A data processing method and device for a wind-solar-hydrogen storage coupling system

Abstract

This specification provides a data processing method and apparatus for wind-solar-hydrogen storage coupled systems. The method involves: acquiring the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupled system; determining the net power sequence based on the operating data; identifying fluctuation states in the net power sequence and determining sliding window parameters based on the identification results; extracting multiple target data segments from the net power sequence according to the window length, and fusing these multiple target data segments with the equipment status parameters according to a window combination method to determine the target feature vector; using a preset power prediction model, determining the net power prediction sequence and prediction confidence level based on the target feature vector; determining the execution power envelope based on the net power prediction sequence, prediction confidence level, and preset operating constraints; and determining the target power of the electrolyzer and the compensation power of the energy storage device based on the execution power envelope. This improves the accuracy of power prediction and safe allocation.

Description

Technical Field

[0001] This specification belongs to the field of new energy power system control technology, and in particular relates to a data processing method and device for a wind-solar-hydrogen storage coupling system. Background Technology

[0002] With the large-scale application of wind-solar-hydrogen storage coupled systems, achieving high real-time power prediction has become crucial for ensuring the dynamic balance and safe operation of multi-energy complementary systems. However, existing technologies mostly employ fixed feature extraction methods, resulting in insufficient power prediction accuracy and response lag under non-stationary operating conditions. Furthermore, they struggle to meet the coordinated control requirements for safe operation and precise power allocation under drastic power fluctuations.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This specification provides a data processing method and apparatus for wind-solar-hydrogen storage coupled systems, which solves the technical problems of insufficient prediction accuracy of wind-solar-hydrogen storage coupled systems due to the use of fixed feature extraction methods in the prior art, and the difficulty in accurately adapting power prediction results to the operating constraints of electrolyzers.

[0005] This specification provides a data processing method and apparatus for a wind-solar-hydrogen storage coupled system, including: Obtain the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupled system; wherein the target wind-solar-hydrogen storage coupled system includes an electrolyzer and an energy storage device; Based on the aforementioned operating data, determine the net power sequence; The net power sequence is subjected to fluctuation state identification, and sliding window parameters are determined based on the identification results; wherein, the sliding window parameters include window length and window combination method; Based on the window length, multiple target data segments are extracted from the net power sequence, and the multiple target data segments are fused with the device status parameters according to the window combination method to determine the target feature vector; wherein, one target data segment corresponds to one time scale; Using a preset power prediction model, the net power prediction sequence and prediction confidence are determined based on the target feature vector; The execution power envelope is determined based on the net power prediction sequence, the prediction confidence level, and the preset operating constraints. Based on the execution power envelope, a safe operating range is determined, and based on the safe operating range, the target power of the electrolyzer and the compensation power of the energy storage device are determined.

[0006] In one embodiment, the method further includes: Obtain the net power measured sequence corresponding to the net power prediction sequence; Calculate the residual between the net power prediction sequence and the net power measured sequence to determine the prediction error; Error attribution is identified based on the prediction error, the target power of the electrolyzer, and the compensation power of the energy storage device, and the parameters of the preset power prediction model are iteratively updated using the identified error attribution results.

[0007] In one embodiment, determining the net power sequence based on the operating data includes: The running data is time-aligned to determine the corresponding running sequence to be processed; Based on the operation sequence to be processed, the target power component is determined; wherein, the target power component includes the actual wind power generation, the actual photovoltaic power generation, the energy storage charging and discharging power, the station load power, and the reserved power parameters; The target power components are subjected to power combining processing to determine the net power sequence.

[0008] In one embodiment, identifying the fluctuation state of the net power sequence and determining the sliding window parameters based on the identification result includes: Using multiple candidate observation windows of preset scales, multi-scale time-domain features are extracted from the net power sequence to determine fluctuation state indices; wherein, the fluctuation state indices include: the first difference, rate of change, and second difference of the net power sequence, as well as the short window range, short window fluctuation intensity, medium window standard deviation, and long window mean corresponding to each of the candidate observation windows; The fluctuation state index is mapped to a preset multi-dimensional feature space for working condition identification to determine the current working condition type; wherein, the current working condition type includes stable working condition, continuous climbing working condition, rapid drop working condition, high frequency oscillation working condition, data abnormal working condition, and equipment disturbance working condition. Based on the current working condition type, a matching window length is determined from the multiple preset scales, and the corresponding window combination method is determined as the sliding window parameter.

[0009] In one embodiment, the step of extracting a target data segment from the net power sequence according to the window length, and fusing the target data segment with the device status parameters according to the window combination method to determine the target feature vector includes: Based on the window length, data sampling points of corresponding duration are extracted from the net power sequence to determine the target data segment; The target data segment is normalized to determine the target temporal feature components; According to the window combination method, the target temporal feature component and the device state parameter are feature aligned and vector concatenated to determine the target feature vector.

[0010] In one embodiment, determining the execution power envelope based on the net power prediction sequence, the prediction confidence level, and preset operational constraints includes: Based on the predicted confidence level and the preset mapping relationship between the confidence level and the adjustment deviation, determine the confidence adjustment margin corresponding to the predicted confidence level; The net power prediction sequence is interval mapped using the confidence adjustment margin to determine the initial power execution interval; Based on the preset operating constraints, the initial power execution range is subjected to amplitude limiting and rate of change constraint processing to determine the execution power envelope; The execution power envelope includes the upper limit and lower limit of the execution power of the target wind-solar-hydrogen storage coupling system at each sampling time within the corresponding preset time period.

[0011] In one embodiment, determining the safe operating range based on the executed power envelope, and determining the target power of the electrolyzer and the compensation power of the energy storage device based on the safe operating range, includes: Based on the upper limit and lower limit of the execution power included in the execution power envelope, the safe operating range of the electrolytic cell at each sampling time within the corresponding preset time period is determined; Based on the safe operating range, the net power prediction sequence is subjected to amplitude limiting mapping to determine the target power of the electrolytic cell; The compensation power of the energy storage device is determined based on the net power prediction sequence and the target power of the electrolyzer.

[0012] This specification provides a data processing device for a wind-solar-hydrogen storage coupled system, comprising: The data acquisition module is used to acquire the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein, the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device; A sequence determination module is used to determine the net power sequence based on the operating data; The parameter determination module is used to identify the fluctuation state of the net power sequence and determine the sliding window parameters based on the identification results; wherein, the sliding window parameters include the window length and the window combination method; The vector determination module is used to extract multiple target data segments from the net power sequence according to the window length, and fuse the multiple target data segments with the device status parameters according to the window combination method to determine the target feature vector; wherein, one target data segment corresponds to one time scale; The power prediction module is used to determine the net power prediction sequence and prediction confidence based on the target feature vector using a preset power prediction model. The power envelope determination module is used to determine the execution power envelope based on the net power prediction sequence, the prediction confidence level, and preset operating constraints. The power determination module is used to determine the safe operating range based on the execution power envelope, and to determine the target power of the electrolyzer and the compensation power of the energy storage device based on the safe operating range.

[0013] This specification also provides a computer-readable storage medium storing computer instructions that, when executed, implement a data processing method for a wind-solar-hydrogen storage coupling system.

[0014] Based on the data processing method for a wind-solar-hydrogen storage coupled system provided in this specification, the following steps are taken: The method acquires the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupled system, which includes an electrolyzer and an energy storage device. A net power sequence is determined based on the operating data. Fluctuation states are identified in the net power sequence, and sliding window parameters are determined based on the identification results. The sliding window parameters include a window length and a window combination method. Multiple target data segments are extracted from the net power sequence according to the window length, and these segments are fused with the equipment status parameters according to the window combination method to determine a target feature vector. Each target data segment corresponds to a time scale. Using a preset power prediction model, a net power prediction sequence and prediction confidence level are determined based on the target feature vector. An execution power envelope is determined based on the net power prediction sequence, the prediction confidence level, and preset operating constraints. An operating safety range is determined based on the execution power envelope, and the target power of the electrolyzer and the compensation power of the energy storage device are determined based on the operating safety range. In this way, by dynamically determining the sliding window parameters and extracting multi-scale target data segments through fluctuation state identification, the problems of insufficient prediction accuracy and response lag caused by fixed feature extraction in existing technologies are solved. Furthermore, by combining prediction confidence and operational constraints to determine the execution power envelope and operational safety range, the coordinated control of the electrolyzer target power and the energy storage equipment compensation power is realized, meeting the requirements for safe operation and accurate allocation under drastic power fluctuations. Attached Figure Description

[0015] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a data processing method for a wind-solar-hydrogen storage coupling system, provided in one embodiment of this specification. Figure 2 This is a schematic diagram of the electronic device structure provided in one embodiment of this specification; Figure 3 This is a schematic diagram of the structural composition of a data processing device for a wind-solar-hydrogen storage coupling system, provided in one embodiment of this specification. Figure 4 This is a schematic diagram of a rolling prediction process for safe regulation of hydrogen production from wind and solar power, provided as an embodiment of this specification. Detailed Implementation

[0017] 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 specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0018] See Figure 1 As shown in the embodiments of this specification, a data processing method for a wind-solar-hydrogen storage coupled system is provided, wherein the method is specifically applied to the server side. In specific implementation, the method may include the following: S101: Obtain the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein, the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device; S102: Determine the net power sequence based on the operating data; S103: Identify the fluctuation state of the net power sequence and determine the sliding window parameters based on the identification results; wherein, the sliding window parameters include the window length and the window combination method; S104: Based on the window length, extract multiple target data segments from the net power sequence, and fuse the multiple target data segments with the device status parameters according to the window combination method to determine the target feature vector; wherein, one target data segment corresponds to one time scale; S105: Using a preset power prediction model, determine the net power prediction sequence and prediction confidence based on the target feature vector; S106: Determine the execution power envelope based on the net power prediction sequence, the prediction confidence level, and the preset operating constraints; S107: Determine the safe operating range based on the execution power envelope, and determine the target power of the electrolytic cell and the compensation power of the energy storage device based on the safe operating range.

[0019] The aforementioned target wind-solar-storage-hydrogen coupling system can refer to an integrated energy system composed of a wind power generation unit, a photovoltaic power generation unit, an energy storage unit (such as a battery), and a water electrolysis hydrogen production unit (electrolyzer).

[0020] The aforementioned operational data can be multi-dimensional raw data generated and collected by the system during real-time operation. It includes, but is not limited to, real-time output of wind turbines and photovoltaics, bus voltage and current, state of charge (SOC) of energy storage systems, real-time load, temperature, and pressure of electrolyzers, and energy flow data between nodes within the system.

[0021] The aforementioned equipment status parameters can refer to parameters that describe the internal physical health and regulation capabilities of electrolyzers and energy storage devices, such as the state of charge (SOC) of the energy storage system, the real-time operating load rate of the electrolyzer, and the electrolyte temperature and pressure status.

[0022] The aforementioned net power sequence can refer to the time sequence of power that is actually available for internal allocation (allocated to electrolyzers or energy storage) after deducting the station load and necessary reserve margin from the sum of the actual generated power of wind power and photovoltaic power.

[0023] The aforementioned fluctuation state identification can be a logic for real-time feature analysis of the net power sequence, aiming to determine the changing attributes of the current energy flow, such as identifying whether it is in a stable operating condition, rapid ramp-up, high-frequency oscillation, or abnormal disturbance state.

