Thermal power frequency modulation-oriented energy storage auxiliary response deviation prediction and precision compensation algorithm system

Abstract

The invention relates to the technical field of automatic power generation control of a power system, and discloses an energy storage auxiliary response deviation prediction and precision compensation algorithm system for thermal power frequency modulation. According to the system, a multi-dimensional topological description is generated by resolving an internal physical structure of the energy storage system, and space-time alignment and feature interleaving are performed on the multi-dimensional topological description and real-time operation information to form a fusion feature spectrum, so that physical constraints are embedded in prediction, and the deviation prediction precision is improved. Based on the prediction result, the system constructs and issues a precision compensation instruction sequence associated with a time sequence, and an actual response track is collected after execution. And generating an iterative optimization quantity by comparing actual and expected responses, and dynamically adjusting a prediction model and an instruction generation rule to form a closed-loop self-optimization mechanism. The system can adapt to battery aging and other long-term changes, continuously maintain high-precision compensation, and enhance the rapidity and stability of fire-storage combined frequency modulation.

Description

Technical Field

[0001] This invention relates to the field of automatic generation control technology for power systems, specifically to an energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for frequency regulation of thermal power plants. Background Technology

[0002] In scenarios where thermal power units are equipped with energy storage systems to jointly participate in grid frequency regulation, achieving precise control of energy storage power is crucial to improving overall frequency regulation quality. Existing technologies primarily rely on statistical analysis of historical power and frequency operating data from outside the energy storage system, or the establishment of simplified equivalent models, to predict potential deviations in its auxiliary response. This typically treats the energy storage system as a single unit or a black box model. Because it fails to consider the differences and coupling relationships in physical topology, response characteristics, and spatiotemporal distribution of internal components such as battery clusters, power conversion modules, and thermal management systems, purely data-driven prediction models are prone to failure when operating conditions are complex and variable, or when the internal state of the system is uneven. The generated static compensation commands are difficult to adapt to the dynamic changes within the system, resulting in insufficient compensation accuracy.

[0003] Existing compensation control strategies are mostly open-loop or semi-closed-loop controls with fixed parameters. After the system generates and executes compensation commands based on preset algorithms, it lacks continuous tracking and in-depth utilization of the actual response effect, and cannot effectively feed back execution deviations to the core logic of prediction and command generation for self-correction. As operating time accumulates, factors such as battery aging and environmental changes can cause system characteristics to drift, the initially calibrated model parameters gradually become inaccurate, and compensation performance systematically declines. To restore accuracy, manual intervention is often required for shutdown testing and readjustment, which not only increases the maintenance burden but also affects the continuity and reliability of frequency regulation services. There is a need for an intelligent compensation system that can embed physical structure knowledge and has the ability to evolve autonomously. Summary of the Invention

[0004] The purpose of this invention is to provide an algorithm system for predicting and compensating for deviations in the auxiliary response of thermal power plant frequency regulation, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an algorithm system for predicting and compensating for energy storage auxiliary response deviations in thermal power frequency regulation, the system comprising: The information perception and analysis module is used to acquire real-time and historical operation perception information from thermal power units and energy storage systems, and to perform multi-level analysis on the operation perception information to generate multi-level perception information containing deviation prediction information. The structural calculation and fusion module is used to calculate the internal physical structure of the energy storage system based on the multi-level sensing information, generate a multi-dimensional topological description of the energy storage physical structure, and perform spatiotemporal alignment and feature interleaving of the multi-level sensing information and the multi-dimensional topological description of the energy storage physical structure to form a fused feature spectrum. The prediction and instruction generation module is used to calculate the expected deviation of the energy storage auxiliary response in the future period based on the fused feature spectrum, and to construct a set of precision compensation instruction sequences with time-series correlation based on the expected deviation. The instruction execution and response acquisition module is used to send the precision compensation instruction sequence to the control execution mechanism of the energy storage system, and to acquire the actual system response trajectory after the control execution mechanism executes the instruction; The iterative optimization module is used to compare the actual response trajectory of the system with the expected response trajectory, generate the iterative optimization quantity of the accuracy compensation strategy, and use the iterative optimization quantity to dynamically adjust the generation logic of the deviation prediction information and the construction rules of the accuracy compensation instruction sequence.

[0006] Preferably, the method for performing multi-level parsing of the operational sensing information to generate multi-level sensing information including deviation prediction information includes: Two types of operational sensing information are collected from the data bus of the thermal power unit control system and the monitoring network of the energy storage management system, respectively. These include real-time power commands, turbine valve openings, and main steam pressure fluctuations on the unit side, as well as battery cluster terminal voltage, in-cluster loop current, and power converter switching status information on the energy storage side. The two types of operational perception information collected are timestamped and outliers are removed. The aligned two types of operational perception information form a synchronous data stream on a unified time axis. The synchronous data stream is parsed at three levels: at the first level, the instantaneous difference between power commands and power outputs at the second-level time scale is extracted to form an instantaneous power tracking error sequence; at the second level, the statistical distribution characteristics of the instantaneous power tracking error sequence at the minute-level time scale are analyzed, including the mean drift and variance expansion coefficient; at the third level, the implicit correlation patterns between the mean drift and variance expansion coefficient and key operating parameters of thermal power units at the hour-level time scale are identified. The instantaneous power tracking error sequence, the mean drift, the variance inflation coefficient, and the implicit correlation pattern are encapsulated and integrated to output structured multi-level sensing information. The core content of the multi-level sensing information is the deviation prediction information of the power tracking error change trend in the next cycle.

[0007] Preferably, the method for calculating the internal physical structure of the energy storage system based on the multi-level sensing information to generate a multi-dimensional topological description of the energy storage physical structure includes: Receive the multi-level sensing information from the previous stage, especially the deviation prediction information therein; Based on the potential weak links in the response indicated by the deviation prediction information, the perception focus on the internal physical structure of the energy storage system is dynamically adjusted. The perception focus refers to the battery module or power circuit area that needs to be calculated in detail. A set of excitation signals with a specific waveform and frequency combination is injected into the energy storage system, and the parameters of the excitation signals are adaptively configured according to the sensing focus. The system receives response signals carrying information about the internal physical structure after the excitation signal propagates within the system via a sensor array distributed on the inner wall of the energy storage system shell. The received response signal is processed by tomographic imaging to calculate the three-dimensional spatial coordinates, electrical connection relationships and thermal distribution gradient of key components inside the energy storage system; Based on the three-dimensional spatial coordinates, the electrical connection relationships, and the thermal distribution gradient, a multi-dimensional topological description of the energy storage physical structure, including spatial, electrical, and thermodynamic dimensions, is constructed. The multi-dimensional topological description identifies the location, connection status, and thermal stability of each component.

[0008] Preferably, the method for spatiotemporally aligning and feature interleaving the multi-level sensing information with the multi-dimensional topological description of the energy storage physical structure to form a fused feature spectrum includes: A time coordinate axis is established for the multi-level sensing information, and a spatial coordinate framework is established for the multi-dimensional topological description of the energy storage physical structure. Establish a unified spatiotemporal reference system, and associate and map the time coordinate axis of the multi-level sensing information with the spatial coordinate framework of the multi-dimensional topological description of the energy storage physical structure to ensure that time events can correspond to spatial locations. Under the unified spatiotemporal reference system, feature vectors that change with time in the multi-level sensing information are extracted, and feature vectors that change with spatial location in the multi-dimensional topological description of the energy storage physical structure are also extracted. Design a feature interleaving network that simultaneously inputs the time-varying feature vector and the spatial location-varying feature vector; Within the feature interleaving network, multiple rounds of feature interaction operations are performed. In each round of operations, the temporal feature vector is modulated according to the spatial feature vector, and the spatial feature vector is also updated according to the temporal feature vector, so that the temporal dynamic information and spatial structural information are deeply coupled. After multiple rounds of feature interaction operations, the feature interleaving network outputs a unified feature representation that integrates temporal dynamics and spatial structure. This unified feature representation is the fused feature spectrum.

[0009] Preferably, the method for calculating the expected deviation of the energy storage auxiliary response in the future time period based on the fused characteristic spectrum includes: The fused feature spectrum is analyzed to identify the feature components that characterize the historical and current response accuracy, as well as the feature components that characterize the internal physical structure state of the energy storage system. A dynamic evolution deviation calculation model is constructed, and the internal parameters of the deviation calculation model are initialized by the feature components in the fused feature spectrum that characterize the internal physical structure state of the energy storage system. The feature components representing the historical and current response accuracy in the fused feature spectrum are input into the initialized deviation calculation model in chronological order. The deviation calculation model simulates the response behavior of the energy storage system under the constraints of its internal physical structure state based on the input feature component sequence, and infers its output in multiple consecutive time slices in the future. The future output derived from the deviation calculation model is compared point by point with the ideal frequency modulation command requirement on the corresponding time slice to calculate the simulated deviation value on each time slice. The simulated deviation values ​​on all future time slices are arranged and combined in chronological order to form a complete distribution sequence of expected deviations for energy storage auxiliary responses in future periods.

