A hydrogen fuel cell power generation system load prediction and health state evaluation method

Abstract

The application provides a hydrogen fuel cell power generation system load prediction and health state evaluation method, and relates to the field of data processing. In the method, the day-ahead reference load prediction result is obtained through VMD-Transformer, and the LightGBM residual correction is used to realize the intraday short-term load prediction; at the same time, the final health state estimation result is generated by combining the grey wolf optimization mechanism identification and the physical constraint TCN-Attention aging residual prediction, so as to generate the coordinated operation control instruction according to the load demand and the stack health state. The technical scheme provided by the application facilitates to improve the dynamic prediction accuracy of strong random non-stationary load and the real-time estimation accuracy of complex aging state of the fuel cell, thereby providing the decision basis for the coordinated operation control considering the load tracking and the life protection.

Description

Technical Field

[0001] This application relates to the technical field of data processing, specifically to a method for load prediction and health status assessment of a hydrogen fuel cell power generation system. Background Technology

[0002] With the development of technology, hydrogen fuel cell power generation systems require multi-timescale coordinated control across load fluctuations, fuel supply, stack safety, and economic dispatch. Load forecasting results affect day-ahead power planning and intraday power point tracking, while fuel cell health status estimation results influence stack permissible output capacity, power ramp-up limits, and lifetime protection strategies. In actual operation, load power data is influenced by factors such as user energy consumption behavior, weather conditions, grid connection status, distributed power source fluctuations, and temporary load connections, typically exhibiting strong randomness, non-stationarity, and multi-period coupling characteristics. If a single time series forecasting model is directly used to process the raw load power data, the model struggles to simultaneously extract long-term trends, periodic variations, and short-term disturbances. Prediction errors tend to accumulate as the prediction time domain lengthens, affecting the accuracy of subsequent coordinated control.

[0003] Meanwhile, fuel cell stacks are affected by factors such as current, gas flow rate, gas pressure, temperature, cooling status, and membrane resistance changes during long-term operation. Their output voltage slowly decays with aging, and is also subject to measurement noise, local spikes, and operational disturbances. Existing health status estimation methods, while relying solely on semi-empirical mechanistic models offer physical interpretability, suffer from difficulties in accurately obtaining model parameters online, and fixed-parameter models struggle to track the dynamic degradation process throughout the entire lifecycle. Data-driven models, while capable of learning nonlinear relationships, tend to ignore the electrochemical constraints of the fuel cell, resulting in predictions that violate physical boundaries, insufficient generalization ability, and weak interpretability. Therefore, achieving both high-precision dynamic prediction of strongly stochastic non-stationary loads and high-precision real-time estimation of the complex aging state of fuel cells has become a critical technical challenge for the coordinated operation of hydrogen fuel cell power generation systems.

[0004] Therefore, in the scenario of coordinated operation of hydrogen fuel cell power generation systems across multiple time scales, there is an urgent need for a processing method that can combine non-stationary load decomposition prediction, intraday residual rolling correction, mechanistic reference modeling, and physical constraint aging residual prediction to improve the dynamic prediction accuracy of strongly random non-stationary loads and the real-time estimation accuracy of complex aging states of fuel cells. This will provide a decision-making basis for coordinated operation control that takes into account both load tracking and lifespan protection. Summary of the Invention

[0005] This application provides a method for load prediction and health status assessment of a hydrogen fuel cell power generation system, which can improve the dynamic prediction accuracy of strongly random non-stationary loads and the real-time estimation accuracy of complex aging states of fuel cells, thereby providing a decision-making basis for coordinated operation control that takes into account both load tracking and life protection.

[0006] The first aspect of this application provides a method for load forecasting and health status assessment of a hydrogen fuel cell power generation system. The method includes: acquiring historical load power data and fuel cell operating status data of the hydrogen fuel cell power generation system; cleaning and normalizing the historical load power data to form a continuous load power sequence, and performing variational mode decomposition on the continuous load power sequence to obtain multiple intrinsic mode components; inputting each intrinsic mode component into a Transformer prediction network to generate mode prediction results, and reconstructing the mode prediction results to form a day-ahead reference load forecast result; forming a historical residual sequence based on the deviation between real-time load power data and the day-ahead reference load forecast result, and then using LightGBM residual analysis. The differential correction model performs rolling predictions to generate intraday short-term load forecast results. The stack output voltage in the fuel cell operating status data is denoised and smoothed, and key operating parameters related to the stack output voltage are selected to form a health state estimation input feature set. Simultaneously, the gray wolf optimization algorithm is used to identify the fuel cell semi-empirical mechanism model, forming a fuel cell reference voltage model. The health state estimation input feature set is input into a TCN-Attention network constrained by physical boundaries to generate fuel cell aging residual prediction results. Based on the fuel cell reference voltage model and the fuel cell aging residual prediction results, a final health state estimation result is generated, which, combined with the intraday short-term load forecast results, generates coordinated operation control commands.

[0007] A second aspect of this application provides a device for load forecasting and health status assessment of a hydrogen fuel cell power generation system. The device includes an acquisition module and a processing module. The acquisition module is used to acquire historical load power data and fuel cell operating status data of the hydrogen fuel cell power generation system. The processing module is used to clean and normalize the historical load power data to form a continuous load power sequence, and to perform variational mode decomposition on the continuous load power sequence to obtain multiple intrinsic mode components. The processing module is further used to input each of the intrinsic mode components into a Transformer prediction network to generate mode prediction results, and to reconstruct the mode prediction results to form a day-ahead reference load prediction result. The processing module is also used to generate historical residuals based on the deviation between the real-time load power data and the day-ahead reference load prediction result. The sequence is used to generate intraday short-term load forecast results through rolling prediction using the LightGBM residual correction model. The processing module is also used to denoise and smooth the stack output voltage in the fuel cell operating status data, and to screen key operating parameters related to the stack output voltage to form a health status estimation input feature set. At the same time, the gray wolf optimization algorithm is used to identify the fuel cell semi-empirical mechanism model to form a fuel cell reference voltage model. The processing module is also used to input the health status estimation input feature set into a TCN-Attention network constrained by physical boundaries to generate fuel cell aging residual prediction results, and to generate a final health status estimation result based on the fuel cell reference voltage model and the fuel cell aging residual prediction results, so as to generate coordinated operation control commands in combination with the intraday short-term load forecast results.

[0008] A third aspect of this application provides an electronic device comprising a processor and a memory; the memory having a stored computer program, wherein the computer program, when executed by the processor, implements a method for load prediction and health status assessment of a hydrogen fuel cell power generation system as described above.

[0009] In a fourth aspect, this application provides a non-transitory computer-readable storage medium storing instructions that, when executed, perform a method for load prediction and health status assessment of a hydrogen fuel cell power generation system as described above.

[0010] In summary, one or more technical solutions provided in this application have at least the following technical effects or advantages: By coordinating load forecasting and health status estimation, the load tracking accuracy and operational safety of hydrogen fuel cell power generation systems can be improved simultaneously. By cleaning, normalizing, and performing variational mode decomposition on historical load power data, strongly stochastic and non-stationary loads can be broken down into intrinsic mode components of different frequency bands, reducing the difficulty of directly predicting the original load using a single model. Then, a Transformer prediction network is used to extract the long-term dependencies of loads in each frequency band, and a LightGBM residual correction model combined with real-time load deviations is used for intraday rolling correction, ensuring that the load forecast results have both day-ahead trend and intraday real-time accuracy. Simultaneously, by denoising the stack output voltage, screening key operating parameters, identifying gray wolf optimization mechanism parameters, and using TCN-Attention aging residual prediction constrained by physical boundaries, both mechanism interpretability and data-driven tracking capabilities can be considered, improving the accuracy of fuel cell health status estimation. Finally, coordinated operation control commands are generated based on the intraday short-term load forecast results and the final health status estimation results, enabling the system to meet load demands while limiting stack overload and accelerated aging. Therefore, it facilitates the improvement of dynamic prediction accuracy for strongly random non-stationary loads and real-time estimation accuracy for complex aging states of fuel cells, thereby providing a decision-making basis for coordinated operation control that takes into account both load tracking and life protection. Attached Figure Description

[0011] Figure 1 A flowchart illustrating a method for load prediction and health status assessment of a hydrogen fuel cell power generation system provided in this application embodiment; Figure 2 A schematic diagram of a load prediction and health status assessment device for a hydrogen fuel cell power generation system provided in this application embodiment; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0012] Explanation of reference numerals in the attached figures: 21. Acquisition module; 22. Processing module; 31. Processor; 32. Communication bus; 33. User interface; 34. Network interface; 35. Memory. Detailed Implementation

[0013] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0014] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0015] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. In addition, the terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0016] To address the aforementioned technical problems, this application provides a method for load prediction and health status assessment of a hydrogen fuel cell power generation system, referring to... Figure 1 , Figure 1 This is a flowchart illustrating a method for load prediction and health status assessment of a hydrogen fuel cell power generation system, provided in an embodiment of this application. The method is applied to a server and includes steps S110 to S160, as follows:

[0017] S110. Obtain historical load power data and fuel cell operating status data of the hydrogen fuel cell power generation system.

[0018] Specifically, when the server collects the active power time series from the load-side power metering points according to a unified sampling clock, it first determines the load-side power metering points in the hydrogen fuel cell power generation system. The load-side power metering point refers to the power collection location that reflects the system's external power supply demand; it can be located on the AC load side, DC bus output side, or grid-connected metering side. The server sends sampling configurations to the load-side power metering points, enabling them to upload active power values, sampling times, and metering point identifiers at fixed sampling intervals. The active power time series is a set of active power values ​​arranged chronologically, used to represent the actual load demand changes faced by the hydrogen fuel cell power generation system during historical operation. To avoid discrepancies between the local clocks of different acquisition devices, the server uses a unified sampling clock as the time reference, mapping each sampling moment to the same time axis, allowing subsequent load forecasting and health status estimation to be correlated based on the same operating time. If multiple instantaneous power sampling values ​​exist within a sampling window, the server can average these values ​​to obtain the active power value corresponding to that sampling moment. The formula for calculating the active power value is:

[0019] in, Indicates the first The active power value corresponding to each sampling moment is obtained by summing the instantaneous power sampling values ​​within the same sampling window; Indicates the first The number of instantaneous power samples actually obtained within each sampling window is a positive integer. The server obtains this number by counting the power sampling records received within the sampling window. Indicates the first Within the sampling window, the first Each instantaneous power sample value is directly collected from the power metering point on the load side; This indicates the sequence number of the instantaneous power sample value within the same sampling window. This formula reduces the impact of single-point acquisition jitter on the active power time series by calculating the average of multiple instantaneous power sample values ​​within the same sampling window, enabling historical load power data to more stably represent the actual load demand.

[0020] When the server uses the active power time series as historical load power data, it not only saves the active power value itself, but also simultaneously saves the sampling time, metering point identifier, power supply mode identifier, and data quality identifier. The sampling time is used to determine the operating time corresponding to each active power value; the metering point identifier is used to distinguish power metering points on different load sides; the power supply mode identifier is used to indicate whether the hydrogen fuel cell power generation system is in grid-connected power supply, off-grid power supply, energy storage-assisted power supply, or load tracking power supply mode; and the data quality identifier is used to record whether there is a data acquisition delay, communication packet loss, metering anomalies, or data completion processing. Through the above processing, the historical load power data is not an isolated power value, but historical load power data with time meaning, equipment source, and operating background, providing basic data for subsequent cleaning, normalization processing, variational mode decomposition, and Transformer prediction network training.

[0021] When the server collects voltage, current, flow rate, pressure, and temperature data for the fuel cell stack, hydrogen supply circuit, air supply circuit, and cooling circuit, it first reads the stack output voltage and output current at the fuel cell stack output terminal according to a unified sampling clock. Then, it reads the hydrogen flow rate, hydrogen pressure, and hydrogen temperature in the hydrogen supply circuit, the air flow rate, air pressure, and air temperature in the air supply circuit, and the cooling temperature and cooling flow rate in the cooling circuit. The fuel cell stack is the core component of a hydrogen fuel cell power generation system that completes the electrochemical reaction and outputs electrical energy. The hydrogen supply circuit provides the fuel cell stack with the hydrogen required for the reaction, the air supply circuit provides the oxygen required for the reaction, and the cooling circuit removes heat from the stack during operation. The stack output voltage directly reflects the performance status of the fuel cell stack, the stack output current reflects the current power generation load level, and the flow rate, pressure, and temperature data reflect the reactant supply and thermal management status. Therefore, these data together constitute the fuel cell operating status data.

