused trajectory V final_traj (t) is calculated with the given prediction time domain and is the main feedback gain of the model predictive control.

[0169] Robustness compensation parameter K robust :According to the uncertainty propagation state vector U propagation (which contains model credibility information) and preset uncertainty boundaries to compensate for model mismatch and external disturbances and enhance the robustness of the system.

[0170] Adaptive control parameter Kadaptive : Update according to the energy transfer matrix Energy described in Embodiment 5 and the system dynamics, and be used to adaptively adjust the intensity of the control action. The final control strategy is an organic integration of these three parameters, ensuring that the control decision is predictive, robust and adaptive. transfer And the system dynamics for adaptive adjustment of the intensity of the control action. The final control strategy is an organic integration of these three parameters, ensuring that the control decision is predictive, robust and adaptive.

[0171] According to one aspect of the present application, step S410 is specifically:

[0172] Parallelly solve the multi-time domain model predictive control (MPC) problem to generate a multi-time domain prediction solution set containing multiple candidate solutions.

[0173] Based on the coupling optimization control strategy generated in the previous step, construct parallel MPC optimization problems for different prediction time domains (for example, five time domains corresponding to N1 to N5). In a specific embodiment, the prediction model for the i-th time domain can be constructed as: x<{ i (k + j|k) = A i j x(k) + ∑ l=1 j (A i j－l B i u i (k + l - 1|k)); where, x i (k + j|k) is the prediction of the state at the future k + j moment at time k, A i , B i are the state space matrices in this time domain, and u i is the control input.

[0174] The corresponding parallel cost function J i can be designed as: J i = ∑ j=1 Ni ||y i (k + j|k) - r i (k + j|k)||2 + λV evolution_penalty ; This function aims to minimize the error between the predicted output y i and the reference trajectory r i , and at the same time, the penalty term λ Vevolution_penalty uses the information of the constraint violation evolution trajectory set to impose penalties on control behaviors that may lead to serious future constraint violations. By parallelly scheduling and managing each solver, a multi-time domain prediction solution set U multi = {u1 * , u2 * ,..., u5 *} and the solution state monitoring matrix are finally obtained.

[0175] According to another aspect of the present application, the construction process of the constraint evolution matrix Aevo is specifically as follows:

[0176] As mentioned above, a 12-dimensional constraint violation state vector V_state is defined.

[0177] The central difference method is used to calculate partial derivatives to improve accuracy. Specifically: A evo [i,j] = d(v i ) / d(v j ) ≈[v i (x j + Δx) - v i (x j - Δx)] / (2Δx);

[0178] The choice of the step size Δx for numerical differentiation is crucial. Δx should be small enough to ensure differentiation accuracy, but should also avoid rounding errors in numerical calculations caused by being too small. Preferably, Δx can be 0.1% to 1% of the rated value of the corresponding physical quantity. evo The physical meaning of [i,j] is the degree to which a unit change in the j-th constraint can cause a change in the ith constraint, and its dimension depends on the specific physical quantities of i and j.

[0179] According to one aspect of the present application, the constraint conservation interpolation algorithm is specifically:

[0180] The conservation of physical constraints between anchor points is defined as the generalized violation of energy conservation, which is mathematically expressed as: E violation =V state T • W • V_state, where W is a diagonal weight matrix used to characterize the severity of different constraint violations. The conservation criterion is that between two anchor points d(E violation ) / dt ≈ 0.

[0181] Based on this conservation criterion, an interpolation function can be obtained by solving a constrained optimization problem to ensure that the trajectory satisfies the endpoint conditions while minimizing the change in its generalized violation energy.

[0182] The boundary condition of the interpolation function is that at the two anchor points t1 and t2, the value of the interpolation trajectory must be equal to the actual measured constraint state value, that is, V interpolated (t1)=V state_actual (t1) and V interpolated (t2)=V state_actual (t2).

[0183] After the blocking is triggered, the system continues to monitor the recovery conditions. The conditions are: consistency _total < k_recov •threshold _adaptive and block _duration > T _min_block Where k _recov is the coefficient of restitution (e.g. 0.8), T _min_block is the minimum blocking duration.

[0184] The recovery process is gradual, and the recovery weight is _weight Using sigmoid function: recovery _weight =sigmoid((t - t _block_start - T _min_block ) / T _recov_scale ). Where T _recov_scale is the time scale parameter of the recovery process.

[0185] If the preset maximum recovery waiting time T _max_wait If the recovery conditions are still not met, the system will trigger alternative strategies, such as abandoning the blocked memory layer (such as the long-term memory layer) and reinitializing it with the historical data of the short-term memory layer.