[0010] Preferably, the method for constructing a set of precision compensation instruction sequences with temporal correlation based on the expected deviation includes: Read the expected deviation distribution sequence, which contains the expected deviation values ​​at each future time point; Set a deviation tolerance threshold, and mark the points in the expected deviation distribution sequence that exceed the deviation tolerance threshold as key compensation points; For each of the key compensation points, a preliminary compensation action command is generated by combining the future time information corresponding to the key compensation point, the spatial structure information corresponding to the fused feature spectrum, and the controllable parameter range of the energy storage system. Analyze the time intervals and logical relationships between the key compensation points, and based on the principles of causal dependence and temporal proximity, insert transitional instructions between adjacent preliminary compensation action instructions to ensure the smoothness and continuity of instruction execution; All the preliminary compensation action instructions and the inserted transitional instructions are arranged and encapsulated in a strict chronological order to form a set of precision compensation instruction sequences that are closely linked in time and logically dependent on each other. Each instruction in the precision compensation instruction sequence is appended with a timestamp tag and a priority tag. The timestamp tag indicates the precise execution time, and the priority tag is used to make a decision in case of instruction conflict.

[0011] Preferably, the method for collecting the actual system response trajectory after the control actuator executes the command includes: The precision compensation command sequence is sent to the local control unit of the energy storage system through an industrial communication protocol. The local control unit drives the power converter and the battery management system to perform specific compensation actions. Simultaneously with the compensation action, a high-precision measuring device deployed on the output bus of the energy storage system starts synchronous data recording. The high-precision measuring device collects instantaneous values ​​of voltage, current, and power at a sampling frequency higher than the command cycle. Vibration and acoustic sensors deployed at key mechanical connections in the energy storage system synchronously collect vibration and acoustic signals that reflect changes in mechanical stress. Read the actual issuance time, execution status code, and execution duration log of each instruction in the precision compensation instruction sequence from the internal register of the local control unit; The instantaneous values ​​of voltage, current and power collected by the high-precision measuring device, the vibration signals and acoustic signals collected by the vibration sensor and acoustic sensor, and the instruction execution log read from the local control unit are strictly aligned and merged in the time domain. The aligned and merged data constitutes a multi-dimensional, high-density dataset that fully records the entire process from the issuance of the command to the generation of a physical response by the system. This process record is the actual response trajectory of the system.

[0012] Preferably, the method for generating iterative optimization quantities for the accuracy compensation strategy by comparing the actual response trajectory of the system with the expected response trajectory includes: The expected response trajectory within the same time interval as the precision compensation instruction sequence is obtained from the upper-level scheduling module of the control system. The expected response trajectory defines the ideal output that the energy storage system should have at each time point. Under the same time reference, the power output component in the actual response trajectory of the system is subtracted from the ideal output component in the expected response trajectory point by point to obtain a series of instantaneous trajectory deviation values; Statistical analysis is performed on the instantaneous trajectory deviation value to calculate its average deviation, standard deviation, and maximum deviation peak value over the entire compensation period. The vibration signal and acoustic signal components in the actual response trajectory of the system are feature extracted to obtain a quantitative index reflecting the degree of change in mechanical stress. The statistical analysis results of the instantaneous trajectory deviation value are weighted and fused with the quantitative index of the degree of mechanical stress change to evaluate the overall compensation accuracy of the precision compensation command sequence issued this time and the stress impact on the equipment. Based on the evaluation results, a quantitative iterative optimization quantity is generated. Specifically, the iterative optimization quantity includes suggested values ​​for the correction direction of relevant parameters in subsequent deviation prediction information, as well as coefficients for adjusting the degree of aggression when constructing subsequent precision compensation instruction sequences.

[0013] Preferably, the method for dynamically adjusting the generation logic of the deviation prediction information and the construction rules of the accuracy compensation instruction sequence using the iterative optimization amount includes: Receive the iterative optimization amount and parse the suggested values ​​for the correction direction of relevant parameters in the deviation prediction information; Based on the suggested values, the internal parameters of the analytical model used in the multi-level analysis of the operational perception information are adjusted. The adjustment of the internal parameters directly affects the instantaneous power tracking error sequence, the mean shift, the variance inflation coefficient, and the extraction tendency and calculation weight of the implicit correlation pattern. The adjusted analytical model will be used to process the operational sensing information collected in the new cycle, thereby changing the content and accuracy of the subsequently generated multi-level sensing information and deviation prediction information; Simultaneously, the coefficients in the iterative optimization quantity that adjust the degree of aggression when constructing the precision compensation instruction sequence are analyzed; Based on the adjusted coefficients, the instruction generation strategy followed in constructing a set of precision compensation instruction sequences with temporal correlation according to the expected deviation is modified. The modification of the instruction generation strategy directly affects the judgment threshold of the key compensation point, the amplitude and rate of the initial compensation action instruction, and the insertion logic of the transitional instruction. The modified instruction generation strategy will be applied to construct subsequent accuracy compensation instruction sequences based on the new expected deviation distribution sequence, thereby forming a closed-loop optimization system that includes information perception, prediction, decision-making, execution, and feedback.

[0014] Preferably, the method for continuous operation of the closed-loop optimization system includes: After the system starts, it continuously executes the complete process from acquiring operational awareness information to generating iterative optimization quantities. The iterative optimizations generated in each iteration will be applied immediately to adjust the internal parameters of the analytical model and the instruction generation strategy in the next iteration. As the loop continues, the internal parameters of the analytical model and the instruction generation strategy are continuously and adaptively updated under the drive of the iterative optimization. The system monitors the magnitude and trend of the iterative optimization quantity in real time. When the iterative optimization quantity is continuously lower than the preset convergence threshold, the system is determined to have entered a quasi-steady state. At this time, the update frequency of the internal parameters of the analytical model and the instruction generation strategy is reduced to save computing resources. When a new operating condition or external disturbance is detected that causes the iterative optimization amount to exceed the preset activation threshold again, the system automatically increases the update frequency and adjustment range of the internal parameters of the analytical model and the instruction generation strategy in order to quickly track the system dynamics. Through the above mechanisms, it is ensured that the energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation can maintain its self-learning and self-optimization capabilities under all operating conditions.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By solving the internal physical structure of the energy storage system, a multi-dimensional topological description is generated. This static structural information is then rigorously spatiotemporally aligned and interwoven with dynamic, multi-level operational sensing information to form a fused feature spectrum. This allows the predictive model to input not only the dynamic sequence of "what is happening" in the system but also the physical constraints and relational networks of the system's structure. Dynamic operational data finds a clear mapping and interpretation path within the static structural framework. This deep fusion alters the cognitive dimension of the predictive model, enabling it to discern complex deviation patterns caused by the heterogeneity of the system's internal physical structure and differences in response lag, thereby improving the accuracy and robustness of predicting future response deviations.

[0016] After calculating the expected deviation based on the fused feature spectrum, a set of precision compensation command sequences with temporal correlation is constructed and executed, and the actual system response trajectory after execution is collected. The actual trajectory is compared with the expected trajectory, and the generated iterative optimization is used to dynamically adjust the generation logic of deviation prediction information and the construction rules of the precision compensation command sequence. This constitutes a complete, data-driven closed-loop optimization circuit. Each command execution and response acquisition becomes a learning experiment for the system's own model, and the generated optimization continuously fine-tunes the internal parameters and rules of the prediction algorithm and command generation strategy. This enables the system to automatically adapt to the effects of long-term factors such as battery aging, environmental changes, and component characteristic drift, and the compensation strategy itself possesses evolutionary capabilities. The system can maintain high-precision compensation performance in long-term operation without frequent manual recalibration, realizing the transformation from fixed algorithm execution to an adaptive intelligent agent. Attached Figure Description

[0017] Figure 1 This is a timing diagram of the energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation as described in this invention. Figure 2 A flowchart for multi-level parsing of perceived information; Figure 3 A flowchart for generating a fused feature spectrum through spatiotemporal alignment and feature interleaving; Figure 4 Comparison chart of deviation index changes during the iterative compensation process of frequency regulation for energy storage-assisted thermal power plants; Figure 5 A graph showing the convergence characteristics of the iterative optimization of the thermal power frequency regulation and energy storage auxiliary response system. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides an algorithm system for predicting and compensating for deviations in the auxiliary response of energy storage for thermal power plant frequency regulation. The system includes: an information perception and analysis module, which collects and processes real-time and historical operational perception information from the thermal power unit control system and the energy storage management system, generating multi-level perception information that includes judgments on future response deviation trends through multi-level analysis; a structure calculation and fusion module, which receives this information and guides the dynamic calculation of the internal physical structure of the energy storage system based on the deviation prediction information, generating a multi-dimensional topological description of the energy storage physical structure; subsequently, it performs spatiotemporal alignment and feature interleaving of the multi-level perception information and the multi-dimensional topological description, outputting a fusion feature spectrum that deeply integrates temporal dynamics and spatial structure; a prediction and command generation module, based on this fusion feature spectrum, calculates the expected deviation of the energy storage auxiliary response within a specific future time period and constructs a set of temporally coherent and logically dependent precision compensation command sequences; and a command execution and response acquisition module, which sends the command sequence to the control actuators of the energy storage system and simultaneously acquires the actual response trajectory of the system during command execution. The iterative optimization module compares the actual system response trajectory collected with the preset expected response trajectory, generates iterative optimization quantities to guide strategy improvement, and feeds these optimization quantities back to the information perception and analysis module and the prediction and instruction generation module to dynamically adjust the generation logic of deviation prediction information and the construction rules of accuracy compensation instruction sequences, thereby achieving continuous self-optimization of the system.

