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

Abstract

The invention discloses a hierarchical optimization scheduling method for an electric power system containing a large-scale water electrolysis hydrogen production device, and relates to the technical field of hierarchical optimization scheduling of the electric power system.A gradient heterogeneous compensation unit fuses covariance alignment and differential privacy encryption, and on the premise that original data does not need to be shared, the optimal scheduling of the electric power system is achieved. The federal learning convergence speed is improved to a practical level; after each hydrogen production station uploads the encryption gradient, a coordinator corrects and eliminates equipment difference interference through statistical distribution, and generates a global consistent frequency modulation strategy, so that commercial sensitive information is protected, and model oscillation failure is avoided; the knowledge distillation unit further compresses the volume of the model to guarantee the real-time response capability of edge nodes, and when the frequency of the power grid changes suddenly, a quick response protocol automatically triggers a dynamic priority distribution mechanism; the regional aggregation layer directly broadcasts a frequency modulation instruction based on a historical performance label, and skips a federated learning iteration process, so that a hydrogen production cluster forms an adjusting resultant force in a second level.

Description

Technical Field

[0001] The present invention relates to the technical field of hierarchical optimization and scheduling of power systems, in particular to a hierarchical optimization and scheduling method for a power system containing a large-scale water electrolysis hydrogen production device. Background Art

[0002] In the current high-proportion new energy power system, the strong volatility of wind power and photovoltaics has led to a sharp increase in the system frequency regulation pressure; large-scale water electrolysis hydrogen production devices with dispersed layouts are regarded as high-quality frequency regulation resources due to their rapid power adjustment characteristics; driven by the dual carbon goals, clusters of hydrogen production stations invested by multiple entities participating in the coordinated frequency regulation of the power grid has become a new trend; however, these hydrogen production stations belong to different operators, involving commercially sensitive data such as hydrogen inventory, equipment life, and hydrogen use contracts, forming a natural information island.

[0003] In order to tap the frequency regulation potential of hydrogen production resources, a federated learning collaborative optimization framework has been proposed recently, such as the 2024 "Active Power Optimization Dispatch in the Electricity Market Environment". This solution allows each hydrogen production station to locally train the scheduling model and only upload the model parameters to the aggregation node to avoid leakage of original data; however, in actual applications, the differences in electrolyzer models and aging levels at each station lead to different local data distributions, and the gradient deviation is aggravated during federated aggregation, causing model oscillation or even divergence; multiple rounds of iterations are required to ensure convergence, which makes it difficult to meet the second-level frequency regulation instruction requirements.

[0004] Although existing solutions have made progress in privacy protection, they have come at the expense of scheduling timeliness and stability; especially when the grid frequency changes suddenly, the hydrogen production cluster cannot quickly form a regulatory force due to low coordination efficiency, and may even trigger chain protection due to local overload. Therefore, there is an urgent need for a lightweight coordination mechanism that adapts to multi-agent games. Summary of the Invention

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

[0006] The present invention provides a hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device to solve the problem that the existing federated learning scheme cannot meet the collaborative needs of a multi-agent hydrogen production cluster due to slow convergence of heterogeneous data and delayed frequency modulation response.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The embodiment of the present invention provides a hierarchical optimization scheduling method for a power system including a large-scale water electrolysis hydrogen production device, which includes:

[0009] At the local dispatching layer, each hydrogen production station is equipped with an edge computing node to collect real-time signals of electrolyzer temperature, hydrogen storage tank pressure, and grid frequency deviation;

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

[0011] At the global optimization layer, the main network dispatching center generates a power allocation coefficient based on the adjustable capacity feedback from the regional layer and issues emergency frequency regulation instructions.

[0012] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, the edge computing node executes:

[0013] Build an LSTM prediction model with frequency deviation, electrolyzer temperature, and tank pressure as inputs to output the hydrogen production power adjustment value;

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

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

[0016] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, the federated learning coordinator includes:

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

[0018] Federal aggregation unit: performs weighted average calculation on the corrected gradient and updates the global frequency modulation decision model;

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

[0020] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, in which: in the gradient heterogeneous compensation unit, the distribution alignment correction process includes:

[0021] Extract site gradient statistics and calculate the nearest N hydrogen production station i i Upload gradient The statistic is calculated as follows:

[0022]

[0023] in, represents the gradient vector uploaded for the kth time, i is the hydrogen station number, k is the sample number, N i is the sample size, μ i is the site mean gradient vector, Σ i is the site covariance matrix;