[0022] When the server uses the collected data as fuel cell operating status data, it needs to ensure that the stack output voltage, stack output current, hydrogen flow rate, air flow rate, hydrogen pressure, air pressure, gas temperature, cooling temperature, and cooling flow rate all correspond to the same sampling time. If there is a slight offset in the upload time of different sensors, the server uses the sampling time of a unified sampling clock as a reference and performs recent time matching or short-time interpolation to complete the data from each acquisition channel, so that the data from each acquisition channel can be combined to form fuel cell operating status data at the same sampling time. If a certain acquisition channel lacks data at the current sampling time, but valid data exists at the sampling times before and after it, the server can perform linear completion based on the valid data before and after it. The formula for calculating the linear completion value is:

[0023] in, Indicates the sampling times that need to be supplemented. The corresponding running status data values ​​are calculated from the valid sampled values ​​before and after. Indicates the sampling time Previous most recent valid sampling time The corresponding operating status data values ​​are actually collected by the corresponding sensors; Indicates the sampling time The most recent valid sampling time The corresponding operating status data values ​​are actually collected by the corresponding sensors; The sampling times that need to be supplemented are determined by a unified sampling clock; This indicates the previous valid sampling time, which is retrieved by the server from the historical sampling records. This indicates the next valid sampling time, retrieved by the server from the sampling records. The formula generates a filler value based on the position of the missing sampling time between the preceding and following valid sampling times, following the changing trend of the valid data. This ensures the continuity of fuel cell operating status data over time and reduces the impact of single-channel missing data on the construction of subsequent health status estimation input feature sets.

[0024] When the server associates historical load power data and fuel cell operating status data at the same sampling time into a system operation record based on time identifiers, it first configures the historical load power data and fuel cell operating status data with the same format of time identifiers. A time identifier is a time field that uniquely represents the sampling time; it can be generated by a unified sampling clock and written synchronously with the collected records. Using the time identifier as an index, the server combines the active power value, stack output voltage, stack output current, hydrogen flow rate, air flow rate, hydrogen pressure, air pressure, gas temperature, cooling temperature, and cooling flow rate at the same sampling time into a single system operation record. The system operation record describes the load demand and fuel cell operating status of the hydrogen fuel cell power generation system at the same operating time, enabling the subsequent load prediction branch to read historical load power data and the health status estimation branch to read fuel cell operating status data, ensuring that both are traceable, comparable, and jointly controllable in time.

[0025] S120. The historical load power data is cleaned and normalized to form a continuous load power sequence, and the continuous load power sequence is subjected to variational mode decomposition to obtain multiple intrinsic mode components.

[0026] Specifically, when identifying abnormal load points in historical load power data, the historical load power data is first sorted according to the sampling time to obtain the original load power sequence. Historical load power data is the active power time series obtained by the server from the load-side power metering points. Abnormal load points refer to power sampling points that significantly deviate from the normal load change pattern, such as power points exceeding the rated operating range of the equipment, power points with excessively large abrupt changes between adjacent sampling times, and power points that maintain the same value for a long time and do not conform to actual load changes. The server can first identify abnormal load points based on the statistical distribution of the original load power sequence, marking sampling points that deviate from the mean by more than a preset range as abnormal load points. The judgment expression for abnormal load points is:

[0027] in, Indicates the first The active power value corresponding to each sampling time is collected by the load-side power metering point according to a unified sampling clock. This indicates the sampling time sequence number, with values ​​increasing sequentially according to time. The average power value represents the original load power sequence, which is obtained by averaging all effective active power values ​​within a set historical time window by the server. The power fluctuation of the original load power sequence is obtained by the server based on the dispersion of all effective active power values ​​within a set historical time window relative to the average power value. The anomaly detection coefficient can be set according to the load fluctuation tolerance of the hydrogen fuel cell power generation system. It is usually a positive number; a larger value indicates a more lenient anomaly detection, while a smaller value indicates a more stringent one. The above expression identifies sampling points that clearly do not conform to the overall load statistics by comparing the deviation of a single active power value from the historical average level, thereby preventing sensor momentary distortion or communication anomalies from entering subsequent continuous load power sequences.

[0028] When identifying missing load points in historical load power data, the server checks the time interval between adjacent sampling times according to a unified sampling clock. A missing load point is a load point where a valid sampling record was not obtained at a sampling time when an active power value should have been present; this is usually caused by communication packet loss, offline acquisition equipment, or data transmission delays. If the time interval between two adjacent valid sampling records is greater than the unified sampling interval, the server fills in the corresponding sampling time between the two adjacent valid sampling records and marks that sampling time as a missing load point. For identified abnormal load points, the server first treats them as invalid sampling points, and then performs the same completion process as for missing load points to ensure that abnormal removal and missing completion use the same time continuity rules.

[0029] When removing and filling in abnormal load points and missing load points, the server first removes the original active power value corresponding to the abnormal load point, and then generates a filler power value based on the nearest effective load point before and after the abnormal or missing load point. An effective load point refers to a sampling point that is not marked as abnormal and has an effective active power value. If effective load points exist before and after the gap, a linear filler method can be used to ensure that the power change after filler is continuously connected to the load trend before and after. The expression for the filler power value is:

[0030] in, Indicates the sampling times that need to be supplemented. The corresponding power compensation value is calculated by the server based on the effective load points before and after. Indicates the sampling time Previous most recent valid sampling time The corresponding active power value is obtained from the valid sampling records in the historical load power data; Indicates the sampling time The most recent valid sampling time The corresponding active power value is obtained from the valid sampling records in the historical load power data; The sampling times that need to be supplemented are generated by a unified sampling clock; Indicates the sampling time The most recent valid sampling time was obtained by the server searching backwards along the timeline; Indicates the sampling time The most recent valid sampling time is then retrieved by the server along the time axis. This expression generates the supplementary power value based on the position of the sampling time to be supplemented between the previous and next valid sampling times, according to the changing trend of the valid power values ​​before and after, so that the load sequence after removing abnormal load points and filling missing load points remains continuous in time.

[0031] When generating the cleaned load power sequence, the server rearranges the original effective active power values ​​and the supplemented power values ​​according to the sampling time, and organizes duplicate sampling times, out-of-order sampling times, and repeatedly uploaded power records, so that only one corresponding active power value is retained for each sampling time. The cleaned load power sequence is a load power sequence that has removed obvious acquisition anomalies and completed missing data supplementation, and it still maintains the temporal variation structure of historical load power data. After generating the cleaned load power sequence, the server continues to perform continuity checks. If there are still time discontinuities between adjacent sampling times that do not conform to the unified sampling interval, the server returns to the missing load point identification and supplementation process until the cleaned load power sequence meets the temporal continuity requirements of subsequent normalization processing and variational mode decomposition.

[0032] When normalizing the cleaning load power sequence, the server converts the active power values ​​at each sampling time to a uniform range based on the maximum and minimum power values ​​in the sequence, forming a continuous load power sequence. Normalization reduces the impact of power amplitude differences under different load scales, operating dates, and power supply modes on subsequent variational mode decomposition and Transformer prediction network training, while preserving the relative relationship of load power changes over time. The normalized expression for the continuous load power sequence is:

[0033] in, Indicates the first The normalized power value corresponding to each sampling time is used as a sequence point in the continuous load power sequence; Indicates the first The cleaned power value corresponding to each sampling time is derived from the original effective active power value or the supplemented power value; This represents the minimum power value in the cleaning load power sequence, which is obtained by the server traversing the cleaning load power sequence. This represents the maximum power value in the cleaning load power sequence, which is obtained by the server traversing the cleaning load power sequence. This represents the sampling time sequence number, and its value is consistent with the time order of the cleaning load power sequence. This expression limits the range of load power variation by using the maximum and minimum power values, mapping load data at different amplitude levels to a uniform scale. This allows the subsequent decomposition process to focus more on the load variation pattern rather than being dominated by the original numerical value.

[0034] If the load power sequence being cleaned contains extreme load days leading to significant differences between the maximum and minimum power values, the server can record the sampling time, power supply mode identifier, and data quality identifier corresponding to the maximum and minimum power values ​​before normalization. This information is useful for subsequent denormalization and anomaly tracing. After normalization, the server arranges the normalized power values ​​into a continuous load power sequence according to the sampling time. This continuous load power sequence maintains the long-term fundamental trend, daily peak-valley variations, and short-term fluctuations of the original load, while also meeting the requirements of continuous input signals for subsequent variational mode decomposition.

[0035] Before performing variational mode decomposition on the continuous load power sequence, the server first sets the number of mode decompositions, the penalty factor, and the convergence tolerance corresponding to the continuous load power sequence. The number of mode decompositions refers to the expected number of intrinsic mode components to be decomposed from the continuous load power sequence, which can be set according to the complexity of long-term trends, daily periodic variations, half-day peak-valley variations, and short-term fluctuations in historical load. The penalty factor is used to constrain the bandwidth of each intrinsic mode component; the larger the penalty factor, the more the intrinsic mode components tend to be frequency-concentrated. The convergence tolerance is used to determine whether two adjacent intrinsic mode components have stabilized during the iteration process; the smaller the convergence tolerance, the more rigorous the decomposition process. The server treats the continuous load power sequence as the load signal to be decomposed and controls the decomposition accuracy and computational stability through the number of mode decompositions, the penalty factor, and the convergence tolerance.

[0036] The goal of variational mode decomposition (VMD) is to ensure that each intrinsic mode component (IMC) can reconstruct a continuous load power sequence, while also achieving a frequency distribution as concentrated as possible around its center frequency. An IMC is a subsequence separated from the continuous load power sequence, possessing relatively independent frequency characteristics and variation patterns. The center frequency is the frequency location corresponding to the concentration of the main fluctuation energy of the IMC. The server achieves the separation of load variation patterns by minimizing the bandwidth of each IMC. The objective expression of VMD is:

[0037] in, This indicates the number of mode decompositions, which is set by the server based on the complexity of load changes before performing variational mode decomposition. Indicates the first Each intrinsic mode component at the sampling time The corresponding sequence values ​​are obtained by iterative solution of variational mode decomposition; Indicates the intrinsic modal component index, with values ​​ranging from... arrive Integers; Indicates the first The center frequency of each intrinsic mode component is obtained by the server during the iteration process based on the spectral energy distribution of that intrinsic mode component; This indicates the continuous load power sequence at the sampling time. The corresponding normalized power value is obtained from the aforementioned normalization process; This represents the unit impulse function, used in the construction of analytic signals; Represents the imaginary unit, used to represent complex numbers in frequency domain processing; This represents the convolution operation, used to generate the analytical form of eigenmode components; This represents the gradient calculated over the sampling time, used to measure the frequency spread of the intrinsic mode components; The square norm is used to measure the bandwidth of each intrinsic mode component. This expression decomposes a strongly random non-stationary load into multiple intrinsic mode components with more concentrated frequencies by constraining the superposition of the intrinsic mode components to equal the continuous load power sequence, while minimizing the frequency bandwidth of each intrinsic mode component.

[0038] When performing variational mode decomposition on a continuous load power sequence based on the number of mode decompositions, penalty factor, and convergence tolerance, the server can employ an iterative approach using alternating direction multipliers. The server first initializes each intrinsic mode component, center frequency, and constraint multiplier. Then, in each iteration, it sequentially updates the intrinsic mode components, center frequency, and constraint multipliers, gradually causing each intrinsic mode component to converge around its corresponding center frequency, and gradually bringing the sum of all intrinsic mode components closer to the continuous load power sequence. During the frequency domain solution process, the... The update expression for each intrinsic mode component is:

[0039] in, Indicates the first The intrinsic mode components in the th eigenmode component at the th The frequency domain representation at the next iteration is calculated by the server in the current iteration; The frequency variable in the frequency domain is obtained by the server after performing a frequency domain transformation on the continuous load power sequence. The frequency domain representation of the continuous load power sequence is provided by the server. Obtained by performing a frequency domain transformation; Indicates the first The intrinsic mode components in the th eigenmode component at the th The frequency domain representation has been updated in this iteration; Indicates the first The intrinsic mode components in the th eigenmode component at the th Frequency domain representation in the next iteration; Indicates the first The frequency domain representation of the constraint multipliers in the next iteration is obtained by the server based on the reconstruction error of each intrinsic mode component. This represents the penalty factor, which is preset by the server and used to control the degree of frequency band concentration. Indicates the first The intrinsic mode components in the th eigenmode component at the th The center frequency at the next iteration; This represents the iteration number, which increments starting from the first iteration after initialization. This expression extracts the frequency components around their center frequency from the remaining load signal by subtracting the frequency domain content already interpreted by other intrinsic mode components when updating the current intrinsic mode component.

[0040] After updating each intrinsic mode component, the server continues to update the corresponding center frequency, ensuring that the center frequency reflects the current concentration of spectral energy for that intrinsic mode component. The expression for updating the center frequency is:

[0041] in, Indicates the first The intrinsic mode components in the th eigenmode component at the th The center frequency after the next iteration is calculated by the server based on the spectral energy distribution of the intrinsic mode component; Represents the frequency variable in the frequency domain; Indicates the first The intrinsic mode components in the th eigenmode component at the th Frequency domain representation at the next iteration; Indicates the first Each intrinsic mode component at frequency The expression represents the spectral energy at a given location; the integral interval represents the statistical distribution of spectral energy within the positive frequency range of the server. This expression determines the dominant frequency positions of the intrinsic mode components by weighting the spectral energy, allowing each intrinsic mode component to gradually form a stable center frequency during the iteration process.