[0020] Example 1: See Figure 2The information perception and analysis module collects operational perception information from the data bus of the thermal power unit control system and the monitoring network of the energy storage management system, respectively, from the unit side and the energy storage side. Unit-side information includes real-time power commands, turbine valve openings, and main steam pressure fluctuations; energy storage-side information includes battery cluster terminal voltage, in-cluster loop current, and power converter switching status. The collected information undergoes timestamp alignment and outlier removal preprocessing to form a synchronous data stream on a unified time axis. This synchronous data stream is then analyzed at three levels. At the first level, the instantaneous difference between the power command and power output at the second-level time scale is calculated, forming an instantaneous power tracking error sequence. At the second level, the statistical characteristics of this instantaneous power tracking error sequence at the minute-level time scale are analyzed, calculating its mean drift and variance expansion coefficient. At the third level, the implicit correlation patterns between the aforementioned mean drift and variance expansion coefficient and key operating parameters of the thermal power unit at the hour-level time scale are identified. Finally, the instantaneous power tracking error sequence, mean drift, variance inflation coefficient, and implicit correlation patterns are encapsulated and integrated to output structured multi-level sensing information, the core of which is deviation prediction information containing the trend of power tracking error changes in the next cycle.

[0021] The structural calculation and fusion module receives the aforementioned multi-level sensing information, particularly the deviation prediction information. Based on the potential weak points indicated by the deviation prediction information, this module dynamically identifies battery modules or power loop regions requiring focused attention as sensing focal points. Subsequently, the module injects a set of excitation signals with waveforms and frequency parameters adaptively configured according to the sensing focal points into the energy storage system. Through a sensor array distributed on the inner wall of the energy storage system's outer shell, it receives response signals carrying structural information generated after the excitation signals propagate within the system. Tomographic imaging processing is applied to these response signals to calculate the three-dimensional spatial coordinates, electrical connections, and thermal distribution gradients of key components within the energy storage system. Based on the calculation results, a multi-dimensional topological description of the energy storage physical structure, encompassing spatial, electrical, and thermodynamic dimensions, is constructed. This description clearly identifies the location, connection status, and thermal stability of each component. In specific implementation, the information sensing and analysis module captures real-time operational sensing information from the unit side through a protocol conversion unit deployed next to the data bus of the thermal power unit control system. The captured information includes real-time power commands, the valve opening of the turbine high-pressure regulating valve, and pressure fluctuation information of the boiler main steam pressure. Simultaneously, by accessing the monitoring network mirror port of the energy storage management system, operational awareness information from the energy storage side is captured. The captured information includes the output voltage of each battery cluster, the loop current of each parallel loop within the cluster, and the switching status information of the insulated-gate bipolar transistors in the bidirectional power converter. The captured operational awareness information of both types is sent to a preprocessing unit with a unified clock source. The preprocessing unit adds a high-precision timestamp to each data point and performs outlier removal according to preset physical limit rules. The two types of operational awareness information, with aligned timestamps and outlier removal, are arranged on a unified time axis, forming a synchronous data stream with continuous time and synchronized data items.

[0022] In some embodiments, the parsing operation of the synchronous data stream is carried out sequentially at three levels. The first level of parsing is measured in seconds. The parsing process calculates the difference between the actual power output value of the thermal power unit and the real-time power command value from the grid dispatching agency within each second, generating one difference data point per second, and continuously calculating to form an instantaneous power tracking error sequence. The second level of parsing is measured in minutes. The parsing process performs statistical analysis on the instantaneous power tracking error sequence within a past time window. The statistical analysis calculates the offset of the arithmetic mean of the sequence relative to zero as the mean drift, and calculates the ratio of the actual variance of the sequence to a benchmark variance as the variance expansion coefficient. The benchmark variance corresponds to the power tracking error variance of the energy storage system under standard test conditions. The third level of parsing is measured in hours. The parsing process applies a pattern recognition algorithm to identify implicit correlation patterns between the mean drift, the variance expansion coefficient, and key operating parameters of the thermal power unit, including boiler combustion efficiency, condenser vacuum, and feedwater temperature. In the third-level analysis process, the specific implementation of the pattern recognition algorithm involves systematic data processing and pattern mining of accumulated mean drift, variance expansion coefficient, and key operating parameters of thermal power units over an hourly time scale. The algorithm first integrates hourly statistics and environmental parameters from the synchronous data stream into multi-dimensional feature vectors, each representing a snapshot of the system state within a time window. Then, pattern recognition techniques are applied to analyze the covariance structure and distribution characteristics among these feature vectors. By calculating the mutual information entropy or similarity matrix between variables, nonlinear implicit correlation patterns are identified. For example, mean drift shows a positive correlation with boiler combustion efficiency, while variance expansion coefficient and condenser vacuum exhibit a phase lag relationship.

[0023] Understandably, the results of the three-level analysis need to be integrated and encapsulated. The encapsulation process packages the four data volumes—instantaneous power tracking error sequence, mean drift, variance inflation coefficient, and implicit correlation pattern—into a predefined structured format. This structured format includes a header, timestamp field, data volume field, and checksum. The resulting data packet is the multi-level sensing information. The data volume field of the multi-level sensing information specifically includes a deviation prediction information sub-segment. The content of this sub-segment is generated by extrapolating the implicit correlation pattern to the next statistical period, and it describes the power tracking error change trend in the next period.

[0024] Optionally, after receiving multi-level sensing information, the structure calculation and fusion module prioritizes reading and parsing the deviation prediction information segment. After parsing the deviation prediction information segment, the structure calculation and fusion module dynamically generates a control command based on the potential weak points in the response indicated by the deviation prediction information segment. The control command is used to adjust the allocation strategy of sensor resources in the subsequent sensing process, concentrating sensing resources on battery module clusters or power converter circuits. These concentrated areas are the sensing focal points. Subsequently, the structure calculation and fusion module controls a programmable signal generator. The signal generator generates a set of excitation signals with specific waveforms and frequencies based on the electrical characteristics and physical layout corresponding to the sensing focal point. The parameters of the excitation signals with specific waveforms and frequencies are uniquely determined by the sensing focal point.

[0025] In practice, the excitation signal is injected into the DC bus or AC output terminal of the energy storage system through a coupling network. Piezoelectric and electromagnetic sensor arrays distributed on the inner wall of the energy storage system's metal casing are activated synchronously to receive the response signals generated after the excitation signal propagates, reflects, and refracts within the system. These response signals carry information about the internal physical structure of the energy storage system. The received response signals are then fed into a tomographic imaging processor. Based on an inversion algorithm using acoustic and electromagnetic waves, the tomographic imaging processor processes the response signals to calculate the three-dimensional spatial coordinates of key components within the energy storage system, the electrical connections between components, and the thermal distribution gradients between different regions.

[0026] In some embodiments, based on the calculated three-dimensional spatial coordinates, electrical connections, and thermal distribution gradients, the structural calculation and fusion module constructs a multi-dimensional topological description of the energy storage physical structure. The construction process first establishes a three-dimensional coordinate system in the software, marking the three-dimensional spatial coordinates of key components within this system, forming a spatial dimension description. Then, based on the electrical connections, directional connection lines are drawn between the spatially marked components; these connection lines represent electrical pathways, forming an electrical dimension description. Simultaneously, based on the thermal distribution gradient data, isotherms or temperature field cloud maps are rendered in different colors on or inside the three-dimensional model; the color gradient represents temperature levels, forming a thermodynamic dimension description. The final output multi-dimensional topological description is a composite data model containing a spatial coordinate layer, an electrical connection layer, and a thermal rendering layer. This data model clearly identifies the position coordinates, connection status, and thermal stability level of each component.

[0027] It is understood that, in the above implementation method, the specific calculation of the variance inflation coefficient can follow the following mathematical relationship: Where: symbol Represents the calculated variance inflation coefficient, with the sign... Represents the actual variance of the instantaneous power tracking error sequence within the second-level analytical time window, with the sign... This represents the baseline variance of the power tracking error of the energy storage system under factory standard test conditions. The baseline variance is a constant that is pre-determined and stored in the system database.

