Multi-site collaborative energy storage centralized control and scheduling method

Abstract

This invention discloses a multi-site collaborative energy storage centralized control and scheduling method, belonging to the field of multi-site collaborative energy storage technology. The method includes the following steps: After identifying ambiguous situations in the scheduling strategy, a response pattern vector is constructed based on the historical scheduling behavior of each site. A response prediction model integrating a multi-head attention mechanism and a behavior trajectory transformation network is used to determine the differences in site response behavior under ambiguous scheduling strategy conditions. Based on the site response behavior differences, a strategy parsing prompt instruction is generated through a behavior deviation value mapping method. This instruction is then combined with the original scheduling strategy to construct a structured scheduling strategy with a consistent parsing path, achieving unified parsing and constraint control of the scheduling strategy execution logic. This invention solves the problem of multi-site response behavior differences caused by ambiguous scheduling strategies, realizes unified parsing and closed-loop control of the scheduling strategy, and ensures the consistency and stability of collaborative scheduling.

Description

Technical Field

[0001] This invention relates to the field of multi-site collaborative energy storage technology, and more specifically to a method for centralized control and scheduling of multi-site collaborative energy storage. Background Technology

[0002] Multi-site collaborative energy storage centralized control and scheduling is a unified scheduling and optimization control technology for distributed energy storage systems. It aims to improve the overall operational efficiency, responsiveness, and economy of the energy storage system by constructing a centralized control platform to achieve real-time monitoring, coordinated control, and optimized energy resource allocation across multiple geographically distributed energy storage sites. In existing technologies, multi-site collaborative energy storage centralized control and scheduling primarily relies on a three-layer scheduling architecture built around a cloud platform, edge computing nodes, and communication networks. Data acquisition terminals and edge controllers are deployed at each energy storage site to upload the real-time operating status of the energy storage devices (such as SOC, voltage, current, and temperature) to the centralized control center. The centralized control platform, based on load forecasting, electricity price information, grid status, and energy storage capacity, uses scheduling algorithms (such as model predictive control, heuristic algorithms, or distributed optimization) to formulate optimal scheduling instructions and distribute the control strategies to each site for execution. The scheduling process typically includes key stages such as data acquisition and preprocessing, site status perception and assessment, scheduling optimization calculation, instruction issuance and execution feedback, and fault diagnosis and fault tolerance handling. High-efficiency system-level operation is achieved through load coordination, energy transfer, and peak shaving / valley filling strategies among multiple sites.

[0003] The existing technology has the following shortcomings:

[0004] When multi-site collaborative energy storage centralized control and dispatching systems execute peak-valley arbitrage dispatching tasks, the dispatching platform typically issues a unified dispatching strategy to multiple sites to guide them in collaboratively participating in peak-shaving responses within a specified time period. However, when the issued dispatching strategy is not clearly defined, using vague statements such as "to respond to the current regional power load trend with a collaborative peak-shaving response" to express the response objective, without specifying whether each site should use the predicted load curve as a reference or respond dynamically based on real-time load fluctuations, different sites will interpret the strategy differently based on their own experience or internal settings, resulting in differences in response behavior during specific execution. This difference will lead to inconsistencies in power output timing between sites, deviations in the execution of peak-shaving objectives, and ultimately undermine the coordination and stability of the overall collaborative dispatching. Existing multi-site collaborative energy storage centralized control and dispatching technologies cannot achieve unified analysis and constraint control of the dispatching strategy execution logic based on the differences in site response behavior under ambiguous dispatching strategies. The fundamental reason is that the dispatching platform has not established a standardized parameter command structure during the strategy issuance process, nor has it provided an ambiguity identification and adaptive adjustment mechanism. Therefore, when scheduling tasks involve coordinated responses from multiple sites, inconsistent response methods can easily lead to problems such as misaligned peak shaving curves, power regulation deviations, and increased grid load fluctuations. In severe cases, this can affect the grid side's overall assessment of the coordinated response task and reduce the reliability and economy of system scheduling.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-site collaborative energy storage centralized control and scheduling method to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a multi-site collaborative energy storage centralized control and scheduling method, specifically including the following steps:

[0008] S1. By constructing a policy field dependency tree that includes response baseline field, timing parameter field and behavior intensity field, perform cross-comparison of policy semantic paths to identify whether there are multiple interpretation paths in the scheduling policy in order to determine whether there is ambiguity.

[0009] S2. After identifying the ambiguity in the scheduling strategy, a response pattern vector is constructed based on the historical scheduling behavior of the site. Then, a response prediction model that integrates a multi-head attention mechanism and a behavior trajectory transformation network is used to determine the differences in the site response behavior when the scheduling strategy is ambiguous.

[0010] S3. Based on the differences in site response behavior, generate policy parsing prompts through behavior deviation value mapping, and combine the policy parsing prompts with the original scheduling policy to construct a structured scheduling policy with consistent parsing paths, thereby achieving unified parsing and constraint control of the scheduling policy execution logic.

[0011] S4. Construct a consistent annotation set based on the structured scheduling strategy, embed the consistent annotation set into the semantic expression of the scheduling strategy, and explicitly replace and reorganize ambiguous fields to form a scheduling strategy structure with a clear parsing path.

[0012] S5. During the execution of the structured scheduling strategy, a dynamic correction loop is established based on the offset function between the actual response trajectory and the expected response trajectory of the station. The explicitness of the annotation and the tolerance range of the behavior are adjusted in real time according to the trajectory offset to achieve consistent closed-loop control of the scheduling strategy execution logic.

[0013] Preferably, S1 specifically includes the following steps:

[0014] S101. Extract the response baseline field, timing parameter field, and behavior intensity field from the scheduling strategy. Based on the causal and constraint relationships of the fields in the logical semantics, establish a multi-level nested field structure and construct a strategy field dependency tree containing the response baseline field, timing parameter field, and behavior intensity field.

[0015] S102. Traverse the policy field dependency tree to generate multiple policy semantic paths. Each policy semantic path contains a response baseline field, a timing parameter field, and a behavior intensity field, and records the field parsing direction and the logical pointer between fields.

[0016] S103. Perform field-level alignment and cross-path reorganization operations on all policy semantic paths, calculate the field cross-rate and semantic overlap between semantic paths. If the field cross-rate of multiple paths is lower than the preset threshold and the semantic overlap is inconsistent, it is determined that there are multiple interpretation paths in the scheduling policy and that there is ambiguity.

[0017] Preferably, S103 is as follows:

[0018] The position index sequence of the response baseline field, the timing parameter field and the behavior intensity field in each policy semantic path is extracted sequentially, and a field-level alignment operation is performed based on the order of field appearance to construct a path field comparison matrix.

[0019] Based on the path field comparison matrix, a cross-path recombination operation is performed to extract the set of deviation paths composed of all inconsistent field combinations. The field cross rate and semantic overlap between each pair of paths are calculated, where the field cross rate is the ratio of the number of co-occurring fields to the total number of fields, and the semantic overlap is the weighted average of the semantic label similarity of the fields.

[0020] Compare the field crossover rate and semantic overlap of all path pairs. If there are path pairs with a field crossover rate lower than a preset threshold and inconsistent semantic overlap, these paths are identified as multiple interpretation paths, and the scheduling strategy is judged to be ambiguous.