[0024] The sliding window parameters mentioned above can be defined by the window length and the window combination method. These parameters are no longer fixed, but are dynamically adjusted according to the results of the above fluctuation state recognition to ensure that the feature extraction can both cover sufficient historical information and respond sensitively to instantaneous changes.

[0025] The window length mentioned above can refer to the number of data sampling points covered by the sliding window on the time axis. Shortening the window length during periods of drastic power fluctuations can eliminate interference from stale data and improve sensitivity; while increasing the window length during periods of stability can smooth out noise and improve prediction stability.

[0026] The aforementioned window combination method refers to how to logically splice or weight multiple windows of different lengths and frequencies under multi-timescale analysis, thereby transforming fragmented sampling points into a data set with rich temporal semantics.

[0027] The aforementioned target data segment can refer to a specific fragment extracted from the original net power sequence according to determined window parameters. These data segments are physical slices of the original signal, carrying the energy evolution trend under the current operating conditions. Each data segment typically corresponds to a specific observation scale (such as an ultra-short-term or short-term sampling interval).

[0028] The aforementioned target feature vector can be a comprehensive feature vector formed by normalizing, reshaping dimensions, and concatenating features of multiple target data segments and device status parameters (such as SOC and temperature).

[0029] The aforementioned net power prediction sequence refers to the power change trend over a predetermined period of time obtained by forward mapping the target feature vector using a power prediction model (such as LSTM, TCN, etc.). It predicts the future energy flow of the system, providing data support for the electrolyzer to predict the adjustment direction in advance and for reserving charging and discharging space for energy storage.

[0030] The aforementioned prediction confidence level can refer to the model's self-assessment index of the accuracy of its prediction results (usually expressed in the form of probability or variance). It reflects the magnitude of prediction risk: the higher the confidence level, the more certain the future trend is, and the more aggressive the allocation strategy can be; the lower the confidence level, the more unpredictable the fluctuations are, and the allocation strategy needs to reserve more adjustment redundancy.

[0031] The aforementioned preset operating constraints can refer to hard limit logic pre-set based on the physical characteristics of the electrolyzer and energy storage, including but not limited to maximum ramp rate, upper and lower limits of power operation, and SOC safety window.

[0032] The aforementioned execution power envelope can refer to a dynamic power boundary calculated by comprehensively considering the prediction sequence, prediction risk (confidence level), and physical constraints. It includes the maximum upper limit and minimum lower limit of execution at each time step, transforming the mathematical prediction expectation into a physically absolutely safe execution range.

[0033] The aforementioned safe operating range can be the final operating domain determined based on the power envelope, further combined with real-time system margin and risk control requirements. It clarifies the ideal load range for the electrolyzer while ensuring safety.

[0034] In some embodiments, acquiring the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system may specifically include: Based on the collected wind power, photovoltaic power, and load power, the original power components of each energy node are determined. The original power components are sampled, synchronized, and physically checked to determine the operating data; The equipment status parameters are determined based on the obtained energy storage state of charge and electrolytic cell status parameters.

[0035] In some embodiments, determining the net power sequence based on the operating data may specifically include: Based on the operational data, timestamp verification, missing value detection, outlier detection, communication delay identification, and physical boundary verification are performed to determine the data quality analysis results; Based on the data quality analysis results, missing points in the running data are filled in, and outliers exceeding the preset physical boundaries are marked to determine the preprocessed data. Resampling and alignment processing is performed based on the timestamp tags and communication delay identifiers in the preprocessed data to obtain the net power sequence.

[0036] Specifically, a multi-dimensional quality assessment is performed on the collected operational data. Timestamp integrity verification logic is used to scan the data stream for logical breakpoints. A physical boundary verification mechanism is employed to retrieve the rated power range of the electrolyzer and the operating limits of the energy storage equipment, identifying outliers where instantaneous range exceedances occur. Combined with a communication delay identification algorithm, the latency in the backhaul process of different communication links is captured, generating multi-dimensional verification tags reflecting data reliability and completeness.

[0037] For identified missing sampling points, mean imputation is performed using steady-state data from adjacent time points, or correlation imputation is performed using actual power output from devices with similar operating characteristics. Short-term trend prediction logic is introduced, performing first-order or higher-order interpolation calculations based on the slope characteristics before and after the missing window. For outlier values ​​marked as exceeding physical boundaries, logical stripping is performed to prevent them from entering the training and update process of the power prediction model, thus avoiding non-physical noise causing model parameter shifts.

[0038] By utilizing timestamp playback logic, data components with communication delays and misalignments are projected onto a unified time reference axis. Resampling is performed on the original features of each energy node according to a preset sampling step size to eliminate sampling phase deviations caused by heterogeneous communication links. By vector summing the calibrated wind and solar power and subtracting station load and energy storage charging / discharging components, a net power sequence with spatiotemporal alignment characteristics is generated, providing a robust data foundation for subsequent feature vector fusion.

[0039] Based on the above embodiments, by using multi-dimensional data detection and timestamp playback alignment, the interference of communication delay and physical noise on the original signal is eliminated, ensuring the temporal consistency and numerical validity of the net power sequence, and guaranteeing the accuracy and reliability of the input features of the subsequent prediction model from the source.

[0040] In some embodiments, the step of using a preset power prediction model to determine the net power prediction sequence and prediction confidence based on the target feature vector may specifically include: Based on the target feature vector, the preset power prediction model is used to perform feature spatiotemporal decoupling to determine the temporal evolution component and the state constraint component. Cross-feature fusion is performed based on the temporal evolution component and the state constraint component to determine the predicted intermediate state features; A two-branch mapping is performed based on the predicted intermediate state features to determine the predicted power and the predicted confidence, respectively.

[0041] Specifically, the target feature vector, aligned along the time axis, is input into a pre-defined power prediction model. The model's built-in encoding layer deconstructs the multi-scale sampled data, extracting the temporal evolution component reflecting the inertia of power changes, and simultaneously extracting the state constraint component reflecting the boundaries of the electrolyzer and energy storage hardware. This spatiotemporal decoupling process ensures the independence of environmental fluctuation characteristics and equipment physical limitations in mathematical representation.

[0042] A multi-head attention operator is used to perform cross-feature fusion on the temporal evolution component and the state constraint component. By calculating the correlation weights between different feature dimensions, the power evolution information at each time scale is logically aligned with the real-time state of charge and load rate of the equipment, producing predictive intermediate features with physical sensing capabilities. This process utilizes the data consistency after time alignment to eliminate feature coupling bias caused by data source latency.

[0043] A decoding network with a dual-output head structure is employed to perform a two-branch mapping on the predicted intermediate state features. The mean regression branch determines the predicted power for future periods by performing linear and nonlinear transformations on the intermediate state features. Simultaneously, the variance estimation branch models the probability distribution of the prediction bias using quantile regression or Gaussian likelihood estimation logic, determining the prediction confidence level that reflects the reliability of the prediction results.

[0044] Based on the above embodiments, by performing feature spatiotemporal decoupling and dual-branch parallel mapping within the model, the synchronous output of predicted power and prediction confidence is achieved. This not only improves the prediction accuracy under fluctuating conditions, but also provides a quantitative risk assessment indicator for the subsequent generation of the safe execution envelope.

[0045] In some embodiments, the preset power prediction model adopts a heterogeneous attention network architecture based on multi-scale spatiotemporal feature decoupling. Its core logic lies in using the Temporal Convolutional Network (TCN) branch to capture the instantaneous high-frequency components of the net power signal under different receptive fields, and combining it in parallel with the Bi-Gated Recurrent Root (Bi-GRU) branch to capture the long-range trend of energy evolution. Furthermore, the spatiotemporal features mentioned above can be nonlinearly fused with the state constraint components reflecting the boundary of device operation through a multi-head cross-attention mechanism to construct prediction intermediate features with physical awareness. Finally, the power distribution is parametrically modeled through the Mixture Density Network (MDN) output layer, and prediction results reflecting the expected value of prediction and prediction confidence that quantitatively characterizes the prediction fluctuation risk are produced simultaneously.

[0046] In some embodiments, regarding the necessity of dynamically adjusting the sliding window length, existing technologies typically employ a fixed sliding window to truncate historical sequences. However, in wind-solar-hydrogen storage coupled systems, net power exhibits strong non-stationarity due to environmental factors. If the window is too long, the model will incorporate too much outdated historical data, leading to a lag in prediction response when power changes abruptly (such as cloud cover or a sudden increase in wind speed), resulting in severe "truncation errors." If the window is too short, the data carries insufficient information and is highly susceptible to random noise interference from sensors, causing frequent, non-physical fluctuations in the prediction results.

[0047] In practice, the window length is dynamically adjusted based on the fluctuation state identification results. When the system is identified as being in a "steady-state operation" condition, the window length is automatically increased to smooth out high-frequency random noise in the original power sequence, and the stability of the prediction is improved by utilizing a longer historical benchmark. When a "sharp slope climb" or "instantaneous drop" condition is detected, the window length is immediately shortened to eliminate historical mean interference before the abrupt change point, enabling the model to focus on the current dramatic evolution trend. This significantly improves the prediction results' ability to quickly capture power abrupt changes, thus reserving sufficient response time for subsequent rapid power regulation of the electrolyzer.

[0048] In some embodiments, regarding the necessity of using multiple windows of different lengths to extract multi-scale features, the fluctuations in wind and solar power exhibit typical multi-timescale characteristics, including both minute-level instantaneous disturbances and hourly-level trend evolutions. Single-dimensional feature extraction methods suffer from the problem of "scale deficiency": short-scale features, while sensitive, lack a global perspective, making it difficult to determine whether the current fluctuation is a transient oscillation or a trend reversal; long-scale features, while possessing a sense of direction, filter out crucial local details, causing the prediction model's perception of the electrolyzer's operating boundary to become blurred, making it difficult to achieve accurate safety allocation.

[0049] This embodiment extracts multiple target data segments in parallel from the net power sequence by pre-setting multiple sliding windows with different lengths. Short-scale windows capture the instantaneous rate of change and high-frequency components of net power, reflecting the current dynamic impact intensity of the system; medium- and long-scale windows capture the mean level and periodic evolution trend of energy flow, reflecting the macroscopic energy background of the system. These target data segments corresponding to different time scales are fused with equipment state parameters (such as energy storage SOC) at the feature level, constructing a target feature vector that simultaneously considers local dynamics and global trends. This coupling of multi-scale features enables the model to identify whether the current power fluctuation is a pulse-type fluctuation requiring rapid replenishment by energy storage, or a trend-type evolution requiring adjustment of the electrolyzer's operating baseline, thereby greatly enhancing the feature vector's descriptive depth for complex operating conditions.

[0050] In some embodiments, the setting of the differential window length and combination method is important because different fluctuation states have fundamentally different requirements for feature extraction. If a single combination mapping logic is used, the model's generalization ability under complex changing conditions will be weakened. For example, under alternating oscillating and stable conditions, if the feature extraction weights cannot be automatically switched, the model will often output incorrect confidence assessments due to feature overload or feature scarcity, leading to inaccurate contraction or expansion of the execution power envelope.

[0051] Based on the fluctuation state identification results, this application determines the window parameters through the following differentiation strategy: Stable operation mode: A single long-period window is set (e.g., L=60 sampling points), and the combination method uses mean aggregation. In this mode, the feature vector focuses on reflecting the statistical stability of power, aiming to provide a smooth operating baseline for the electrolyzer and reduce unnecessary operation frequency of the equipment.