[0028] Example 2: See Figure 3 After generating a multi-dimensional topological description of the energy storage physical structure, the structural solution and fusion module performs feature fusion. The module establishes a time coordinate axis for the multi-level sensing information and a spatial coordinate frame for the multi-dimensional topological description. By establishing a unified spatiotemporal reference system, the time coordinate axis and the spatial coordinate frame are mapped together, enabling a correspondence between temporal events and spatial locations. Under this spatiotemporal reference system, feature vectors that change with time in the multi-level sensing information are extracted, as are feature vectors that change with spatial location in the multi-dimensional topological description. A feature interleaving network is designed to simultaneously input the temporal and spatial feature vectors. Within the feature interleaving network, multiple rounds of feature interaction operations are performed. In each round, the temporal feature vector is modulated according to the spatial feature vector, and the spatial feature vector is updated according to the temporal feature vector, thereby achieving deep coupling between temporal dynamic information and spatial structural information. After multiple rounds of interaction operations, the feature interleaving network outputs a unified feature representation that integrates temporal and spatial information; this unified feature representation is the fused feature spectrum.

[0029] In practical implementation, after completing the construction of the multi-dimensional topological description of the energy storage physical structure, the structural calculation and fusion module initiates the spatiotemporal alignment and feature interleaving process. The structural calculation and fusion module establishes a time coordinate axis based on absolute time for the multi-level sensing information from the information perception and analysis module. The scale of the time coordinate axis is strictly aligned with the time scale of data acquisition. Simultaneously, the structural calculation and fusion module establishes a spatial coordinate frame with the geometric center of the energy storage system entity as the origin for the multi-dimensional topological description of the energy storage physical structure. This spatial coordinate frame adopts a three-dimensional Cartesian coordinate system. The structural calculation and fusion module then establishes a unified spatiotemporal reference system. This unified spatiotemporal reference system uses a mapping function to associate each timestamp on the time coordinate axis of the multi-level sensing information with one or more spatial coordinate regions in the spatial coordinate frame of the multi-dimensional topological description of the energy storage physical structure. The mapping function is determined based on the system layout configuration file and the signal propagation delay model, ensuring that, for example, a power fluctuation event observed at a certain moment corresponds to the three-dimensional coordinates of the specific battery module that generated the fluctuation in space.

[0030] In some embodiments, under a unified spatiotemporal reference frame, the structure calculation and fusion module performs feature extraction operations. For multi-level sensing information, the extraction operation extracts numerical features of vectors from data fields such as instantaneous power tracking error sequence, mean drift, and variance expansion coefficient for each sampling point on the time coordinate axis, forming a set of feature vectors with fixed dimensions that vary with time. Each element of the time-varying feature vector corresponds to an operational status index. For the multi-dimensional topological description of the energy storage physical structure, the extraction operation traverses predefined gridded spatial units in the spatial coordinate frame, calculates statistics from the three-dimensional spatial coordinate data, electrical connection relationship data, and thermal distribution gradient data contained in each spatial unit, forming a set of feature vectors with fixed dimensions that vary with spatial location. Each element of the time-varying feature vector corresponds to the structural or state attribute of the spatial unit.

[0031] Optionally, a feature interleaving network is deployed within the structure calculation and fusion module. This network employs an architecture combining graph neural networks and attention mechanisms. It has two feature input ports: one receives a sequence of feature vectors that change over time, and the other receives a set of feature vectors that change with spatial location. Within the network, both time-varying and spatially-varying feature vectors are input in parallel to multiple interactive computation layers. In each round of interactive computation, the time-varying feature vectors are treated as a whole to query and modulate the spatially-varying feature vectors. The modulation process weights and aggregates spatial features based on the importance of temporal features. Simultaneously, the updated spatially-varying feature vectors serve as context to update and correct the time-varying feature vectors. The correction process adjusts the values ​​of temporal features based on the constraints of spatial structure on temporal dynamics. Through this bidirectional modulation and correction, temporal dynamics and spatial structure information are deeply coupled at the feature level.

[0032] In practice, the feature interleaving network performs multiple rounds of feature interaction operations. The number of rounds is a configurable hyperparameter. The output of each round serves as the input for the next round. After a preset number of iterations, the feature representation output by the feature interleaving network tends to stabilize. Ultimately, the feature interleaving network outputs a unified feature representation that integrates temporal dynamics and spatial structure. This unified feature representation is a fixed-dimensional tensor, also known as the fused feature spectrum. This fused feature spectrum is then passed to subsequent prediction and instruction generation modules for further processing.

[0033] In some embodiments, the update of the spatial feature vector by the k-th round of interaction operation in the feature interleaving network follows a mathematical rule, which can be formally described by the following formula: Where: symbol This represents the set state of all feature vectors that change with spatial location after the k-th round of interactive computation, denoted by [symbol]. This represents the set of all feature vectors that change with spatial location after the (k-1)th round of interactive computation, which is also the input of this round of computation. (Symbol: ) Represents the sequence state of all time-varying feature vectors after the (k-1)th round of interactive computation, denoted by [symbol]. This represents an attention function, which is used to compute a sequence of feature vectors that change over time. For each feature vector that varies with spatial location Importance weight distribution, sign This represents a feature update function, which is based on the attention function. The calculated weights are applied to the input feature vector that varies with spatial location. Weighted aggregation and nonlinear transformation are performed to generate an updated set of spatial feature vectors. .

[0034] It is understandable that during the operation of the feature interleaving network, the feature vectors that change over time are also updated synchronously. The update process adopts symmetrical logic, that is, the spatial feature vectors are used to adjust the temporal feature vectors during the update. Finally, after multiple rounds of interactive operation, the feature vectors in both time and space are no longer independent, but permeate each other and jointly constitute the mathematical basis of the fused feature spectrum.

[0035] Example 3: The prediction and command generation module calculates the expected deviation based on the received fused feature spectrum. The module analyzes the fused feature spectrum, identifying the feature components representing historical and current response accuracy, as well as the feature components representing the internal physical structure state of the energy storage system. A dynamically evolving deviation calculation model is constructed, whose internal parameters are initialized by the feature components representing the physical structure state in the fused feature spectrum. The feature components representing historical and current response accuracy in the fused feature spectrum are input into the initialized deviation calculation model in chronological order. Based on the input feature component sequence, the model simulates the response behavior of the energy storage system under the constraints of the internal physical structure state, and extrapolates its output over future consecutive time slices. The future output extrapolated by the model is compared point-by-point with the ideal frequency regulation command requirements for the corresponding time slice, calculating the simulated deviation value for each time slice. The simulated deviation values ​​for all future time slices are arranged in chronological order, forming a distribution sequence of expected deviations for the energy storage auxiliary response in future periods.

[0036] The prediction and command generation module constructs a precision compensation command sequence based on the expected deviation distribution sequence. The module reads this sequence and sets a deviation tolerance threshold, marking points exceeding the threshold as critical compensation points. For each critical compensation point, a preliminary compensation action command is generated by combining its corresponding time information, spatial structure information in the fused feature spectrum, and the controllable parameter range of the energy storage system. The time intervals and logical relationships between the critical compensation points are analyzed. Based on causal dependence and temporal proximity principles, transitional commands are inserted between adjacent preliminary compensation action commands to ensure smooth and continuous command execution. All preliminary compensation action commands and inserted transitional commands are arranged and encapsulated in strict chronological order, forming a set of precision compensation command sequences that are tightly linked in time and logically dependent. Each command in this sequence is appended with a timestamp and a priority tag. The timestamp indicates the precise execution time, and the priority tag is used for adjudication in case of command conflicts.

[0037] In practical implementation, the prediction and command generation module receives the fused feature spectrum from the structure calculation and fusion module. The prediction and command generation module then initiates the analysis of the fused feature spectrum. The analysis process identifies feature components in the fused feature spectrum that characterize historical and current response accuracy. These feature components include, for example, the statistical moments of historical power tracking errors and the slope of the changing trend. Simultaneously, the analysis process identifies feature components in the fused feature spectrum that characterize the internal physical structure state of the energy storage system. These feature components include, for example, the battery cluster internal resistance distribution vector, the power loop connection impedance matrix, and the principal components of the thermal distribution gradient field. The prediction and command generation module constructs a dynamically evolving deviation calculation model. The deviation calculation model employs a recurrent neural network architecture with gated recurrent units. The internal parameters of the deviation calculation model, including the weight matrix and bias vector of the recurrent neural network, are initialized by the feature components in the fused feature spectrum that characterize the internal physical structure state of the energy storage system. The initialization process maps the physical structure state feature components to the initial values ​​of the internal state of the recurrent neural network through a fully connected layer. The initialized deviation calculation model receives a sequence of feature components representing the historical and current response accuracy in chronological order. The deviation calculation model performs calculations in the internal state space based on the input feature component sequence to simulate the dynamic response behavior of the energy storage system under specific internal physical structure state constraints and to deduce its power output value in multiple consecutive future time slices. Each future time slice corresponds to a predefined control period.