[0042] After each iteration, the server determines whether convergence has been reached based on the degree of change in the intrinsic mode components obtained from two adjacent iterations. If the degree of change is greater than the convergence tolerance, it indicates that the intrinsic mode components are still unstable, and the server continues to execute the next iteration; if the degree of change is not greater than the convergence tolerance, the iteration stops, and the current intrinsic mode component is determined as the decomposition result. The convergence criterion expression is:

[0043] in, Indicates the number of modal decompositions; Indicates the first The intrinsic mode components in the th eigenmode component at the th Frequency domain representation at the next iteration; Indicates the first The intrinsic mode components in the th eigenmode component at the th Frequency domain representation at the next iteration; Indicates the first iteration in two consecutive iterations. The degree of change of each intrinsic modal component; Indicates the first The energy scale of each intrinsic mode component in the previous iteration is used to relativize the degree of change; This represents the convergence tolerance, pre-set by the server, used to determine whether the iteration has reached a stable state. This expression determines whether the variational mode decomposition has stabilized by comparing the relative changes of all intrinsic mode components in two adjacent iterations, avoiding premature stopping that could cause mode aliasing, and also avoiding excessive iteration that could waste computational resources.

[0044] After obtaining multiple intrinsic mode components (IMCs) with different center frequencies, the server transforms each IMC from the frequency domain back to the time domain and establishes a mode order according to the center frequency, either from high to low or from low to high. Low-frequency IMCs typically correspond to long-term load trends, mid-frequency IMCs typically correspond to daily periodic variations and semi-daily peak-valley variations, and high-frequency IMCs typically correspond to short-term disturbances, random energy consumption behavior, and distributed power source fluctuations. The server binds each IMC with its sampling time, center frequency, and mode order to form the mode sequence data used by the subsequent Transformer prediction network. Through this processing, the variations at different time scales in the continuous load power sequence are separated into multiple IMCs with clearer variation patterns, providing a data foundation for subsequently predicting each IMC separately and reconstructing the day-ahead baseline load forecast results.

[0045] S130. Input each intrinsic mode component into the Transformer prediction network to generate mode prediction results, and reconstruct the mode prediction results to form the day-ahead baseline load prediction results.

[0046] Specifically, when the server constructs modal sequence data based on the sampling time, center frequency, and modal order corresponding to each intrinsic modal component, it first reads each intrinsic modal component output by variational mode decomposition and retains its sampling time consistent with the continuous load power sequence. The sampling time is the time position generated by the unified sampling clock, used to ensure that different intrinsic modal components can be aligned on the time axis; the center frequency is the main fluctuation frequency of the corresponding intrinsic modal component, used to represent the load change frequency band corresponding to that intrinsic modal component; the modal order is the sequence number obtained by the server according to the center frequency from low to high or from high to low, used to subsequently distinguish the reconstruction positions of different intrinsic modal components. The server binds the sequence value, sampling time, center frequency, and modal order of the same intrinsic modal component at consecutive sampling times to form modal sequence data, so that each intrinsic modal component has clear time and frequency band attributes. The expression for the modal sequence data is:

[0047] in, Indicates the first The mode sequence data corresponding to each intrinsic mode component is constructed by the server based on the output results of variational mode decomposition. Indicates the first Each sampling moment is generated by a unified sampling clock and saved along with historical load power data; Indicates the first The intrinsic mode components in the th eigenmode component at the th The sequence values ​​at each sampling time point are obtained by variational mode decomposition; Indicates the first The center frequencies corresponding to each intrinsic mode component are obtained by iterative calculation using variational mode decomposition. The modal order is determined by the server based on the frequency of each center. This represents the total number of sampling times, obtained by the server counting the number of consecutive sampling records participating in the day-ahead prediction training. This expression binds the temporal variation values ​​of the intrinsic mode components to the center frequency and mode order, enabling the subsequent Transformer prediction network to handle load variation patterns in different frequency bands separately.

[0048] When the server generates modal training samples based on modal sequence data, it slides and truncates each modal sequence data according to a preset historical window length and a day-ahead prediction time domain. The historical window length refers to the number of continuous sampling points used to describe past load changes in a single training iteration, set by the server based on the sampling interval and load cycle length. The day-ahead prediction time domain refers to the number of future sampling points to be predicted, typically corresponding to the control requirements of day-ahead load forecasting. The server uses continuous sequence values ​​of the same intrinsic modal component within the historical window length as historical mode segments, and uses the immediately following continuous sequence values ​​within the day-ahead prediction time domain as prediction targets, combining the two to form modal training samples. The expression for constructing modal training samples is:

[0049] in, Indicates the first The intrinsic mode components in the th eigenmode component at the th The historical modal segments formed by the sliding starting points are continuously extracted by the server from the modal sequence data; This represents the prediction target corresponding to the historical mode segment, which is extracted by the server from continuous sampling points following the historical mode segment; Indicates the first The intrinsic mode components in the th eigenmode component at the th The sequence values ​​at each sampling time point; This indicates the length of the historical window, which is set by the server based on the historical span required for load prediction. This represents the number of sampling points corresponding to the day-ahead prediction time domain, which is calculated by the server based on the day-ahead prediction duration and sampling interval. This represents the starting point index of the sliding motion, which is obtained by the server moving step by step according to the sampling time sequence. This expression converts continuous modal sequence data into trainable modal training samples through a sliding truncation method, enabling the Transformer prediction network to learn the mapping relationship between historical modal segments and future modal changes.

[0050] Before feeding the modal training samples into the Transformer prediction network, the server first performs feature mapping on the historical modal segments, converting single sequence values ​​into modal feature representations of a unified dimension within the network. These modal feature representations serve as intermediate data for the Transformer prediction network's attention computation, expanding the payload sequence values ​​to multiple feature channels. Because the Transformer prediction network employs a parallel computing structure, it doesn't inherently recognize the temporal order of samples; therefore, the server needs to overlay temporal position markers onto the modal feature representations. These temporal position markers are encoded information representing the temporal position of a sample value within a historical modal segment, ensuring that the network retains the chronological order of different sampling times within the same intrinsic modal component. The expression for the modal feature representation is:

[0051] in, Indicates the first The intrinsic mode components in the th eigenmode component at the th The initial modal feature representations corresponding to each sliding start point are generated by the server before entering the Transformer prediction network; This represents historical modal fragments, provided by the modality training samples; This represents the feature mapping matrix, which is learned by the Transformer prediction network during training. This represents a time location marker, generated by the server based on the historical window length and network feature dimensions. This expression adds the numerical features of historical modal segments to the time location marker, enabling the Transformer prediction network to still identify the temporal order of samples during parallel processing.

[0052] Temporal location markers can be constructed using alternating sine and cosine functions, giving different encodings to different sampling locations while maintaining a continuous relationship between adjacent sampling locations. The server generates temporal location markers based on the position index of the sampled value within the historical modality segment and the feature dimensions of the Transformer prediction network, where even-numbered dimensions use sine functions and odd-numbered dimensions use cosine functions. The expression for the temporal location marker is:

[0053] in, Indicates the first The position in the The time location marker values ​​on each feature dimension are generated by the server based on a sine function; Indicates the first The position in the The time location marker values ​​on each feature dimension are generated by the server based on the cosine function; This indicates the position number of the sampled value in the historical modal segment, and the value is numbered by the server from front to back according to the length of the historical window. This represents half of the feature dimension index, and its value is determined by the feature dimension of the Transformer prediction network. This represents the feature dimension of the Transformer prediction network, set by the network structure parameters. This expression uses sine and cosine variations of different frequencies on different feature dimensions to enable the time location markers to distinguish both the absolute location of sampling moments and the relative positional relationships between different sampling moments.

[0054] When the server extracts long-term dependencies corresponding to each intrinsic mode component using the Transformer prediction network, it first converts the mode feature representation into query features, key features, and value features in each layer of the network. Query features indicate which historical locations need to be considered at the current sampling position, key features represent the temporal features that can be matched at each historical location, and value features carry the load change information that needs to be conveyed at each historical location. Long-term dependencies refer to the change associations that still exist between sampling times that are relatively long apart for the same intrinsic mode component, such as the continuation of low-frequency trends, the repetition of periodic peaks and troughs, and the attenuation of high-frequency fluctuations. The expressions for query features, key features, and value features are:

[0055] in, Indicates the first The query features corresponding to each attention head are obtained by multiplying the current layer modality feature representation with the query mapping matrix; Indicates the first The key features corresponding to each attention head are obtained by multiplying the current layer modality feature representation with the key mapping matrix; Indicates the first The value features corresponding to each attention head are obtained by multiplying the current layer modality feature representation with the value mapping matrix; Indicates the first The intrinsic mode components in the th eigenmode component at the th The first sliding start point, the Transformer prediction network's [number]th ... The modal feature representation of a layer is obtained from the output of the previous layer of the network; , and They represent the first The query mapping matrix, key mapping matrix, and value mapping matrix of each attention head are all obtained by training the Transformer prediction network; This represents the attention head number, set by the network structure parameters. This expression extracts multiple related perspectives from the same modality feature representation through different mapping matrices, providing a foundation for subsequent multi-head self-attention computation.

[0056] When a server computes long-term dependencies using multi-head self-attention, it matches query features with key features to obtain the degree of attention between different sampling times. Then, it uses this degree of attention to weight and summarize the value features, resulting in the temporal correlation output for each attention head. Multiple attention heads can learn the temporal correlation of the same intrinsic mode component from different perspectives. For example, one attention head might focus on long-term trend continuation, another on peak-valley repetition, and yet another on short-term fluctuation decay. The expression for multi-head self-attention is:

[0057] in, Indicates the first The temporal correlation features of each attention head output are calculated by the server during the forward propagation of the Transformer prediction network; Indicates the first The query features corresponding to each attention head; Indicates the first The key feature transpose corresponding to each attention head is obtained by the server performing a matrix transpose on the key features; Indicates the first The value features corresponding to each attention head; The feature dimension of a single attention head is derived from the feature dimension of the Transformer prediction network and the number of attention heads. This represents the normalization function, used to convert the matching results between different sampling times into levels of interest. This expression determines the strength of the influence of historical sampling times on the current prediction process by comparing the feature similarity between different sampling times, thereby extracting the long-term dependencies corresponding to the intrinsic mode components.

[0058] After multiple attention heads output temporal correlation features, the server concatenates these features and obtains a comprehensive temporal correlation feature through an output mapping matrix. Subsequently, the Transformer prediction network uses residual connections, normalization, and a feedforward network to stably propagate and nonlinearly transform the comprehensive temporal correlation feature, enabling long-term dependencies to be strengthened layer by layer. The expression for the comprehensive temporal correlation feature is:

[0059] in, This represents the comprehensive temporal correlation features, obtained by fusing the temporal correlation features of multiple attention heads; This indicates a concatenation operation, used to combine temporal correlation features output by different attention heads along the feature dimension; They represent the first The first to the second Temporal correlation features of the attention head output; This indicates the number of attention heads, which is set by the structural parameters of the Transformer prediction network. This represents the output mapping matrix, obtained by training the Transformer prediction network. This expression fuses the long-term dependencies extracted by different attention heads, enabling the Transformer prediction network to simultaneously utilize multiple temporally related perspectives to predict future changes in intrinsic mode components.

[0060] When the server generates mode prediction results for each intrinsic mode component based on long-term dependencies, it feeds the mode feature representation output from the last layer of the Transformer prediction network into the prediction output layer to obtain the mode prediction value corresponding to each prediction time in the day-ahead prediction domain. The mode prediction result refers to the prediction sequence of a certain intrinsic mode component in the future day-ahead prediction domain, which still retains the corresponding mode order and prediction time. The expression for the mode prediction result is:

[0061] in, Indicates the first The intrinsic mode components in the th eigenmode component at the th The modal prediction results corresponding to each sliding start point are obtained from the output of the Transformer prediction network. This represents the prediction mapping function corresponding to the Transformer prediction network, which is determined by the network parameters. Sure; This represents all learnable parameters in the Transformer prediction network, including the feature mapping matrix, attention mapping matrix, output mapping matrix, and feedforward network parameters, which are updated during the training process based on modality training samples. Represents historical modal fragments; This represents the time location marker. The expression indicates that the Transformer prediction network learns the future changes of intrinsic mode components in the day-ahead prediction time domain based on historical mode fragments and time location markers, and outputs the corresponding mode prediction results.

[0062] When aligning the modal prediction results according to modal order and prediction time, the server first reads the modal order and prediction time corresponding to each modal prediction result, and then places different modal prediction results at the same prediction time into the same reconstruction set. The prediction time refers to a future time point within the day-ahead prediction time domain, generated by the server based on the prediction start point, sampling interval, and day-ahead prediction time domain. The reconstruction set is a set of modal prediction values ​​from different intrinsic modal components at the same prediction time. The server ensures that each prediction time includes corresponding low-frequency, mid-frequency, and high-frequency modal prediction values ​​through modal order, avoiding misalignment and superposition of a certain intrinsic modal component at an incorrect prediction time.