[0021] Preferably, S2 specifically includes the following steps:

[0022] S201. After identifying the ambiguous situation in the scheduling strategy, extract the behavior records of multiple energy storage sites in the historical scheduling tasks for the ambiguous scheduling strategy. Construct a response mode vector based on the response time offset, power adjustment trajectory, load reference type and feedback adjustment cycle before and after the strategy is triggered. Each response mode vector includes the behavior start time difference, power response function sequence and strategy field weight factor group.

[0023] S202. Input all response mode vectors into the response prediction model that integrates the multi-head attention mechanism and the behavior trajectory transformation network. Calculate the attention weight distribution of each policy field to the response behavior through the multi-head attention mechanism. Use the behavior trajectory transformation network to perform time-series remapping on the power response function sequence and output the predicted response behavior trajectory of each energy storage site under the current ambiguous policy conditions.

[0024] S203. Based on the predicted response behavior trajectory of each energy storage site, perform time dimension alignment and numerical normalization processing, calculate the difference indicators between trajectories in response direction, response rhythm and output intensity, and construct a site response behavior difference matrix to represent the degree and type of difference in response behavior between different sites when there is ambiguity in the scheduling strategy.

[0025] Preferably, S202 specifically refers to:

[0026] For each response pattern vector, expand the field dimensions, extract the initial weight values ​​corresponding to the response baseline field, time series parameter field and behavior intensity field, and input the initial weight values ​​and power response function sequence into the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network to complete the input vector format standardization and dimension alignment.

[0027] In the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network, each attention head calculates the attention weight of each policy field to the behavior at different time periods in the power response function sequence, forming multiple weight distribution vectors. All weight distribution vectors are then weighted and synthesized to generate a unified field attention weight distribution matrix.

[0028] The field attention weight distribution matrix is ​​input into the behavior trajectory transformation network, and a time-series remapping operation is performed on the power response function sequence. Through time-domain interpolation and feature smoothing, the predicted response behavior trajectory of each energy storage site under the current ambiguous strategy condition is obtained, which serves as a high-order representation of the site response mode.

[0029] Preferably, S203 is as follows:

[0030] For the predicted response behavior trajectory of each energy storage site, a time dimension alignment operation is performed to uniformly map the time axis in the trajectory to a standard time reference frame, and the power value sequence is subjected to maximum and minimum normalization processing to ensure that all trajectories are comparable under a unified time scale and power range.

[0031] Based on the normalized predicted response behavior trajectory, the trajectories between any two energy storage sites are compared sequentially. By calculating the consistency of power change direction, the synchronization error of response rhythm, and the mean deviation of output intensity, the difference values ​​of response direction, response rhythm, and output intensity between each pair of trajectories are obtained.

[0032] A three-dimensional response behavior difference matrix is ​​constructed by combining the three types of difference values ​​between all pairs of energy storage sites. Each element in the matrix represents the degree and type of difference in the behavioral response of two sites when there is ambiguity in the scheduling strategy, which is used for the adjustment and control of subsequent scheduling parsing logic.

[0033] Preferably, S3 is as follows:

[0034] Based on the response direction difference value, response rhythm difference value, and output intensity difference value in the site response behavior difference matrix, a behavior deviation value mapping rule set is constructed. The difference values ​​of different dimensions are mapped to field parsing offset weights. Field groups with significant response deviations are extracted through field parsing offset weights, and strategy parsing prompts containing field identifiers, recommended parsing methods, and adjustment levels are generated.

[0035] Perform field-level structure mapping operations on the policy parsing prompts and the original scheduling policy. Based on the field identifiers in the prompts, perform structural anchoring and semantic association on the corresponding fields in the original scheduling policy, and rewrite the field parsing path according to the recommended parsing method to construct a structured scheduling policy with consistent parsing paths.

[0036] Based on the structured scheduling strategy, adjustment level parameters are introduced from the strategy parsing prompt instructions to construct field-level parsing boundary constraints and behavioral response strength limits. By unifying the field parsing logic and behavioral response specifications through boundary constraints and strength limits, the unified parsing and constraint control of the scheduling strategy execution logic is achieved.

[0037] Preferably, S4 is as follows:

[0038] Based on the semantic parsing path, parsing offset weight and field boundary constraint parameters of each field in the structured scheduling strategy, the semantic combination relationship of the response benchmark field, the timing parameter field and the behavior intensity field is extracted, a field interpretation logic index table is constructed, and annotation information entries containing field labels, interpretation scope, behavior response mode and parsing prompts are generated according to the semantic dependency and logical order between fields, forming a consistent annotation set.

[0039] The consistent annotation set is embedded into the semantic expression of the scheduling strategy. The field position index in the structured scheduling strategy is identified through the field label matching mechanism. The corresponding annotation information entries are inserted into the scheduling strategy according to the logical hierarchy using the semantic embedding operation. The semantic structure tree is reconstructed so that each field has a clear parsing boundary and response expectation, realizing the semantic fusion of annotation content and strategy structure.

[0040] Based on the semantic structure tree of the scheduling strategy, ambiguous field nodes are identified. By comparing the interpretation range and response mode of the fields in the consistent annotation set, the field rewriting operation is performed to replace the original ambiguous fields with explicit semantic fields and adjust the semantic order and nesting level to form a scheduling strategy structure with a unique parsing path and a clear structural hierarchy. This structure is used for unified parsing of the scheduling strategy execution logic and behavioral response specification control.

[0041] Preferably, S5 is as follows:

[0042] During the execution of the structured scheduling strategy, the actual response trajectory of each energy storage site is collected and aligned with the expected response trajectory in terms of time and power dimensions. A multi-dimensional offset function is constructed based on the response start time difference, power offset value and response rhythm error. The offset function is used to reflect the dynamic difference between the actual response trajectory and the expected response trajectory.

[0043] Based on the time offset component, power offset component and rhythm difference component of the multidimensional offset function, the annotation explicitness adjustment parameter and the behavior tolerance range adjustment parameter are calculated respectively. The adjustment parameters are mapped to the field parsing control factor and the behavior response threshold control factor to construct the annotation control model and the tolerance control model.

[0044] The field parsing control factor is applied to the annotation display level of each field in the structured scheduling strategy to achieve real-time adjustment of the annotation explicitness. At the same time, the behavior response threshold control factor is used to dynamically adjust the behavior tolerance range parameter, establishing a dynamic correction loop that coordinates the parsing logic adjustment and behavior specification correction, and completing the consistent closed-loop control of the scheduling strategy execution logic.

[0045] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0046] 1. This invention achieves automatic identification of potential ambiguities in scheduling policies by constructing a policy field dependency tree and performing cross-comparison of policy semantic paths, avoiding the problem of misunderstandings among multiple stations in collaborative scheduling tasks caused by policy semantic ambiguity. Based on the historical scheduling behavior of stations, a response pattern vector is constructed, and a response prediction model integrating a multi-head attention mechanism and a behavior trajectory transformation network is combined to effectively capture the differences in response behavior of each station under ambiguous policy conditions. Furthermore, a response behavior difference matrix is ​​generated through time alignment and normalization analysis, clearly quantifying the specific deviations of different stations in response direction, rhythm, and intensity. This makes the policy ambiguity problem resolvable and quantifiable, providing a precise basis for policy adjustment.