[0038] The prediction and command generation module compares the future output sequence derived from the deviation calculation model point-by-point with the ideal frequency regulation command demand sequence for the corresponding time slice obtained from the upper-level energy management system. The comparison process calculates the algebraic difference between the simulated output value and the ideal demand value for each time slice; this algebraic difference is the simulated deviation value for that time slice. All simulated deviation values ​​for future time slices are arranged and combined according to their chronological order, forming a complete expected deviation distribution sequence for the energy storage auxiliary response within a specific future period. This expected deviation distribution sequence is an ordered list of timestamps and corresponding deviation values. The prediction and command generation module reads this expected deviation distribution sequence. Internally, the module sets a deviation tolerance threshold, determined based on the performance evaluation standards of the frequency regulation service and the safe operating boundaries of the energy storage equipment. Points in the expected deviation distribution sequence where the absolute value of the deviation exceeds the deviation tolerance threshold are marked as critical compensation points, each accompanied by its timestamp and deviation amplitude information. For each critical compensation point, the prediction and command generation module combines the future time information corresponding to the critical compensation point, the spatial structure information associated with that moment in the fusion feature spectrum, and the controllable parameter range of the energy storage system to generate a preliminary compensation action command. The preliminary compensation action command includes the target power adjustment amount, the power change rate limit, and the start time of execution.

[0039] In some embodiments, the prediction and instruction generation module analyzes the time intervals and logical relationships between key compensation points. These logical relationships include causal dependencies and temporal proximity. Based on the principles of causal dependencies and temporal proximity, the prediction and instruction generation module inserts transitional instructions between adjacent preliminary compensation action instructions. The process of inserting transitional instructions calculates a smooth power transition trajectory based on the power target values ​​and time intervals of the two preceding and following preliminary compensation action instructions. This smooth power transition trajectory is generated through polynomial interpolation or a Bézier curve. The transitional instructions are the discrete power instruction point sequences corresponding to this smooth trajectory. The prediction and instruction generation module arranges and encapsulates all preliminary compensation action instructions and inserted transitional instructions in a strict chronological order, forming a set of precision compensation instruction sequences that are tightly linked in time and logically dependent. The prediction and instruction generation module adds a timestamp and priority tag to each instruction in the precision compensation instruction sequence. The timestamp indicates the precise execution time when the instruction is issued to the actuator, and the priority tag is used to adjudicate when multiple instructions arrive within the same control cycle due to system delays or resource conflicts; higher-priority instructions will be executed first.

[0040] It is understandable that the above deviation calculation model involves approximating the system dynamics during the derivation of simulated deviation values ​​on future time slices. The derivation process can be conceptually represented by a simplified difference equation, which describes the evolution of the output deviation under the influence of internal states: Where: symbol The prediction and instruction generation module deduces the future... Simulated deviation value at time, sign Represents a time slice step size, symbol The eigenvector sequence input to the deviation calculation model at time t represents the system history and current state, and the symbol is... The symbol represents the set of internal parameters of the deviation calculation model obtained by initializing the physical structure state feature components in the fused feature spectrum. The nonlinear mapping function represents the deviation calculation model and is learned through model training.

[0041] Optionally, the generation rules for transitional instructions can be configured when constructing the precision compensation instruction sequence. For example, the generation rules could stipulate that a transitional instruction must be inserted when the time interval between two key compensation points is less than a critical value, where the critical value is determined based on the minimum stable operating cycle of the power converter. The prediction and instruction generation module outputs the final generated precision compensation instruction sequence to the communication buffer, awaiting invocation from the instruction execution and response acquisition module.

[0042] Example 4: The instruction execution and response acquisition module sends the generated precision compensation instruction sequence to the local control unit of the energy storage system via an industrial communication protocol. The local control unit drives the power converter and battery management system to execute specific compensation actions. Simultaneously with the compensation actions, a high-precision measurement device deployed on the output bus of the energy storage system initiates synchronous data recording, acquiring instantaneous values ​​of voltage, current, and power at a sampling frequency higher than the instruction cycle. Vibration sensors and acoustic sensors deployed at key mechanical connections synchronously acquire vibration and acoustic signals reflecting changes in mechanical stress. The actual issuance time, execution status code, and execution duration log of each instruction in the precision compensation instruction sequence are read from the internal register of the local control unit. The instantaneous electrical values ​​acquired by the high-precision measurement device, the physical signals acquired by the vibration and acoustic sensors, and the instruction execution log read from the local control unit are strictly aligned and merged in the time domain. The aligned and merged multi-dimensional, high-density dataset completely records the entire process from instruction issuance to the system's physical response; this process record constitutes the actual response trajectory of the system.

[0043] The iterative optimization module obtains the desired response trajectory within the same time interval as the precision compensation command sequence from the upper-level scheduling module of the control system. This trajectory defines the ideal output that the energy storage system should have at each time point. Under the same time reference, the power output component in the actual response trajectory of the system is subtracted from the ideal output component in the desired response trajectory point by point to obtain a series of instantaneous trajectory deviation values. These instantaneous trajectory deviation values ​​are statistically analyzed to calculate their average deviation, standard deviation, and maximum peak deviation during the compensation period. Feature extraction is performed on the vibration and acoustic signal components in the actual response trajectory of the system to obtain quantitative indicators reflecting the degree of mechanical stress change. The statistical analysis results of the instantaneous trajectory deviation values ​​are weighted and fused with the quantitative indicators of the degree of mechanical stress change to comprehensively evaluate the overall compensation accuracy and stress impact on the equipment caused by the issued precision compensation command sequence. Based on the evaluation results, a quantitative iterative optimization quantity is generated. This optimization quantity includes suggested values ​​for the correction direction of relevant parameters in subsequent deviation prediction information, as well as coefficients for adjusting the aggressiveness of subsequent precision compensation command sequences.

[0044] In practical implementation, the instruction execution and response acquisition module transmits the accuracy compensation instruction sequence issued by the prediction and instruction generation module to the local control unit of the energy storage system via the industrial Ethernet communication protocol. The local control unit is typically a programmable logic controller or an embedded controller. After receiving the accuracy compensation instruction sequence, the local control unit drives the gate drive circuit of the insulated gate bipolar transistor of the power converter and the equalization control circuit of the battery management system to execute specific compensation actions according to the timestamp tag and execution logic in the instruction. During the entire time window of the compensation action execution, high-precision measurement devices deployed on the AC or DC output bus of the energy storage system start synchronous data recording. The high-precision measurement devices include voltage transformers, current transformers, and power analyzers with a bandwidth of not less than 10kHz. The high-precision measurement devices continuously acquire the instantaneous value sequence of output voltage, output current, and output power at a sampling frequency one order of magnitude higher than the instruction cycle. Simultaneously, vibration sensors deployed at the mounting bolts of the power converter radiator and the copper busbar bolts connecting the battery clusters in the energy storage system, along with broadband acoustic sensors deployed within the battery compartment, are activated. The vibration and acoustic sensors synchronously acquire the raw waveforms of vibration and acoustic signals reflecting changes in mechanical stress and the acoustic emission characteristics of the internal electric arc. The instruction execution and response acquisition module accesses the shared memory or internal registers of the local control unit to read the actual issuance time, execution status code, and execution time log from instruction parsing to action completion for each instruction in the precision compensation instruction sequence.

[0045] The command execution and response acquisition module distributes a high-precision synchronous clock source signal to the high-precision measuring device, vibration sensor, acoustic sensor, and local control unit, ensuring that all data have a unified time scale. The module then initiates a data alignment and merging process. This process uses the sampling clock of the high-precision measuring device as a reference, strictly aligning the instantaneous value sequences of voltage, current, and power, the envelope sequences of vibration and acoustic signals after feature extraction, and the discrete command execution logs read from the local control unit in the time domain. The alignment process uses an interpolation algorithm to align non-uniformly sampled log data points to uniform sampling time points. The aligned data is then merged into a multi-column time series dataset. Each row in the dataset corresponds to the same sampling time, and each column represents an observed variable. This multi-dimensional, high-density dataset completely records the entire process from the issuance of the precision compensation command sequence to the physical response of the energy storage system. This process record is defined as the actual response trajectory of the system. Refer to Table 1 for the key parameters and data sources involved in the acquisition process.