[0186] According to one aspect of the present application, when constructing the short-term memory layer, in addition to calculating F short (k), and also calculate its variance Var short (k) and confidence Conf short (k). Variance reflects the volatility of short-term memory, and confidence is a value between [0,1] calculated based on the variance, which is used for subsequent weight distribution. Similarly, the construction of the long-term memory layer also includes its trend slope Slope long (k) and Stability long (k) is calculated. These complete layer status information are integrated for more refined system status assessment.

[0187] According to one aspect of the present application, the calculation formula for short-term-long-term consistency is: _SL = ||F _short (k) - F _long (k)|| / max(||F _short (k)||,||F _long (k)||). Normalization eliminates the influence of the feature amplitude itself, making the consistency index more comparable.

[0188] According to one aspect of the present application, the conflict resolution and repair process is specifically as follows:

[0189] By calculating the Euclidean distance and vector angle between solution vectors, conflicting solution pairs with significant differences are identified.

[0190] For the identified conflicting solution pairs, the constraint violation evolution trajectory set V is used evolution Perform forward-looking evaluations to quantify the severity of constraint violations that each solution may cause in the future and prioritize conflicts.

[0191] For high-priority conflicts, the gradient grad (V evolution ) calculates the repair direction, adaptively determines the repair step size, and constructs the repair mapping function.

[0192] Bring the repaired solution into the system model to verify whether it meets all hard constraints and evaluate its performance loss. Among all feasible repair solutions, select the solution with the best comprehensive performance as the final u optimal .

[0193] For example, at t=0.1s, a voltage drop of 30% occurs in the system.

[0194] t=0.11s: Collect data and calculate the multi-scale fault feature matrix F _multi .

[0195] The three-layer memory architecture calculates F _short (k)=[...], F _long (k)=[...].

[0196] Calculate consistency according to the complete formula _SL =0.85.

[0197] At this time, the adaptive threshold _adaptive =0.7. Because 0.85 > 0.7, a short-term to long-term transmission blockade is triggered.

[0198] Generate time domain probability distribution P _final =[0.7, 0.2, 0.1, 0, 0] T .

[0199] t=0.12s: Based on P _final 5 trajectories are deduced in parallel.

[0200] According to the actual measured value V _state_actual , calculate the trajectory confidence Conf _traj =[0.9, 0.6, 0.3, 0.1, 0.1] T .

[0201] Using Conf _traj Feedback enhancement, get P _enhancedBased on P _enhanced Fusion gets the final trajectory V _final_traj .

[0202] According to one aspect of the present application, if all repair solutions fail to pass feasibility verification during the solution repair process, the system will abandon the repair and select the solution with the least constraint violation among all candidate solutions as the suboptimal solution, and trigger an alarm to ensure the safe operation of the system.

[0203] In summary, in order to solve the problem of coupling between uncertainty propagation and constraint evolution under fault transients, which existing technologies cannot effectively deal with, a complete closed-loop control concept integrating perception-prediction-correction-coupling and its specific implementation architecture are proposed.

[0204] Specifically, this invention constructs an adaptive control system in which a probabilistic model (i.e., the time-domain probability distribution) and a physical model (i.e., the constraint violation evolution trajectory) mutually supervise and co-evolve. In the prior art, the construction of the probabilistic model and the prediction of the physical model are typically two independent processes or have only a one-way dependency. This results in both failing simultaneously when the grid operating conditions change drastically, creating a vicious cycle. This invention groundbreakingly establishes a bidirectional information channel between the two: on the one hand, the prediction accuracy of the physical model is quantified through trajectory confidence, and feedback is used to correct and enhance the understanding of the probabilistic model; on the other hand, the more credible probabilistic model after correction serves as a weight to guide the fusion of multiple physical prediction trajectories, thereby generating a more reliable single prediction. This symbiotic evolutionary design concept enables the entire control system to self-correct and co-optimize in the face of uncertainty.

[0205] For example, the hierarchical uncertainty propagation blocking mechanism is essentially different from the gating regulation in existing networks such as LSTM. The propagation blocking of the present invention is an active risk isolation triggered by real-time working conditions and physical consistency judgment. Its purpose is to prevent erroneous information pollution when a violent conflict between model cognition and physical reality is detected, thereby providing a stable and reliable probabilistic model foundation for subsequent bidirectional coupling. Similarly, the trajectory anchor mechanism is not a conventional filtering correction. It is driven by a specific event of physical mode switching in the system, and is combined with a physical constraint conservation interpolation algorithm between anchor points. It is specifically used to solve the trajectory divergence problem under nonlinear and multi-mode constraints, ensuring that the physical model can continue to provide meaningful feedback information.

[0206] Therefore, the various technical features of this invention share a close internal logic and synergistic benefits. Without a propagation blocking mechanism, the bidirectional coupling input would be contaminated; without a trajectory anchor mechanism, the trajectory confidence feedback to the probabilistic model would quickly become ineffective. It is precisely because of the synergistic effect of these various elements that this invention achieves significant overall technical benefits. Test data demonstrates an improvement in comprehensive prediction accuracy of up to 65%.