[0052] Continuous climbing / falling conditions: Set up a combination of short-period windows with a stepped distribution (e.g., L_1=5, L_2=15). The combination method uses differential fusion logic, focusing on extracting the first and second derivative features of power changes. In this mode, the feature vectors can keenly sense the acceleration of power evolution, guiding the model to quickly adjust the expected prediction value.

[0053] High-frequency oscillation condition: Set multi-scale overlapping windows (e.g., L_1=5, L_2=30, L_3=60). The combination method adopts weighted stacking processing, and the instantaneous fluctuation component is used as a "disturbance term" to logically map with the long-term trend component. In this mode, the redundant information of multiple time scales is used to lock the oscillation center, ensuring that the prediction confidence can truly reflect the current random risk, thereby defining a reasonable safety range for the electrolyzer.

[0054] Based on the above embodiments, by dynamically adjusting the window length and multi-scale feature combination method according to the fluctuation state, the problems of prediction response lag and scale perception lack caused by fixed sampling logic are solved, enabling the target feature vector to be deeply coupled with instantaneous dynamics and global trends, thus ensuring the accuracy of power allocation and the safety of system operation from the feature source.

[0055] In some embodiments, real-time fluctuation state identification is performed on the net power sequence to address the non-stationary characteristics of wind and solar energy input. When the identification result indicates that the current condition is "high-frequency oscillation," the sliding window parameters are adaptively adjusted. Specifically, the window length is reduced to improve the accuracy of capturing instantaneous power jumps, and "multi-scale differential hierarchical coupling" is determined as the window combination method. At this time, short-scale data segments (e.g., 10 seconds) representing instantaneous evolution trends and medium-scale data segments (e.g., 60 seconds) representing local energy averages are extracted in parallel from the net power sequence.

[0056] In determining the target feature vector, this process not only utilizes the truncated multi-scale target data segments but also deeply couples with equipment state parameters (such as electrolyte temperature and operating current density) that reflect the real-time response capability of the electrolyzer. By concatenating the normalized multi-scale time-series features with the physical constraint parameters of the electrolyzer, the generated target feature vector can simultaneously characterize the dynamic disturbances of external energy and the load-bearing boundaries of internal equipment. This mechanism, based on dynamically adjusting the sampling depth and topology according to the fluctuation state, solves the feature dilution and phase lag problems that exist in traditional fixed sampling windows during sudden changes in operating conditions, significantly enhancing the perception depth of subsequent prediction models for nonlinear fluctuations in the system.

[0057] In some embodiments, while obtaining the net power prediction sequence, a preset power prediction model simultaneously produces a prediction confidence level reflecting prediction uncertainty, which serves as the probability weight for subsequent safety controls. When extreme weather causes the prediction confidence level to decrease (e.g., the confidence level falls below a preset threshold of 0.8), this strategy implements conservative control logic by compressing the width of the execution power envelope. Specifically, by combining the minimum load limit (operational constraint) of the electrolyzer with the current low confidence interval, the determined execution power envelope converges towards the center, aiming to avoid frequent start-ups and shutdowns of the electrolyzer or crossing of the electrochemical safety boundary due to prediction errors.

[0058] Subsequently, the operating safety range is determined based on the converged execution power envelope, serving as the decision-making basis for the coordinated scheduling of the electrolyzer and energy storage equipment. When the predicted sequence shows a power surplus but low confidence, the determined operating safety range will prioritize locking the electrolyzer's target power within its rated efficiency range, while allocating the fluctuation pressure caused by the prediction error to the compensation power of the energy storage equipment. The energy storage equipment performs rapid power compensation based on the lower limit of the operating safety range to absorb excess instantaneous energy or make up for the power gap. Through this progressive logic of "confidence-driven envelope, envelope-defined range," passive protection for the long-life operation of the electrolyzer under complex weather conditions is achieved, and the flexible compensation capability of the energy storage equipment ensures the logical robustness of the entire coupled system in the event of power prediction distortion.

[0059] In some embodiments, the method may further include the following: S1: Obtain the net power measured sequence corresponding to the net power prediction sequence; S2: Calculate the residual value between the net power prediction sequence and the net power measured sequence to determine the prediction error; S3: Based on the prediction error, the target power of the electrolyzer, and the compensation power of the energy storage device, error attribution is identified, and the parameters of the preset power prediction model are iteratively updated using the identified error attribution results.

[0060] Specifically, at a preset sampling node after each prediction cycle, real-time power data fed back from the sensor is retrieved to form a net power measurement sequence. By aligning the net power prediction sequence and the net power measurement sequence in the time domain, the residual distribution of the two at each sampling point is calculated, thereby determining the prediction error for the current cycle. This error not only reflects the mathematical deviation of the algorithm but also contains coupling information between the randomness of wind and solar fluctuations and the non-ideal nature of hardware execution response.

[0061] When executing the error attribution identification logic, the prediction error is correlated with the target power of the electrolyzer and the compensation power of the energy storage device. If the prediction error increases significantly at the moment when the electrolyzer's power is limited, it is determined to be a power deviation caused by physical constraints on the execution side; if the error is accompanied by nonlinear fluctuations in the state of charge (SOC) of the energy storage device, it is determined to be an equilibrium failure caused by energy storage response lag. By constructing an error feature mapping matrix, the prediction error is deconstructed into multiple dimensions such as data quality interference, sudden changes in operating conditions, model structure shift, and execution response deviation, thus establishing accurate attribution results.

[0062] Based on the identified error attribution results, a targeted differentiated correction strategy is initiated to achieve rolling iteration. When the error is identified as being caused by sensor data drift or communication packet loss, a data reconstruction algorithm is triggered to smooth the input features using the historical trends of measured values. If the error originates from a sudden and drastic change in wind and solar power output, a window switching command is immediately executed. By dynamically adjusting the step size and depth of the sliding window, the model's sensitivity to capturing features under abrupt changes in operating conditions is improved.

[0063] For systematic errors with constant bias characteristics, linear or nonlinear bias correction operators are used to compensate for the prediction baseline of the next cycle. When it is found that the energy storage device fails to reach the compensation power target value within the preset time, the energy storage compensation correction strategy is executed, and the inertial lag at the execution level is offset by adjusting the adjustment coefficient of the charging and discharging power. When the residual value continuously exceeds the preset confidence threshold and the attribution to the environmental distribution drifts, the model incremental update is initiated. The latest measured samples are used to perform small-step online gradient descent on the weight parameters of the prediction model to ensure that the prediction logic always remains dynamically synchronized with the current physical evolution trend.

[0064] Based on the above embodiments, by introducing a rolling iteration mechanism based on error attribution identification, the prediction model is transformed from "static mapping" to "dynamic evolution". It can accurately select adaptive strategies such as data reconstruction or incremental model update according to error attribution, which significantly reduces the cumulative prediction deviation under complex and fluctuating operating conditions and ensures the coordinated response accuracy of electrolyzers and energy storage equipment during long-term operation.

[0065] In some embodiments, the method for determining the net power sequence based on the operating data may further include the following: S1: Perform time alignment processing on the running data to determine the corresponding running sequence to be processed; S2: Determine the target power component based on the operation sequence to be processed; wherein, the target power component includes the actual wind power generation, the actual photovoltaic power generation, the energy storage charging and discharging power, the station load power, and the reserved power parameters; S3: Perform power combining processing on the target power components to determine the net power sequence.

[0066] Specifically, refined spatiotemporal governance is performed on the acquired operational data. Timestamp integrity verification logic is used to identify logical breakpoints in the data stream, and a communication delay identification algorithm is combined to capture the lag duration during the backhaul process of different physical links. For identified missing data points, mean-based filling is performed using steady-state data from adjacent time points, or short-term trend completion is performed based on the power evolution slope before and after the missing window. For sampling phase misalignment caused by communication delays, timestamp playback logic is used to project the original sampled values ​​of each energy node onto a unified time reference axis, and resampling alignment is performed to eliminate asynchronous interference caused by heterogeneous communication links, thereby producing a time-consistent operational sequence to be processed.

[0067] Based on the operational sequence to be processed, target power components in various dimensions are deconstructed according to energy flow direction. These target power components encompass the actual wind and solar power generation on the supply side, and are also deeply coupled with the energy storage charging and discharging power on the equipment side. Specifically, a reserved power parameter is introduced to characterize the DC bus stability margin or the power gap required by grid limiting instructions, and the station load power consumed by the substation distribution circuits is acquired simultaneously. By decomposing the components of the operational sequence to be processed at each moment, it is ensured that each power dimension has a clear physical attribute and source basis.

[0068] Power combining is performed to generate the final net power sequence. Based on the energy conservation principle, the actual wind power generation, actual photovoltaic power generation, and dischargeable energy storage power are vector-summed, and the energy storage charging power, station load power, and reserved power parameters are subtracted. The resulting net power sequence physically represents the net energy flow that can be safely absorbed by the electrolyzer under the current operating conditions. When the electrolyzer is defined as the primary load to be absorbed, this sequence reflects the remaining capacity for renewable energy; when grid connection coordination is involved, this sequence characterizes the adjustable power boundary after deducting grid connection plans.

[0069] Based on the above embodiments, by using timestamp playback alignment and multi-source power component synthesis, data distortion caused by cross-protocol communication delays is eliminated, and accurate quantification of available power accuracy in complex energy flow environments is achieved, providing a deterministic input benchmark for subsequent load adjustment of electrolyzers within safety boundaries.

[0070] In some embodiments, the method of identifying fluctuation states in the net power sequence and determining sliding window parameters based on the identification results may further include the following: S1: Using multiple candidate observation windows of preset scales, multi-scale time-domain features are extracted from the net power sequence to determine fluctuation state indicators; wherein, the fluctuation state indicators include: the first difference, rate of change and second difference of the net power sequence, as well as the short window range, short window fluctuation intensity, medium window standard deviation and long window mean corresponding to each candidate observation window respectively; S2: Map the fluctuation state index to a preset multi-dimensional feature space for working condition identification and determine the current working condition type; wherein, the current working condition type includes stable working condition, continuous climbing working condition, rapid drop working condition, high frequency oscillation working condition, data abnormal working condition, and equipment disturbance working condition. S3: Based on the current working condition type, determine the matching window length from the multiple preset scales, and determine the corresponding window combination method as the sliding window parameter.

[0071] Among them, the short-scale window utilizes the range and fluctuation intensity to construct a highly sensitive "sensing probe," aiming to capture extreme jumps in power sequences at the sub-second level. The range index can more directly and quickly identify instantaneous energy drops caused by cloud cover or wind turbine shutdown than the standard deviation, avoiding the dilution of abrupt changes in characteristics by statistical averaging under small sample sizes. Combined with the relative quantity index of fluctuation intensity, it can logically and accurately distinguish between trend components during a smooth ramp-up process and high-frequency pulses under drastic operating conditions, thus providing the most direct physical basis for rapid early warning of electrolyzer operating boundaries.

[0072] The mesoscale window uses standard deviation as the core indicator, focusing on assessing the overall consistency and stability of the power distribution. Compared to the range, which is susceptible to single outliers, standard deviation quantifies the dispersion of power points relative to the mean over a continuous period, thus accurately answering whether the power flow is "stable" in the current time domain. This extraction of distribution characteristics effectively filters out stray noise interference, providing statistical support for determining whether the system has entered a quasi-steady-state evolution phase, and guiding electrolyzers to avoid unnecessary frequent adjustments in fluctuating environments.