[0063] When the server reconstructs the modal prediction results for the same prediction time, it superimposes the modal prediction values ​​of different intrinsic mode components at the same prediction time to obtain the day-ahead baseline load prediction value corresponding to that prediction time. These values ​​are then arranged consecutively according to the prediction times within the day-ahead prediction time domain to form the day-ahead baseline load prediction result. The day-ahead baseline load prediction result is a total load prediction sequence reconstructed from multiple modal prediction results. It simultaneously includes the long-term fundamental trend corresponding to the low-frequency intrinsic mode components, the periodic changes corresponding to the mid-frequency intrinsic mode components, and the short-term fluctuations corresponding to the high-frequency intrinsic mode components. The reconstruction expression for the day-ahead baseline load prediction result is:

[0064] in, Indicates the predicted time The corresponding day-ahead baseline load forecast is obtained by reconstructing the modal forecast values ​​at the same forecast time from the server. Indicates the number of predictions within the time domain in the previous period. Each prediction time point is generated by the server based on the prediction start point and sampling interval; Indicates the first Each intrinsic mode component at the prediction time The corresponding modal prediction values ​​are obtained from the modal prediction results output by the Transformer prediction network; This indicates the number of intrinsic modal components, i.e., the number of modal decompositions set in variational mode decomposition; This indicates the modal order. This expression reconstructs the total load forecast at the same forecast time by superimposing the modal forecast values ​​of different frequency bands. This allows the day-ahead baseline load forecast to take into account the long-term trend, periodic changes, and short-term disturbances of the load.

[0065] After generating the day-ahead baseline load forecast, the server binds the forecast result with the forecast start point, forecast time, forecast time domain, modal order, and model version identifier. The model version identifier records the Transformer forecast network version used when generating the day-ahead baseline load forecast, facilitating subsequent forecast result tracking and model updates. The day-ahead baseline load forecast will serve as the benchmark data for subsequent intraday short-term load forecasts. When the server obtains real-time load power data, it can retrieve the day-ahead baseline load forecast result at the same time based on the real-time sampling time and further generate a historical residual sequence. Thus, the modal prediction process of the intrinsic modal components is continuously linked to the rolling prediction process of the subsequent LightGBM residual correction model.

[0066] S140. Based on the deviation between real-time load power data and the day-ahead baseline load forecast results, a historical residual sequence is formed, and the LightGBM residual correction model is used for rolling forecasting to form the intraday short-term load forecast results.

[0067] Specifically, when acquiring real-time load power data according to the sampling interval consistent with the day-ahead baseline load forecast, the server first reads the forecast timetable corresponding to the day-ahead baseline load forecast and determines the adjacent forecast time intervals in the forecast timetable as the intraday rolling correction sampling interval. Real-time load power data refers to the active power data collected in real-time by the load-side power metering point during the actual intraday operation of the hydrogen fuel cell power generation system. The real-time sampling time refers to the acquisition time corresponding to the real-time load power data. The server continuously receives real-time load power data according to the sampling interval and retrieves the baseline load forecast value at the same time from the day-ahead baseline load forecast based on the real-time sampling time. If the real-time sampling time is completely consistent with the forecast time in the day-ahead baseline load forecast, a direct match is performed; if there is a slight time offset, the server determines the forecast time closest to the real-time sampling time and within the allowable offset range as the matching time, thereby ensuring that the real-time load power data and the day-ahead baseline load forecast are compared at the same time position. The matching time can be determined by the following expression:

[0068] in, The predicted time that matches the real-time sampling time is retrieved by the server from the predicted time table corresponding to the day-ahead baseline load prediction results; This indicates the real-time sampling time, which is synchronously uploaded by the load-side power metering point along with the real-time load power data. This indicates the candidate forecast time in the day-ahead baseline load forecast results, which is generated by the server based on the day-ahead forecast start point and sampling interval. This represents the set of all forecast times corresponding to the day-ahead baseline load forecast results, which is saved by the server when the day-ahead baseline load forecast results are generated; This expression selects the prediction time with the smallest time deviation from the set of candidate prediction times. By using the minimum time deviation retrieval method, it establishes a correspondence between real-time load power data and the closest day-ahead baseline load prediction result, avoiding distortion in subsequent load predictions due to minor sampling clock deviations.

[0069] When comparing real-time load power data with the day-ahead baseline load forecast at the same time, the server subtracts the day-ahead baseline load forecast value from the real-time load power value at the same matching time to obtain the load forecast deviation. The load forecast deviation refers to the deviation of the actual load during the day from the day-ahead forecast load, and is used to characterize the forecast error of the day-ahead baseline load forecast in actual operation during the day. If the load forecast deviation is positive, it indicates that the actual load is higher than the day-ahead baseline load forecast, and the hydrogen fuel cell power generation system may need to increase its output power or call up compensation power; if the load forecast deviation is negative, it indicates that the actual load is lower than the day-ahead baseline load forecast, and the hydrogen fuel cell power generation system may need to reduce its output power or adjust its power supply strategy. The expression for the load forecast deviation is:

[0070] in, Indicates the real-time sampling time The corresponding load forecast deviation is obtained by comparing real-time load power data with the day-ahead baseline load forecast results from the server. Indicates the real-time sampling time The corresponding real-time load power value is obtained in real time from the power metering point on the load side; Indicates the matching prediction time. The corresponding day-ahead baseline load forecast value is obtained by the server from the day-ahead baseline load forecast results; Indicates the real-time sampling time; This represents the predicted time that matches the real-time sampling time. This expression quantifies intraday load changes not covered by the day-ahead forecast by using the difference between the actual load and the day-ahead predicted load at the same time location, providing a basis for subsequently constructing historical residual sequences.

[0071] When accumulating multiple load forecast deviations in chronological order to form a historical residual sequence, the server arranges the load forecast deviations corresponding to multiple consecutive real-time sampling times in chronological order and retains the load forecast deviations within the most recent period according to a preset rolling window length. The historical residual sequence is not a newly acquired load power sequence, but rather an error sequence formed by the continuous deviation between real-time load power data and the day-ahead baseline load forecast result. It is used to describe the changes in the day-ahead baseline load forecast result during intraday operation, such as systematic overestimation, systematic underestimation, short-term abrupt changes, continuous shifts, and gradual declines. The expression for the historical residual sequence is:

[0072] in, Indicates the time up to the real-time sampling time The resulting historical residual sequence is obtained by accumulating multiple load forecast deviations in chronological order by the server. to These represent the load prediction deviations corresponding to each real-time sampling moment within the rolling window, calculated from the aforementioned deviations; This indicates the number of sampling points corresponding to the length of the rolling window, which is set by the server based on the historical residual span required for intraday rolling forecasting. This indicates the earliest real-time sampling time within the scrolling window; This represents the latest real-time sampling time within the rolling window. By retaining the load forecast deviation over a recent period, this expression enables the LightGBM residual correction model to make rolling forecasts based on the latest error changes, rather than relying on a fixed, unchanging daily error distribution.

[0073] When constructing residual feature samples based on historical residual sequences, the server extracts the current residual value, residual change direction, residual duration, residual peak position, residual fallback status, and baseline load level from the historical residual sequences. The current residual value refers to the load forecast deviation corresponding to the latest real-time sampling moment within the rolling window, used to represent the immediate deviation of the current actual load from the day-ahead forecast load; the residual change direction refers to the increasing or decreasing trend of the current residual relative to the residual at the previous real-time sampling moment, used to determine whether the deviation is widening or narrowing; the residual duration refers to the length of time the load forecast deviation remains in the same deviation direction, used to distinguish between short-term disturbances and sustained offsets; the residual peak position refers to the sampling moment with the largest absolute value of the load forecast deviation within the rolling window, used to represent the time position of the deviation impact; the residual fallback status refers to the degree to which the residual recovers from the peak to the normal level, used to determine whether the intraday load disturbance is weakening; the baseline load level refers to the day-ahead baseline load forecast value at the corresponding moment, used to assist the model in judging the differences in residual changes under different load levels. The residual feature sample can be represented as:

[0074] in, Indicates the real-time sampling time The corresponding residual feature samples are constructed by the server based on historical residual sequences and day-ahead baseline load forecast results; This represents the current residual value, obtained by subtracting the day-ahead baseline load forecast from the real-time load power value; The residual change value corresponding to the direction of residual change is obtained by subtracting the residual value at the previous real-time sampling time from the current residual value; The duration of the residual is determined by the number of sampling points where the load forecast deviation remains continuously positive or negative, combined with the sampling interval, based on the server's statistical count of the number of sampling points. This indicates the location of the residual peak, which is obtained by the server finding the sampling moment with the largest absolute value of the load forecast deviation within the current scrolling window; This indicates the residual decline state, which is obtained by the server based on the degree of decrease of the current residual value relative to the peak residual; This represents the baseline load level, provided by the day-ahead baseline load forecast at the same matching forecast time. This expression transforms the simple historical residual sequence into a structured sample that reflects the magnitude, trend, persistence, impact location, and degree of decline of the residuals, enabling the LightGBM residual correction model to learn more stable intraday residual variation patterns.

[0075] The direction of residual change can be represented by the difference between the current residual value and the residual value at the previous real-time sampling time. The expression for the residual change value corresponding to the direction of residual change is as follows: in, Indicates the real-time sampling time The corresponding residual change value is calculated by the server based on the load prediction deviation between two adjacent real-time sampling times; This represents the current residual value; This indicates the load forecast deviation corresponding to the previous real-time sampling time, provided by the historical residual sequence; Indicates that it is located at The expression compares the residual changes between adjacent real-time sampling times to determine whether the load forecast deviation is increasing, decreasing, or remaining stable.

[0076] The residual fallback state can be represented by the degree of closeness between the current residual value and the peak residual within the scrolling window. The expression for the residual fallback state is: in, Indicates the real-time sampling time The corresponding residual fallback status is calculated by the server based on the load forecast deviation within the current rolling window; This represents the current residual value; Indicates the sampling time within the scrolling window The corresponding load forecast bias is provided by the historical residual sequence; This represents the set of sampling moments within the current scrolling window, determined by the server based on the scrolling window length. This represents the maximum absolute value of the load forecast deviation within the current scrolling window, obtained by the server iterating through the historical residual sequence. This represents a small constant to prevent the denominator from being zero, and is preset to a positive number by the server. This expression determines whether the residual disturbance has fallen from its peak state by comparing the current residual amplitude with the peak residual amplitude within the rolling window. The larger the value, the more obvious the fall; the smaller the value, the closer it is to the peak disturbance state.

[0077] When feeding residual feature samples into the LightGBM residual correction model for rolling prediction, the server first uses the LightGBM residual correction model to bin the continuous features in the residual feature samples, and then learns the mapping relationship between residual features and future residuals step by step through multiple decision trees. The LightGBM residual correction model is a residual prediction model based on gradient boosting decision trees. It constructs weak prediction trees through multiple iterations and accumulates the results of multiple weak prediction trees to form the residual prediction result. Rolling prediction refers to the server updating the historical residual sequence and residual feature samples after obtaining new real-time load power data, and predicting the residual changes in the short-term prediction domain. The expression for the residual prediction result is:

[0078] in, Indicates the time of real-time sampling The generated residual prediction results in the future short-term prediction time domain are obtained from the output of the LightGBM residual correction model; This represents the prediction mapping function corresponding to the LightGBM residual correction model, derived from the model parameters. Sure; This indicates that the structural parameters, leaf node output values, and splitting rules of each decision tree in the LightGBM residual correction model are learned from historical training data. Indicates the real-time sampling time The corresponding residual feature samples; to These represent the predicted residual values ​​at each future prediction time point; This indicates the number of prediction steps corresponding to the short-term forecast time domain, determined by the server based on the lead time and sampling interval required for intraday coordinated control. This expression represents how the LightGBM residual correction model predicts potential deviations in the day-ahead baseline load forecast results within the short-term forecast time domain, based on the residual change status within the current rolling window.

[0079] Within the LightGBM residual correction model, the predicted residual value at a future prediction time can be represented as the sum of the output values ​​of multiple decision trees. The expression for the predicted residual value is: in, Indicates the future number The predicted residual values ​​for each prediction time point are output by the LightGBM residual correction model. This represents the prediction step number within the short-term prediction time domain, with values ​​ranging from... arrive ; This indicates the number of decision trees in the LightGBM residual correction model, which is determined during the model training process. Indicates the first The prediction function corresponding to each decision tree is determined by the splitting rule of the decision tree and the output value of the leaf nodes; This represents the current residual feature sample; This represents the decision tree number. This expression accumulates the judgment results of multiple decision trees on the current residual feature sample, enabling the residual prediction value to simultaneously reflect the impact of multiple residual features on future bias changes.