[0207] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A control method for a cascaded H-bridge inverter, characterized in that: include: Process the original sampling data of the three-phase voltage and current of the power grid and the state parameters of each unit of the cascaded H-bridge to generate a multi-scale fault feature matrix and a system state parameter set; Based on the multi-scale fault feature matrix, probability reconstruction is performed through hierarchical uncertainty propagation to obtain the time domain probability distribution matrix and uncertainty propagation state vector; In response to the time-domain probability distribution matrix, the uncertainty propagation state vector, and the system state parameter set, a spatiotemporal coupled trajectory prediction and optimization is performed to form a coupled optimization control strategy and a constraint violation evolution trajectory set; And generate control instructions for each level of cascade units accordingly; Performing spatiotemporal coupled trajectory prediction and optimization includes: Analyze the time domain probability distribution matrix and assign a corresponding prediction depth to each probability component; For each assigned prediction depth, multiple independent constraint violation evolution trajectories are deduced in parallel in combination with the constraint evolution matrix derived from the system state parameter set. All the deduced constraint violation evolution trajectories are collected to construct a multi-time domain trajectory prediction set; Compare each constraint violation evolution trajectory in the multi-time domain trajectory prediction set with the actual measured constraint violation state in real time; Based on the prediction deviation generated by the comparison, a corresponding trajectory confidence is quantitatively generated for each evolution trajectory; Using the confidence of all trajectories, feedback enhancement is performed on the time domain probability distribution matrix to reconstruct an enhanced time domain probability matrix that has been verified by trajectories. Probabilistic reconstruction via hierarchical uncertainty propagation includes: Based on the multi-scale fault feature matrix, an instantaneous propagation layer reflecting the current fault information is constructed; Fuse the current state of the instantaneous propagation layer with the previous state of the short-term memory layer, and iteratively update the current state of the short-term memory layer; The current state of the short-term memory layer is coupled with the state of the long-term memory layer at the previous moment, evolving into the current state of the long-term memory layer; The feature data of the instantaneous propagation layer, short-term memory layer and long-term memory layer are integrated into a hierarchical fault feature set.

2. The method according to claim 1, characterized in that Perform spatiotemporal coupled trajectory prediction and optimization, including: Real-time monitoring of system state parameter sets to identify mode switching points where constraint violation modes undergo qualitative changes; Establishing the identified mode switching points as constraint violation trajectory anchor points to construct a constraint violation anchor point set; Taking the anchor point as the benchmark, periodic realignment correction is performed on the prediction process of the constraint-violated evolution trajectory to suppress the accumulation of prediction errors.

3. The method according to claim 2, characterized in that The process of constructing a constraint violation evolution trajectory segment connecting any two consecutive anchor points defined by the constraint violation anchor point set includes: Establish interpolation criteria based on the conservation of physical constraints; According to the interpolation criterion, the constraint states at the two anchor points are fused with the constraint evolution matrix derived from the system state parameter set to generate trajectory segments.

4. The method according to claim 1, wherein Probabilistic reconstruction via hierarchical uncertainty propagation also includes: Calculate the feature differences between each memory layer in the layered fault feature set to generate a memory consistency vector; When the difference represented by the memory consistency vector exceeds the propagation blocking threshold, the propagation blocking state matrix is ​​triggered and generated; According to the propagation blocking state matrix, the information propagation path between layers is selectively cut off or maintained, and the layered fault feature set is reconstructed into the layered feature set after blocking.

5. The method according to claim 4, characterized in that The propagation blocking threshold is determined by the following process: Dynamically maintain time window data containing historical memory consistency vectors; Analyze the statistical distribution characteristics of the time window data and extract its mean and variance; Based on the mean and variance, the propagation blocking threshold is calculated to achieve dynamic adaptive adjustment of the threshold.

6. The method according to claim 1, characterized in that Executing spatiotemporal coupled trajectory prediction and optimization also includes: Extract each probability component in the enhanced time domain probability matrix as the fusion weight corresponding to each evolution trajectory in the multi-time domain trajectory prediction set; And for all the evolving trajectories in the multi-time domain trajectory prediction set, weighted fusion based on fusion weights is implemented to condense them into a single fusion trajectory prediction.

7. The method according to claim 1, characterized in that The process of generating control instructions for each cascade unit based on this includes: Solve multi-horizon model predictive control problems in parallel and generate a multi-horizon prediction solution set containing multiple candidate solutions; Detect conflicting solutions in a set of multi-temporal prediction solutions; And using the constraint violation evolution trajectory set, a solution repair mapping is constructed to guide the repair direction of the conflict solution, thereby generating the optimal solution after repair; Based on the optimal solution after repair, the control instructions of each cascade unit are determined and output.

Images