Layered optimization scheduling method for power system containing large-scale water electrolysis hydrogen production device

Abstract

The application discloses a layered optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device, relates to the technical field of layered optimization scheduling of power systems, and the gradient heterogeneous compensation unit of the application fuses covariance alignment and differential privacy encryption, and under the premise that original data is not shared, the convergence speed of federal learning is improved to a practical level; after each hydrogen production station uploads an encrypted gradient, the coordinator eliminates equipment difference interference by statistical distribution correction to generate a globally consistent frequency modulation strategy, which protects commercial sensitive information and avoids model shock failure; the knowledge distillation unit further compresses the model volume to guarantee the real-time response capability of edge nodes; when the power grid frequency suddenly changes, a dynamic priority allocation mechanism is automatically triggered by a fast response protocol; the regional aggregation layer directly broadcasts frequency modulation instructions based on historical performance labels, skips the federal learning iteration process, and enables the hydrogen production cluster to form an adjusting resultant force within seconds.

Description

Technical Field

[0001] This invention relates to the field of hierarchical optimization scheduling technology for power systems, and in particular to a hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production units. Background Technology

[0002] In current high-proportion renewable energy power systems, the strong volatility of wind and solar power has led to a surge in system frequency regulation pressure. Large-scale water electrolysis hydrogen production units with decentralized layouts are considered high-quality frequency regulation resources due to their rapid power adjustment characteristics. Driven by dual carbon targets, the participation of hydrogen production station clusters invested by multiple entities in grid-coordinated frequency regulation has become a new trend. However, these hydrogen production stations belong to different operators and involve commercially sensitive data such as hydrogen inventory, equipment lifespan, and hydrogen use contracts, forming natural information silos.

[0003] To tap the frequency regulation potential of hydrogen production resources, a federated learning collaborative optimization framework has recently been proposed, such as the "Active Power Optimization Dispatch under the Electricity Market Environment" in 2024. This scheme allows each hydrogen production station to train its scheduling model locally and only uploads the model parameters to the aggregation node, thus avoiding the leakage of raw data. However, in practical applications, differences in the model and aging of electrolyzers at each station lead to different local data distributions, which exacerbates gradient bias during federated aggregation, causing model oscillations or even divergence. To ensure convergence, multiple iterations are required, which is difficult to meet the requirements of second-level frequency regulation commands.

[0004] While existing solutions have made progress in privacy protection, they come at the cost of scheduling timeliness and stability. In particular, when the grid frequency changes abruptly, hydrogen production clusters are unable to quickly form a regulatory force due to low coordination efficiency, and may even trigger cascading protection due to local overload. Therefore, there is an urgent need for a lightweight collaborative mechanism that can adapt to multi-stakeholder game dynamics. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] This invention provides a hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices, which solves the problem that existing federated learning schemes cannot meet the collaborative needs of multi-entity hydrogen production clusters due to slow convergence of heterogeneous data and sluggish frequency regulation response.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] This invention provides a hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices, comprising:

[0009] At the local scheduling layer, each hydrogen production station is equipped with edge computing nodes to collect electrolyzer temperature, hydrogen storage tank pressure and power grid frequency deviation signals in real time.

[0010] In the regional aggregation layer, a federated learning coordinator is deployed to receive encrypted gradient parameters uploaded by each edge node;

[0011] At the global optimization layer, the main network dispatch center generates power allocation coefficients based on the adjustable capacity feedback from the regional layer and issues emergency frequency adjustment commands.

[0012] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, wherein: the edge computing nodes perform:

[0013] An LSTM prediction model is constructed with frequency deviation, electrolyzer temperature, and storage tank pressure as inputs, and the output is the hydrogen production power adjustment amount;

[0014] Gradient encryption operation: Add Laplace noise perturbation to the local model gradient to generate encrypted gradients that satisfy differential privacy;

[0015] The encrypted gradient is transmitted to the regional aggregation layer via a blockchain encrypted channel.