[0073] The long-scale window performs deep filtering of high-frequency clutter by calculating the mean, aiming to pinpoint the current "energy baseline" and quasi-steady-state benchmark of the power station. Since the dynamic evolution characteristics in the short time domain have been decoupled in multiple dimensions within the short and medium-scale windows, the long-scale window does not require redundant fluctuation indicators. Instead, it extracts the central axis of power evolution through averaging. This design achieves dimensional orthogonality of the feature space, enabling the final generated feature vector to simultaneously consider the sensitivity of micro-pulses and the stability of macro-energy trends, ensuring that subsequent control logic always follows the true energy center of gravity of the system.

[0074] Specifically, before performing fluctuation state identification, a set of candidate observation windows with physical time significance are pre-defined as perception benchmarks. These include short-scale windows (10 to 30 seconds) for capturing micro-perturbations, medium-scale windows (60 to 120 seconds) for capturing evolutionary trends, and long-scale windows (300 to 900 seconds) for identifying the operating baseline. Using these pre-defined observation scales, point-to-point dynamic evolution features and surface-to-surface statistical distribution indicators are extracted in parallel from the net power sequence. The "velocity" of the fluctuation is quantified by calculating the first-order difference and the rate of change, and the "acceleration" of the fluctuation is characterized by the second-order difference. Simultaneously, the range, fluctuation intensity, standard deviation, and mean corresponding to each candidate observation window are extracted to construct a fluctuation state index that can holographically describe the energy evolution characteristics of the system. Since the candidate observation windows have a defined physical duration before identification, it ensures that the feature extraction logic has a stable referential basis at different data resolutions.

[0075] The extracted fluctuation state indicators are projected onto a preset multi-dimensional feature space, and operating condition identification is performed by evaluating the topological distribution of each indicator in the space. Within this space, if the first-order difference continuously exceeds a preset slope threshold and the second-order difference tends to zero, it is identified as a continuous climbing condition or a rapid drop condition; if the fluctuation intensity in the short window increases sharply and the standard deviation in the medium window increases significantly, it is identified as a high-frequency oscillation condition. For data jumps caused by sensor drift or communication packet loss, physical boundary verification logic is used to determine them as abnormal data conditions. This feature space-based mapping mechanism achieves accurate classification of complex variable operating conditions and can effectively isolate nonlinear disturbances caused by random environmental fluctuations and equipment actions.

[0076] Based on the identified current operating condition type, the target sliding window parameters for prediction are determined through a pre-defined operating condition-parameter mapping matrix. When the operating condition is identified as stable, a long-scale window is matched from multiple pre-defined scales, and a mean-based combination method is used to filter out random noise through long-term smoothing effects. If the operating condition is identified as a rapidly changing condition such as a rapid drop, a short-scale window is immediately matched, and a differential fusion-based combination method is switched to eliminate interference from outdated historical means, enabling the feature vector to instantly pinpoint the turning point in power evolution. For high-frequency oscillating operating conditions, a multi-scale feature stacking strategy is implemented, weighting and splicing the instantaneous peaks and valleys captured by the short window with the oscillation centers captured by the medium window. This "pre-defined detection, then state activation" logic ensures that the sliding window parameters are deeply coupled with the current physical evolution situation.

[0077] Based on the above embodiments, by introducing a candidate observation window of a preset scale and a multi-dimensional feature space mapping logic, the referential ambiguity problem in the process of determining sliding window parameters is eliminated, and accurate classification of complex fluctuations is achieved in a heterogeneous sampling resolution environment. This ensures that the target feature vector can reconstruct the sampling depth and combined weights in real time according to the evolution characteristics of the working conditions, and significantly improves the response sensitivity and perception depth of the prediction logic under dynamic changing working conditions.

[0078] In some embodiments, when a stable operating condition is identified, a long-scale window is matched, and the corresponding window combination method is a single-scale averaging combination. Under this condition, the power average within a long-scale window of 300 to 900 seconds is extracted, and the smoothing effect of the long time domain is used to filter out high-frequency random ripples, providing a stable energy evolution benchmark for the electrolyzer.

[0079] When the condition is identified as a continuous uphill climb, a short-scale window and a medium-scale window are matched, and the corresponding window combination method is a cascaded combination of trend features. The short-scale window of 10 to 30 seconds is used to capture the current instantaneous uphill rate, and the medium-scale window of 60 to 120 seconds is combined to verify the consistency of the slope. By concatenating the features of the two scales, the feature vector is ensured to have both response sensitivity and trend stability.

[0080] When a rapid drop condition is identified, a short-scale window is matched, and the corresponding window combination method is a single-scale real-time update combination. At this time, only the sampled data within the 10-second short-scale window is retained. By forcibly removing old and stable features in the medium and long-scale windows, the historical mean is prevented from smoothing and diluting the current drastic signal, so that the feature vector can accurately pinpoint the physical turning point of the power drop.

[0081] When the signal is identified as a high-frequency oscillation, a short-scale window and a medium-scale window are matched, and the corresponding window combination method is a multi-scale statistical fusion combination. The instantaneous extreme value features extracted by the short-scale window and the standard deviation features extracted by the medium-scale window are weighted and fused, so that the instantaneous peaks and troughs of the fluctuation and the dispersion of energy distribution are reflected in the same feature vector, thereby achieving decoupled representation of complex non-stationary signals.

[0082] When an abnormal data condition is identified, a long-scale window is matched, and the corresponding window combination method is a historical mean compensation combination. When sensor jumps or communication packet loss are detected, the real-time inputs of short-scale and medium-scale are logically blocked, and the historical mean of a long-scale window of more than 300 seconds is matched as the substitute feature for the current moment to ensure the continuity of the prediction logic.

[0083] When the condition is identified as equipment disturbance, a short-scale window is matched, and the corresponding window combination method is an asymmetric weighted combination. For step disturbances caused by energy storage unit tripping or electrolyzer load command jumps, a 10-second short-scale window is matched, and higher weight coefficients are assigned to the sampling data of the nearest time. The feature depth is quickly reconstructed to adapt to the new operating equilibrium point after the equipment action.

[0084] In some embodiments, the step of identifying the fluctuation state of the net power sequence and determining the sliding window parameters based on the identification result may further include: S1: Using multiple observation windows of preset scales, multi-time-domain features are extracted from the net power sequence to determine fluctuation state indices; wherein, the fluctuation state indices include first-order difference, rate of change and second-order difference representing instantaneous evolution trend, as well as short window range, short window fluctuation intensity, medium window standard deviation and long window mean corresponding to each observation window; S2: Map the fluctuation state index to a preset multi-dimensional feature space to identify the working condition and determine the current working condition type; S3: Based on the current operating condition type, use change point detection logic to identify the energy evolution inflection point in the net power sequence, and perform dynamic backtracking on the starting boundary of the target sliding window based on the energy evolution inflection point; S4: Determine the dynamic scaling factor of the window length based on the fluctuation state index, and calculate the window length and window combination method in the sliding window parameters in combination with the initial boundary.

[0085] Specifically, before performing fluctuation state identification, a set of observation windows with differentiated time-series scales are pre-configured as sensory benchmarks to perform initial multi-dimensional feature deconstruction on the net power sequence. Using these pre-defined observation scales, the rate of change and first-order difference characterizing the instantaneous intensity of energy evolution are extracted in parallel, and the second-order difference is used to capture the acceleration characteristics of power fluctuations. Simultaneously, the short-window range, short-window fluctuation intensity, medium-window standard deviation, and long-window mean corresponding to each observation window are calculated to construct a fluctuation state index that can take into account both micro-pulses and macro-trends. Since the observation windows serve as pre-defined static scales, they provide a definite referential basis for index calculation, eliminating logical redundancy in parameter setting under unknown operating conditions.

[0086] The acquired fluctuation state indicators are projected onto a pre-defined multi-dimensional feature space, and precise operating condition identification is performed by evaluating the clustering distribution of feature vectors in the space. When the spatial coordinates show that the first-order difference continuously exceeds the slope threshold and is accompanied by a significant shift in the second-order difference, it is determined that the current condition is either a continuous climb or a rapid drop. If the range and fluctuation intensity within a short window show a non-linear surge, it is identified as a high-frequency oscillation condition. This feature space-based identification mechanism can effectively decouple the random fluctuations on the new energy side from the non-stationary disturbances on the equipment side, providing operating condition context for the dynamic optimization of subsequent execution parameters.

[0087] For the identified variable operating condition types, a change point detection algorithm is invoked to scan the net power sequence for instantaneous nodes where the energy evolution law undergoes a fundamental shift, identifying these as energy evolution inflection points. The algorithm uses these inflection points as physical alignment benchmarks to dynamically backtrack the starting boundary of the target sliding window. If the current power drop is detected to have started 15 seconds prior, the starting point of the target sliding window is automatically aligned to that inflection point, rather than mechanically using a preset level. This ensures that the data sequence entering the prediction model completely and purely contains the current physical evolution process, physically avoiding the dilution interference of outdated historical data on abrupt changes.

[0088] Based on the established initial boundary, a dynamic scaling factor for the window length is calculated using the energy gradient and dispersion in the fluctuation state index. This scaling factor, as a continuous variable, can adjust the sampling depth in real time, achieving dynamic resonance between the window duration and the power fluctuation frequency. Combined with the dynamically backtracked initial boundary, the sliding window parameters are finally calculated and generated. This asymmetric feature sampling logic guided by a variable point ensures that the target feature vector is locked within the most valuable physical evolution interval, guaranteeing sub-second-level precise alignment between the feature extraction process and the true physical dynamics of the coupled architecture.

[0089] Based on the above embodiments, by introducing a boundary dynamic backtracking and scaling factor adjustment mechanism driven by change point detection, deep coupling between sliding window parameters and power evolution physical process is achieved. This solves the problems of feature truncation and time-domain perception lag that exist in traditional fixed-level windows at the moment of operating condition switching. It significantly enhances the accuracy of feature vectors in characterizing the non-stationary features of the system under complex variable operating conditions, and provides a physically consistent data underlying layer for the accurate distribution of the safe power envelope of the electrolytic cell.

[0090] Furthermore, considering the extremely strong non-stationarity of the net power sequence in the wind-solar-hydrogen storage coupling architecture, the effectiveness of the feature extraction logic depends on whether the sampling window can accurately cover the current physical evolution cycle. After determining the operating condition type, in order to eliminate the feature truncation and phase lag caused by the traditional fixed-gear window, this embodiment performs dynamic reconstruction of the window boundary based on physical change point driven.

[0091] A real-time scan of the net power sequence is performed using change point detection algorithms (such as cumulative sum CUSUM or log-likelihood ratio test). By capturing abrupt changes in the mean or variance of the statistical distribution, the algorithm accurately identifies instantaneous nodes where the power evolution logic undergoes a fundamental shift, defining them as energy evolution inflection points. Using this inflection point as the physical axis, the algorithm dynamically backtracks the starting boundary of the target sliding window. If the identification result shows that the current power drop began 14.5 seconds before the current moment, the starting position of the target sliding window is automatically backtracked 14.5 seconds into the historical time domain, ensuring that the captured data segment physically and semantically completely contains only this wave of energy evolution. This asymmetric sampling strategy avoids the dilution of current abrupt changes by stale steady-state data from the source, giving the target feature vector an extremely high signal-to-noise ratio.