[0080] When compensating the residual forecast results to the day-ahead baseline load forecast results at the corresponding forecast time, the server first matches each residual forecast value in the future short-term forecast time domain with the day-ahead baseline load forecast value at the same forecast time, and then adds the two together to obtain the intraday short-term load forecast value at the corresponding forecast time. The intraday short-term load forecast result refers to the load forecast result after superimposing the intraday rolling residual correction on the day-ahead baseline load forecast result. It retains both the long-term trend and periodic changes in the day-ahead baseline load forecast result and reflects the intraday deviation changes revealed by the real-time load power data. The expression for the intraday short-term load forecast value is:

[0081] in, Indicates the future number The intraday short-term load forecast value corresponding to each forecast time is obtained by the server adding the day-ahead baseline load forecast value and the residual forecast value; Indicates the future number The day-ahead baseline load forecast value corresponding to each forecast time is obtained by the server from the day-ahead baseline load forecast results; Indicates the future number The predicted residual values ​​for each prediction time are obtained from the output of the LightGBM residual correction model. Represents relative to the real-time sampling time The next Each prediction time is generated by the server based on the sampling interval; This indicates the prediction step number within the short-term forecast time domain. This expression compensates the future residual forecast value for the corresponding day-ahead baseline load forecast value, allowing the day-ahead forecast result to be dynamically corrected by the intraday real-time error pattern, thereby forming an intraday short-term load forecast result that is closer to the actual load demand.

[0082] After generating the intraday short-term load forecast, the server arranges multiple consecutive intraday short-term load forecast values ​​within the future short-term forecast time domain according to the forecast time, and binds them to a rolling window identifier, a model version identifier, a residual correction status, and a data quality identifier. The rolling window identifier records the range of historical residual sequences used to generate the intraday short-term load forecast; the model version identifier records the version of the LightGBM residual correction model; the residual correction status indicates whether the intraday short-term load forecast has completed residual compensation; and the data quality identifier indicates whether there is any incompleteness, delay, or anomaly in the real-time load power data. Through this processing, the intraday short-term load forecast results can be directly transmitted to the coordinated operation control link and used to determine the direction and magnitude of power demand changes within the future short-term forecast time domain.

[0083] S150. The stack output voltage in the fuel cell operating status data is denoised and smoothed, and key operating parameters related to the stack output voltage are selected to form a set of input features for health status estimation. At the same time, the semi-empirical mechanism model of the fuel cell is identified through the gray wolf optimization algorithm to form a fuel cell reference voltage model.

[0084] Specifically, when extracting the stack output voltage and corresponding operating parameters from the fuel cell operating status data according to the sampling time, the server first reads the timestamp from each fuel cell operating status data entry and arranges the stack output voltage, stack output current, hydrogen flow rate, air flow rate, hydrogen pressure, air pressure, gas temperature, cooling temperature, and cooling flow rate on the same operating time axis according to the timestamp. The stack output voltage is the voltage value output by the fuel cell stack under the current electrochemical reaction state, which can directly reflect the performance degradation state of the fuel cell stack. The operating parameters are the operating condition data that affect the change of the stack output voltage, used to characterize reactant supply, load intensity, and thermal management status. If a certain operating parameter is missing at the same sampling time, the server first marks the corresponding data quality status and fills it in under the condition that the effective data before and after the operating parameter is continuous, so that the stack output voltage and operating parameters maintain a corresponding relationship in time.

[0085] When supplementing abnormal voltage points in the fuel cell stack output voltage, the server first identifies abnormal voltage points based on the degree of local fluctuation in the fuel cell stack output voltage. Abnormal voltage points refer to abnormal voltage sampling points caused by acquisition jitter, communication packet loss, instantaneous sensor distortion, or electromagnetic interference. They typically manifest as short-term spikes, momentary drops, or long-term fixed values. The server can identify abnormal voltage points based on the voltage change amplitude between adjacent sampling times. The expression for determining abnormal voltage points is:

[0086] in, Indicates the first The stack output voltage corresponding to each sampling moment is obtained from the voltage acquisition channel at the output end of the fuel cell stack; Indicates the first The output voltage of the fuel cell at each sampling moment is obtained by the same voltage acquisition channel at the previous sampling moment; This indicates the sampling time sequence number, which is incremented by the server according to the time identifier. This represents the voltage anomaly judgment coefficient, which is preset by the server based on the allowable instantaneous voltage fluctuation range of the fuel cell stack. A larger value indicates a more lenient anomaly judgment, while a smaller value indicates a more stringent anomaly judgment. This represents the median absolute deviation of the fuel cell stack output voltage within a set historical window. The server first calculates the median of the fuel cell stack output voltage within that historical window, and then calculates the median of the absolute values ​​of the differences between the output voltage of each fuel cell stack and that median. This expression filters out short-term abnormal voltage points that clearly do not belong to the aging trend by comparing the voltage fluctuations at adjacent sampling times with the historical normal fluctuation levels.

[0087] When filling in the identified abnormal voltage points, the server sets the original voltage value corresponding to the abnormal voltage point to an invalid value and generates a filled voltage value based on the nearest valid voltage point before and after the abnormal voltage point. A valid voltage point refers to a sampling point that is not marked as abnormal and has a valid voltage value. If valid voltage points exist before and after the abnormal voltage point, the server uses a linear filling method to recover the stack output voltage at that sampling moment. The expression for the filled voltage value is:

[0088] in, Indicates the sampling time The corresponding supplementary voltage value is calculated by the server based on the effective voltage points before and after. Indicates the sampling time Previous most recent valid sampling time The corresponding fuel cell output voltage is actually obtained from the voltage acquisition channel; Indicates the sampling time The next most recent valid sampling time The corresponding fuel cell output voltage is actually obtained from the voltage acquisition channel; The sampling times that need to be supplemented are determined by a unified sampling clock; Indicates the location at the sampling time The most recent valid sampling time was obtained by the server searching backwards along the timeline; Indicates the location at the sampling time The most recent valid sampling time is retrieved by the server along the time axis. This expression recovers the voltage value based on the position of the sampling time to be supplemented between the previous and subsequent valid sampling times, along the trend of changes in the previous and subsequent valid voltage points. This ensures that the output voltage of the fuel cell stack remains continuous in time after supplementation and reduces the interference of abnormal voltage points on the subsequent aging state estimation.

[0089] When normalizing the padded fuel cell output voltage, the server converts the padded output voltage to a uniform value range based on the maximum and minimum voltage values ​​within a set sample range. Normalization reduces voltage amplitude differences between different operating batches, fuel cell states, and acquisition intervals, allowing subsequent wavelet denoising and correlation analysis to focus more on voltage change trends. The expression for the normalized fuel cell output voltage is:

[0090] in, Indicates the first The normalized voltage value corresponding to each sampling time is obtained by the server through normalization of the padded stack output voltage; Indicates the first The padded output voltage of the stack corresponding to each sampling time comes from the original effective voltage value or the padded voltage value. This represents the minimum voltage value within the set sample range, obtained by the server iterating through and completing the output voltage of the fuel cell stack. This indicates the maximum voltage value within the set sample range, which is obtained by the server iterating through and completing the output voltage of the fuel cell stack. This indicates the sampling time sequence number. This expression defines the voltage variation range by using the maximum and minimum voltage values, mapping different voltage levels to a uniform scale, thus facilitating subsequent analysis of voltage decay trends within the same numerical range.

[0091] When performing wavelet denoising on the normalized fuel cell stack output voltage, the server decomposes the normalized output voltage into a low-frequency degradation component and a high-frequency disturbance component. Wavelet denoising refers to the process of reconstructing the voltage sequence by performing multi-level decomposition of the voltage sequence using wavelet basis functions and weakening high-frequency noise. The low-frequency degradation component characterizes the slow performance degradation of the fuel cell stack over operating time; the high-frequency disturbance component mainly corresponds to measurement noise, local spikes, and short-term operating condition disturbances. The server performs threshold shrinkage on the high-frequency disturbance component and retains the low-frequency degradation component reflecting the aging trend to obtain the processed fuel cell stack output voltage. The wavelet denoising reconstruction expression is:

[0092] in, Indicates the sampling time The corresponding processed stack output voltage is obtained by the server performing wavelet denoising reconstruction on the normalized stack output voltage. This indicates the number of wavelet decomposition levels, which is set by the server based on the voltage sequence sampling length and noise frequency characteristics. Indicates the first The low-frequency coefficients of the layer are obtained by the server performing wavelet decomposition on the normalized stack output voltage, and are used to preserve the slow degradation trend. Indicates the first The layer scaling function is determined by the wavelet basis selected by the server. Indicates the first Layer A high-frequency coefficient, obtained by the server performing wavelet decomposition on the normalized stack output voltage, is used to describe high-frequency disturbances; This represents the threshold shrinkage function, which is determined by the server based on a threshold. Reduce the high-frequency coefficients; Indicates the first The layer wavelet function is determined by the selected wavelet basis; Indicates the wavelet decomposition level number; This indicates the coefficient number under the corresponding level. This expression preserves the low-frequency degradation part and compresses the high-frequency disturbance part, so that the processed stack output voltage retains the aging degradation trend while reducing the impact of acquisition noise and local spikes.

[0093] When performing correlation analysis between operating parameters and the processed stack output voltage, the server pairs each operating parameter with the processed stack output voltage according to the sampling time and calculates the correlation between the two. Correlation analysis is a processing method that evaluates the strength of the cooperative changes of two sequences on the same time axis; key operating parameters refer to operating parameters whose relationship with the processed stack output voltage changes meets preset conditions, reflecting operating conditions that have a strong impact on the health status of the fuel cell. The server can use the Pearson correlation coefficient to determine the degree of correlation, and the expression for the degree of correlation is:

[0094] in, Indicates the first The correlation between each operating parameter and the processed stack output voltage is calculated by the server based on paired data at the same sampling time. Indicates the first The operating parameter in the first The parameter values ​​corresponding to each sampling time are obtained from the fuel cell operating status data; Indicates the first The average value of each operating parameter within a set sample range is obtained by averaging all valid parameter values ​​of that operating parameter by the server. Indicates the first The processed stack output voltage corresponding to each sampling time; This represents the average value of the processed fuel cell output voltage within a set sample range, which is obtained by averaging the output voltages of all processed fuel cells by the server. This indicates the number of sampling points participating in the correlation analysis, which is obtained by the server counting the valid paired data; This indicates the sequence number of the operating parameter. This expression determines the strength of the linear correlation between the operating parameter and the voltage decay state by comparing whether the processed stack output voltage deviates synchronously from its average level when the operating parameter deviates from its average level.

[0095] When determining operational parameters whose relevance meets preset conditions as key operational parameters, the server compares the relevance of each operational parameter with a preset relevance threshold. If the absolute value of the relevance of a given operational parameter is not less than the preset relevance threshold, then that operational parameter is determined as a key operational parameter. The preset relevance threshold is set by the server based on sample size, the number of operational parameters, and model input dimensionality limitations. A higher value results in fewer key operational parameters being selected, but with stronger relevance; a lower value results in more key operational parameters being selected, but may introduce invalid operating condition disturbances. The expression for determining key operational parameters is:

[0096] in, This represents a set of key operating parameters, which are selected by the server based on their relevance. Indicates the first Each operating parameter corresponds to one of the following: fuel cell output current, hydrogen flow rate, air flow rate, hydrogen pressure, air pressure, gas temperature, cooling temperature, or cooling flow rate. Indicates the first The degree of correlation between each operating parameter and the processed fuel cell stack output voltage; This represents a preset relevance threshold, which is pre-set by the server or determined based on the validation results of training samples. This expression filters data by using a relevance threshold, ensuring that the subsequent health status estimation input feature set retains only operating parameters more closely related to changes in the stack output voltage, thus reducing the impact of redundant data on network training and health status estimation.

[0097] When combining the processed stack output voltage and key operating parameters according to the sampling time, the server uses the sampling time as an index to combine the processed stack output voltage and key operating parameters at the same sampling time into a health state feature vector, and arranges them in chronological order to form the health state estimation input feature set. This health state estimation input feature set is the temporal feature set used by the subsequent TCN-Attention network to learn the fuel cell aging residuals. The processed stack output voltage represents the direct change in the fuel cell health state, while the key operating parameters represent the operating conditions that affect this change. The expression for the health state feature vector is:

[0098] in, Indicates the first The health status feature vector corresponding to each sampling time is obtained by the server according to the sampling time. Indicates the first The processed stack output voltage corresponding to each sampling time; to They represent the first The parameter values ​​of each key operating parameter at each sampling time point were extracted from the fuel cell operating status data and screened through correlation analysis. This indicates the number of key operating parameters, which are obtained by the server based on preset relevant thresholds. This indicates the sampling time sequence number. This expression binds the voltage degradation state with the corresponding operating condition factors at the same sampling time, enabling subsequent aging residual prediction to utilize both the voltage change trajectory and the operating parameter change trajectory simultaneously.