[0016] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, the federated learning coordinator includes:

[0017] Gradient heterogeneous compensation unit: performs distribution alignment correction on the gradients uploaded by each hydrogen production station;

[0018] Federation Aggregation Unit: Performs weighted averaging calculation on the corrected gradient to update the global frequency modulation decision model;

[0019] Knowledge distillation unit: Compresses the global model into a lightweight micro-model and distributes it to edge nodes.

[0020] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, wherein: in the gradient heterogeneous compensation unit, the distribution alignment correction process includes:

[0021] Extract site gradient statistics and calculate hydrogen production station recent Next upload gradient The statistic is calculated using the following formula:

[0022] ,

[0023] in, Indicates the first The gradient vector uploaded next time. Number the hydrogen production station. The sample number. For the sample size, The gradient vector of the site mean. The site covariance matrix;

[0024] The federal coordinator builds a global reference based on statistics from each site, specifically:

[0025] ,

[0026] in, The total number of sites that participated in the upload. The global gradient mean vector, The global gradient covariance matrix;

[0027] Generate a mapping matrix using the square root and inverse square root of a matrix:

[0028] ,

[0029] in, For the site The homogeneous mapping matrix, The positive semi-definite square root of the global covariance. To find the inverse square root of the site covariance, define the matrix. have ,

[0030] In the formula,

[0031] ,

[0032] in, The eigenvector matrix, These are the square roots and inverse square roots of the corresponding eigenvalue matrices;

[0033] For this round of uploading gradient Perform whitening and translation:

[0034] ,

[0035] in, For the current gradient vector, The intermediate gradient after whitening and centroid alignment;

[0036] The formula for measuring the difference between a site and the global distribution is as follows:

[0037] ,

[0038] in, As an index of heterogeneity, Represents the Frobenius norm;

[0039] Generate dynamic compensation coefficients:

[0040] ,

[0041] in, For dynamic compensation coefficients, To compensate for the intensity constant, ranging from 0 to 1, dimensionless. is a dimensionless isomerism threshold constant.

[0042] Further smoothing is achieved using logical functions:

[0043] ,

[0044] in, This is the slope adjustment coefficient;

[0045] when Compensation will be initiated at the appropriate time.

[0046] ,

[0047] in, This is the aligned gradient vector sent to the federated aggregation unit.

[0048] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, the operation of the gradient heterogeneous compensation unit includes:

[0049] Match the preset aging correction factor according to the equipment model code;

[0050] Dynamic compensation coefficient generation: Real-time compensation coefficients are generated by combining the power grid frequency fluctuation amplitude and the aging correction factor;

[0051] The original gradient is corrected based on the real-time compensation coefficient.

[0052] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, the emergency frequency regulation command response process includes:

[0053] The regional aggregation layer invokes a pre-defined fast response protocol to skip the federated learning iteration process;

[0054] Dynamic priority allocation: Priority tags are generated based on historical frequency modulation response delay rates and broadcast to edge nodes;

[0055] The edge node is triggered to output power adjustment commands in real time based on the priority label.

[0056] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, it further includes a constraint guarantee mechanism:

[0057] Start-stop count monitoring: An electrolytic cell start-stop counter is embedded in the edge node to accumulate the number of start-stop cycles in real time;

[0058] Constraint activation condition: When the cumulative number of times reaches 90% of the preset threshold, the power smoothing filter is automatically activated;

[0059] Penalty term adjustment: Hard constraint guarantees are achieved by adjusting the coefficient of the constraint penalty term in the LSTM loss function.

[0060] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, the operation of the power smoothing filter includes:

[0061] Slope limiting module: applies a preset limit to the rate of change of power regulation commands;

[0062] Time delay compensation module: dynamically compensates for command delay based on the electrolytic cell thermal inertia model;

[0063] The smoothed power command is output to the hydrogen production unit's execution unit.

[0064] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, wherein: in the time delay compensation module, the dynamic compensation command delay method is as follows:

[0065] Establish a thermal inertia model for the electrolytic cell:

[0066] ,

[0067] in, This indicates the internal temperature of the electrolytic cell. For the rate of temperature change, The electrothermal conversion coefficient, Instantaneous input power, For heat dissipation coefficient, For ambient temperature, Equivalent heat capacity;

[0068] ,

[0069] in, The time constant is the thermal inertia.