[0047] 2. This invention generates policy parsing prompts by mapping behavioral deviation values, and combines this with field-level anchoring, semantic reconstruction, and annotation fusion to structurally enhance the original scheduling policy, establishing a unified scheduling policy structure with unique parsing paths and clear field boundaries. Furthermore, it constructs a multi-dimensional offset function and a dynamic correction loop during policy execution to achieve real-time calibration of response behavior and adaptive adjustment of annotation explicitness, thereby building a closed-loop control system encompassing policy identification, behavior modeling, parsing reconstruction, and execution correction. This solution not only improves the consistency of multi-site collaborative response behavior and the stability of scheduling execution, but also enhances the system's adaptive handling capability for inconsistent policy expressions under complex operating conditions, significantly improving the reliability and accuracy of centralized control scheduling. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0049] Figure 1 This is a flowchart illustrating the multi-site collaborative energy storage centralized control and scheduling method of the present invention. Detailed Implementation

[0050] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0051] This invention provides, for example Figure 1 The multi-site collaborative energy storage centralized control and scheduling method shown includes the following steps:

[0052] S1. By constructing a policy field dependency tree that includes response baseline field, timing parameter field and behavior intensity field, perform cross-comparison of policy semantic paths to identify whether there are multiple interpretation paths in the scheduling policy in order to determine whether there is ambiguity.

[0053] In this embodiment, S1 specifically includes the following steps:

[0054] S101. Extract the response baseline field, timing parameter field, and behavior intensity field from the scheduling strategy. Based on the causal and constraint relationships of the fields in the logical semantics, establish a multi-level nested field structure and construct a strategy field dependency tree containing the response baseline field, timing parameter field, and behavior intensity field.

[0055] In the process of extracting the response baseline field, time-series parameter field, and behavior intensity field from the scheduling strategy and constructing the policy field dependency tree, it is first necessary to define the semantic role and logical hierarchy of the three core fields. The response baseline field is a data unit used to determine the reference basis for site response behavior, commonly including target load type, load forecasting method, or real-time power source; the time-series parameter field is used to characterize the time-dimensional features of scheduling behavior, such as the time window, execution cycle, and adjustment step size; the behavior intensity field reflects the power adjustment magnitude, load transfer ratio, or response level that the site should take in response behavior, belonging to the behavior quantification control layer. By parsing the natural language statements or structured policy templates in the scheduling strategy, these three types of fields and their original values ​​are extracted, and a semantic causal graph is established based on their order of action and mutual constraints in policy execution. Based on the graph, the fields are constructed into a nested structure according to the semantic flow of "response baseline field leading—time-series parameter field controlling—behavior intensity field implementation," where each response baseline field node connects to a set of time-series parameter fields, and each time-series parameter field corresponds to one or more behavior intensity fields, ultimately forming a tree structure. This hierarchically nested multi-layered field structure can fully represent the parsing path from the response object to the behavior implementation of the strategy intent, thereby constructing a strategy field dependency tree containing response baseline fields, temporal parameter fields, and behavior intensity fields. This provides a structured foundation for subsequent semantic path generation and ambiguity identification. The construction process can employ an automatic annotation model based on domain semantic ontology, combined with a rule-based field alignment algorithm to achieve efficient extraction and structure building.

[0056] S102. Traverse the policy field dependency tree to generate multiple policy semantic paths. Each policy semantic path contains a response baseline field, a timing parameter field, and a behavior intensity field, and records the field parsing direction and the logical pointer between fields.

[0057] In the process of traversing the policy field dependency tree to generate multiple policy semantic paths, it is first necessary to clarify the parsing hierarchy and logical connection relationships between the nodes in the policy field dependency tree. Each policy semantic path should start from a response baseline field, sequentially associate with a specific time-series parameter field, and then connect with a behavior intensity field to form a complete semantic path chain. The response baseline field is usually used to identify the reference type of the scheduling target, such as whether it is based on predicted load, real-time power, or external signal sources. The time-series parameter field corresponds to the time positioning and adjustment rhythm of the scheduling behavior, such as the response start and end time, scheduling frequency, or window interval. The behavior intensity field points to the execution magnitude, multiplier, or operation gradient of the response behavior. During the path generation process, a depth-first traversal can be used to recursively extract all path combinations that conform to the structural logic from the root node downwards. At the same time, the parsing direction between each pair of fields is recorded during the path generation process, such as "depends on," "controls," or "triggers," to clarify the upstream and downstream relationships between fields. In addition, a logical pointing linked list needs to be constructed for each path to indicate the constraint form between field nodes, such as linear dependency, logical inclusion, or conditional activation. The implementation can automatically identify the set of paths that conform to structural rules by combining graph structure traversal algorithms with a semantic annotation rule base, and generate a structured semantic path dataset as the input basis for subsequent comparison operations. This mechanism for generating policy semantic paths can effectively decompose the multiple parsing possibilities of scheduling policies and establish a standard field combination pattern to support subsequent ambiguity judgment and unified policy parsing.

[0058] S103. Perform field-level alignment and cross-path reorganization operations on all policy semantic paths, calculate the field cross-rate and semantic overlap between semantic paths. If the field cross-rate of multiple paths is lower than the preset threshold and the semantic overlap is inconsistent, it is determined that there are multiple interpretation paths in the scheduling policy and that there is ambiguity.

[0059] Performing field-level alignment and cross-path reorganization operations, and calculating the field cross-rate and semantic overlap between policy semantic paths, the core purpose is to identify the possibility of multiple interpretations of the scheduling policy before policy issuance, thereby proactively detecting the risk of collaborative scheduling deviations caused by semantic ambiguity. In multi-site collaborative energy storage scenarios, each site may independently parse the policy text based on its own understanding. If there are inconsistencies in field combination logic, deviations in field meaning interpretation, or unclear execution benchmarks in the scheduling policy, it will lead to discrepancies in site response behavior at the timing or power levels. Structural-level field alignment can reveal the degree of consistency in field ordering and hierarchy across different paths; while calculating semantic overlap can further quantify the similarity of field meanings, thereby comprehensively judging whether there are fundamental interpretation conflicts between paths. The significance of this operation lies in its proactive detection based on the structure and semantics of the policy itself, rather than relying on the actual response behavior of the sites, avoiding the exposure of problems only during the scheduling execution phase, and improving the accuracy of policy design and the reliability of collaborative control.

[0060] In this embodiment, S103 specifically refers to:

[0061] The position index sequence of the response baseline field, the timing parameter field and the behavior intensity field in each policy semantic path is extracted sequentially, and a field-level alignment operation is performed based on the order of field appearance to construct a path field comparison matrix.

[0062] In the process of performing field-level alignment and constructing a path field comparison matrix for policy semantic paths, the positional information of the response benchmark field, timing parameter field, and behavior intensity field needs to be extracted sequentially from each policy semantic path, and a unique index is assigned to each type of field within the path. The response benchmark field is usually located at the beginning of the path, identifying the power reference type upon which the scheduling behavior depends; the timing parameter field is used to represent the scheduling time range and execution rhythm, located in the middle of the path; the behavior intensity field reflects the execution intensity or control level, usually located at the end of the path. For any two semantic paths, the fields of the same category are compared one-to-one by aligning the field indices, identifying the structural correspondence of the fields in semantic expression, and recording whether there are matches, offsets, or missing values ​​between the fields. The comparison results are recorded in the form of triples, including field pairs, positional differences, and matching identifiers, to generate the path field comparison matrix. The rows and columns of this matrix correspond to the field sequences in the two policy semantic paths, and the cells store the similarity measure between the fields, usually scored using tag word embedding cosine similarity or semantic distance indicators. Higher values ​​in the matrix indicate more consistent semantics and more corresponding structures between the fields. By constructing a path field alignment matrix, a precise structural foundation can be provided for subsequent cross-path recombination and field cross-rate analysis. This process can be automated using structured field matching algorithms and semantic tag vector encoding models, ensuring the accuracy and consistency of field comparisons.