[0092] While determining the initial boundary, a dynamic scaling factor for the window length is calculated using the second-order difference and short-window fluctuation intensity in the fluctuation state index through a preset nonlinear mapping function. This scaling factor represents the "elasticity" of the sampling window in the time domain. When an increase in the intensity of high-frequency random oscillations is detected, the scaling factor drives the window length to continuously evolve in the shortening direction to enhance the accuracy of capturing instantaneous peaks; when the power enters a quasi-steady-state evolution, the scaling factor guides the window to smoothly lengthen. Combining the initial boundary determined by dynamic backtracking and the sampling depth calculated by the dynamic scaling factor, the final sliding window parameters are generated. In terms of window combination, non-uniform time weighting coefficients are assigned according to the distance of backtracking, giving higher feature weights to data near the change point. Through this synergistic logic of "initial boundary optimization" and "continuous scaling of sampling depth," sub-second-level locking of the feature extraction process and the dynamic physical processes of wind, solar, and hydrogen storage is achieved, significantly enhancing the sensitivity of subsequent prediction models to the operating boundaries of the electrolyzer.

[0093] Based on the above embodiments, by introducing dynamic backtracking of the starting boundary driven by the energy evolution inflection point and continuous scaling factor adjustment, the problems of physical phase lag and incomplete feature representation in the process of determining sliding window parameters are solved, and the precise physical alignment of feature sampling boundary and power evolution starting point is achieved, which greatly improves the accuracy of the target feature vector in describing the non-stationary and nonlinear fluctuation characteristics of the system.

[0094] In some embodiments, determining the matching window length from the plurality of preset scales according to the current working condition type, and determining the corresponding window combination method as the sliding window parameter, may specifically include: Based on the current working condition type, a target duration scale corresponding to the current working condition type is selected from the multiple preset scales, and the window length is determined. Based on the current working condition type, using the preset working condition-parameter mapping relationship, a target fusion strategy matching the current working condition type is determined, and the window combination method is determined; The parameters of the sliding window are determined by performing parameter encapsulation based on the window length and the window combination method.

[0095] Specifically, during parameter matching, the architecture selects a target duration scale from multiple pre-defined candidate observation scales based on the identified current operating condition type, choosing one with specific time-domain representation capabilities. When the identification result indicates a stable operating condition, the matching logic locks a pre-defined scale (e.g., 300 to 900 seconds) with long-term smoothness characteristics as the determined window length, aiming to extract a quasi-steady-state energy benchmark by covering a longer time span. If the identification result switches to a rapid drop condition, the matching logic immediately shrinks the sampling scale, selecting a highly sensitive short-scale window (e.g., 10 to 20 seconds) as the window length. This dynamic selection of duration scale based on physical evolution characteristics ensures that the subsequently extracted data segments are physically aligned with the frequency characteristics of the current power fluctuations at the sub-second level.

[0096] While locking the window length, a target fusion strategy that matches the physical characteristics of the current operating condition is determined using a pre-defined operating condition-parameter mapping relationship, thereby determining the corresponding window combination method. For high-frequency oscillation operating conditions, the mapping logic activates a "multi-scale feature cascade" target fusion strategy, requiring the spatial topological combination of pulse features extracted from short-scale windows and distribution stability features extracted from medium-scale windows. For operating conditions with data anomalies, the mapping logic switches to a "historical mean compensation" strategy, forcibly ignoring damaged real-time window data by adjusting the weight allocation ratio of each observation window. This mechanism, where the fusion strategy reconstruction is directly driven by operating condition semantics, transforms the window combination method from a static format definition into a data assembly protocol that can dynamically evolve with the external environment.

[0097] Ultimately, the architecture encapsulates the determined window length (reflecting sampling depth) and window combination method (reflecting logical structure) into sliding window parameters. These parameters serve as the execution benchmark for subsequent data extraction and feature reconstruction, encompassing not only specific time-series numerical information but also topology reconfiguration instructions for specific operating conditions. Through this step-by-step mapping chain, the ambiguous fluctuation characteristics of the new energy side are transformed into algorithmic control instructions with high physical consistency. This solves the technical bottleneck of balancing feature integrity and real-time prediction in complex coupled scenarios with heterogeneous data, providing precise logical guidance for the subsequent generation of high-dimensional target feature vectors.

[0098] In some embodiments, the method of extracting a target data segment from the net power sequence according to the window length and fusing the target data segment with the device status parameters according to the window combination method to determine the target feature vector may further include the following: S1: Based on the window length, extract data sampling points of corresponding duration from the net power sequence to determine the target data segment; S2: Normalize the target data segment to determine the target temporal feature components; S3: Based on the window combination method, the target temporal feature component and the device state parameter are feature aligned and vector concatenated to determine the target feature vector.

[0099] Specifically, based on the window length corresponding to each identification condition, the historical time segment in the net power sequence is accurately located. For each sliding window, data sampling points of the corresponding sampling step size are extracted from the sequence to form a target data segment with a physical time scale. In scenarios with multi-scale window combinations, downsampling or interpolation alignment is performed on data segments of different lengths to ensure that the sampling point sets at each time scale have a unified descriptive benchmark in the logical dimension, thereby producing a target data segment that reflects a multi-dimensional slice of power evolution.

[0100] Dimensionless preprocessing is performed to eliminate dimensional differences between heterogeneous energy components. Min-Max Scaling is used to map the numerical magnitudes of the target data segments to the [0,1] interval, ensuring the stability of gradient descent during model training. Tensor shape transformation is performed on the input normalization of the deep learning model, reconstructing the one-dimensional time series into a matrix representation, where the number of channels corresponds to the number of truncated target data segments. By introducing multi-dimensional tensor reshaping logic, the original time sampling points are transformed into target temporal feature components with spatial semantics.

[0101] Cross-modal feature alignment and vector concatenation logic are executed. Given that the sampling frequency of device state parameters such as electrolyte temperature, internal pressure, and state of charge of energy storage devices is typically lower than that of the power sequence, broadcast mapping or zero-order hold logic is used to extend the low-frequency acquired device state vectors to the same time step as the target temporal feature components. The aligned state parameters are then used as enhancement dimensions and laterally pressed into the channel axis of the target temporal feature components using a vector concatenation operator. This fusion process logically locks the dynamic energy evolution characteristics with static or low-frequency physical boundary constraints, producing a target feature vector that holographically describes the current operational status of the coupled architecture.

[0102] Based on the above embodiments, by performing tensor reconstruction and cross-modal alignment and splicing on multi-scale data segments, the sampling frequency deviation and dimensional differences between heterogeneous data are eliminated, and the deep coupling of power evolution characteristics and device physical constraints in the feature space is realized. This significantly enhances the semantic expression depth of input features and provides a standardized data layer for the prediction model to accurately perceive the physical operating boundary.

[0103] In some embodiments, during feature reconstruction, the "window combination method" acts as the physical topology blueprint and dimensionality protocol for generating the target feature vector. Taking the identified high-frequency oscillation condition as an example, if the determined window combination method is a multi-scale spatiotemporal stacking combination, this method directly defines the internal organization of the feature vector, requiring an asymmetric mapping between the pulse component extracted by the short-scale window and the distribution discreteness component extracted by the medium-scale window. Under this constraint, the subsequent feature alignment action is not simply clock synchronization, but rather, based on the spatiotemporal correlation logic preset by this combination method, the time-series feature components after multi-scale fusion are normalized and mapped to the sampling resolution of the real-time collected DC bus voltage, energy storage SOC, and other device state parameters. The final vector splicing process strictly follows the feature priority order defined by this blueprint, cascading the multi-scale components reflecting new energy fluctuations with the state parameters reflecting the physical constraints of the controlled equipment. Through this fusion mechanism guided by the window combination method, the generated target feature vector can transform discrete time-series segments into semantic tensors with physical causal relationships, ensuring accurate alignment between the feature expression depth at the input of the prediction model and the evolutionary characteristics of the current operating condition.

[0104] In some embodiments, the step of aligning the target temporal feature components with the device state parameters and concatenating their vectors according to the window combination method to determine the target feature vector may specifically include: Based on the window combination method, determine the corresponding feature alignment granularity and vector topology; Using the aforementioned feature alignment granularity, resampling matching of the target temporal feature components and the device state parameters is performed at the sampling frequency to determine the alignment feature matrix; According to the vector topology, the alignment feature matrix is ​​subjected to dimensional mapping and concatenation of multidimensional features to determine the target feature vector.

[0105] Specifically, in determining the target feature vector, the window combination method acts as the logical protocol and spatiotemporal blueprint for heterogeneous data fusion. During execution, based on the feature alignment granularity determined by the window combination method, dynamic matching of the sampling frequency is performed between the target time-series feature components from the new energy side and the equipment status parameters from the electrolyzer side. When the window combination method corresponds to a single-scale instantaneous update combination under rapid power drop conditions, the feature alignment granularity is set to high-precision millisecond level. By performing linear interpolation on the equipment status parameters, sub-second physical alignment is ensured between parameters such as DC voltage and current reflecting the equipment operating boundary and the time-series features of instantaneous power drop.

[0106] Based on the determined alignment feature matrix, a pre-defined vector topology structure using window combination is employed to perform dimensional mapping of multi-dimensional features. The vector topology structure defines the priority and spatial weights of the physical quantities in the feature space. For high-frequency oscillation conditions, this structure places the range component reflecting local fluctuations at the beginning of the vector and cascades the temperature parameter reflecting equipment thermal stability with the pressure parameter reflecting energy level according to the topology protocol. This ordered concatenation is not a simple data stacking, but rather, through the constraints of the topology structure, transforms the originally discrete sampling points into input tensors with physical semantic relationships, enabling the subsequent prediction model to directly deconstruct the "fluctuation-response" coupling logic from the topological distribution of the vector.

[0107] The final target feature vector serves as the direct driving force of the prediction model, and its internal structure deeply integrates the operating context implied by the window combination method. Because the feature stitching process strictly adheres to the dimension space defined by the vector topology, it ensures that the generated target feature vector not only contains multi-scale fluctuation information from the new energy side but also dynamically couples the physical constraints of the controlled equipment at specific operating points. Through this alignment and stitching mechanism guided by the combination method, a logical leap from the original heterogeneous data flow to feature vectors with high physical consistency is achieved, providing a standardized feature layer for the prediction model to accurately lock the system's safe operation envelope under dynamic and varying operating conditions.

[0108] In some embodiments, the method for determining the execution power envelope based on the net power prediction sequence, the prediction confidence level, and preset operating constraints may further include the following: S1: Based on the predicted confidence level and the preset mapping relationship between the confidence level and the adjustment deviation, determine the confidence adjustment margin corresponding to the predicted confidence level; S2: Use the confidence adjustment margin to perform interval mapping on the net power prediction sequence to determine the initial power execution interval; S3: Based on the preset operating constraints, the initial power execution range is subjected to amplitude limiting and rate of change constraint processing to determine the execution power envelope; The execution power envelope includes the upper limit and lower limit of the execution power of the target wind-solar-hydrogen storage coupling system at each sampling time within the corresponding preset time period.