[0070] Estimate the real-time inertial constant:

[0071] ,

[0072] in, For the first The inertial estimate at each sampling time. For static compensation bias, For dynamic adjustment coefficients, and These represent the temperatures at two adjacent moments. The sampling period;

[0073] Generate discrete feedforward compensation coefficients:

[0074] ,

[0075] in, These are discrete feedforward compensation coefficients;

[0076] The formula for calculating the lead time of a command is:

[0077] ,

[0078] in, The power command after advance compensation. and These are the current and previous raw power commands, respectively;

[0079] Output delay compensation:

[0080] ,

[0081] in, This is the final power command sent to the slope limiting module.

[0082] As a preferred embodiment of the hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production devices described in this invention, wherein: the knowledge distillation unit performs:

[0083] Feature alignment loss calculation: constrain the output difference between the miniature model and the global model at the key frequency feature layer;

[0084] Distillation temperature parameter adjustment: The softmax temperature parameter is dynamically adjusted according to the current power grid frequency fluctuation variance;

[0085] Generate a micro-frequency modulation decision model suitable for edge computing resources.

[0086] The beneficial effects of this invention are as follows: The gradient heterogeneous compensation unit of this invention integrates covariance alignment and differential privacy encryption, which improves the convergence speed of federated learning to a practical level without sharing the original data; After each hydrogen production station uploads the encrypted gradient, the coordinator eliminates equipment difference interference through statistical distribution correction and generates a globally consistent frequency modulation strategy, which protects commercially sensitive information and avoids model oscillation failure; The knowledge distillation unit further compresses the model volume and ensures the real-time response capability of edge nodes.

[0087] When the grid frequency changes abruptly, the fast response protocol automatically triggers a dynamic priority allocation mechanism; the regional aggregation layer directly broadcasts frequency modulation commands based on historical performance tags, skipping the federated learning iteration process, enabling the hydrogen production cluster to form a regulatory force within seconds; priority tags are dynamically updated according to the site response latency rate, ensuring that high-reliability sites take priority in undertaking critical regulation tasks, completely eliminating the risk of frequency modulation failure caused by command lag in traditional solutions; the edge nodes of this invention have built-in start-stop counters and power smoothing filters, and monitor the equipment status in real time through an LSTM loss function with embedded physical constraints; when the electrolyzer approaches the start-stop limit, the slope limiting and thermal inertia compensation modules are automatically activated, dynamically adjusting the power command change rate and timing advance to ensure that the command conforms to the equipment thermodynamic constraints; this mechanism avoids equipment overload caused by frequent rescheduling from the source and extends the life of critical components. Attached Figure Description

[0088] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0089] Figure 1 This is a flowchart illustrating the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device, as described in Example 1. Detailed Implementation

[0090] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0091] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0092] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0093] Example 1, referring to Figure 1 This embodiment provides a hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production units, including:

[0094] At the local scheduling layer, each hydrogen production station is equipped with edge computing nodes to collect electrolyzer temperature, hydrogen storage tank pressure and power grid frequency deviation signals in real time.

[0095] Edge computing node execution:

[0096] An LSTM prediction model is constructed with frequency deviation, electrolyzer temperature, and storage tank pressure as inputs, and the output is the hydrogen production power adjustment amount;

[0097] Gradient encryption operation: Add Laplace noise perturbation to the local model gradient to generate encrypted gradients that satisfy differential privacy;

[0098] The encrypted gradient is transmitted to the regional aggregation layer via a blockchain encrypted channel;

[0099] In the regional aggregation layer, a federated learning coordinator is deployed to receive encrypted gradient parameters uploaded by each edge node;

[0100] The federated learning coordinator includes:

[0101] Gradient heterogeneous compensation unit: performs distribution alignment correction on the gradients uploaded by each hydrogen production station;

[0102] Federation Aggregation Unit: Performs weighted averaging calculation on the corrected gradient to update the global frequency modulation decision model;