[0063] Based on the path field comparison matrix, a cross-path recombination operation is performed to extract the set of deviation paths composed of all inconsistent field combinations. The field cross rate and semantic overlap between each pair of paths are calculated, where the field cross rate is the ratio of the number of co-occurring fields to the total number of fields, and the semantic overlap is the weighted average of the semantic label similarity of the fields.

[0064] In the process of performing cross-path recombination and calculating field cross-rate and semantic overlap using the path field comparison matrix, it is first necessary to perform a field-level structural cross-comparison of any two policy semantic paths to identify field combinations that differ in field order, type, or semantics, and classify them into the deviation path set. The field cross-rate represents the ratio between the number of fields in the two paths that have positional matching and semantic consistency and the total number of fields in the two paths. The calculation method depends on the ratio of the number of field pairs marked as highly similar matches in the path field comparison matrix to the total number of field nodes. For example, if two paths each contain five fields, and three of these fields have the same name and function and appear in the same position, then the field cross-rate is three divided by five. Semantic overlap measures the semantic consistency of two paths at the field level. It typically uses the semantic label similarity of each pair of fields in the comparison matrix as the scoring basis, combined with the weighted position of the fields in the path. Specifically, a pre-trained word embedding model can be used to map field labels into vectors, and then the semantic similarity can be calculated using the angular distance between the vectors. If path A uses "real-time load response" and path B uses "predicted load alignment," although their field positions correspond, their semantic labels differ significantly, resulting in a substantial decrease in semantic overlap. When the field crossover rate between two paths is below a preset lower limit and there is a significant difference in semantic overlap, it indicates that the paths differ in both structure and meaning. This can be identified as a multi-interpretation path, providing a basis for subsequent ambiguity identification. This process typically combines a path comparison matrix structure, a label vector model, and a segmented scoring mechanism.

[0065] Compare the field crossover rate and semantic overlap of all path pairs. If there are path pairs with a field crossover rate lower than a preset threshold and inconsistent semantic overlap, these paths are identified as multiple interpretation paths, and the scheduling strategy is judged to be ambiguous.

[0066] In comparing the field crossover rate and semantic overlap of execution path pairs, a systematic screening of all pairwise combinations of policy semantic paths is required. Potential sources of ambiguity are identified by analyzing the consistency of field structure and semantic similarity. The field crossover rate represents the similarity of two paths in field structure, while semantic overlap indicates the consistency of path fields in semantic understanding. When the field crossover rate of a path pair is lower than the set structural consistency threshold, and the semantic overlap shows significant inconsistency, it indicates that the two paths differ significantly in both structural arrangement and semantic interpretation. In this case, the path pair can be identified as multiple interpretation paths, meaning that the policy semantics may be interpreted as two or more execution methods in different paths. Multiple interpretation paths mean that the scheduling policy lacks a unique execution logic, which can easily lead to differences in response behavior in multi-site collaborative scenarios, thus affecting the consistency of collaborative scheduling. The preset threshold is usually obtained by statistically analyzing the reasonable structural difference tolerance range in historical policy parsing data to ensure the judgment criteria have engineering practicality. During implementation, clustering algorithms can be used to group and classify the crossover rate and semantic overlap of path pairs, marking path pairs falling into abnormal distribution intervals, thereby accurately determining whether ambiguity exists. This allows for proactive identification of semantic inconsistency risks before the strategy is implemented, providing a safeguard for subsequent unified parsing and control.

[0067] S2. After identifying the ambiguity in the scheduling strategy, a response pattern vector is constructed based on the historical scheduling behavior of the site. Then, a response prediction model that integrates a multi-head attention mechanism and a behavior trajectory transformation network is used to determine the differences in the site response behavior when the scheduling strategy is ambiguous.

[0068] In this embodiment, S2 specifically includes the following steps:

[0069] S201. After identifying the ambiguous situation in the scheduling strategy, extract the behavior records of multiple energy storage sites in the historical scheduling tasks for the ambiguous scheduling strategy. Construct a response mode vector based on the response time offset, power adjustment trajectory, load reference type and feedback adjustment cycle before and after the strategy is triggered. Each response mode vector includes the behavior start time difference, power response function sequence and strategy field weight factor group.

[0070] After identifying ambiguous scheduling strategies, behavioral data from multiple energy storage sites during past executions of tasks with ambiguous strategies can be collected to extract scheduling response characteristics before and after strategy execution. The response pattern vector is constructed as follows: First, the strategy issuance time is marked on the timeline, and the time difference between the strategy trigger point and the response initiation point is calculated based on the actual response behavior of the energy storage sites, defined as the behavior initiation time difference value. Second, the power adjustment change trajectory exhibited by the site during the response process is extracted to form a power response function sequence, used to describe its dynamic response process. Then, the load type (e.g., real-time load or predicted load) and the periodicity of feedback adjustments referenced by the site during the response process are identified. Response baseline fields, time series parameter fields, and behavior intensity fields are extracted from the strategy text. A strategy field weight factor group is generated by assigning values ​​based on the correspondence between fields and behaviors, reflecting the site's attention to the interpretation logic of different fields. Finally, the behavior initiation time difference value, the power response function sequence, and the strategy field weight factor group are concatenated into a unified response pattern vector to characterize the historical response patterns of the site under ambiguous strategy conditions, providing a high-dimensional behavioral representation foundation for subsequent differential prediction modeling.

[0071] S202. Input all response mode vectors into the response prediction model that integrates the multi-head attention mechanism and the behavior trajectory transformation network. Calculate the attention weight distribution of each policy field to the response behavior through the multi-head attention mechanism. Use the behavior trajectory transformation network to perform time-series remapping on the power response function sequence and output the predicted response behavior trajectory of each energy storage site under the current ambiguous policy conditions.

[0072] S203. Based on the predicted response behavior trajectory of each energy storage site, perform time dimension alignment and numerical normalization processing, calculate the difference indicators between trajectories in response direction, response rhythm and output intensity, and construct a site response behavior difference matrix to represent the degree and type of difference in response behavior between different sites when there is ambiguity in the scheduling strategy.

[0073] Inputting response pattern vectors and constructing predicted response behavior trajectories addresses the issue of response behavior deviations that may arise from inconsistent parsing logic among different energy storage sites when scheduling strategies are ambiguous. By integrating the response pattern vector input with a multi-head attention mechanism and a behavior trajectory transformation network into a response prediction model, accurate modeling of the correlation between different strategy fields and behavior trajectories can be achieved, enabling a clear prediction of the potential response trend of each site under the attention weight of specific strategy fields. Furthermore, time-dimensional alignment and numerical normalization of the predicted response behavior trajectories eliminate dimensional differences in time-series references and power amplitudes among different sites, ensuring comparability between trajectories. Subsequently, by constructing multi-dimensional indicators such as response direction differences, response rhythm differences, and output intensity differences, the behavioral deviations caused by differences in the understanding of the strategy by each energy storage site under the same scheduling strategy can be comprehensively characterized. This ultimately forms a structured site response behavior difference matrix, used to quantify the inconsistency in coordinated responses caused by strategy ambiguity. This process not only provides data support for subsequent adjustments and unified control of the scheduling strategy parsing logic but also lays the foundation for the scheduling platform to identify, dynamically compensate for, and intelligently correct behavioral differences in ambiguous environments.