[0024] The federation coordinator builds a global reference based on the statistics of each site, specifically:

[0025]

[0026] Where M is the total number of sites participating in the upload, μ0 is the global gradient mean vector, and Σ0 is the global gradient covariance matrix;

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

[0028]

[0029] Among them, A i is the homogeneous mapping matrix of site i, is the positive semidefinite square root of the global covariance, is the inverse square root of the site covariance, and the definition matrix Σ i have

[0030] Where,

[0031]

[0032] Among them, Q0,Q i is the eigenvector matrix, are the square root and inverse square root of the corresponding eigenvalue matrix;

[0033] Upload the gradient g for this round i Perform whitening and translation:

[0034]

[0035] Among them, g i is the current gradient vector, It is the middle gradient after whitening and centroid alignment;

[0036] Calculate the difference measure between the site and the global distribution, the formula is:

[0037] D i =‖Σ i -Σ0‖ F ,

[0038] Among them, D i is the heterogeneity index, ‖·‖ F represents the Frobenius norm;

[0039] Generate motion compensation coefficients:

[0040]

[0041] Among them, c iis the dynamic compensation coefficient, η is the compensation strength constant, ranging from 0 to 1, dimensionless, θ is the heterogeneity threshold constant, dimensionless;

[0042] Further smoothing is performed using a logistic function:

[0043]

[0044] Where λ is the slope adjustment coefficient;

[0045] When D i Compensation starts when ≥0.1θ,

[0046]

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

[0048] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present 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: Combines the grid frequency fluctuation amplitude and aging correction factor to generate real-time compensation coefficient;

[0051] A correction operation is performed on the original gradient based on the real-time compensation coefficient.

[0052] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, the emergency frequency modulation instruction response process includes:

[0053] The regional aggregation layer calls the preset fast response protocol and skips the federated learning iteration process;

[0054] Dynamic priority allocation: Generates priority labels based on historical FM response delay rates and broadcasts them to edge nodes;

[0055] Trigger edge nodes to instantly output power adjustment instructions based on priority labels.

[0056] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, it also includes a constraint guarantee mechanism:

[0057] Start and stop frequency monitoring: embed electrolytic cell start and stop counters in edge nodes to accumulate start and stop times 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 protection is achieved by adjusting the constraint penalty term coefficient in the LSTM loss function.

[0060] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, the operation of the power smoothing filter includes:

[0061] Slope limit module: imposes preset limits on the rate of change of power regulation commands;

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

[0063] Output the smoothed power command to the hydrogen production device execution unit.

[0064] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, in which: in the time delay compensation module, the dynamic compensation instruction delay mode is:

[0065] Establish the electrolytic cell thermal inertia model:

[0066]

[0067] Among them, T e (t) represents the internal temperature of the electrolytic cell, dT e (t) / dt is the temperature change rate, γ e is the electrothermal conversion coefficient, P in (t) is the instantaneous input power, h e is the heat dissipation coefficient, T env is the ambient temperature, C e is the equivalent heat capacity;

[0068]

[0069] Among them, τ e is the thermal inertia time constant,

[0070] Estimate the real-time inertia constant:

[0071]

[0072] in, is the inertia estimation value at the nth sampling moment, κ1 is the static compensation bias, κ2 is the dynamic adjustment coefficient, and are the temperatures at two adjacent moments, Δt s is the sampling period;

[0073] Generate discrete feedforward compensation coefficients:

[0074]

[0075] in, is the discrete feedforward compensation coefficient;

[0076] Calculate the instruction advance, the formula is:

[0077]

[0078] in, is the power instruction after advance compensation, and They are the current and last original power instructions respectively;

[0079] Output delay compensation:

[0080]

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

[0082] As a preferred solution of the hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device described in the present invention, the knowledge distillation unit performs:

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

[0084] Distillation temperature parameter adjustment: Dynamically adjust the softmax temperature parameter according to the current grid frequency fluctuation variance;

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

[0086] The beneficial effects of the present invention are as follows: the gradient heterogeneous compensation unit of the present invention integrates covariance alignment and differential privacy encryption, and 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 the interference of equipment differences through statistical distribution correction and generates a globally consistent frequency modulation strategy, which not only protects commercial sensitive information but also avoids model oscillation failure; the knowledge distillation unit further compresses the model volume to ensure the real-time response capability of the edge node.