[0046] Table 1: Comparison Table of Key Parameters for Command Execution and Data Source Parameters / Signal Type Measuring device / data source Sampling frequency / precision physical dimensions instantaneous value of output voltage High-precision voltage transformer for busbars 100kHz / 0.1% Volt (V) Instantaneous value of output current High-precision current transformer for busbars 100kHz / 0.2% Ampere (A) Instantaneous value of output power Power analyzer (calculated) 10kHz / 0.2% kilowatts (kW) Mechanical vibration signal piezoelectric vibration sensor 20kHz / 16-bit <![CDATA[m per second squared (m / s 2 )]]> Acoustic emission signal Wideband acoustic sensors 200kHz / 24-bit Pascal Command issuance time Local control unit system clock 1 microsecond Seconds (s) Instruction execution status Local Control Unit Status Register Event triggered No unit code In some embodiments, the iterative optimization module obtains the expected response trajectory within the same time interval as the recently executed precision compensation instruction sequence from the instruction buffer of the upper-level scheduling module. The expected response trajectory is calculated by the upper-level scheduling module based on the grid frequency regulation requirements and the planned output of thermal power units. The expected response trajectory defines the ideal power value that the energy storage system should output at each time point. Under the same time reference, the iterative optimization module performs point-by-point subtraction between the power output component in the actual system response trajectory and the ideal power output component in the expected response trajectory, generating a series of instantaneous trajectory deviation values ​​corresponding to the time axis. The iterative optimization module performs statistical analysis on the instantaneous trajectory deviation value sequence, calculating the arithmetic mean of the instantaneous trajectory deviation values ​​over the entire compensation execution period as the average deviation, calculating the standard deviation of the instantaneous trajectory deviation values ​​as the deviation standard deviation, and identifying the maximum absolute value of the instantaneous trajectory deviation values ​​as the maximum deviation peak value.

[0047] The iterative optimization module simultaneously extracts time-frequency domain features from the vibration and acoustic signal components in the actual system response trajectory. Feature extraction includes calculating the root mean square value of the vibration signal during command execution as a vibration intensity index, and calculating the cumulative energy of the acoustic signal in a specific high-frequency band as an arc noise index. The vibration intensity index and the arc noise index together constitute a quantitative index reflecting the degree of change in mechanical and electrical stress. The iterative optimization module weights and fuses the three statistical results—average deviation, standard deviation, and peak deviation—with the two quantitative indices: vibration intensity and arc noise. Before weighted fusion, each index needs to be normalized. Normalization involves dividing the current observed value of each index by its corresponding historical benchmark or safety threshold, converting each index into a dimensionless ratio. The weighting coefficients for weighted fusion are pre-configured based on the system's preference for tracking accuracy and equipment lifespan. Weighted fusion is used to evaluate the overall compensation accuracy of the issued accuracy compensation command sequence and its potential stress impact on the equipment.

[0048] It is understandable that, based on the comprehensive score and various sub-indicators mentioned above, the iterative optimization module generates a quantified iterative optimization quantity. The iterative optimization quantity is a structured data package. Specifically, it includes suggested values ​​for correcting relevant parameters in subsequent deviation prediction information, such as suggesting adjustments to the time window length for calculating the mean drift or the baseline variance value in the variance inflation coefficient. The iterative optimization quantity also includes coefficients for adjusting the aggressiveness of the subsequent precision compensation instruction sequence; the aggressiveness coefficient is used to linearly scale the power adjustment amplitude in the precision compensation instruction sequence. The quantization rule for generating the iterative optimization quantity is represented by a multi-objective weighted function.

[0049] Optionally, the calculation of iterative optimization quantities can be based on the following multi-objective evaluation function framework: Where: symbol The symbol represents the comprehensive evaluation value calculated to generate the iterative optimization quantity. The serial number and symbol represent the evaluation indicators. These represent the current observed values ​​of five indicators: mean deviation, standard deviation, peak value of maximum deviation, vibration intensity index, and arc noise index, respectively. These represent preset benchmark values ​​or safety thresholds corresponding to the five indicators mentioned above, with symbols... These represent preset weighting coefficients corresponding to the five normalized ratios mentioned above, and the sum of these weighting coefficients is 1. The iterative optimization module optimizes based on the comprehensive evaluation value. The different numerical ranges that the values ​​fall into are mapped to generate specific parameter correction suggestions and aggression level adjustment coefficients.

[0050] See Figure 4In the evaluation of the accuracy compensation strategy by the iterative optimization module, this figure illustrates the dynamic changes of the core deviation indicators under multiple rounds of compensation iterations. Specifically, the bars in the figure correspond to the average deviation (blue bars), standard deviation of deviation (orange bars), and peak value of maximum deviation (green bars) after each iteration, while the black dashed trend line shows the decreasing trajectory of the average deviation with each iteration. The data distribution shows that as the number of compensation iterations increases, all three types of deviation indicators exhibit a continuous convergence trend, reflecting the effectiveness of the iterative optimization in adjusting the deviation prediction logic and the rules for constructing compensation instructions. During parameter configuration, the statistical window for the deviation indicators is the complete time period of a single round of compensation execution, and the peak value of maximum deviation is selected as the maximum absolute value of the instantaneous trajectory deviation within that time period.

[0051] Example 5: The iterative optimization module dynamically adjusts the logic and rules of the upstream module using the generated iterative optimization quantities. The module receives the iterative optimization quantities and analyzes the suggested values ​​for correcting relevant parameters in the deviation prediction information. Based on these suggested values, it adjusts the internal parameters of the analytical model used by the information perception and analysis module during multi-level analysis. The adjustment of these internal parameters directly affects the instantaneous power tracking error sequence, mean drift, variance inflation coefficient, and the extraction tendency and calculation weight of implicit correlation patterns. The adjusted analytical model will be used to process the operational perception information collected in the new cycle, thereby changing the content and accuracy of the subsequently generated multi-level perception information and deviation prediction information. Simultaneously, it analyzes the coefficients in the iterative optimization quantities that adjust the aggressiveness of the precision compensation command sequence. Based on these coefficients, it modifies the command generation strategy followed by the prediction and command generation module in constructing the precision compensation command sequence. The modification of the command generation strategy directly affects the judgment threshold for key compensation points, the amplitude and rate of the initial compensation action command, and the insertion logic of transitional commands. The modified command generation strategy will be applied to construct subsequent precision compensation command sequences based on the new expected deviation distribution sequence.

[0052] The continuous operation of the closed-loop optimization system is manifested in the cyclical linkage of its modules. After system startup, it cyclically executes the complete process from acquiring operational awareness information to generating iterative optimization quantities. The iterative optimization quantities generated in each loop are immediately applied to adjust the internal parameters of the analytical model and the instruction generation strategy for the next loop. As the loop continues, the internal parameters of the analytical model and the instruction generation strategy are continuously and adaptively updated under the drive of the iterative optimization quantities. The system monitors the magnitude and trend of the iterative optimization quantities in real time. When the iterative optimization quantities remain below a preset convergence threshold, the system is determined to have entered a quasi-steady state. At this point, the update frequency of the internal parameters of the analytical model and the instruction generation strategy is reduced to conserve computational resources. When a new operating condition or external disturbance is detected that causes the iterative optimization quantities to exceed the preset activation threshold again, the system automatically increases the update frequency and adjustment magnitude of the internal parameters of the analytical model and the instruction generation strategy to quickly track system dynamics.

[0053] In practical implementation, the iterative optimization module receives iterative optimization data from the upstream process. It then analyzes the suggested values ​​for correcting relevant parameters in the deviation prediction information from these data. These suggested values ​​exist in key-value pairs or vectors, such as specific instructions like "increase the mean drift calculation window length by 10%" or "multiply the variance inflation coefficient baseline value by 0.95." Based on these suggested values, the iterative optimization module generates parameter adjustment instructions, which are sent to the information perception and analysis module via the system's internal bus. The information perception and analysis module adjusts the internal parameters of the analytical model used in the multi-level analysis of the operational perception information according to the received parameter adjustment instructions. These internal parameters include the sliding time window length used to calculate the mean drift, the baseline variance value used as the denominator when calculating the variance inflation coefficient, and the feature weight matrix of the pattern recognition algorithm used to identify implicit correlation patterns. The adjusted analytical model will be used to process the operational sensing information collected in the new cycle, thereby changing the filtering characteristics of the subsequently generated instantaneous power tracking error sequence, the numerical results of the mean drift and variance inflation coefficient, and the tendency to extract implicit correlation patterns, ultimately affecting the content and accuracy of multi-level sensing information and deviation prediction information. Simultaneously, the iterative optimization module analyzes the coefficients in the iterative optimization quantities that adjust the aggressiveness of the constructed precision compensation command sequence. The aggressiveness adjustment coefficient is a scalar coefficient between 0.5 and 2.0. Based on the aggressiveness adjustment coefficient, the iterative optimization module generates a strategy update command, which is sent to the prediction and command generation module. The prediction and command generation module, based on the strategy update command, modifies the command generation strategy followed in constructing a set of temporally correlated precision compensation command sequences based on the expected deviation. The modification of the command generation strategy directly affects the judgment threshold of key compensation points, the power adjustment amplitude and rate of change limits of the initial compensation action commands, and the insertion logic and smoothness of transitional commands. The modified instruction generation strategy will be applied to construct subsequent accuracy compensation instruction sequences based on the new expected deviation distribution sequence, thereby forming a closed-loop optimization system between information perception, deviation prediction, compensation decision-making, instruction execution and effect feedback.