[0074] In this embodiment, S202 specifically refers to:

[0075] For each response pattern vector, expand the field dimensions, extract the initial weight values ​​corresponding to the response baseline field, time series parameter field and behavior intensity field, and input the initial weight values ​​and power response function sequence into the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network to complete the input vector format standardization and dimension alignment.

[0076] When processing each response pattern vector, the vector structure first needs to be expanded in terms of field dimensions. This involves decomposing the composite fields within the response pattern vector into three sub-vectors in a predetermined order, corresponding to the response baseline field, time series parameter field, and behavioral intensity field, respectively. Key behavioral elements within each field, such as response reference source, start time offset, adjustment slope, and maximum output value, are represented as independent dimensions and assigned initial weight values ​​in numerical or embedded vector form, quantifying the field's influence in response decision-making. Next, the initial weight values ​​corresponding to these three fields are concatenated with the power response function sequence to form a unified input structure. The power response function sequence describes the temporal power change pattern of the site during the actual response process, sampled at fixed time intervals to form a one-dimensional time series vector. To ensure consistency in model input dimensions, all input vectors need to undergo format standardization and dimension alignment operations, including numerical normalization, dimension padding or pruning, and unified vector channel structure. After the above processing is completed, the unified input vector that integrates behavioral weight information and response behavior sequence features is finally input into the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network, for subsequent response behavior prediction and difference recognition modeling.

[0077] In the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network, each attention head calculates the attention weight of each policy field to the behavior at different time periods in the power response function sequence, forming multiple weight distribution vectors. All weight distribution vectors are then weighted and synthesized to generate a unified field attention weight distribution matrix.

[0078] In the response prediction model integrating a multi-head attention mechanism and a behavior trajectory transformation network, to achieve dynamic perception of power response behavior by policy fields, the preprocessed input vector is first fed into multiple independent attention heads in the multi-head attention structure. Each attention head operates in parallel in different parameter spaces, constructing query vectors for the response baseline field, temporal parameter field, and behavior intensity field, and using the feature representation of each time step in the power response function sequence as the key-value vector. Each attention head measures the attention level of different policy fields at each time point using the dot product attention calculation method, thereby generating a set of corresponding weight distribution vectors. These weight distribution vectors represent the time period and behavioral feature differences that each type of policy field focuses on throughout the entire response behavior process. Subsequently, for the output weight distribution vector of each attention head, a score is evaluated based on its strength of influence and prediction accuracy in the model structure, and corresponding fusion weights are assigned. All weight distribution vectors are weighted and superimposed to finally synthesize a unified field attention weight distribution matrix. This matrix describes the attention focus of policy fields in the temporal dimension in a two-dimensional form, providing high-resolution semantic guidance for temporal reconstruction of behavior trajectories and analysis of response behavior differences.

[0079] The field attention weight distribution matrix is ​​input into the behavior trajectory transformation network, and a time-series remapping operation is performed on the power response function sequence. Through time-domain interpolation and feature smoothing, the predicted response behavior trajectory of each energy storage site under the current ambiguous strategy condition is obtained, which serves as a high-order representation of the site response mode.

[0080] After inputting the field attention weight distribution matrix into the behavior trajectory transformation network, to achieve the temporal remapping operation of the power response function sequence, the time periods that need to be retained or enhanced in the response behavior are first determined based on the weight of each time step in the field attention weight distribution matrix. Based on this, a time-domain interpolation function is constructed to compress the time sampling interval of low-weight regions and enhance the temporal resolution of high-weight regions, enabling the power response function sequence to undergo nonlinear remapping on the time axis. Simultaneously, to avoid abrupt changes or oscillations in the response trajectory during the remapping process, a high-order feature smoothing kernel is used to perform multi-layer convolutional filtering on the remapping results, ensuring the continuity and physical rationality of the power curve. The final predicted response behavior trajectory for each energy storage site under ambiguous policy conditions will include dynamic adjustments to the semantic attention regions of the policy fields, possessing high expressiveness and temporal adaptability. It can serve as a high-order representation input for the site's response pattern under complex policy environments, used for subsequent behavior difference modeling and scheduling constraint generation.

[0081] In this embodiment, S203 specifically refers to:

[0082] For the predicted response behavior trajectory of each energy storage site, a time dimension alignment operation is performed to uniformly map the time axis in the trajectory to a standard time reference frame, and the power value sequence is subjected to maximum and minimum normalization processing to ensure that all trajectories are comparable under a unified time scale and power range.

[0083] When performing time-dimensional alignment on the predicted response behavior trajectory of each energy storage site, a unified standard time reference frame must first be selected. This reference frame is typically defined by the start and end times of the scheduling task and divided into time axes with fixed time granularity. Then, the time axes of each energy storage site's response trajectory are mapped to the corresponding time points in the standard time reference frame using linear interpolation or dynamic time warping methods, thus achieving consistent alignment of response behavior in the time dimension. After time mapping, the power response value sequence in each trajectory undergoes max-min normalization. The maximum and minimum values ​​in the sequence are calculated, and all power values ​​are linearly mapped to the interval between zero and one, ensuring comparability of response intensity between different sites. The core of time-dimensional alignment is ensuring that the response behavior of all sites is within a unified temporal context. The goal of max-min normalization is to eliminate deviations caused by differences in site power levels, thereby providing a consistent and comparable trajectory expression for subsequent behavior difference measurement. The entire process emphasizes temporal uniformity and numerical normalization to ensure the consistency and discriminability of the analysis results of the behavior trajectories.

[0084] Based on the normalized predicted response behavior trajectory, the trajectories between any two energy storage sites are compared sequentially. By calculating the consistency of power change direction, the synchronization error of response rhythm, and the mean deviation of output intensity, the difference values ​​of response direction, response rhythm, and output intensity between each pair of trajectories are obtained.

[0085] Based on the normalized predicted response behavior trajectory, when comparing the behavior trajectories of any two energy storage sites, the power change trend sequence of the two trajectories on a unified time reference frame is first extracted. To calculate the consistency of power change direction, it is necessary to determine the direction of power value change (increasing, decreasing, or unchanged) between adjacent time points, and to count the ratio between the number of time periods in which the two trajectories have the same change direction within the same time period and the total number of time periods, thereby generating a response direction difference value. The response rhythm synchronization error is calculated by analyzing the time offset of the two trajectories at the major power change inflection points. For example, if the power increase of one site occurs at the tenth minute, while that of the other site occurs at the thirteenth minute, this time difference is the local rhythm offset, which is accumulated and averaged across the entire sequence to form the response rhythm difference value. The output intensity mean deviation is calculated by comparing the difference in average power values ​​of the two trajectories over the entire time period. For example, if the overall response power mean of one site is 0.6 and that of the other is 0.4, the deviation is 0.2, which is the output intensity difference value. These three difference values ​​can comprehensively reflect the degree of difference in response direction, behavior rhythm and adjustment intensity of different sites when executing the same ambiguous scheduling strategy, thereby supporting the decision-making basis for the subsequent construction of the difference matrix and the adjustment of strategy logic.