[0087] When the grid frequency suddenly changes, the fast response protocol automatically triggers the dynamic priority allocation mechanism; the regional aggregation layer directly broadcasts the frequency regulation instructions based on the historical performance tags, skipping the federated learning iterative process, so that the hydrogen production cluster can form a regulation force within seconds; the priority tags are dynamically updated according to the site response delay rate to ensure that high-reliability sites take priority in key regulation tasks, completely eliminating the risk of frequency regulation failure caused by instruction lag in traditional solutions; the edge node of the present invention has a built-in start and stop counter and power smoothing filter, which monitors the equipment status in real time through the LSTM loss function embedded with physical constraints; when the electrolyzer approaches the start and stop limit, the slope limit and thermal inertia compensation module are automatically activated to dynamically adjust the power instruction change rate and time advance to ensure that the instructions comply with the equipment thermodynamic constraints; this mechanism avoids equipment overload caused by frequent rescheduling from the source and extends the life of key components. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0089] Figure 1 Schematic diagram of the process of hierarchical optimization scheduling of a power system containing a large-scale water electrolysis hydrogen production device in Example 1. DETAILED DESCRIPTION

[0090] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0091] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0092] Secondly, the term "one embodiment" or "embodiment" 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 various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0093] Example 1, with reference to Figure 1 This embodiment provides a hierarchical optimization scheduling method for a power system including a large-scale water electrolysis hydrogen production device, comprising:

[0094] At the local dispatching layer, each hydrogen production station is equipped with an edge computing node to collect real-time signals of electrolyzer temperature, hydrogen storage tank pressure, and grid frequency deviation;

[0095] Edge computing nodes execute:

[0096] Build an LSTM prediction model with frequency deviation, electrolyzer temperature, and tank pressure as inputs to output the hydrogen production power adjustment value;

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

[0098] Transmit the encrypted gradient to the regional aggregation layer through the blockchain encrypted channel;

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

[0100] The Federated Learning Coordinator includes:

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

[0102] Federal aggregation unit: performs weighted average calculation on the corrected gradient and updates the global frequency modulation decision model;

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

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

[0105] Extract site gradient statistics and calculate the nearest N hydrogen production station i i Upload gradient The statistic is calculated as follows:

[0106]

[0107] in, represents the gradient vector uploaded for the kth time, i is the hydrogen station number, k is the sample number, N i is the sample size, μ i is the site mean gradient vector, Σ i is the site covariance matrix;

[0108] The federation coordinator builds a global reference based on the statistics of each site, specifically:

[0109]

[0110] Where M is the total number of sites participating in the upload, μ0 is the global gradient mean vector, and Σ0 is the global gradient covariance matrix;

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

[0112]

[0113] Among them, A i is the homogeneous mapping matrix of site i, is the positive semidefinite square root of the global covariance, is the inverse square root of the site covariance, and the definition matrix Σ i have

[0114] Where,

[0115]

[0116] Among them, Q0,Q i is the eigenvector matrix, are the square root and inverse square root of the corresponding eigenvalue matrix;

[0117] Upload the gradient g for this round i Perform whitening and translation:

[0118]

[0119] Among them, g i is the current gradient vector, It is the middle gradient after whitening and centroid alignment;

[0120] Calculate the difference measure between the site and the global distribution, the formula is:

[0121] D i =‖Σ i -Σ0‖ F ,

[0122] Among them, D i is the heterogeneity index, ‖·‖ F represents the Frobenius norm;

[0123] Generate motion compensation coefficients:

[0124]

[0125] Among them, c i is the dynamic compensation coefficient, η is the compensation strength constant, ranging from 0 to 1, dimensionless, θ is the heterogeneity threshold constant, dimensionless;

[0126] Further smoothing is performed using a logistic function:

[0127]

[0128] Where λ is the slope adjustment coefficient;

[0129] When D i Compensation starts when ≥0.1θ,

[0130]

[0131] in, is 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 biases through mean shift. The heterogeneity index uses a single norm to measure high-dimensional differences. The subsequent compensation coefficient varies continuously with the heterogeneity to avoid oscillations caused by step-by-step correction. The logical function form accelerates convergence in large heterogeneous scenarios and automatically attenuates the compensation amount in small heterogeneous scenarios to maintain information integrity. The entire mapping does not rely on the physical parameters of the equipment and can work in real time under aging and load fluctuation environments, effectively reducing the number of convergence rounds and taking into account the requirements of real-time scheduling and differential privacy.