[0109] Specifically, a preset probability deviation mapping operator is used to quantify and transform the output prediction confidence level. A nonlinear mapping function is established between the confidence level and the adjustment deviation. When the prediction confidence level is in a high range, the confidence level adjustment margin is reduced to reflect the lower prediction risk under the current operating conditions, allowing for a narrower power fluctuation range. When the prediction confidence level drops due to severe wind and solar power fluctuations, the confidence level adjustment margin is automatically expanded to offset the prediction uncertainty by increasing power redundancy. This mapping mechanism transforms the mathematical probability distribution into a physical control margin, providing a dynamic risk boundary for subsequent interval mapping.

[0110] Based on this, the predicted power is used as the central value, and symmetrical or asymmetric interval mapping is performed by adjusting the margin according to the determined confidence level. For each predicted point in the future, the margin value is used to construct the corresponding power coverage bandwidth, thereby determining the initial power execution interval. This interval forms a "power tunnel" in the time domain that dynamically scales with prediction fluctuations and changes in confidence level, initially outlining the theoretical range within which the electrolyzer can operate while ensuring the absorption rate.

[0111] Furthermore, the physical characteristics of the electrolyzer and energy storage hardware are introduced as preset operational constraints to refine the initial power execution range. The minimum operating load, rated power upper limit, and maximum allowable ramp rate of the electrolyzer are retrieved to perform point-by-point limiting and rate-of-change constraints on the initial power execution range. Rigid truncation is applied to portions of the range exceeding the physical range, and slope smoothing is performed on portions of the range exceeding the equipment response speed, thereby determining the final execution power envelope. This execution power envelope clearly defines the absolute upper and lower limits of the execution power at each sampling time, ensuring dynamic alignment between the issued control commands and the physical boundaries.

[0112] Based on the above embodiments, by introducing a collaborative mapping logic of confidence adjustment margin and physical operating constraints, the transformation of prediction uncertainty into a safe operating domain is realized. This not only enhances the system's ability to withstand wind and solar fluctuation risks, but also ensures that the electrolyzer operates within the physical boundaries that meet electrochemical characteristics, effectively avoiding the risk of frequent equipment overruns and lifespan degradation caused by prediction deviations.

[0113] In some embodiments, the method of determining a safe operating range based on the execution power envelope, and determining the target power of the electrolyzer and the compensation power of the energy storage device based on the safe operating range, may further include the following in specific implementations: S1: Based on the upper limit and lower limit of the execution power included in the execution power envelope, determine the safe operating range of the electrolytic cell at each sampling time within the corresponding preset time period; S2: Perform amplitude limiting mapping processing on the net power prediction sequence based on the safe operating range to determine the target power of the electrolytic cell; S3: Determine the compensation power of the energy storage device based on the net power prediction sequence and the target power of the electrolyzer.

[0114] Specifically, by using defined upper and lower limits of execution power, a specific operational range is delineated for each sampling point of the electrolyzer within a future time period. By mapping the envelope values ​​to the corresponding time axis, the safe operating range for each sampling moment is determined. This range, at the physical level, defines the extreme values ​​of load fluctuations that the electrolyzer is allowed to perform under the current health state, thermal balance parameters, and load rate constraints, providing a rigid physical constraint benchmark for subsequent power allocation.

[0115] The net power prediction sequence is logically aligned point-by-point with the safe operating range at each sampling time. A limiting mapping process is performed: when the net power prediction sequence value is within the safe operating range, the predicted value is retained as the basis for electrolyzer operation; when the predicted value exceeds the upper limit of the execution power or falls below the lower limit, the command value at that time is forcibly corrected to the corresponding boundary extreme value, thereby determining the target power of the electrolyzer. This limiting mapping process ensures that the issued load command is always locked within the preset safe execution tunnel, eliminating the risk of electrolyzer operation exceeding limits due to algorithm prediction deviations.

[0116] To address power fluctuations that electrolyzers cannot fully absorb due to physical limitations, an energy storage device is introduced to perform residual mitigation. The net power prediction sequence and the target power of the electrolyzer are subtracted in real time to determine the power deviation. This deviation is then used as input to determine the compensation power of the energy storage device. Through a coordinated control logic that prioritizes the electrolyzer's response to fluctuations within its safe range and the energy storage device's response to residuals outside its range, precise power balance is achieved under conditions of severe power fluctuations in the coupled architecture. This ensures the safe operation of the electrolyzer while maximizing the local absorption efficiency of new energy sources.

[0117] Based on the above embodiments, by constructing a safe operating range and executing limiting mapping and power difference logic, deep collaboration between the electrolyzer and the energy storage device under physical constraints is achieved. This ensures that the electrolyzer always operates within the safe execution envelope and utilizes the energy storage device to accurately offset the prediction residuals, significantly improving the operational safety level and power allocation accuracy of the multi-energy complementary architecture under dynamic fluctuation conditions.

[0118] As can be seen from the above, the data processing method for a wind-solar-hydrogen storage coupling system provided in this specification acquires the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device; based on the operating data, a net power sequence is determined; the net power sequence is subjected to fluctuation state identification, and a sliding window parameter is determined based on the identification result; wherein the sliding window parameter includes a window length and a window combination method; based on the window length, multiple target data segments are extracted from the net power sequence, and the multiple target data segments are fused with the equipment status parameters according to the window combination method to determine a target feature vector; wherein one target data segment corresponds to one time scale; using a preset power prediction model, a net power prediction sequence and prediction confidence are determined based on the target feature vector; based on the net power prediction sequence, the prediction confidence, and preset operating constraints, an execution power envelope is determined; based on the execution power envelope, an operating safety range is determined, and based on the operating safety range, the target power of the electrolyzer and the compensation power of the energy storage device are determined. In this way, by dynamically determining the sliding window parameters and extracting multi-scale target data segments through fluctuation state identification, the problems of insufficient prediction accuracy and response lag caused by fixed feature extraction in existing technologies are solved. Furthermore, by combining prediction confidence and operational constraints to determine the execution power envelope and operational safety range, the coordinated control of the electrolyzer target power and the energy storage equipment compensation power is realized, meeting the requirements for safe operation and accurate allocation under drastic power fluctuations.

[0119] See Figure 2 As shown in the embodiments of this specification, a specific electronic device is also provided, wherein the electronic device includes a network communication port 201, a processor 202 and a memory 203, and the above structures are connected by internal cables so that the various structures can perform specific data interaction.

[0120] Specifically, the network communication port 201 can be used to acquire the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device.

[0121] The processor 202 can specifically be used to: determine a net power sequence based on the operating data; identify fluctuation states in the net power sequence and determine sliding window parameters based on the identification results; wherein the sliding window parameters include window length and window combination method; extract multiple target data segments from the net power sequence according to the window length, and fuse the multiple target data segments with the device state parameters according to the window combination method to determine a target feature vector; wherein one target data segment corresponds to one time scale; determine a net power prediction sequence and prediction confidence based on the target feature vector using a preset power prediction model; determine an execution power envelope based on the net power prediction sequence, the prediction confidence, and preset operating constraints; determine an operating safety range based on the execution power envelope, and determine the target power of the electrolyzer and the compensation power of the energy storage device based on the operating safety range.

[0122] The memory 203 can be used to store the corresponding instruction program.

[0123] Based on the above method, the relevant structural performance of electronic devices can be effectively utilized to improve the data processing speed of electronic devices and efficiently realize a data processing method for wind-solar-hydrogen storage coupling systems.

[0124] In this embodiment, the network communication port 201 can be a virtual port bound to different communication protocols, thereby enabling the sending or receiving of different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. Furthermore, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM or CDMA; it can also be a Wi-Fi chip; or it can be a Bluetooth chip.

[0125] In this embodiment, the processor 202 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification is not limiting.

[0126] In this embodiment, the memory 203 may include a hierarchy. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0127] This specification also provides a computer-readable storage medium based on the above-described data processing method for a wind-solar-hydrogen storage coupling system, which acquires the operating data and equipment status parameters of a target wind-solar-hydrogen storage coupling system. The target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device. Based on the operating data, a net power sequence is determined. Fluctuation state identification is performed on the net power sequence, and sliding window parameters are determined based on the identification results. The sliding window parameters include a window length and a window combination method. Based on the window length, multiple target data segments are extracted from the net power sequence, and the multiple target data segments are fused with the equipment status parameters according to the window combination method to determine a target feature vector. Each target data segment corresponds to a time scale. Using a preset power prediction model, a net power prediction sequence and prediction confidence level are determined based on the target feature vector. An execution power envelope is determined based on the net power prediction sequence, the prediction confidence level, and preset operating constraints. An operating safety range is determined based on the execution power envelope, and the target power of the electrolyzer and the compensation power of the energy storage device are determined based on the operating safety range.

[0128] In this embodiment, the storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured according to standards specified in the communication protocol for network connection communication.

[0129] In this embodiment, the specific functions and effects implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other embodiments, and will not be repeated here.

[0130] See Figure 3 At the software level, this specification also provides a data processing device for a wind-solar-hydrogen storage coupling system, which may specifically include the following structural modules: The data acquisition module 301 is used to acquire the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein, the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device; The sequence determination module 302 is used to determine the net power sequence based on the operating data; The parameter determination module 303 is used to identify the fluctuation state of the net power sequence and determine the sliding window parameters based on the identification result; wherein, the sliding window parameters include window length and window combination method; The vector determination module 304 is used to extract multiple target data segments from the net power sequence according to the window length, and fuse the multiple target data segments with the device status parameters according to the window combination method to determine the target feature vector; wherein, one target data segment corresponds to one time scale; The power prediction module 305 is used to determine the net power prediction sequence and prediction confidence based on the target feature vector using a preset power prediction model. The power envelope determination module 306 is used to determine the execution power envelope based on the net power prediction sequence, the prediction confidence level, and preset operating constraints. The power determination module 307 is used to determine the safe operating range based on the execution power envelope, and to determine the target power of the electrolyzer and the compensation power of the energy storage device based on the safe operating range.

[0131] In some embodiments, the above-described apparatus further includes: acquiring a net power measured sequence corresponding to the net power prediction sequence; calculating the residual value between the net power prediction sequence and the net power measured sequence to determine the prediction error; identifying the error attribution based on the prediction error, the target power of the electrolyzer, and the compensation power of the energy storage device; and using the identified error attribution results to iteratively update the parameters of the preset power prediction model.

[0132] In some embodiments, the sequence determination module 302 performs time alignment processing on the running data to determine the corresponding running sequence to be processed; and determines the target power component based on the running sequence to be processed; wherein the target power component includes wind power generation, photovoltaic power generation, energy storage charging and discharging power, station load power, and reserved power parameters; and performs power synthesis processing on the target power component to determine the net power sequence.

[0133] In some embodiments, the parameter determination module 303, in specific implementation, performs multi-scale time-domain feature extraction on the net power sequence to determine fluctuation state indicators; wherein, the fluctuation state indicators include first-order difference, rate of change, second-order difference, short-window range, short-window fluctuation intensity, medium-window standard deviation, and long-window mean; the fluctuation state indicators are mapped to a preset multi-dimensional feature space for operating condition identification to determine the current operating condition type; wherein, the current operating condition type includes stable operating condition, continuous climbing operating condition, rapid drop operating condition, high-frequency oscillation operating condition, data anomaly operating condition, or equipment disturbance operating condition; according to the current operating condition type, using a preset mapping relationship, the window length and window combination method matching the current operating condition type are determined as the sliding window parameters.