[0086] A three-dimensional response behavior difference matrix is ​​constructed by combining the three types of difference values ​​between all pairs of energy storage sites. Each element in the matrix represents the degree and type of difference in the behavioral response of two sites when there is ambiguity in the scheduling strategy, which is used for the adjustment and control of subsequent scheduling parsing logic.

[0087] When constructing the three-dimensional response behavior difference matrix, the first step is to extract the corresponding response direction difference value, response rhythm difference value, and output intensity difference value for the response behavior between any two energy storage sites, and organize these three types of difference values ​​into a difference vector with a fixed order. Then, using a double-nested traversal approach, the difference vector generation operation is performed on all energy storage site combinations, filling the corresponding positions of the three-dimensional matrix with the difference vectors of each pair of sites as matrix elements, forming a response behavior difference matrix indexed by site pairs and using the three types of difference indicators as dimensions. Each element in this three-dimensional matrix represents the fine-grained difference in response behavior between two energy storage sites when there is ambiguity in the scheduling strategy, quantitatively characterizing multiple dimensions such as behavior direction, rhythm synchronization, and output intensity. This matrix can be efficiently stored and retrieved using tensor data structures, and combined with graph computing or clustering algorithms to identify the distribution patterns of behavior, it enables the visual identification and grouping of site response consistency levels, providing a precise data foundation for subsequent analytical path adjustment and constraint control of the scheduling strategy execution logic.

[0088] S3. Based on the differences in site response behavior, generate policy parsing prompts through behavior deviation value mapping, and combine the policy parsing prompts with the original scheduling policy to construct a structured scheduling policy with consistent parsing paths, thereby achieving unified parsing and constraint control of the scheduling policy execution logic.

[0089] In this embodiment, S3 specifically refers to:

[0090] Based on the response direction difference value, response rhythm difference value, and output intensity difference value in the site response behavior difference matrix, a behavior deviation value mapping rule set is constructed. The difference values ​​of different dimensions are mapped to field parsing offset weights. Field groups with significant response deviations are extracted through field parsing offset weights, and strategy parsing prompts containing field identifiers, recommended parsing methods, and adjustment levels are generated.

[0091] To construct a set of rules for mapping behavioral deviation values, the response direction difference, response rhythm difference, and output intensity difference values ​​in the site response behavior difference matrix are first categorized by feature processing, and their attribute classification in terms of behavioral deviation type is marked, such as rhythm misalignment, response direction contradiction, or intensity distribution imbalance. Then, based on these deviation types, field parsing offset weight generation rules are defined to establish mapping relationships, converting the causal offset degree between the three types of difference values ​​and policy fields into a field-level parsing offset weight vector. Subsequently, field groups with significant response deviations are selected according to the intensity ranking of offset weights, and the differences in the interpretation methods of these fields across different sites are analyzed using a historical scheduling semantic pattern library. Based on this analysis, a field identifier, recommended parsing method, and parsing adjustment level are assigned to each field group, ultimately generating a policy parsing prompt instruction. This instruction has structured expression capabilities and can directly participate in the construction of structured scheduling policies.

[0092] The response direction difference value indicates the degree of deviation in the power change trend of a site, used to identify whether there is reverse adjustment behavior; the response rhythm difference value indicates the asynchrony between the start time of the behavior and the adjustment speed, and is a key indicator reflecting the time step and tracking capability; the output intensity difference value measures the relative difference in the cumulative output capability of each site within the target time period. Field parsing offset weight is a quantitative indicator that measures the degree of influence of a field on behavioral deviation in the current semantic environment. Its value comes from the attribution backtracking analysis of multi-dimensional behavioral differences to the strategy field. The field identifier in the strategy parsing prompt is used to uniquely locate the field to be adjusted. The recommended parsing method represents the standard semantic interpretation path that should be preferentially adopted for this field. The adjustment level determines the offset correction strength and parsing constraint level to be applied in the parsing logic. These three elements together support the consistent construction of the strategy parsing path.

[0093] Perform field-level structure mapping operations on the policy parsing prompts and the original scheduling policy. Based on the field identifiers in the prompts, perform structural anchoring and semantic association on the corresponding fields in the original scheduling policy, and rewrite the field parsing path according to the recommended parsing method to construct a structured scheduling policy with consistent parsing paths.

[0094] To achieve field-level structural mapping between policy parsing prompts and the original scheduling policy, the field identifiers contained in the policy parsing prompts are first structurally matched with all fields in the original scheduling policy. Structural anchoring is used to identify their specific positions within the policy field dependency structure. Next, based on the recommended parsing method provided in the prompts, the semantic interpretation path for each field is determined, and its path in the policy dependency tree is rewritten using a field parsing path reconstruction algorithm. This path rewriting process includes re-marking the parsing starting point, semantic filtering of intermediate nodes, logical updates to the path direction, and connection to the behavioral output at the end of the path, ensuring that the field interpretation process no longer introduces ambiguous expressions. After rewriting all target field paths, the newly generated field path information is integrated to construct a structured scheduling policy with consistent parsing paths, thus providing a unified policy understanding foundation for multi-site execution behavior.

[0095] Field-level structure mapping refers to the process of mapping the structured parameters in the policy parsing prompts to the original scheduling policy field structure. Its key lies in accurate field anchoring and path adaptation capabilities. Field anchoring is used to find the field node corresponding to the field identifier in the policy field dependency structure. Semantic association is used to understand the contextual relationship of the field in the original scheduling semantics, ensuring that the original policy logic is not broken during path rewriting. The recommended parsing method defines the target interpretation path direction of the field, such as whether scheduling is based on predicted load curves or real-time load fluctuations. Rewriting the field parsing path essentially modifies the edge connection method and information flow in the policy field dependency tree, ensuring consistency in the semantic path used by all sites when executing the scheduling policy. The structured scheduling policy constructed in this way has clear paths, unified semantics, and ambiguity avoidance capabilities.

[0096] Based on the structured scheduling strategy, adjustment level parameters are introduced from the strategy parsing prompt instructions to construct field-level parsing boundary constraints and behavioral response strength limits. By unifying the field parsing logic and behavioral response specifications through boundary constraints and strength limits, the unified parsing and constraint control of the scheduling strategy execution logic is achieved.

[0097] To achieve unified parsing and constraint control of the scheduling policy execution logic, an adjustment level parameter is first extracted from the policy parsing prompts. This parameter indicates the allowable range of behavioral response adjustments and the flexibility of the parsing path during the parsing of each field. Based on this parameter, field-level parsing boundary constraints are constructed. By setting the allowable semantic offset range within the field's interpretation range, the degree of deviation of the field interpretation path during policy execution is limited. Simultaneously, corresponding behavioral response intensity constraint rules are generated based on the adjustment level mapping, defining the upper and lower limits of the output range and fluctuation gradient for each response field. This ensures that each station, when facing the same scheduling policy, will not deviate from the expected response rhythm and power level due to excessively strong or weak responses. Finally, the field-level parsing boundaries and behavioral response intensity constraints are bound to the structured scheduling policy, achieving a unified standard for the policy semantic parsing path, execution magnitude, and behavioral response, eliminating execution discrepancies between stations caused by different policy interpretations.