[0133] The operations of the gradient heterogeneous compensation unit include:

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

[0135] Dynamic compensation coefficient generation: Combines the grid frequency fluctuation amplitude and aging correction factor to generate real-time compensation coefficient;

[0136] Performing a correction operation on the original gradient based on the real-time compensation coefficient;

[0137] The knowledge distillation unit performs:

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

[0139] Distillation temperature parameter adjustment: Dynamically adjust the softmax temperature parameter according to the current 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 grid dispatch center generates power allocation coefficients based on the adjustable capacity fed back from the regional layer and issues emergency frequency regulation instructions;

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

[0143] The regional aggregation layer calls the preset fast response protocol and skips the federated learning iteration process;

[0144] Dynamic priority allocation: Generates priority labels based on historical FM response delay rates and broadcasts them to edge nodes;

[0145] Trigger edge nodes to instantly output power adjustment instructions based on priority labels;

[0146] It also includes constraint guarantee mechanisms:

[0147] Start and stop frequency monitoring: embed electrolytic cell start and stop counters in edge nodes to accumulate start and stop times 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 guarantee is achieved by adjusting the constraint penalty term coefficient in the LSTM loss function;

[0150] The operations of the power smoothing filter include:

[0151] Slope limit module: imposes preset limits on the rate of change of power regulation commands;

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

[0153] Outputting the smoothed power command to the execution unit of the hydrogen production device;

[0154] In the time delay compensation module, the dynamic compensation instruction delay method is:

[0155] Establish the electrolytic cell thermal inertia model:

[0156]

[0157] Among them, T e (t) represents the internal temperature of the electrolytic cell, dT e (t) / dt is the temperature change rate, γ e is the electrothermal conversion coefficient, P in (t) is the instantaneous input power, h e is the heat dissipation coefficient, T env is the ambient temperature, C e is the equivalent heat capacity;

[0158]

[0159] Among them, τ e is the thermal inertia time constant,

[0160] Estimate the real-time inertia constant:

[0161]

[0162] in, is the inertia estimation value at the nth sampling moment, κ1 is the static compensation bias, κ2 is the dynamic adjustment coefficient, and are the temperatures at two adjacent moments, Δt s is the sampling period;

[0163] Generate discrete feedforward compensation coefficients:

[0164]

[0165] in, is the discrete feedforward compensation coefficient;

[0166] Calculate the instruction advance, the formula is:

[0167]

[0168] in, is the power instruction after advance compensation, and They are the current and last original power instructions respectively;

[0169] Output delay compensation:

[0170]

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

[0172] Specifically, the process uses thermoelectric coupling differential equations to characterize the temperature rise mechanism of the electrolytic cell, converts the heat dissipation and heat storage parameters into time constants, and then uses the temperature slope to adaptively estimate the inertia size to avoid the fixed coefficient from failing under different load conditions. The discrete feedforward link injects an equal-amplitude advance into the original power instruction, so that the equivalent transfer function approximates a unit-gain pure delay, thereby suppressing the energy lag caused by thermal inertia. The compensation coefficient converges rapidly with the temperature change rate, reducing power overshoot in high-speed load mutation 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 the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device, characterized in that: include, At the local dispatching layer, each hydrogen production station is equipped with an edge computing node to collect real-time signals of electrolyzer temperature, hydrogen storage tank pressure, and grid frequency deviation; At the regional aggregation layer, a federated learning coordinator is deployed to receive the encrypted gradient parameters uploaded by each edge node; At the global optimization layer, the main network dispatching center generates a power allocation coefficient based on the adjustable capacity feedback from the regional layer and issues emergency frequency regulation instructions.

2. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 1, characterized in that: The edge computing node performs: Build an LSTM prediction model with frequency deviation, electrolyzer temperature, and tank pressure as inputs to output the hydrogen production power adjustment value; Gradient encryption operation: Add Laplace noise perturbation to the local model gradient to generate encrypted gradients that meet differential privacy requirements; The encrypted gradient is transmitted to the regional aggregation layer through the blockchain encrypted channel.

3. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 1, characterized in that: The federated learning coordinator includes: Gradient heterogeneity compensation unit: performs distribution alignment correction on the gradients uploaded by each hydrogen production station; Federal aggregation unit: performs weighted average calculation on the corrected gradient and updates the global frequency modulation decision model; Knowledge distillation unit: compresses the global model into a lightweight micro-model and sends it to the edge nodes.

4. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 3, characterized in that: In the gradient heterogeneous compensation unit, the distribution alignment correction process includes: Extract site gradient statistics and calculate the nearest N hydrogen production station i i Upload gradient The statistic is calculated as follows: in, represents the gradient vector uploaded for the kth time, i is the hydrogen station number, k is the sample number, N i is the sample size, μ i is the site mean gradient vector, Σ i is the site covariance matrix; The federation coordinator builds a global reference based on the statistics of each site, specifically: Where M is the total number of sites participating in the upload, μ0 is the global gradient mean vector, and Σ0 is the global gradient covariance matrix; Generate a mapping matrix using matrix square root and inverse square root: Among them, A i is the homogeneous mapping matrix of site i, is the positive semidefinite square root of the global covariance, is the inverse square root of the site covariance, and the matrix Σ is defined i have Where, Among them, Q0,Q i is the eigenvector matrix, are the square root and inverse square root of the corresponding eigenvalue matrix; Upload the gradient g for this round i Perform whitening and translation: Among them, g i is the current gradient vector, It is the middle gradient after whitening and centroid alignment; Calculate the difference measure between the site and the global distribution, the formula is: D i =‖S i -Σ0‖ F , Among them, D i is the heterogeneity index, ‖·‖ F represents the Frobenius norm; Generate motion compensation coefficients: Among them, c i is the dynamic compensation coefficient, η is the compensation strength constant, ranging from 0 to 1, dimensionless, θ is the heterogeneity threshold constant, dimensionless; Further smoothing is performed using a logistic function: Where λ is the slope adjustment coefficient; When D i Compensation starts when ≥0.1θ, in, is the aligned gradient vector sent to the federated aggregation unit.

5. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 3, characterized in that: The operation of the gradient isomerization compensation unit includes: Match the preset aging correction factor according to the equipment model code; Dynamic compensation coefficient generation: Combines the grid frequency fluctuation amplitude and aging correction factor to generate real-time compensation coefficient; A correction operation is performed on the original gradient based on the real-time compensation coefficient.

6. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 1, characterized in that: The emergency frequency modulation instruction response process includes: The regional aggregation layer calls the preset fast response protocol and skips the federated learning iteration process; Dynamic priority allocation: Generates priority labels based on historical FM response delay rates and broadcasts them to edge nodes; Trigger edge nodes to instantly output power adjustment instructions based on priority labels.

7. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 1, characterized in that: It also includes constraint guarantee mechanisms: Start and stop frequency monitoring: embed electrolytic cell start and stop counters in edge nodes to accumulate start and stop times 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 protection is achieved by adjusting the constraint penalty term coefficient in the LSTM loss function.

8. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 7, characterized in that: The operation of the power smoothing filter includes: Slope limit module: imposes preset limits on the rate of change of power regulation commands; Time delay compensation module: dynamically compensates for instruction delay based on the electrolytic cell thermal inertia model; Output the smoothed power command to the hydrogen production device execution unit.

9. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 8, characterized in that: In the time delay compensation module, the dynamic compensation instruction delay method is: Establish the electrolytic cell thermal inertia model: Among them, T e (t) represents the internal temperature of the electrolytic cell, dT e (t) / dt is the temperature change rate, γ e is the electrothermal conversion coefficient, P in (t) is the instantaneous input power, h e is the heat dissipation coefficient, T env is the ambient temperature, C e is the equivalent heat capacity; Among them, τ e is the thermal inertia time constant, Estimate the real-time inertia constant: in, is the inertia estimation value at the nth sampling moment, κ1 is the static compensation bias, κ2 is the dynamic adjustment coefficient, and are the temperatures at two adjacent moments, Δt s is the sampling period; Generate discrete feedforward compensation coefficients: in, is the discrete feedforward compensation coefficient; Calculate the instruction advance, the formula is: in, is the power instruction after advance compensation, and They are the current and last original power instructions respectively; Output delay compensation: in, is the final power command sent to the slope limit module.

10. A hierarchical optimization scheduling method for a power system containing a large-scale water electrolysis hydrogen production device according to claim 3, characterized in that: The knowledge distillation unit performs: Feature alignment loss calculation: constraining the output difference between the micro model and the global model at the key frequency feature layer; Distillation temperature parameter adjustment: Dynamically adjust the softmax temperature parameter according to the current grid frequency fluctuation variance; Generate a micro-frequency modulation decision model suitable for edge computing resources.

Images