[0103] Knowledge distillation unit: Compresses the global model into a lightweight micro-model and distributes it to edge nodes;

[0104] In the gradient heterogeneous compensation unit, the distribution alignment correction process includes:

[0105] Extract site gradient statistics and calculate hydrogen production station recent Next upload gradient The statistic is calculated using the following formula:

[0106] ,

[0107] in, Indicates the first The gradient vector uploaded next time. Number the hydrogen production station. The sample number. For the sample size, The gradient vector of the site mean. The site covariance matrix;

[0108] The federal coordinator builds a global reference based on statistics from each site, specifically:

[0109] ,

[0110] in, The total number of sites that participated in the upload. The global gradient mean vector, The global gradient covariance matrix;

[0111] Generate a mapping matrix using the square root and inverse square root of a matrix:

[0112] ,

[0113] in, For the site The homogeneous mapping matrix, The positive semi-definite square root of the global covariance. To find the inverse square root of the site covariance, define the matrix. have ,

[0114] In the formula,

[0115] ,

[0116] in, The eigenvector matrix, These are the square roots and inverse square roots of the corresponding eigenvalue matrices;

[0117] For this round of uploading gradient Perform whitening and translation:

[0118] ,

[0119] in, For the current gradient vector, The intermediate gradient after whitening and centroid alignment;

[0120] The formula for measuring the difference between a site and the global distribution is as follows:

[0121] ,

[0122] in, As an index of heterogeneity, Represents the Frobenius norm;

[0123] Generate dynamic compensation coefficients:

[0124] ,

[0125] in, For dynamic compensation coefficients, To compensate for the intensity constant, ranging from 0 to 1, dimensionless. is a dimensionless isomerism threshold constant.

[0126] Further smoothing is achieved using logical functions:

[0127] ,

[0128] in, This is the slope adjustment coefficient;

[0129] when Compensation will be initiated at the appropriate time.

[0130] ,

[0131] in, The aligned gradient vector sent to the federated aggregation unit;

[0132] Specifically, the process first unifies the gradient scale through covariance whitening, then removes static bias through mean migration, uses a single norm to measure high-dimensional differences, and the subsequent compensation coefficient changes continuously with the heterogeneity to avoid oscillations caused by step correction. The logistic function form accelerates convergence in large heterogeneous scenarios and automatically decays the compensation amount in small heterogeneous scenarios to maintain information integrity. The entire mapping does not depend on the physical parameters of the device and can work in real time under aging and load fluctuation environments, effectively reducing the number of convergence rounds and taking into account real-time scheduling and differential privacy requirements.

[0133] The operation of the gradient heterogeneous compensation unit includes:

[0134] Match the preset aging correction factor according to the equipment model code;

[0135] Dynamic compensation coefficient generation: Real-time compensation coefficients are generated by combining the power grid frequency fluctuation amplitude and the aging correction factor;

[0136] The original gradient is corrected based on the real-time compensation coefficient.

[0137] Knowledge distillation unit execution:

[0138] Feature alignment loss calculation: constrain the output difference between the miniature model and the global model at the key frequency feature layer;

[0139] Distillation temperature parameter adjustment: The softmax temperature parameter is dynamically adjusted according to the current power grid frequency fluctuation variance;

[0140] Generate a micro-frequency modulation decision model suitable for edge computing resources;

[0141] At the global optimization layer, the main network dispatch center generates power allocation coefficients based on the adjustable capacity feedback from the regional layer and issues emergency frequency adjustment commands.