[0098] The adjustment level parameter is a scalar or vector used to identify the parsing sensitivity and behavioral intensity control requirements of each field in the scheduling strategy. Its value reflects the adjustment flexibility tolerated by the strategy. Field-level parsing boundary constraints are a mechanism to limit the allowed offset range of strategy fields in the semantic interpretation path, preventing logical deviations in field interpretation. Behavioral response intensity limits ensure that sites do not exhibit behavioral overreach or uncontrolled regulation after the strategy is parsed by limiting the maximum response rate, response duration, and maximum power output of energy storage devices. The core purpose of their synergy is to precisely control the parsing method of each field and the physical behavioral parameters of each response action, enabling the strategy to form a consistent and controllable execution loop across different sites after its issuance, thereby improving the stability and predictability of multi-site collaborative scheduling.

[0099] S4. Construct a consistent annotation set based on the structured scheduling strategy, embed the consistent annotation set into the semantic expression of the scheduling strategy, and explicitly replace and reorganize ambiguous fields to form a scheduling strategy structure with a clear parsing path.

[0100] In this embodiment, S4 specifically refers to:

[0101] Based on the semantic parsing path, parsing offset weight and field boundary constraint parameters of each field in the structured scheduling strategy, the semantic combination relationship of the response benchmark field, the timing parameter field and the behavior intensity field is extracted, a field interpretation logic index table is constructed, and annotation information entries containing field labels, interpretation scope, behavior response mode and parsing prompts are generated according to the semantic dependency and logical order between fields, forming a consistent annotation set.

[0102] To ensure consistency and clarity of scheduling strategies during multi-site collaborative execution, the semantic parsing path, parsing offset weight, and field boundary constraint parameters of each field in the structured scheduling strategy can be analyzed to extract the semantic combination relationships between response baseline fields, timing parameter fields, and behavior intensity fields. Based on this, a field interpretation logic index table is constructed, recording the semantic position, priority relationship, and logical connection method of each field in the scheduling logic. Subsequently, according to the semantic dependencies and logical order between fields, annotation information entries containing field labels, interpretation scope, corresponding behavior response patterns, and recommended parsing methods are generated one by one, and these entries are combined into a consistent annotation set. The core of this process is to make the implicit or ambiguous semantic relationships in the original strategy structure explicit and structured, so that subsequent scheduling strategy execution has a unified semantic interpretation standard and behavioral norm constraints, improving multi-site scheduling coordination and system stability.

[0103] The semantic parsing path refers to the logical connection trajectory of a field from the root node to the current field node in the structured scheduling policy semantic tree structure, reflecting the semantic source and action path of the field in the execution logic. The parsing offset weight is a priority or interpretation strength adjustment factor used to indicate the potential for ambiguity in the field's parsing process. Field boundary constraint parameters define the semantic boundaries and allowed response behavior ranges of the field, used to constrain behavioral offsets during execution. Response baseline fields typically involve reference standards for scheduling behavior, such as predicted load and real-time load. Time-series parameter fields define the time attributes of response behavior, including response start time, duration, and feedback cycle. Behavior strength fields describe the power amplitude, regulation capacity, or maximum / minimum output range of the site response. The field interpretation logic index table forms a logical matrix through the combination patterns of the above fields. Annotation information entries generate structured metadata based on field position, function, and contextual relationships, achieving full-domain annotation of the field parsing logic. By constructing a consistent annotation set, pre-annotation verification and semantic consistency optimization can be achieved before policy issuance, building a policy preprocessing mechanism oriented towards ambiguity avoidance.

[0104] The consistent annotation set is embedded into the semantic expression of the scheduling strategy. The field position index in the structured scheduling strategy is identified through the field label matching mechanism. The corresponding annotation information entries are inserted into the scheduling strategy according to the logical hierarchy using the semantic embedding operation. The semantic structure tree is reconstructed so that each field has a clear parsing boundary and response expectation, realizing the semantic fusion of annotation content and strategy structure.

[0105] To achieve a unified understanding and controllable execution of structured scheduling strategies, a consistent set of annotations can be embedded into the semantic expression of the scheduling strategy to enhance the semantic interpretability of scheduling fields. First, a field label matching mechanism identifies the location index of each field node in the structured scheduling strategy, ensuring that each annotation entry can be accurately located to its corresponding target field. Then, a semantic embedding operation is used to insert annotation entries into the strategy fields in a logical hierarchical structure, merging them with the original semantic path of the fields. This operation not only preserves the original logical structure of the strategy but also adds information such as parsing boundaries, response expectations, and recommended behavior patterns, thus achieving a deep integration of the strategy structure and annotation semantics. Finally, a complete semantic structure tree is reconstructed, giving each field clear semantic boundaries and behavioral guidance, ensuring that the scheduling platform and site terminals form a consistent interpretation logic during strategy execution, effectively reducing response deviations and ambiguity risks.

[0106] The field label matching mechanism precisely pairs label fields in the annotation set based on the naming, semantic type, and position identifier of fields in the structured strategy, ensuring a one-to-one correspondence between annotation content and field content. The field position index represents the path node number of the field in the strategy structure tree, guiding the specific location where the annotation is inserted. Semantic embedding is a structured information fusion method that models the logical relationship between fields and annotation content, embedding annotation content below field nodes or in the field attribute area while maintaining the original execution order and logical direction of the fields. The reconstructed semantic structure tree is an enhanced strategy parsing graph, where each field not only has original scheduling parameters but also includes multi-dimensional auxiliary information such as its parsing boundaries, behavior tolerance, and response mode labels. After semantic fusion of annotation content and strategy structure, different scheduling execution units can automatically extract parsing rules and response requirements based on this structure, avoiding execution discrepancies caused by ambiguous interpretations from the source and improving the reliability and accuracy of collaborative response tasks.

[0107] Based on the semantic structure tree of the scheduling strategy, ambiguous field nodes are identified. By comparing the interpretation range and response mode of the fields in the consistent annotation set, the field rewriting operation is performed to replace the original ambiguous fields with explicit semantic fields and adjust the semantic order and nesting level to form a scheduling strategy structure with a unique parsing path and a clear structural hierarchy. This structure is used for unified parsing of the scheduling strategy execution logic and behavioral response specification control.

[0108] To eliminate ambiguity in scheduling policy execution, based on the semantic structure tree of the scheduling policy, potential ambiguities in field nodes can be identified and semantic reconstruction performed. First, a semantic consistency check is performed on all field nodes in the semantic structure tree. By comparing the interpretation range and response pattern of the field in the consistency annotation set, field nodes with overlapping interpretation ranges or conflicting response patterns are identified. After identifying ambiguous field nodes, a field rewriting operation is performed, replacing the original ambiguous field with an explicit semantic field that is semantically clear and has well-defined boundaries in the consistency annotation set. Simultaneously, the semantic order and nesting level of fields in the policy structure are reorganized to make the field combination logic more linear and unambiguous, thereby constructing a scheduling policy structure with a unique resolution path and standardized response guidance. The purpose of this operation is to ensure that the scheduling policy achieves the same source and same interpretation across different execution terminals, improving the stability and consistency of the scheduling execution logic.

[0109] The scheduling policy semantic structure tree is a hierarchical tree model where nodes represent scheduling policy fields and edges represent semantic dependencies between fields. Ambiguous field nodes are identified within this structure by comparing their interpretation scope and response mode labels in a consistent annotation set. The interpretation scope reflects the logical boundary of a field in behavioral response, while the response mode indicates the specific way the field guides terminal behavior. When multiple fields have overlapping interpretation scopes or inconsistent response modes, they are considered semantically ambiguous. Field rewriting replaces ambiguous fields with explicit semantic fields, which possess clear boundaries, a single point of reference, and no ambiguity. Semantic order adjustment ensures clear logical order of field parsing, while nesting level adjustment optimizes and reconstructs the parent-child structure of fields, giving the final policy a hierarchical structure and a linearly interpretable parsing path. This entire process improves the expressive accuracy and execution compatibility of the policy structure, laying the foundation for subsequent behavioral response specification control.