[0054] In some embodiments, the continuous operation of the closed-loop optimization system is manifested in the automatic, cyclical execution of the information perception and analysis module, structure calculation and fusion module, prediction and instruction generation module, instruction execution and response acquisition module, and iterative optimization module after system power-on initialization, from acquiring operational perception information to generating iterative optimization quantities. The iterative optimization quantities generated in each cycle are immediately applied to adjust the internal parameters of the analytical model of the information perception and analysis module and the instruction generation strategy of the prediction and instruction generation module for the next cycle. As the cycle continues, the internal parameters of the analytical model and the instruction generation strategy are continuously and adaptively updated under the drive of the iterative optimization quantities. The system monitors the magnitude and trend of the iterative optimization quantities in real time, and the value of the iterative optimization quantities is represented by the comprehensive evaluation value calculated by the iterative optimization module. When the comprehensive evaluation value remains below a preset convergence threshold for several consecutive control cycles, the system determines that it has entered a quasi-steady-state operation phase. At this time, the system reduces the update frequency of the internal parameters of the analytical model of the information perception and analysis module and the instruction generation strategy of the prediction and instruction generation module, from once per cycle to once every ten cycles, to save computational resources. When the system detects new operating conditions or external disturbances that cause the comprehensive evaluation value to exceed the preset activation threshold again by monitoring sudden changes in external scheduling instructions or abnormal fluctuations in internal sensor data, the system automatically restores the update frequency of the internal parameters of the analytical model and the instruction generation strategy to each cycle, and allows the magnitude of a single adjustment to be increased proportionally according to the size of the iterative optimization, so as to quickly track the system dynamics.

[0055] It is understandable that the adjustment process of the internal parameters of the analytical model needs to follow the principle of stability to avoid excessive parameter jumps. Parameter adjustment can be formally described as a constrained incremental update process. The suggested values ​​of parameter correction directions regarding the bias prediction information in the iterative optimization are quantified as an adjustment vector, which acts on the parameter space of the analytical model. The parameter adjustment process can be formally described by the following constrained update rules: Where: symbol The parsing model in the information perception and parsing module will be used in the next cycle. The internal parameter vector to be used, symbol This represents the analytical model in the current cycle. The internal parameter vector used, symbol The learning rate factor represents the model parameters. It takes a smaller value when the system is in a quasi-steady state and a larger value when the number of iterations exceeds the activation threshold. The sign... This represents the suggested value vector for parameter correction directions extracted from the iterative optimization quantities, with the symbol... This represents the sign function, used to obtain the direction and sign of the suggestion value vector. Element-wise multiplication of vectors, sign Represents the function that takes the minimum value element by element, symbol This represents the element-wise upper limit vector set for the adjustment range of the internal parameters of the analytical model, used to prevent system instability caused by excessive single adjustments. The adjustment of the instruction generation strategy follows a similar logic: the aggression level adjustment coefficient is directly multiplied by the critical compensation point determination threshold, and then multiplied by the power adjustment range of the initial compensation action instruction.

[0056] Optionally, the continuous operation mechanism of the closed-loop optimization system is managed by an independent monitoring thread. This thread periodically reads the iterative optimization data, parses the version hash values ​​of the model's internal parameters, and the configuration identifiers of the instruction generation strategy. The monitoring thread maintains a system state machine, which switches between "rapid learning," "steady-state operation," and "perturbation tracking" states based on preset convergence thresholds, activation thresholds, and the trend of historical iterative optimization data. Different parameter update frequencies and learning rate factors are configured for each state.

[0057] See Figure 5 In the closed-loop optimization operation of the energy storage auxiliary response deviation prediction and accuracy compensation system for thermal power frequency regulation, this figure depicts the dynamic change of the comprehensive evaluation value of the iterative optimization quantity with the convergence cycle and the fitting characteristics of the convergence trend. The blue curve in the figure represents the real-time change of the actual iterative optimization quantity of the system. Its value is generated by weighted fusion of the statistical results of the instantaneous deviation between the actual response trajectory and the expected response trajectory of the system and the quantitative index of mechanical stress change, directly reflecting the current optimization effect of the accuracy compensation strategy. The red dashed line is the exponential fitting line of the actual iterative optimization quantity. The R-squared value of the fitting result reaches 0.872, indicating that the decay process of the iterative optimization quantity of the system has a significant exponential convergence law, which verifies the effectiveness of the logic of "iterative optimization quantity driving the adaptive update of analytical model parameters / instruction generation strategy" in the closed-loop optimization mechanism. The green dashed line in the figure corresponds to the system's preset convergence threshold. This threshold is the judgment benchmark for the system to enter quasi-steady-state operation. When the actual iterative optimization quantity is lower than this threshold for several consecutive control cycles, the system will reduce the update frequency of the internal parameters of the analytical model and the instruction generation strategy to achieve efficient utilization of computing resources. Judging from the curve change characteristics, the amount of iterative optimization decays rapidly in the initial stage and then fluctuates slightly around the convergence threshold, which is consistent with the dynamic state switching mechanism of the system of "rapid learning-disturbance tracking-steady-state operation", reflecting the adaptability and stability of the system's closed-loop optimization.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, 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 process, method, article, or apparatus.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A prediction and accuracy compensation algorithm system for energy storage auxiliary response deviation for frequency regulation in thermal power plants, characterized in that, The system includes the following modules: The information perception and analysis module is used to acquire real-time and historical operation perception information from thermal power units and energy storage systems, and to perform multi-level analysis on the operation perception information to generate multi-level perception information containing deviation prediction information. The structural calculation and fusion module is used to calculate the internal physical structure of the energy storage system based on the multi-level sensing information, generate a multi-dimensional topological description of the energy storage physical structure, and perform spatiotemporal alignment and feature interleaving of the multi-level sensing information and the multi-dimensional topological description of the energy storage physical structure to form a fused feature spectrum. The prediction and instruction generation module is used to calculate the expected deviation of the energy storage auxiliary response in the future period based on the fused feature spectrum, and to construct a set of precision compensation instruction sequences with time-series correlation based on the expected deviation. The instruction execution and response acquisition module is used to send the precision compensation instruction sequence to the control execution mechanism of the energy storage system, and to acquire the actual system response trajectory after the control execution mechanism executes the instruction; The iterative optimization module is used to compare the actual response trajectory of the system with the expected response trajectory, generate the iterative optimization quantity of the accuracy compensation strategy, and use the iterative optimization quantity to dynamically adjust the generation logic of the deviation prediction information and the construction rules of the accuracy compensation instruction sequence.

2. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 1, characterized in that, The method for performing multi-level parsing of the operational sensing information to generate multi-level sensing information containing deviation prediction information includes: Two types of operational sensing information are collected from the data bus of the thermal power unit control system and the monitoring network of the energy storage management system, respectively. These include real-time power commands, turbine valve openings, and main steam pressure fluctuations on the unit side, as well as battery cluster terminal voltage, in-cluster loop current, and power converter switching status information on the energy storage side. The two types of operational perception information collected are timestamped and outliers are removed. The aligned two types of operational perception information form a synchronous data stream on a unified time axis. The synchronous data stream is parsed at three levels: at the first level, the instantaneous difference between power commands and power outputs at the second-level time scale is extracted to form an instantaneous power tracking error sequence; at the second level, the statistical distribution characteristics of the instantaneous power tracking error sequence at the minute-level time scale are analyzed, including the mean drift and variance expansion coefficient; at the third level, the implicit correlation patterns between the mean drift and variance expansion coefficient and key operating parameters of thermal power units at the hour-level time scale are identified. The instantaneous power tracking error sequence, the mean drift, the variance inflation coefficient, and the implicit correlation pattern are encapsulated and integrated to output structured multi-level sensing information. The core content of the multi-level sensing information is the deviation prediction information of the power tracking error change trend in the next cycle.

3. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 2, characterized in that, The method for calculating the internal physical structure of the energy storage system based on the multi-level sensing information and generating a multi-dimensional topological description of the energy storage physical structure includes: Receive the multi-level sensing information from the previous stage, especially the deviation prediction information therein; Based on the potential weak links in the response indicated by the deviation prediction information, the perception focus on the internal physical structure of the energy storage system is dynamically adjusted. The perception focus refers to the battery module or power loop area that needs to be calculated in detail. A set of excitation signals with a specific waveform and frequency combination is injected into the energy storage system, and the parameters of the excitation signals are adaptively configured according to the sensing focus. The system receives response signals carrying information about the internal physical structure after the excitation signal propagates within the system via a sensor array distributed on the inner wall of the energy storage system shell. The received response signal is processed by tomographic imaging to calculate the three-dimensional spatial coordinates, electrical connection relationships and thermal distribution gradient of key components inside the energy storage system; Based on the three-dimensional spatial coordinates, the electrical connection relationships, and the thermal distribution gradient, a multi-dimensional topological description of the energy storage physical structure, including spatial, electrical, and thermodynamic dimensions, is constructed. The multi-dimensional topological description identifies the location, connection status, and thermal stability of each component.

4. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 3, characterized in that, The method for spatiotemporally aligning and feature-interleaving the multi-level sensing information with the multi-dimensional topological description of the energy storage physical structure to form a fused feature spectrum includes: A time coordinate axis is established for the multi-level sensing information, and a spatial coordinate framework is established for the multi-dimensional topological description of the energy storage physical structure. Establish a unified spatiotemporal reference system, and associate and map the time coordinate axis of the multi-level sensing information with the spatial coordinate framework of the multi-dimensional topological description of the energy storage physical structure to ensure that time events can correspond to spatial locations. Under the unified spatiotemporal reference system, feature vectors that change with time in the multi-level sensing information are extracted, and feature vectors that change with spatial location in the multi-dimensional topological description of the energy storage physical structure are also extracted. Design a feature interleaving network that simultaneously inputs the time-varying feature vector and the spatial location-varying feature vector; Within the feature interleaving network, multiple rounds of feature interaction operations are performed. In each round of operations, the temporal feature vector is modulated according to the spatial feature vector, and the spatial feature vector is also updated according to the temporal feature vector, so that the temporal dynamic information and spatial structural information are deeply coupled. After multiple rounds of feature interaction operations, the feature interleaving network outputs a unified feature representation that integrates temporal dynamics and spatial structure. This unified feature representation is the fused feature spectrum.

5. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 4, characterized in that, The method for calculating the expected deviation of the energy storage-assisted response in the future time period based on the fused characteristic spectrum includes: The fused feature spectrum is analyzed to identify the feature components that characterize the historical and current response accuracy, as well as the feature components that characterize the internal physical structure state of the energy storage system. A dynamic evolution deviation calculation model is constructed, and the internal parameters of the deviation calculation model are initialized by the feature components in the fused feature spectrum that characterize the internal physical structure state of the energy storage system. The feature components representing the historical and current response accuracy in the fused feature spectrum are input into the initialized deviation calculation model in chronological order. The deviation calculation model simulates the response behavior of the energy storage system under the constraints of its internal physical structure state based on the input feature component sequence, and infers its output in multiple consecutive time slices in the future. The future output derived from the deviation calculation model is compared point by point with the ideal frequency modulation command requirement on the corresponding time slice to calculate the simulated deviation value on each time slice. The simulated deviation values ​​on all future time slices are arranged and combined in chronological order to form a complete distribution sequence of expected deviations for energy storage auxiliary responses in future periods.

6. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 5, characterized in that, The method for constructing a set of precision compensation instruction sequences with temporal correlation based on the expected deviation includes: Read the expected deviation distribution sequence, which contains the expected deviation values ​​at each future time point; Set a deviation tolerance threshold, and mark the points in the expected deviation distribution sequence that exceed the deviation tolerance threshold as key compensation points; For each of the key compensation points, a preliminary compensation action command is generated by combining the future time information corresponding to the key compensation point, the spatial structure information corresponding to the fused feature spectrum, and the controllable parameter range of the energy storage system. Analyze the time intervals and logical relationships between the key compensation points, and based on the principles of causal dependence and temporal proximity, insert transitional instructions between adjacent preliminary compensation action instructions to ensure the smoothness and continuity of instruction execution; All the preliminary compensation action instructions and the inserted transitional instructions are arranged and encapsulated in a strict chronological order to form a set of precision compensation instruction sequences that are closely linked in time and logically dependent on each other. Each instruction in the precision compensation instruction sequence is appended with a timestamp tag and a priority tag. The timestamp tag indicates the precise execution time, and the priority tag is used to make a decision in case of instruction conflict.

7. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 6, characterized in that, The method for collecting the actual system response trajectory after the control actuator executes the command includes: The precision compensation command sequence is sent to the local control unit of the energy storage system through an industrial communication protocol. The local control unit drives the power converter and the battery management system to perform specific compensation actions. Simultaneously with the compensation action, a high-precision measuring device deployed on the output bus of the energy storage system starts synchronous data recording. The high-precision measuring device collects instantaneous values ​​of voltage, current, and power at a sampling frequency higher than the command cycle. Vibration and acoustic sensors deployed at key mechanical connections in the energy storage system synchronously collect vibration and acoustic signals that reflect changes in mechanical stress. Read the actual issuance time, execution status code, and execution duration log of each instruction in the precision compensation instruction sequence from the internal register of the local control unit; The instantaneous values ​​of voltage, current and power collected by the high-precision measuring device, the vibration signals and acoustic signals collected by the vibration sensor and acoustic sensor, and the instruction execution log read from the local control unit are strictly aligned and merged in the time domain. The aligned and merged data constitutes a multi-dimensional, high-density dataset that fully records the entire process from the issuance of the command to the generation of a physical response by the system. This process record is the actual response trajectory of the system.

8. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for frequency regulation of thermal power plants according to claim 7, characterized in that, The method for generating iterative optimization quantities for the accuracy compensation strategy by comparing the actual response trajectory of the system with the expected response trajectory includes: The expected response trajectory within the same time interval as the precision compensation instruction sequence is obtained from the upper-level scheduling module of the control system. The expected response trajectory defines the ideal output that the energy storage system should have at each time point. Under the same time reference, the power output component in the actual response trajectory of the system is subtracted from the ideal output component in the expected response trajectory point by point to obtain a series of instantaneous trajectory deviation values; Statistical analysis is performed on the instantaneous trajectory deviation value to calculate its average deviation, standard deviation, and maximum deviation peak value over the entire compensation period. The vibration signal and acoustic signal components in the actual response trajectory of the system are feature extracted to obtain a quantitative index reflecting the degree of change in mechanical stress. The statistical analysis results of the instantaneous trajectory deviation value are weighted and fused with the quantitative index of the degree of mechanical stress change to evaluate the overall compensation accuracy of the precision compensation command sequence issued this time and the stress impact on the equipment. Based on the evaluation results, a quantitative iterative optimization quantity is generated. Specifically, the iterative optimization quantity includes suggested values ​​for the correction direction of relevant parameters in subsequent deviation prediction information, as well as coefficients for adjusting the degree of aggression when constructing subsequent precision compensation instruction sequences.

9. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 8, characterized in that, The method for dynamically adjusting the generation logic of the deviation prediction information and the construction rules of the accuracy compensation instruction sequence using the iterative optimization amount includes: Receive the iterative optimization amount and parse the suggested values ​​for the correction direction of relevant parameters in the deviation prediction information; Based on the suggested values, the internal parameters of the analytical model used in the multi-level analysis of the operational perception information are adjusted. The adjustment of the internal parameters directly affects the instantaneous power tracking error sequence, the mean shift, the variance inflation coefficient, and the extraction tendency and calculation weight of the implicit correlation pattern. The adjusted analytical model will be used to process the operational sensing information collected in the new cycle, thereby changing the content and accuracy of the subsequently generated multi-level sensing information and deviation prediction information; Simultaneously, the coefficients in the iterative optimization quantity that adjust the degree of aggression when constructing the precision compensation instruction sequence are analyzed; Based on the adjusted coefficients, the instruction generation strategy followed in constructing a set of precision compensation instruction sequences with temporal correlation according to the expected deviation is modified. The modification of the instruction generation strategy directly affects the judgment threshold of the key compensation point, the amplitude and rate of the initial compensation action instruction, and the insertion logic of the transitional instruction. The modified instruction generation strategy will be applied to construct subsequent accuracy compensation instruction sequences based on the new expected deviation distribution sequence, thereby forming a closed-loop optimization system that includes information perception, prediction, decision-making, execution, and feedback.

10. The energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation according to claim 9, characterized in that, The continuous operation method of the closed-loop optimization system includes: After the system starts, it continuously executes the complete process from acquiring operational awareness information to generating iterative optimization quantities. The iterative optimizations generated in each iteration will be applied immediately to adjust the internal parameters of the analytical model and the instruction generation strategy in the next iteration. As the loop continues, the internal parameters of the analytical model and the instruction generation strategy are continuously and adaptively updated under the drive of the iterative optimization. The system monitors the magnitude and trend of the iterative optimization quantity in real time. When the iterative optimization quantity is continuously lower than the preset convergence threshold, the system is determined to have entered a quasi-steady state. At this time, the update frequency of the internal parameters of the analytical model and the instruction generation strategy is reduced to save computing resources. When a new operating condition or external disturbance is detected that causes the iterative optimization amount to exceed the preset activation threshold again, the system automatically increases the update frequency and adjustment range of the internal parameters of the analytical model and the instruction generation strategy in order to quickly track the system dynamics. Through the above mechanisms, it is ensured that the energy storage auxiliary response deviation prediction and accuracy compensation algorithm system for thermal power frequency regulation can maintain its self-learning and self-optimization capabilities under all operating conditions.

Images