[0099] When constructing a semi-empirical mechanistic model for a fuel cell to output a reference voltage result, the server establishes a voltage response relationship based on the electrochemical reaction mechanism of the fuel cell stack. A fuel cell semi-empirical mechanistic model is a model that combines mechanistic terms such as theoretical voltage, activation loss, ohmic loss, and concentration loss with empirical parameters to output a reference voltage result under given operating conditions. The reference voltage result is a voltage estimate obtained by the fuel cell semi-empirical mechanistic model under the current operating conditions and mechanistic parameters, providing a mechanistic benchmark that conforms to electrochemical laws. The expression for the fuel cell semi-empirical mechanistic model is:

[0100] in, The reference voltage result is calculated by the fuel cell semi-empirical mechanism model based on the current operating conditions. The reversible voltage is calculated from the reactant partial pressures corresponding to the fuel cell operating temperature, hydrogen pressure, and air pressure. This represents activation loss, caused by the resistance to the initiation of the electrochemical reaction, and is affected by the stack output current, operating temperature, and the physical parameters to be identified. This represents ohmic loss, caused by the equivalent resistance of the proton membrane, electrodes, and connecting components, and is affected by the stack output current, operating temperature, and the physical parameters to be identified. This represents concentration loss, caused by insufficient reactant transport under high load conditions, and is affected by the stack output current, limiting current, and the physical parameters to be identified. This expression yields a reference voltage result consistent with the electrochemical mechanism of a fuel cell by subtracting different loss terms from the theoretical reversible voltage.

[0101] When identifying the physical parameters to be identified in the semi-empirical mechanism model of a fuel cell as unknown parameters, the server first reads the operating conditions that can be directly obtained by sensors, such as stack output current, gas temperature, hydrogen pressure, air pressure, and gas flow rate. Then, parameters that cannot be directly measured online but affect the voltage response are classified as unknown parameters. The physical parameters to be identified may include activation loss coefficient, ohmic resistance coefficient, concentration loss coefficient, limiting current correction coefficient, or membrane impedance correction coefficient. The server combines the unknown parameters into an unknown parameter vector to facilitate unified optimization by the Gray Wolf optimization algorithm. The expression for the unknown parameter vector is:

[0102] in, The unknown parameter vector is obtained by the server by combining the physical parameters to be identified in the semi-empirical mechanism model of the fuel cell. to These represent different physical parameters to be identified, specifically derived from the coefficients of mechanistic terms in the semi-empirical mechanism model of fuel cells that cannot be directly measured; The number of unknown parameters is determined by the structure of the semi-empirical mechanism model of the fuel cell. This expression unifies the physical parameters to be identified, which are scattered across different voltage loss terms, enabling the Grey Wolf optimization algorithm to find parameter combinations that make the model voltage closely approximate the actual voltage within the same search space.

[0103] When iteratively optimizing unknown parameters using the Gray Wolf Optimization Algorithm, the server treats each candidate unknown parameter vector as a gray wolf individual position and compares the model voltage output by the fuel cell semi-empirical mechanism model under that candidate unknown parameter vector with the actual voltage obtained from the polarization experiment. The Gray Wolf Optimization Algorithm is a swarm intelligence optimization algorithm simulating the hunting behavior of a gray wolf pack. It guides other individuals to update their positions through the optimal, second-best, and third-best individuals. The model voltage is the voltage value output by the fuel cell semi-empirical mechanism model under the candidate unknown parameter vector; the actual voltage is the voltage value measured under the corresponding current condition in the polarization experiment. The server evaluates the quality of each gray wolf individual position through the fitting error, which is expressed as:

[0104] in, Represents the candidate unknown parameter vector The corresponding fitting error is calculated by the server based on the polarization experimental data and the model output data; This indicates the number of sampling points in the polarization experiment, which is obtained by reading the polarization experiment records from the server. Indicates the sampling point number of the polarization experiment; Indicates the first The actual voltage corresponding to each polarization experiment sampling point is obtained by polarization experiment measurement; Indicates the first Each polarization experiment sampling point in the candidate unknown parameter vector The corresponding model voltage is calculated from the semi-empirical mechanism model of the fuel cell. This expression measures the fitting accuracy of the candidate unknown parameter vector to the fuel cell voltage response by statistically analyzing the mean square difference between the model voltage and the actual voltage. The smaller the fitting error, the closer the candidate unknown parameter vector is to the true mechanism parameter state.

[0105] In each iteration of the gray wolf optimization, the server sorts all gray wolf individual positions according to their fitting errors from smallest to largest, and selects the three gray wolf individual positions with the smallest fitting errors as the optimal, second-best, and third-best individual positions. Subsequently, the server guides the updating of other gray wolf individual positions based on these three positions, causing the candidate unknown parameter vector to gradually move towards the region with smaller fitting errors. The expression for updating the gray wolf individual positions is:

[0106] in, This indicates that a certain gray wolf individual is in the [number]th [year]. The individual position after the next iteration corresponds to the updated candidate unknown parameter vector; This indicates that the individual gray wolf was in the [number]th [year]. The individual position at the next iteration corresponds to the candidate unknown parameter vector before the update; Indicates the first The optimal individual position in the next iteration is selected by the server based on the minimum fitting error; Indicates the first The suboptimal individual position in the next iteration is selected by the server based on the second smallest fitting error; Indicates the first The position of the third best individual in the next iteration is selected by the server based on the third smallest fitting error; , , These represent the search step size coefficients affected by the convergence factor and the random number, respectively, which are generated by the server in each iteration; , , These represent the perturbation coefficients, which are generated by the server in each iteration; This indicates the number of iterations. This expression, by simultaneously referencing the positions of three good candidate unknown parameter vectors, balances local convergence and global search, allowing the individual gray wolf to gradually approach the unknown parameter region that reduces fitting error.

[0107] When determining whether the fitting error meets the convergence condition, the server can simultaneously use a fitting error threshold and the maximum number of iterations for control. If the fitting error corresponding to the current optimal individual position is not greater than the preset error threshold, or the decrease in fitting error over adjacent iterations is less than the preset change threshold, the server determines that the Grey Wolf optimization algorithm has reached convergence. If the convergence condition is not met and the number of iterations has not reached the maximum number of iterations, the Grey Wolf individual position continues to be updated. The expression for the convergence condition is:

[0108] in, Indicates the first The fitting error corresponding to the optimal individual position in the next iteration is calculated from the fitting error expression. This indicates the preset error threshold, which is set by the server based on the accuracy requirements of the polarization experiment fitting. Indicates the first The fitting error corresponding to the optimal individual position in the next iteration; This represents a preset threshold for change, set by the server based on iterative stability requirements. This expression determines whether the optimization of the unknown parameters has reached a stable state by judging whether the fitting error is sufficiently small or whether the fitting error has essentially stopped decreasing.

[0109] When the unknown parameters satisfying the convergence condition are determined as the identified set of mechanistic parameters, the server extracts the candidate unknown parameter vector corresponding to the optimal individual position when convergence is achieved, and determines this candidate unknown parameter vector as the identified set of mechanistic parameters. Subsequently, the server substitutes the identified set of mechanistic parameters into the fuel cell semi-empirical mechanistic model, fixing the physical parameters to be identified in the model, enabling the fuel cell semi-empirical mechanistic model to output a stable reference voltage result under the current stack structure and operating conditions, thus forming the fuel cell reference voltage model. The expression of the fuel cell reference voltage model is:

[0110] in, Indicates the sampling time The corresponding reference voltage result is output by the fuel cell reference voltage model; This represents a semi-empirical mechanism model of a fuel cell after substituting the identified set of mechanism parameters. Indicates the sampling time The corresponding stack output current is obtained from the fuel cell operating status data; Indicates the sampling time The corresponding gas temperature or stack operating temperature is obtained from fuel cell operating status data; Indicates the sampling time The corresponding hydrogen reaction partial pressure is obtained by converting hydrogen pressure and hydrogen supply status. Indicates the sampling time The corresponding oxygen reaction partial pressure is calculated from the air pressure and the air supply status. This represents the set of identified mechanistic parameters, output by the Grey Wolf optimization algorithm when the convergence condition is met. This expression applies the real-time operating conditions and the identified set of mechanistic parameters to the semi-empirical mechanistic model of the fuel cell, enabling the fuel cell reference voltage model to provide a reference voltage result with electrochemical mechanism constraints, and providing a mechanistic benchmark for subsequent fuel cell aging residual prediction and final health state estimation results.

[0111] S160. Input the health status estimate into the feature set and input it into the TCN-Attention network constrained by the physical boundary to generate the fuel cell aging residual prediction result. Then, generate the final health status estimate result based on the fuel cell reference voltage model and the fuel cell aging residual prediction result. Combine the intraday short-term load prediction result to generate coordinated operation control commands.

[0112] Specifically, when organizing the health state estimation input feature set into health state time-series samples according to the sampling time, the server first reads the processed stack output voltage and key operating parameters corresponding to each sampling time in the health state estimation input feature set, and arranges them in chronological order. The health state estimation input feature set is a data set consisting of the processed stack output voltage and key operating parameters related to the processed stack output voltage; the health state time-series sample is a time-series sample extracted from the health state estimation input feature set according to a continuous time window, used to describe the voltage decay trajectory and operating condition change trajectory of the fuel cell stack during continuous operation. The server continuously extracts the health state estimation input feature set according to a preset health state window length, and arranges multiple health state feature vectors within each window into a health state time-series sample, the expression of which is:

[0113] in, Indicates the time up to the sampling point The resulting health status time series samples are extracted by the server from the health status estimation input feature set according to a time window; Indicates the sampling time The corresponding health status feature vector is obtained by combining the processed stack output voltage and key operating parameters; The length of the health status window is set by the server based on the rate of aging of the fuel cell, the sampling interval, and the sensing range of the TCN-Attention network. This represents the current sampling time, determined by a unified sampling clock. This expression preserves the temporal evolution information of the fuel cell's operating state through continuous time windows, enabling the subsequent TCN-Attention network to learn the decay relationship across sampling times, rather than estimating the health state based solely on a single sampling time.

[0114] When determining the physical boundary constraints to limit the network output based on the fuel cell operating conditions, the server determines the theoretical upper limit of the fuel cell stack under the current operating conditions based on the gas temperature, hydrogen pressure, air pressure, and reactant supply status at the current sampling time, and uses this theoretical upper limit as the core boundary of the physical boundary constraints. Physical boundary constraints refer to the constraints that prevent the TCN-Attention network output from violating the electrochemical laws of the fuel cell; the network output refers to the voltage correction correlation results obtained by the TCN-Attention network based on the time-series samples of the healthy state. For a fuel cell, the actual output voltage should not exceed the theoretical upper limit allowed under the current temperature and reactant partial pressure conditions. Therefore, the server constructs physical boundary constraints based on the current operating conditions, so that the TCN-Attention network is constrained by physical boundary constraints during both training and prediction. The expression for the theoretical upper limit of the physical boundary constraints is:

[0115] in, Indicates the sampling time The corresponding theoretical upper limit of voltage is calculated by the server based on the fuel cell operating conditions; The reference reaction potential is determined by the fuel cell stack model, reaction system, and calibration parameters. This represents the temperature correction factor, determined by the fuel cell electrochemical model or calibration experiment. Indicates the sampling time The corresponding gas temperature or stack operating temperature is obtained from fuel cell operating status data. This represents the reference temperature, determined by the model calibration conditions. This represents the reactant pressure correction factor, determined by the fuel cell electrochemical model or calibration experiment. Indicates the sampling time The corresponding hydrogen reaction partial pressure is obtained by converting hydrogen pressure and hydrogen supply status. Indicates the sampling time The corresponding oxygen reaction partial pressure is calculated from air pressure and air supply conditions. This expression describes the relationship between the theoretical upper limit of the fuel cell voltage and operating conditions through temperature and reactant partial pressure terms, allowing the physical boundary constraints to change dynamically with real-time operating conditions, rather than using a fixed threshold to limit the network output.

[0116] After feeding the health state time-series samples into the TCN-Attention network, the server first extracts the temporal decay features from the health state time-series samples through the TCN part. The TCN part is a temporal convolutional network used to process time-series data through causal convolution and dilated convolution. Temporal decay features refer to the slow voltage decrease, local fluctuation propagation, and cumulative effects of operating condition disturbances exhibited by the fuel cell stack at continuous sampling times. Causal convolution is used to ensure that feature extraction at the current sampling time only depends on the current sampling time and historical sampling times, avoiding the leakage of future sampling information; dilated convolution is used to expand the temporal perception range, enabling the network to cover a longer operating history with fewer layers. The temporal convolution expression of the TCN part is:

[0117] in, Indicates the first The layer TCN part at the sampling time The extracted temporal decay features are calculated by the server during network forward propagation; This represents a nonlinear activation function used to enhance the TCN part's ability to express nonlinear aging relationships; This indicates the kernel length, which is set by the TCN-Attention network structure. Indicates the first The first in the layer The convolution weights corresponding to each convolution position are obtained through network training. Indicates the first Layer at historical sampling time The corresponding feature representation, when The time is provided by the health status feature vector in the health status time series sample; Indicates the first The layer dilation coefficient, set by the network structure, is used to control the time interval spanned by convolution; Indicates the first Layer bias terms are obtained through network training; This indicates the position number within the convolution kernel. This expression enables the TCN part to simultaneously capture short-term operating condition disturbances and long-term voltage degradation trends by reviewing historical sampling features at different intervals at the current sampling time.