[0134] In some embodiments, the vector determination module 304, in its specific implementation, extracts data sampling points of corresponding duration from the net power sequence according to the window length to determine the target data segment; performs normalization processing on the target data segment to determine the target time-series feature component; and performs feature alignment and vector concatenation with the device state parameters according to the window combination method to determine the target feature vector.

[0135] In some embodiments, the power envelope determination module 306, in its specific implementation, determines a confidence adjustment margin corresponding to the predicted confidence level based on the predicted confidence level and a preset mapping relationship between confidence level and adjustment deviation; uses the confidence adjustment margin to perform interval mapping processing on the net power prediction sequence to determine the initial power execution interval; and performs amplitude limiting and rate of change constraint processing on the initial power execution interval based on the preset operating constraints to determine the execution power envelope; wherein, the execution power envelope includes the upper limit and lower limit of the execution power of the target wind-solar-hydrogen storage coupling system at each sampling time within the corresponding preset time period.

[0136] In some embodiments, the power determination module 307, in its specific implementation, uses the upper limit of the execution power and the lower limit of the execution power to perform amplitude limiting processing on the net power prediction sequence to determine the target power of the electrolyzer; and determines the compensation power of the energy storage device based on the net power prediction sequence and the target power of the electrolyzer.

[0137] It should be noted that the units, devices, or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described by dividing them into various modules according to their functions. Of course, in implementing this specification, the functions of each module can be implemented in the same software and / or hardware, or modules that implement the same function can be implemented by a combination of sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. For example, units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection between the devices or units shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0138] As can be seen from the above, a data processing device for a wind-solar-hydrogen storage coupling system provided in the embodiments of this specification acquires the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device; based on the operating data, a net power sequence is determined; the net power sequence is subjected to fluctuation state identification, and a sliding window parameter is determined based on the identification result; wherein the sliding window parameter includes a window length and a window combination method; based on the window length, multiple target data segments are extracted from the net power sequence, and the multiple target data segments are fused with the equipment status parameters according to the window combination method to determine a target feature vector; wherein one target data segment corresponds to one time scale; using a preset power prediction model, a net power prediction sequence and prediction confidence are determined based on the target feature vector; based on the net power prediction sequence, the prediction confidence, and preset operating constraints, an execution power envelope is determined; based on the execution power envelope, an operating safety range is determined, and based on the operating safety range, the target power of the electrolyzer and the compensation power of the energy storage device are determined.

[0139] In a specific scenario example, the data processing method and apparatus for wind-solar-hydrogen storage coupling systems provided in this specification can be applied. This solves the technical problems of insufficient prediction accuracy in wind-solar-hydrogen storage coupling systems due to the use of fixed feature extraction methods in existing technologies, and the difficulty in accurately adapting power prediction results to the operational constraints of electrolyzers. The specific implementation process may include the following; please refer to [the documentation / reference]. Figure 4 As shown.

[0140] S1: Near-term real-time data collection.

[0141] At the current predicted time t0, the station's SCADA system, energy storage management system, electrolyzer control system, and DC bus monitoring device collect near-term operation data according to the preset sampling period.

[0142] The operational data includes at least: actual wind power generation Pw(t); actual photovoltaic power generation Ppv(t); energy storage charging and discharging power Pes(t), where discharging is positive and charging is negative; current operating power of the electrolytic cell Pel(t); DC bus voltage Udc(t); electrolytic cell temperature, pressure, start-up and shutdown status, load rate or health status parameters; station load or auxiliary power Paux(t).

[0143] The sampling period can be 1 second, 5 seconds, 10 seconds, or a sub-minute period determined according to communication and control conditions.

[0144] S2: Data quality inspection and time alignment.

[0145] The collected data undergoes timestamp verification, missing data detection, outlier detection, communication delay identification, and physical boundary verification.

[0146] Missing data handling: Complete the missing data based on adjacent time points, data from similar devices, or short-term trends.

[0147] Outlier handling: Data that clearly exceeds the physical boundaries of the device is marked and not directly involved in model updates.

[0148] Communication delay handling: For data misalignment caused by communication delay, resampling and alignment are performed according to the timestamp playback method.

[0149] S3: Net power sequence construction.

[0150] Construct a net power sequence based on the energy flow direction of wind, solar and hydrogen storage systems. In some implementations, the net power sequence satisfies:

[0151] in, For energy storage, dischargeable power, For energy storage charging power, This is a power margin reserved for system backup requirements, power grid power rationing requirements, or DC bus stability requirements.

[0152] Specifically, when the system needs to use the electrolyzer as the main load absorber, This indicates the net power of new energy sources that can be safely absorbed by the electrolytic cell; when the system needs to consider grid-connected power as well. This indicates the adjustable power after deducting grid connection plans, station loads, and energy storage constraints.

[0153] S4: Fluctuation state identification and adaptive window construction.

[0154] This step is used to identify the current fluctuation state of the wind-solar-hydrogen storage system before making predictions, and to determine the time length and data resolution used for subsequent feature construction based on the fluctuation state. This specification does not use a fixed-length sliding window as the sole input, but instead uses a method of "uniform data resolution + multi-scale candidate window + adaptive selection of operating conditions" to construct the prediction input.

[0155] Here, data resolution refers to the time interval between two adjacent sampling points. Preferably, data such as actual wind power generation, actual photovoltaic power generation, energy storage charging and discharging power, electrolyzer operating power, DC bus voltage, electrolyzer temperature, pressure, and energy storage SOC are collected at a sampling period of 1 second or 5 seconds. If the on-site SCADA system, edge acquisition device, or communication link supports second-level data transmission, a data resolution of 1 second / point is preferred; if limited by equipment interfaces, communication protocols, or data stability, a data resolution of 5 seconds / point or 10 seconds / point is used. Data from different sources are resampled to the same timeline before entering this step, for example, to 1 second / point or 5 seconds / point, to avoid feature misalignment due to inconsistent timestamps among wind power, photovoltaic, energy storage, and electrolyzer data.

[0156] This specification defines multiple candidate time windows, which are defined according to time length, not just a fixed number of data points. The window length for each candidate time includes: Short time window: The length is preferably 10 to 30 seconds, used to capture rapid dynamic processes such as sudden changes in wind speed, cloud cover, rapid switching of energy storage, and sudden changes in electrolyzer load; Medium time window: The length is preferably 60 to 120 seconds, used to identify general short-term fluctuation trends, continuous climbing trends and minute-level power change states; Long-term window: The length is preferably 300 to 900 seconds, used to identify slowly changing power baselines, equipment operating references, long-term system drift, and model bias trends.

[0157] When the data resolution is 1 second / point, the 60-second time window contains 60 sampling points; when the data resolution is 5 seconds / point, the 60-second time window contains 12 sampling points. Therefore, this invention, by defining the window by time length, allows the same technical solution to be adapted to field systems with different sampling frequencies.

[0158] Fluctuation state indices are calculated based on the net power sequence at near-term times. These indices include, but are not limited to: first-order difference; rate of change; second-order difference, used to represent the change in the rate of change of net available power; short-window range; short-window fluctuation intensity; medium-window mean; medium-window standard deviation; long-window mean; number of consecutive unidirectional changes; voltage fluctuation index; energy storage SOC change rate; and electrolyzer load change rate.

[0159] Based on the above indicators, the current operating conditions are divided into: stable operating conditions, continuous climbing operating conditions, rapid drop operating conditions, high-frequency oscillation operating conditions, abnormal data operating conditions, or equipment disturbance operating conditions.

[0160] In some embodiments, when the rate of change exceeds a preset rate of change threshold continuously within a short time window, and the number of consecutive changes in the same direction reaches a preset number, it is determined to be a continuous climbing condition or a rapid falling condition. When the positive and negative rates of change frequently alternate within a short time window and the standard deviation of the short window exceeds the preset fluctuation threshold, it is determined to be a high-frequency oscillation condition. When the fluctuation intensity of the short time window, medium time window, and long time window are all lower than the corresponding threshold, it is judged as a stable operating condition. When a sudden change in net available power is accompanied by an abnormal DC bus voltage, a sudden change in electrolytic cell load command, or a switch in energy storage charging and discharging state, it is determined to be an equipment disturbance condition. When power data exceeds physical boundaries, timestamps are misaligned, or there are consecutive missing measurements, it is determined to be an abnormal data condition.

[0161] Adaptive window selection strategies include: When identified as a stable operating condition, a medium or long time window is used as the main feature window, such as using 120 to 300 seconds of data, to improve prediction stability and reduce the impact of instantaneous noise. When the condition is identified as a continuous uphill or rapid drop, a short time window is used as the main feature window, such as 10 to 30 seconds of data, to improve the prediction response speed. When the condition is identified as a high-frequency oscillation, a combination of short and medium time windows is used. The short time window is used to capture instantaneous oscillations, while the medium time window is used to identify the oscillation center and the overall trend. When an abnormal data condition is identified, the online update of the model is frozen, and only data repair or backup model prediction is performed. When the equipment is identified as under disturbance conditions, the electrolytic cell load change rate, energy storage action amount, and DC bus voltage are introduced as correction features.

[0162] S5: Target feature vector generation.

[0163] The input target feature vector is constructed based on the adaptive window determined by S4. The target feature vector includes: Original sequence and statistical characteristics: net power sequence within the adaptive window, and mean, variance, and range.

[0164] Dynamic characteristics: power difference, rate of change, second-order difference, and fluctuation state coding.

[0165] Physical characteristics: energy storage SOC and charge / discharge margin, adjustable / lower margin of electrolytic cell, DC bus voltage fluctuation.

[0166] Confidence feature: Data quality confidence.

[0167] S6: Net power prediction sequence prediction.

[0168] Input the target feature vector into the preset power prediction model, and output the prediction sequence for the next H time steps:

[0169] The preset power prediction model can employ linear regression, GBDT, TCN, or a multi-model fusion structure. The model synchronously outputs prediction confidence or prediction interval.

[0170] S7: Generation of safe power envelope for electrolytic cell.

[0171] By combining the predicted sequence with physical constraints, the execution power envelope for the next H time steps is generated: Envelope(tk)=[Pel_lower(tk),Pel_upper(tk)] When the prediction confidence level is lower than the preset threshold, the power envelope is automatically reduced; when the electrolytic cell health is poor or the DC bus voltage fluctuation exceeds the threshold, the safety margin Psafe is increased and the load regulation range is reduced.

[0172] S8: Synergistic response of electrolyzer and energy storage.

[0173] Based on the execution power envelope generated by S7, determine the target power of the electrolyzer and the energy storage compensation power.

[0174] When the net available power forecast is within the power envelope and the confidence level is high, the electrolyzer load is adjusted first to absorb new energy power. When the predicted net available power is higher than the safe limit of the electrolyzer, the electrolyzer is prioritized to be upgraded to the safe limit, and the remaining power is absorbed by energy storage or a power limiting strategy is implemented. When the predicted net available power is lower than the safety lower limit of the electrolyzer, the energy storage discharge will be used to compensate first, so that the electrolyzer will not fall out of the minimum stable load range. When the prediction confidence is low and the fluctuation is a rapid drop, reserve the energy storage discharge margin in advance and limit the electrolyzer's up-adjustment speed; When the prediction confidence is low and the fluctuation is in a rapid ramp-up state, reserve the energy storage charging margin in advance and limit the frequent start-up and shutdown of the electrolyzer.