[0142] The emergency frequency modulation command response process includes:

[0143] The regional aggregation layer invokes a pre-defined fast response protocol to skip the federated learning iteration process;

[0144] Dynamic priority allocation: Priority tags are generated based on historical frequency modulation response delay rates and broadcast to edge nodes;

[0145] The edge node is triggered to output power adjustment commands in real time based on the priority label;

[0146] It also includes constraint and protection mechanisms:

[0147] Start-stop count monitoring: An electrolytic cell start-stop counter is embedded in the edge node to accumulate the number of start-stop cycles in real time;

[0148] Constraint activation condition: When the cumulative number of times reaches 90% of the preset threshold, the power smoothing filter is automatically activated;

[0149] Penalty term adjustment: Hard constraint guarantees are achieved by adjusting the coefficient of the constraint penalty term in the LSTM loss function;

[0150] The operation of a power smoothing filter includes:

[0151] Slope limiting module: applies a preset limit to the rate of change of power regulation commands;

[0152] Time delay compensation module: dynamically compensates for command delay based on the electrolytic cell thermal inertia model;

[0153] The smoothed power command is output to the hydrogen production unit's execution unit;

[0154] In the time delay compensation module, the dynamic compensation command delay method is as follows:

[0155] Establish a thermal inertia model for the electrolytic cell:

[0156] ,

[0157] in, This indicates the internal temperature of the electrolytic cell. For the rate of temperature change, The electrothermal conversion coefficient, Instantaneous input power, For heat dissipation coefficient, For ambient temperature, Equivalent heat capacity;

[0158] ,

[0159] in, The time constant is the thermal inertia.

[0160] Estimate the real-time inertial constant:

[0161] ,

[0162] in, For the first The inertial estimate at each sampling time. For static compensation bias, For dynamic adjustment coefficients, and These represent the temperatures at two adjacent moments. The sampling period;

[0163] Generate discrete feedforward compensation coefficients:

[0164] ,

[0165] in, These are discrete feedforward compensation coefficients;

[0166] The formula for calculating the lead time of a command is:

[0167] ,

[0168] in, The power command after advance compensation. and These are the current and previous raw power commands, respectively;

[0169] Output delay compensation:

[0170] ,

[0171] in, This is the final power command sent to the slope limiting module;

[0172] Specifically, this process uses thermoelectric coupling differential equations to characterize the temperature rise mechanism of the electrolytic cell, converts heat dissipation and heat storage parameters into time constants, and then uses temperature slope to adaptively estimate the magnitude of inertia to avoid the failure of fixed coefficients under different load conditions. The discrete feedforward stage injects an equal-amplitude advance into the original power command, making the equivalent transfer function approximately a unity-gain pure delay, thereby suppressing the energy lag caused by thermal inertia. The compensation coefficient converges rapidly with the rate of temperature change, reducing power overshoot in high-speed load change scenarios, and automatically converges to zero in the low-speed steady-state stage to maintain energy balance.

[0173] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A hierarchical optimal scheduling method for power systems containing large-scale water electrolysis hydrogen production units, characterized in that, include, At the local scheduling layer, each hydrogen production station is equipped with edge computing nodes to collect electrolyzer temperature, hydrogen storage tank pressure and power grid frequency deviation signals in real time. In the regional aggregation layer, a federated learning coordinator is deployed to receive encrypted gradient parameters uploaded by each edge node. The edge computing node performs local LSTM training based on the collected frequency deviation, temperature and pressure signals to generate local gradients, and applies Laplacian noise to the local gradients to satisfy differential privacy before uploading them to the regional aggregation layer via a blockchain encrypted channel. At the global optimization layer, the main network dispatch center generates power allocation coefficients based on the adjustable capacity feedback from the regional layer and issues emergency frequency adjustment commands. In the gradient heterogeneous compensation unit, distribution alignment correction is performed, and the process includes: Extract site gradient statistics and calculate hydrogen production station recent Next upload gradient The statistic is calculated using the following formula: ， in, Indicates the first The gradient vector uploaded next time. Number the hydrogen production station. The sample number. For the sample size, The gradient vector of the site mean. The site covariance matrix; The federal coordinator builds a global reference based on statistics from each site, specifically: ， in, The total number of sites that participated in the upload. The global gradient mean vector, The global gradient covariance matrix; Generate a mapping matrix using the square root and inverse square root of a matrix: ， in, For the site The homogeneous mapping matrix, The positive semi-definite square root of the global covariance. To find the inverse square root of the site covariance, define the matrix. have , In the formula, ， in, The eigenvector matrix, These are the square roots and inverse square roots of the corresponding eigenvalue matrices; For this round of uploading gradient Perform whitening and translation: ， in, For the current gradient vector, The intermediate gradient after whitening and centroid alignment; The formula for measuring the difference between a site and the global distribution is as follows: ， in, As an index of heterogeneity, Represents the Frobenius norm; Generate dynamic compensation coefficients: ， in, For dynamic compensation coefficients, To compensate for the intensity constant, ranging from 0 to 1, dimensionless. is a dimensionless isomerism threshold constant. Further smoothing is achieved using logical functions: ， in, This is the slope adjustment coefficient; when Compensation will be initiated at the appropriate time. ， in, This is the aligned gradient vector sent to the federated aggregation unit.

2. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production equipment as described in claim 1, characterized in that, The federated learning coordinator includes: Gradient heterogeneous compensation unit: performs distribution alignment correction on the gradients uploaded by each hydrogen production station; Federation Aggregation Unit: Performs weighted averaging calculation on the corrected gradient to update the global frequency modulation decision model; Knowledge distillation unit: Compresses the global model into a lightweight micro-model and distributes it to edge nodes.

3. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production equipment as described in claim 2, characterized in that, The operation of the gradient heterogeneous compensation unit includes: Match the preset aging correction factor according to the equipment model code; Dynamic compensation coefficient generation: Real-time compensation coefficients are generated by combining the power grid frequency fluctuation amplitude and the aging correction factor; The original gradient is corrected based on the real-time compensation coefficient.

4. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production equipment as described in claim 1, characterized in that, The emergency frequency modulation command response process includes: The regional aggregation layer invokes a pre-defined fast response protocol to skip the federated learning iteration process; Dynamic priority allocation: Priority tags are generated based on historical frequency modulation response delay rates and broadcast to edge nodes; The edge node is triggered to output power adjustment commands in real time based on the priority label.

5. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production units as described in claim 1, characterized in that, It also includes constraint and protection mechanisms: Start-stop count monitoring: An electrolytic cell start-stop counter is embedded in the edge node to accumulate the number of start-stop cycles in real time; Constraint activation condition: When the cumulative number of times reaches 90% of the preset threshold, the power smoothing filter is automatically activated; Penalty term adjustment: Hard constraint guarantees are achieved by adjusting the coefficient of the constraint penalty term in the LSTM loss function.

6. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production equipment as described in claim 5, characterized in that, The operation of the power smoothing filter includes: Slope limiting module: applies a preset limit to the rate of change of power regulation commands; Time delay compensation module: dynamically compensates for command delay based on the electrolytic cell thermal inertia model; The smoothed power command is output to the hydrogen production unit's execution unit.

7. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production equipment as described in claim 6, characterized in that, In the time delay compensation module, the dynamic compensation command delay method is as follows: Establish a thermal inertia model for the electrolytic cell: ， in, This indicates the internal temperature of the electrolytic cell. For the rate of temperature change, The electrothermal conversion coefficient, For instantaneous input power, For heat dissipation coefficient, For ambient temperature, Equivalent heat capacity; ， in, The time constant is the thermal inertia. Estimate the real-time inertial constant: ， in, For the first The inertial estimate at each sampling time. For static compensation bias, For dynamic adjustment coefficients, and These represent the temperatures at two adjacent moments. The sampling period; Generate discrete feedforward compensation coefficients: ， in, These are discrete feedforward compensation coefficients; The formula for calculating the lead time of a command is: ， in, The power command after advance compensation. and These are the current and previous raw power commands, respectively; Output delay compensation: ， in, This is the final power command sent to the slope limiting module.

8. The hierarchical optimization scheduling method for power systems containing large-scale water electrolysis hydrogen production equipment as described in claim 2, characterized in that, The knowledge distillation unit performs the following: Feature alignment loss calculation: constrain the output difference between the miniature model and the global model at the key frequency feature layer; Distillation temperature parameter adjustment: The softmax temperature parameter is dynamically adjusted according to the current power grid frequency fluctuation variance; Generate a micro-frequency modulation decision model suitable for edge computing resources.

Images