[0118] When extracting degradation-sensitive relationships using the TCN-Attention network, the server feeds the temporal degradation features output from the TCN part into the Attention part, and assigns attention to the feature contributions corresponding to different sampling times and different key operating parameters. Degradation-sensitive relationships refer to the sensitive correlation between the processed stack output voltage degradation and changes in key operating parameters, such as the impact of continuously increasing stack output current, insufficient air pressure, fluctuating hydrogen flow, and cooling temperature deviations on the voltage aging residual. The Attention part assigns different levels of attention to different time segments, giving higher weight to time segments with stronger relationships to fuel cell aging residuals during prediction. The attention weight expression is:

[0119] in, Indicates the sampling time The corresponding attention weights are calculated by the Attention component based on the temporal decay characteristics; Indicates the sampling time The corresponding temporal decay characteristics are output by the TCN part; This represents the attention score vector, obtained through network training. This represents the attention mapping matrix, obtained through network training. This represents the attention bias term, obtained through network training; This represents a nonlinear transformation function used to enhance the expressive power of attention scores for complex degradation relationships; This represents the set of sampling times covered by the current health status time series samples, which is determined by the server based on the health status window length; This represents the candidate sampling time in the set of sampling times. This expression compares the contribution of temporal decay characteristics at different sampling times to the prediction of aging residuals, enabling the Attention component to highlight time segments that are more sensitive to fuel cell degradation.

[0120] After obtaining the attention weights, the server performs weighted fusion of the temporal degradation features within the healthy state time series samples to obtain degradation-sensitive features. The network output layer then generates fuel cell aging residual prediction results based on these features. The fuel cell aging residual prediction results refer to the dynamic voltage decay portion that the fuel cell reference voltage model fails to explain, compensating for the difficulty in tracking long-term aging changes with fixed mechanism parameters. The fuel cell aging residual prediction results are not a replacement for the reference voltage results, but rather a dynamic correction of them. The expressions for the degradation-sensitive features and the fuel cell aging residual prediction results are:

[0121] in, Indicates the sampling time The corresponding degradation-sensitive features are obtained by the server through a weighted fusion of attention weights and temporal decay features; Indicates the sampling time The corresponding attention weights; Indicates the sampling time Corresponding temporal decay characteristics; This represents the set of sampling times covered by the current health status time series samples; Indicates the sampling time The corresponding fuel cell aging residual prediction results are obtained from the output of the TCN-Attention network; These represent the residual output weights, obtained through network training. This represents the residual output bias term, obtained through network training. This expression first aggregates time-series segments that are more sensitive to aging, and then maps the aggregated degradation-sensitive features to the fuel cell aging residual prediction results, enabling the network to output dynamic correction quantities related to the long-term degradation process of the fuel cell.

[0122] During the training of the TCN-Attention network, the server incorporates physical boundary constraints into the network optimization objective, ensuring that the network output fits the actual aging residuals while not exceeding the voltage upper limit allowed by the electrochemical laws of the fuel cell. The server first calculates the aging residual labels based on the reference voltage output from the fuel cell reference voltage model and the actual voltage during the training phase. Then, it uses the TCN-Attention network to predict the fuel cell aging residual predictions, and further constrains the fused voltage state results through physical boundary constraints. The training loss expression is as follows:

[0123] in, This represents the training loss of the TCN-Attention network, which is calculated by the server during network training and used to update network parameters. This represents the number of time-series health status samples participating in the current training batch, which is obtained by the server from the training data; Indicates the first The prediction results of fuel cell aging residuals corresponding to each training sample are output by the TCN-Attention network; Indicates the first The aging residual label corresponding to each training sample is obtained from the difference between the actual voltage and the reference voltage result; This represents the physical boundary constraint penalty coefficient, which is set by the server based on the data fitting accuracy and physical consistency requirements. Indicates the first The reference voltage result corresponding to each training sample is output by the fuel cell reference voltage model; Indicates the first The theoretical voltage upper limit corresponding to each training sample is calculated from the physical boundary constraints; This indicates that a penalty is incurred only when the fused voltage state exceeds the theoretical voltage limit. The first part of the expression ensures that the predicted fuel cell aging residuals closely approximate the actual aging residuals, while the second part ensures that the fused voltage state does not violate physical boundary constraints, thus balancing data fitting ability and electrochemical rationality.

[0124] When fusing the reference voltage result output by the fuel cell reference voltage model and the fuel cell aging residual prediction result, the server first calls the fuel cell reference voltage model according to the current operating conditions to obtain the reference voltage result corresponding to the current sampling time. Then, the fuel cell aging residual prediction result output by the TCN-Attention network is superimposed on the reference voltage result to obtain the corrected voltage state result. The reference voltage result provides a basic voltage estimate that conforms to the mechanism, while the fuel cell aging residual prediction result provides dynamic corrections caused by long-term degradation and operating condition coupling. Together, they form the final health state estimation result. The expression for the final health state estimation result is:

[0125] in, Indicates the sampling time The corresponding final health state estimation results are used to characterize the corrected voltage state corresponding to the current health level of the fuel cell stack. Indicates the sampling time The corresponding reference voltage result is output by the fuel cell reference voltage model based on the current operating conditions; Indicates the sampling time The corresponding fuel cell aging residual prediction results are output by the TCN-Attention network based on time-series health state samples. This expression, through the fusion of mechanistic benchmarks and data-driven residuals, ensures that the final health state estimation results retain electrochemical interpretability while also possessing the ability to track long-term dynamic aging processes.

[0126] When determining the direction and magnitude of power demand changes within the future short-term forecast time domain based on intraday short-term load forecast results, the server reads the intraday short-term load forecast values ​​corresponding to adjacent forecast times within the future short-term forecast time domain and compares the power change between the current forecast time and the next forecast time. The direction of power demand change refers to the trend of load demand rising, falling, or remaining stable relative to the current state within the future short-term forecast time domain; the magnitude of power demand change refers to the magnitude of the load demand change within the future short-term forecast time domain. The expression for power demand change is:

[0127] in, Indicates the predicted time in the future Relative to the current sampling time The power demand range is calculated by the server based on the intraday short-term load forecast and the current actual power. Indicates the predicted time in the future The corresponding intraday short-term load forecast is provided by the intraday short-term load forecast results; Indicates the current sampling time The corresponding actual power supply is obtained from the load-side power metering point or the system power controller; This represents the prediction step number within the short-term prediction time domain. This expression determines the power change target required for coordinated operation control commands to increase, decrease, or maintain output by comparing the future predicted load with the current actual power supply.

[0128] When determining the permissible power variation range for a fuel cell based on the final health status estimate, the server compares the final health status estimate with a health status threshold and combines this with the fuel cell's rated power, permissible ramp-up capability, and aging risk limits to determine the power variation range. The power variation range refers to the range of output power adjustments that the fuel cell is allowed to handle at its current health level, including the maximum permissible increase in power, the maximum permissible decrease in power, and the permissible ramp-up speed. If the final health status estimate indicates a good fuel cell health status, the server allows a larger power variation range; if the final health status estimate indicates an aging risk to the fuel cell, the server narrows the power variation range. The expression for the power variation range is:

[0129] in, Indicates the sampling time The corresponding range of power variation that the fuel cell can withstand is calculated by the server based on the final health status estimation results; This indicates the upper limit of the rated power variation that a fuel cell can be allowed to undergo when it is in a healthy and normal state. It is determined by the rated power of the fuel cell stack, the power converter capability, and the system control strategy. Indicates the sampling time The corresponding final health status estimate; The voltage threshold representing the aging risk is determined by fuel cell life testing, operation and maintenance strategies, or safety control standards. The healthy and safe voltage threshold is determined by the normal operating range of the fuel cell. This represents a limiting function, used to restrict the value within the parentheses to between 0 and 1. This expression dynamically adjusts the permissible power variation range of the fuel cell based on the position of the final health state estimate between the aging risk voltage threshold and the health safety voltage threshold, ensuring that the permissible power variation decreases as the health state deteriorates.

[0130] When generating coordinated operation control commands based on the direction, magnitude, and range of power demand changes, the server first determines whether the power demand magnitude exceeds the power change range. If the power demand magnitude does not exceed the power change range, the server generates coordinated operation control commands whereby the fuel cell undertakes the main power adjustment; if the power demand magnitude exceeds the power change range, the server limits the fuel cell output power adjustment to within the power change range and allocates the excess to energy storage units, grid-connected power regulation, or adjustable load management. The expression for the fuel cell output power adjustment is:

[0131] in, Indicates the predicted time in the future The corresponding fuel cell output power adjustment amount is determined by the server when generating the coordinated operation control command; The sign function is used to preserve the direction of power demand change; it is positive when power demand increases and negative when power demand decreases. Indicates the predicted time in the future The corresponding power demand range; Represents the absolute value of the power demand magnitude; This indicates the range of power variation that the fuel cell can handle at the current sampling time; This function takes the smaller value and limits the adjustment of the fuel cell output power to within the allowable range under healthy conditions. This expression satisfies load tracking requirements while constraining the fuel cell output power adjustment within the allowable range under healthy conditions, thereby preventing fuel cell overload or accelerated aging due to sudden changes in load tracking.

[0132] When coordinating operation control commands to adjust fuel cell output power, hydrogen supply, air supply, and cooling conditions, the server generates corresponding power control commands based on the fuel cell output power adjustment, and simultaneously generates hydrogen supply adjustment commands, air supply adjustment commands, and cooling condition adjustment commands. The hydrogen supply adjustment command adjusts hydrogen flow rate and pressure to match reactant supply with the target output power; the air supply adjustment command adjusts air flow rate and pressure to ensure oxygen supply meets the electrochemical reaction requirements; and the cooling condition adjustment command adjusts cooling flow rate or cooling temperature to maintain the fuel cell stack in a safe thermal state. If the final health status estimate indicates a risk of fuel cell aging, the server further restricts fuel cell power ramp-up when generating the coordinated operation control commands. The power ramp-up constraint expression is:

[0133] in, Indicates the predicted time in the future The corresponding target output power of the fuel cell is obtained by superimposing the current output power of the fuel cell and the adjustment amount of the fuel cell output power; Indicates the current sampling time The corresponding target output power of the fuel cell is obtained from the current control state record; Indicates the prediction step number within the short-term prediction time domain; The sampling interval is determined by a unified sampling clock; Indicates the sampling time The corresponding allowable power ramp-up rate is determined by the server based on the final health status estimate, temperature conditions, and aging risk level. This expression limits the variation in the target output power of the fuel cell per unit time, preventing the fuel cell from experiencing excessively rapid load changes under aging risk conditions. This ensures that coordinated operation control commands simultaneously meet the requirements of load tracking, reactant supply matching, thermal management safety, and lifespan protection.

[0134] In this embodiment, the Transformer prediction network is used for day-ahead load forecasting. Its training data is not the original historical load power data, but rather multiple intrinsic mode components obtained from variational mode decomposition. First, mode sequence data is constructed for each intrinsic mode component according to the sampling time. Then, historical mode segments are extracted as training samples, and the actual mode changes in the subsequent day-ahead forecast time domain are used as training labels. During training, the Transformer prediction network maintains the sampling order through time position markings, learns the long-term dependencies of each intrinsic mode component through a multi-head self-attention mechanism, and outputs the corresponding mode prediction results. After training, the mode prediction results are reconstructed to obtain the day-ahead baseline load forecast results.

[0135] The LightGBM residual correction model is used for intraday short-term load forecasting. Its training data comes from the deviation between real-time load power data and the day-ahead baseline load forecast. The server first accumulates multiple load forecast deviations to form a historical residual sequence, then extracts the current residual value, residual change direction, residual duration, residual peak position, residual decline status, and baseline load level to form residual feature samples. During training, the LightGBM residual correction model learns the relationship between residual feature samples and future residuals through multiple decision trees. During online operation, the model predicts future short-term residuals based on the latest residual feature samples and compensates the residual prediction results for the day-ahead baseline load forecast, forming the intraday short-term load forecast.

[0136] A semi-empirical mechanistic model for fuel cells is used to describe the voltage response of fuel cell stacks under different operating conditions. This model is constructed based on the electrochemical mechanism of fuel cells, expressing the stack output voltage as the theoretical voltage minus activation losses, ohmic losses, and concentration losses. Since some physical parameters are difficult to measure directly, the server treats these unidentified physical parameters as unknown parameters and uses polarization experimental data for parameter identification.

[0137] The fuel cell reference voltage model is used to output reference voltage results. It is not a separately retrained model, but rather formed by substituting the identified set of mechanistic parameters into the fuel cell semi-empirical mechanistic model. During online operation, the server inputs the current stack output current, temperature, hydrogen pressure, and air pressure, among other operating conditions, into the fuel cell reference voltage model to obtain reference voltage results that conform to the electrochemical mechanism, providing a mechanistic benchmark for subsequent health state estimation.