[0175] S9: Prediction error attribution and hierarchical correction.

[0176] When the prediction error exceeds a preset threshold, instead of directly updating the overall model parameters, error attribution is first identified. The error attribution results include: Data quality errors: caused by missing data, bad pixels, communication delays, and timestamp misalignment; New energy sudden change type error: caused by sudden changes in wind speed, cloud cover, and rapid changes in sunlight; Energy storage operation errors: caused by deviations of energy storage charging and discharging power from the plan; Electrolytic cell disturbance-related errors: caused by electrolytic cell load adjustment, start-up and shutdown, or DC side disturbances; Model drift error: caused by seasonal changes, equipment degradation, or long-term operating condition changes.

[0177] For data quality errors, perform data reconstruction but do not update the prediction model; For errors caused by sudden changes in new energy sources, shorten the sliding window and switch to the climbing or falling model; For errors related to energy storage operations, the energy storage compensation characteristics and SOC availability margin are corrected. For electrolytic cell disturbance-related errors, correct the electrolytic cell load change rate, DC bus voltage, and safe power envelope; For model drift errors, only update the model bias term or lightweight parameters; when multiple consecutive windows are identified as model drift, then trigger incremental training or offline retraining.

[0178] S10: Rolling iteration.

[0179] The latest actual data is continuously incorporated according to the preset rolling step size, and S1 to S9 are executed repeatedly to achieve high-frequency prediction and safe collaborative response.

[0180] In some embodiments, taking a wind-solar hybrid hydrogen production power plant as an example, this architecture integrates a wind power generation unit, a photovoltaic power generation unit, an energy storage system, multiple alkaline electrolyzers, a DC bus, and a central control platform. The electrolyzers are allowed to be set to a load range of 30% to 100% of their rated power, have a second-level load adjustment capability, and their physical characteristics require avoiding large and frequent load fluctuations.

[0181] At the current time t_0, the control platform collects operational data and equipment status parameters from the past 60 seconds, covering actual wind power generation, actual photovoltaic power generation, energy storage charging and discharging power, electrolytic cell load, DC bus voltage, and electrolytic cell temperature and pressure data. Obvious anomalies are removed through a data quality detection mechanism, and timestamp playback logic is used to time-align communication delay data, thereby producing a foundation of time-consistent data for processing.

[0182] A net power sequence is constructed and fluctuation state identification is performed. If a continuous increase in net power is detected within the past 10 seconds and the rate of change exceeds a set threshold, it is identified as a continuous climbing condition. The window length in the sliding window parameters is dynamically shortened from 60 seconds to 20 seconds, and the weights of the rate of change feature and the second-order difference feature in the target feature vector are increased simultaneously. If the net power exhibits a high-frequency fluctuation trend, a 10-second short window and a 60-second medium window are simultaneously enabled through multi-scale fusion logic to extract composite time-domain features.

[0183] The system uses a pre-set power prediction model to output the prediction results and prediction confidence levels for the next 10 seconds. The control platform combines the current load of the electrolyzer, the adjustable / lower margins, the DC bus voltage, and the state of charge (SOC) of the energy storage to generate an execution power envelope reflecting the physical execution boundary. When the prediction result rises rapidly but the prediction confidence level is low, the operating safety range will automatically shrink according to the confidence level adjustment margin, allowing the electrolyzer to determine the target power only within a safe ramp rate, while reserving charging space for the energy storage equipment as compensation power to absorb prediction deviations. If the prediction result shows a rapid downward trend, the control platform will reduce the upward adjustment command for the electrolyzer in advance, ensuring that the electrolyzer load does not fall below the minimum stable operating boundary of 30% by reserving energy storage discharge capacity.

[0184] After the actual power is generated, if the prediction error exceeds a preset threshold, error attribution identification is performed. If the error is accompanied by sudden changes in DC bus voltage and changes in electrolytic cell load commands, it is determined to be an electrolytic cell disturbance error. In this case, the electrolytic cell load change rate feature in the target feature vector is corrected and the safety margin of the power envelope is adjusted, and the parameter update of the prediction model is not triggered temporarily. If the error is accompanied by a rapid decrease in photovoltaic power and normal energy storage response, it is determined to be a new energy sudden change error. The window length is shortened and the model is switched to a fast drop model for adaptation. If the error persists for multiple windows and cannot be explained by data quality or equipment behavior, it is determined to be a model drift error, and the lightweight parameter incremental update of the prediction model is triggered.

[0185] The above embodiments achieve closed-loop rolling control from sensing near-real-time data to ensuring safe operation of the electrolyzer. While maximizing the local consumption rate of new energy, this effectively reduces the operational risks and lifespan losses faced by the electrolyzer due to drastic load fluctuations.

[0186] While this specification provides the steps of operation for the methods described in the embodiments or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or client product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0187] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0188] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of this specification can essentially be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of this specification.

[0189] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations and modifications are possible without departing from the spirit of this specification, and it is intended that the appended claims cover such variations and modifications without departing from the spirit of this specification.

Claims

1. A data processing method for a wind-solar-hydrogen storage coupled system, characterized in that, include: Obtain the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupled system; wherein the target wind-solar-hydrogen storage coupled system includes an electrolyzer and an energy storage device; Based on the aforementioned operating data, determine the net power sequence; The net power sequence is subjected to fluctuation state identification, and sliding window parameters are determined based on the identification results; wherein, the sliding window parameters include window length and window combination method; Based on the window length, multiple target data segments are extracted from the net power sequence, and the multiple target data segments are fused with the device status parameters according to the window combination method to determine the target feature vector; wherein, one target data segment corresponds to one time scale; Using a pre-defined power prediction model, the net power prediction sequence and prediction confidence are determined based on the target feature vector; The execution power envelope is determined based on the net power prediction sequence, the prediction confidence level, and the preset operating constraints. Based on the execution power envelope, a safe operating range is determined, and based on the safe operating range, the target power of the electrolyzer and the compensation power of the energy storage device are determined. The step of identifying the fluctuation state of the net power sequence and determining the sliding window parameters based on the identification result includes: Using multiple candidate observation windows of preset scales, multi-scale time-domain features are extracted from the net power sequence to determine fluctuation state indices; wherein, the fluctuation state indices include: the first difference, rate of change, and second difference of the net power sequence, as well as the short window range, short window fluctuation intensity, medium window standard deviation, and long window mean corresponding to each of the candidate observation windows; The fluctuation state index is mapped to a preset multi-dimensional feature space for working condition identification to determine the current working condition type; wherein, the current working condition type includes stable working condition, continuous climbing working condition, rapid drop working condition, high frequency oscillation working condition, data abnormal working condition, and equipment disturbance working condition. Based on the current working condition type, a matching window length is determined from the multiple preset scales, and the corresponding window combination method is determined as the sliding window parameter.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the net power measured sequence corresponding to the net power prediction sequence; Calculate the residual between the net power prediction sequence and the net power measured sequence to determine the prediction error; Error attribution is identified based on the prediction error, the target power of the electrolyzer, and the compensation power of the energy storage device, and the parameters of the preset power prediction model are iteratively updated using the identified error attribution results.

3. The method according to claim 2, characterized in that, The step of determining the net power sequence based on the operational data includes: The running data is time-aligned to determine the corresponding running sequence to be processed; Based on the operation sequence to be processed, the target power component is determined; wherein, the target power component includes the actual wind power generation, the actual photovoltaic power generation, the energy storage charging and discharging power, the station load power, and the reserved power parameters; The target power components are subjected to power combining processing to determine the net power sequence.

4. The method according to claim 3, characterized in that, The step of extracting a target data segment from the net power sequence according to the window length, and fusing the target data segment with the device status parameters according to the window combination method to determine the target feature vector includes: Based on the window length, data sampling points of corresponding duration are extracted from the net power sequence to determine the target data segment; The target data segment is normalized to determine the target temporal feature components; According to the window combination method, the target temporal feature component and the device state parameter are feature aligned and vector concatenated to determine the target feature vector.

5. The method according to claim 4, characterized in that, The step of determining the execution power envelope based on the net power prediction sequence, the prediction confidence level, and preset operating constraints includes: Based on the predicted confidence level and the preset mapping relationship between the confidence level and the adjustment deviation, determine the confidence adjustment margin corresponding to the predicted confidence level; The net power prediction sequence is interval mapped using the confidence adjustment margin to determine the initial power execution interval; Based on the preset operating constraints, the initial power execution range is subjected to amplitude limiting and rate of change constraint processing to determine the execution power envelope; The execution power envelope includes the upper limit and lower limit of the execution power of the target wind-solar-hydrogen storage coupling system at each sampling time within the corresponding preset time period.

6. The method according to claim 2, characterized in that, The step of determining the safe operating range based on the executed power envelope, and determining the target power of the electrolyzer and the compensation power of the energy storage device based on the safe operating range, includes: Based on the upper limit and lower limit of the execution power included in the execution power envelope, the safe operating range of the electrolytic cell at each sampling time within the corresponding preset time period is determined; Based on the safe operating range, the net power prediction sequence is subjected to amplitude limiting mapping to determine the target power of the electrolytic cell; The compensation power of the energy storage device is determined based on the net power prediction sequence and the target power of the electrolyzer.

7. A data processing device for a wind-solar-hydrogen storage coupling system, characterized in that, include: The data acquisition module is used to acquire the operating data and equipment status parameters of the target wind-solar-hydrogen storage coupling system; wherein, the target wind-solar-hydrogen storage coupling system includes an electrolyzer and an energy storage device; A sequence determination module is used to determine the net power sequence based on the operating data; The parameter determination module is used to identify the fluctuation state of the net power sequence and determine the sliding window parameters based on the identification results; wherein, the sliding window parameters include the window length and the window combination method; The vector determination module is used to extract multiple target data segments from the net power sequence according to the window length, and fuse the multiple target data segments with the device status parameters according to the window combination method to determine the target feature vector; wherein, one target data segment corresponds to one time scale; The power prediction module is used to determine the net power prediction sequence and prediction confidence based on the target feature vector using a preset power prediction model. The power envelope determination module is used to determine the execution power envelope based on the net power prediction sequence, the prediction confidence level, and preset operating constraints. The power determination module is used to determine the safe operating range based on the execution power envelope, and to determine the target power of the electrolyzer and the compensation power of the energy storage device based on the safe operating range. The step of identifying the fluctuation state of the net power sequence and determining the sliding window parameters based on the identification result includes: Using multiple candidate observation windows of preset scales, multi-scale time-domain features are extracted from the net power sequence to determine fluctuation state indices; wherein, the fluctuation state indices include: the first difference, rate of change, and second difference of the net power sequence, as well as the short window range, short window fluctuation intensity, medium window standard deviation, and long window mean corresponding to each of the candidate observation windows; The fluctuation state index is mapped to a preset multi-dimensional feature space for working condition identification to determine the current working condition type; wherein, the current working condition type includes stable working condition, continuous climbing working condition, rapid drop working condition, high frequency oscillation working condition, data abnormal working condition, and equipment disturbance working condition. Based on the current working condition type, a matching window length is determined from the multiple preset scales, and the corresponding window combination method is determined as the sliding window parameter.

8. An electronic device, characterized in that, It includes a processor and a memory for storing processor-executable instructions, wherein the processor, when executing the instructions, implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores computer instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 6.

Images