[0138] A physically boundary-constrained TCN-Attention network is used to predict fuel cell aging residuals. Its training data comes from a health state estimation input feature set, which consists of the processed stack output voltage and key operating parameters. During training, a reference voltage result is first output using a fuel cell reference voltage model, and then the difference between the actual voltage and the reference voltage result is used as the aging residual label. The TCN part extracts temporal decay features during continuous operation, while the Attention part identifies time segments and key operating parameters that have a significant impact on the aging residuals and outputs the predicted fuel cell aging residuals. Simultaneously, physical boundary constraints ensure that the network output does not violate the electrochemical laws of the fuel cell.

[0139] This application also provides a load prediction and health status assessment device for a hydrogen fuel cell power generation system, characterized in that, referring to... Figure 2 , Figure 2 This is a schematic diagram of a load forecasting and health status assessment device for a hydrogen fuel cell power generation system provided in an embodiment of this application. The device is a server, which includes an acquisition module 21 and a processing module 22. The acquisition module 21 is used to acquire historical load power data and fuel cell operating status data of the hydrogen fuel cell power generation system. The processing module 22 is used to clean and normalize the historical load power data to form a continuous load power sequence, and to perform variational mode decomposition on the continuous load power sequence to obtain multiple intrinsic mode components. The processing module 22 is also used to input each intrinsic mode component into a Transformer prediction network to generate mode prediction results, and to reconstruct the mode prediction results to form a day-ahead baseline load prediction result. The processing module 22 is also used to compare the real-time load power data with the day-ahead baseline load prediction result. The deviations form a historical residual sequence, which is then rolled and predicted using the LightGBM residual correction model to generate intraday short-term load forecast results. Processing module 22 is also used to denoise and smooth the stack output voltage in the fuel cell operating status data, and to screen key operating parameters related to the stack output voltage to form a health status estimation input feature set. At the same time, the gray wolf optimization algorithm is used to identify the fuel cell semi-empirical mechanism model to form a fuel cell reference voltage model. Processing module 22 is also used to input the health status estimation input feature set into a TCN-Attention network constrained by physical boundaries to generate fuel cell aging residual prediction results, and to generate the final health status estimation results based on the fuel cell reference voltage model and the fuel cell aging residual prediction results, so as to generate coordinated operation control commands in combination with intraday short-term load forecast results.

[0140] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0141] This application also provides an electronic device, with reference to... Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: at least one processor 31, at least one network interface 34, a user interface 33, a memory 35, and at least one communication bus 32.

[0142] The communication bus 32 is used to enable communication between these components.

[0143] The user interface 33 may include a display screen and a camera. Optionally, the user interface 33 may also include a standard wired interface and a wireless interface.

[0144] The network interface 34 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0145] The processor 31 may include one or more processing cores. The processor 31 connects to various parts of the server via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in the memory 35, and calling data stored in the memory 35 to perform various server functions and process data. Optionally, the processor 31 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 31 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 31 and may be implemented as a separate chip.

[0146] The memory 35 may include random access memory (RAM) or read-only memory. Optionally, the memory 35 may include a non-transitory computer-readable storage medium. The memory 35 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 35 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 35 may also be at least one storage device located remotely from the aforementioned processor 31. Figure 3 As shown, the memory 35, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a method of load prediction and health status assessment of a hydrogen fuel cell power generation system.

[0147] exist Figure 3 In the electronic device shown, the user interface 33 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 31 can be used to call the application program stored in the memory 35, which is a method for load prediction and health status assessment of a hydrogen fuel cell power generation system. When executed by one or more processors, the electronic device performs one or more methods as described in the above embodiments.

[0148] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0149] This application also provides a non-transitory computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0150] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0151] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0152] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0153] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0154] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0155] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of other embodiments of this disclosure upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for load prediction and health status assessment of a hydrogen fuel cell power generation system, characterized in that, The method includes: Acquire historical load power data and fuel cell operating status data of hydrogen fuel cell power generation systems; The historical load power data is cleaned and normalized to form a continuous load power sequence, and the continuous load power sequence is subjected to variational mode decomposition to obtain multiple intrinsic mode components. Each of the intrinsic modal components is input into the Transformer prediction network to generate modal prediction results, and the modal prediction results are reconstructed to form the day-ahead baseline load prediction results; Based on the deviation between real-time load power data and the day-ahead baseline load forecast results, a historical residual sequence is formed, and a short-term intraday load forecast result is generated through rolling forecasting using the LightGBM residual correction model. The stack output voltage in the fuel cell operating status data is denoised and smoothed, and key operating parameters related to the stack output voltage are selected to form a health status estimation input feature set. At the same time, the semi-empirical mechanism model of the fuel cell is identified by the gray wolf optimization algorithm to form a fuel cell reference voltage model. The health status estimate is input into the feature set and then into a TCN-Attention network constrained by physical boundaries to generate fuel cell aging residual prediction results. Finally, a health status estimate is generated based on the fuel cell reference voltage model and the fuel cell aging residual prediction results. This is then combined with the intraday short-term load forecast results to generate coordinated operation control commands.

2. The method for load prediction and health status assessment of a hydrogen fuel cell power generation system according to claim 1, characterized in that, The acquisition of historical load power data and fuel cell operating status data of the hydrogen fuel cell power generation system specifically includes: Active power time series of load-side power metering points are collected according to a unified sampling clock, and the active power time series is used as the historical load power data. Collect voltage, current, flow rate, pressure, and temperature data corresponding to the fuel cell stack, hydrogen supply circuit, air supply circuit, and cooling circuit, and use the collected data as the operating status data of the fuel cell. The historical load power data and the fuel cell operating status data at the same sampling time are associated as a system operation record according to the time identifier.

3. The method for load prediction and health status assessment of a hydrogen fuel cell power generation system according to claim 1, characterized in that, The process of cleaning and normalizing the historical load power data to form a continuous load power sequence, and then performing variational mode decomposition on the continuous load power sequence to obtain multiple intrinsic mode components, specifically includes: Identify abnormal load points and missing load points in the historical load power data, and remove and fill in the abnormal load points and missing load points to form a cleaned load power sequence; The cleaning load power sequence is normalized to form the continuous load power sequence; The number of mode decompositions, the penalty factor, and the convergence tolerance corresponding to the continuous load power sequence are set, and variational mode decomposition is performed on the continuous load power sequence according to the number of mode decompositions, the penalty factor, and the convergence tolerance to obtain multiple intrinsic mode components with different center frequencies.

4. The method for load prediction and health status assessment of a hydrogen fuel cell power generation system according to claim 1, characterized in that, The process of inputting each of the intrinsic mode components into the Transformer prediction network to generate mode prediction results, and reconstructing the mode prediction results to form the day-ahead baseline load prediction results, specifically includes: Modal sequence data is constructed based on the sampling time, center frequency, and modal order corresponding to each intrinsic modal component, and modal training samples are formed based on the modal sequence data; The modal training samples are input into the Transformer prediction network, and the temporal order of each sampled value in the modal training samples is determined by time position marking. The Transformer prediction network extracts the long-term dependencies corresponding to each intrinsic mode component, and generates the mode prediction results corresponding to each intrinsic mode component based on the long-term dependencies. The prediction results of each mode are aligned according to the modal order and prediction time, and the prediction results of each mode at the same prediction time are reconstructed to form the day-ahead baseline load prediction result.

5. The method for load prediction and health status assessment of a hydrogen fuel cell power generation system according to claim 1, characterized in that, The process of forming a historical residual sequence based on the deviation between real-time load power data and the day-ahead baseline load forecast result, and then using the LightGBM residual correction model for rolling forecasting, to generate intraday short-term load forecast results, specifically includes: The real-time load power data is obtained at a sampling interval consistent with the day-ahead reference load forecast result, and the day-ahead reference load forecast result at the same time is retrieved according to the real-time sampling time. The real-time load power data is compared with the day-ahead baseline load forecast result at the same time to obtain the load forecast deviation, and multiple load forecast deviations are accumulated in chronological order to form the historical residual sequence; Based on the historical residual sequence, residual feature samples are constructed, including the current residual value, residual change direction, residual duration, residual peak position, residual fallback status, and baseline load level; The residual feature samples are input into the LightGBM residual correction model for rolling prediction, generating residual prediction results in the future short-term prediction time domain. The residual prediction result is compensated to the day-ahead baseline load prediction result at the corresponding prediction time to form the intraday short-term load prediction result.

6. The method for load prediction and health status assessment of a hydrogen fuel cell power generation system according to claim 1, characterized in that, The process involves denoising and smoothing the stack output voltage in the fuel cell operating status data, filtering key operating parameters related to the stack output voltage to form a health state estimation input feature set, and simultaneously identifying a semi-empirical mechanism model of the fuel cell using the Grey Wolf optimization algorithm to form a fuel cell reference voltage model. Specifically, this includes: According to the sampling time, the stack output voltage and corresponding operating parameters are extracted from the fuel cell operating status data, and the stack output voltage is subjected to abnormal voltage point completion, normalization processing and wavelet denoising processing to form the processed stack output voltage. The correlation analysis between the operating parameters and the processed stack output voltage was performed, and the operating parameters whose correlation met the preset conditions were determined as key operating parameters. The processed stack output voltage and the key operating parameters are combined according to the sampling time to form the health status estimation input feature set; A semi-empirical mechanism model for fuel cells is constructed to output reference voltage results, and the physical parameters to be identified in the semi-empirical mechanism model for fuel cells are determined as unknown parameters. The unknown parameters are iteratively optimized using the Grey Wolf optimization algorithm so that the fitting error between the model voltage output by the fuel cell semi-empirical mechanism model and the actual voltage obtained from the polarization experiment meets the convergence condition. The unknown parameters that satisfy the convergence condition are identified as the set of mechanism parameters, and the set of identified mechanism parameters is substituted into the fuel cell semi-empirical mechanism model to form the fuel cell reference voltage model.

7. The method for load prediction and health status assessment of a hydrogen fuel cell power generation system according to claim 1, characterized in that, The process involves inputting the health status estimate into a TCN-Attention network constrained by physical boundaries to generate fuel cell aging residual prediction results. Based on the fuel cell reference voltage model and the fuel cell aging residual prediction results, a final health status estimate is generated. This, combined with the intraday short-term load forecast results, generates coordinated operation control commands. Specifically, this includes: The health state estimation input feature set is organized into health state time series samples according to the sampling time, and physical boundary constraints for limiting the network output results are determined based on the fuel cell operating conditions. The health state time series samples are input into the TCN-Attention network, and the TCN-Attention network is used to extract the time series decay features and degradation sensitivity relationships in the health state time series samples to generate the fuel cell aging residual prediction results. The reference voltage result output by the fuel cell reference voltage model and the fuel cell aging residual prediction result are fused to generate the final health status estimate. Based on the intraday short-term load forecast results, determine the direction and magnitude of power demand changes in the future short-term forecast time domain, and determine the range of power changes that the fuel cell can withstand based on the final health status estimation results; The coordinated operation control command is generated based on the direction of power demand change, the magnitude of power demand change, and the range of power change, so that the coordinated operation control command is used to adjust the fuel cell output power, hydrogen supply, air supply and cooling conditions, and to limit the fuel cell power ramp-up when the final health state estimation result indicates that the fuel cell has an aging risk.

8. A device for load prediction and health status assessment of a hydrogen fuel cell power generation system, characterized in that, The apparatus is used to execute the load prediction and health status assessment method for a hydrogen fuel cell power generation system as described in any one of claims 1 to 7, the apparatus comprising an acquisition module and a processing module, wherein... The acquisition module is used to acquire historical load power data and fuel cell operating status data of the hydrogen fuel cell power generation system; The processing module is used to clean and normalize the historical load power data to form a continuous load power sequence, and to perform variational mode decomposition on the continuous load power sequence to obtain multiple intrinsic mode components. The processing module is further configured to input each of the intrinsic modal components into the Transformer prediction network, generate modal prediction results, and reconstruct the modal prediction results to form the day-ahead baseline load prediction results; The processing module is also used to form a historical residual sequence based on the deviation between the real-time load power data and the day-ahead baseline load forecast result, and to form an intraday short-term load forecast result by rolling forecasting through the LightGBM residual correction model. The processing module is also used to denoise and smooth the stack output voltage in the fuel cell operating status data, and to filter key operating parameters related to the stack output voltage to form a health status estimation input feature set. At the same time, it uses the gray wolf optimization algorithm to identify the fuel cell semi-empirical mechanism model and form a fuel cell reference voltage model. The processing module is further configured to input the health status estimation input feature set into a TCN-Attention network constrained by physical boundaries to generate fuel cell aging residual prediction results, and generate final health status estimation results based on the fuel cell reference voltage model and the fuel cell aging residual prediction results, so as to generate coordinated operation control commands in combination with the intraday short-term load prediction results.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements a method for load prediction and health status assessment of a hydrogen fuel cell power generation system as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores instructions that, when executed, perform a method for load prediction and health status assessment of a hydrogen fuel cell power generation system as described in any one of claims 1 to 7.

Images