[0110] S5. During the execution of the structured scheduling strategy, a dynamic correction loop is established based on the offset function between the actual response trajectory and the expected response trajectory of the station. The explicitness of the annotation and the tolerance range of the behavior are adjusted in real time according to the trajectory offset to achieve consistent closed-loop control of the scheduling strategy execution logic.

[0111] In this embodiment, S5 specifically refers to:

[0112] During the execution of the structured scheduling strategy, the actual response trajectory of each energy storage site is collected and aligned with the expected response trajectory in terms of time and power dimensions. A multi-dimensional offset function is constructed based on the response start time difference, power offset value and response rhythm error. The offset function is used to reflect the dynamic difference between the actual response trajectory and the expected response trajectory.

[0113] During the execution of the structured scheduling strategy, the actual response trajectories generated in real time by each energy storage site are collected to obtain multi-dimensional time series data containing timestamp sequences and corresponding power output values. By constructing a unified time reference frame, the actual response trajectories of each site are aligned with the expected response trajectories in the time dimension, making each time node comparable. Subsequently, based on the power output differences at each time point, the time difference between response initiation points, and the frequency characteristic differences of the entire response trajectory, the response initiation time difference, power offset value, and response rhythm error are calculated respectively. These three types of difference parameters are mapped to continuous time functions to construct a multi-dimensional offset function, which is used to describe the dynamic deviation relationship between the actual trajectory and the expected trajectory in different dimensions. This method can capture subtle changes in the site response behavior in real time and provide a quantitative basis for subsequent strategy adjustments and behavior corrections.

[0114] The response start time difference measures whether the actual response is earlier or later than the expected response, reflecting the timing accuracy of the site's response. The power offset value represents the difference between the actual power and the target power at each time point, used to assess the degree of deviation in response intensity. The response rhythm error is calculated based on the power change rate and fluctuation frequency, used to characterize the rhythmic consistency of the overall response behavior. These three technical characteristics together constitute the core indicators of the multidimensional offset function, representing the adaptation of scheduling behavior in terms of time, amplitude, and rhythm, respectively. By functionalizing these indicators, real-time monitoring and dynamic evaluation can be achieved without interfering with the original scheduling execution, ensuring that subsequent correction logic is targeted and accurate.

[0115] Based on the time offset component, power offset component and rhythm difference component of the multidimensional offset function, the annotation explicitness adjustment parameter and the behavior tolerance range adjustment parameter are calculated respectively. The adjustment parameters are mapped to the field parsing control factor and the behavior response threshold control factor to construct the annotation control model and the tolerance control model.

[0116] In the consistent closed-loop control process of the scheduling strategy execution logic, based on the output of the multi-dimensional offset function, the time offset component, power offset component, and rhythm difference component are extracted respectively, and used as inputs to calculate the annotation explicitness adjustment parameter and the behavior tolerance range adjustment parameter. The time offset component represents the delay in response initiation, the power offset component characterizes the numerical deviation of power output, and the rhythm difference component reflects the synchronicity deviation of power fluctuations during the response process. By setting a linear interpolation or piecewise function mapping mechanism, the numerical inputs of these three dimensions are mapped to the variation range of the analytical control factor and the threshold control factor. For example, when the time offset is too large, the annotation explicitness can be increased to prompt the strategy analytical logic to strengthen the timing annotation; when the power offset value continuously exceeds the expected window range, the behavior tolerance boundary can be tightened to force the response behavior to remain within a stricter control range. Through this calculation and mapping process, the closed-loop feedback transformation of the difference quantity into the control strategy is completed.

[0117] The field parsing control factor controls the granularity of annotations and the strength of parsing constraints for each field in the scheduling strategy. As the degree of deviation increases, it enhances the semantic indicativeness and constraint weight of the field, thereby guiding the scheduling logic closer to the standard interpretation path. The behavior response threshold control factor dynamically adjusts the allowable fluctuation tolerance range of response behavior in the power and time dimensions. By increasing or decreasing the boundary range, it achieves a dynamic balance between the flexibility and stability of scheduling control. The annotation control model tracks the change trajectory of the field control factors and adjusts the visualization weight and constraint hint level of the annotation content in the semantic structure of the scheduling strategy in real time. The tolerance control model dynamically converges or widens the abnormal range of the behavior response based on the fluctuation of the difference between the response behavior and the control factor. Together, they constitute the core adjustment engine of the dynamic correction mechanism, ensuring that scheduling execution maintains structural consistency and behavioral stability in continuous response feedback.

[0118] The field parsing control factor is applied to the annotation display level of each field in the structured scheduling strategy to achieve real-time adjustment of the annotation explicitness. At the same time, the behavior response threshold control factor is used to dynamically adjust the behavior tolerance range parameter, establishing a dynamic correction loop that coordinates the parsing logic adjustment and behavior specification correction, and completing the consistent closed-loop control of the scheduling strategy execution logic.

[0119] During the execution of the scheduling strategy, the field parsing control factor is used to dynamically adjust the annotation display level of each field in the structured scheduling strategy. By numerically mapping the control factor to the explicit labels of annotation information entries, the intensity of field annotations at different semantic levels can be controlled. Specifically, the numerical range of the field parsing control factor can be divided into multiple annotation level segments, each level corresponding to a set of annotation granularity and semantic constraint templates. This allows for the automatic activation of more explicit and restrictive annotation content when the control factor increases. Simultaneously, the behavior response threshold control factor is used to dynamically adjust the tolerance range parameters of the response behavior. The adjustment range includes the allowable value for response start delay, the power output offset window, and the rhythm synchronization error limit. Through the linkage mapping mechanism between the control factor and the behavior response model, the response boundaries can be compressed or relaxed in real time, enabling the scheduling execution to self-adjust when response behavior offsets occur.

[0120] The field parsing control factor and the behavior response threshold control factor together constitute a two-way feedback mechanism for parsing logic adjustment and behavior specification correction. Driven by a dynamic offset function, they form a synergistic relationship within the policy execution feedback loop. The field parsing control factor improves the clarity of field interpretation, leading to more consistent site behavior, while the behavior response threshold control factor ensures the stability of execution results by narrowing the range of behavior offsets. The combined effect of these two control factors establishes a dynamic correction loop. That is, during scheduling policy execution, the system can autonomously strengthen or weaken the parsing and response control parameters based on the actual degree of deviation in site behavior, thereby continuously correcting execution deviations and ensuring consistent execution of the policy meaning across multiple site environments. This achieves adaptive adjustment throughout the entire process of scheduling logic closed-loop control.

[0121] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0122] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0123] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0127] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-site collaborative energy storage centralized control and scheduling method, characterized in that, Specifically, the following steps are included: S1. By constructing a policy field dependency tree that includes response baseline field, timing parameter field and behavior intensity field, perform cross-comparison of policy semantic paths to identify whether there are multiple interpretation paths in the scheduling policy in order to determine whether there is ambiguity. S2. After identifying the ambiguity in the scheduling strategy, a response pattern vector is constructed based on the historical scheduling behavior of the site. Then, a response prediction model that integrates a multi-head attention mechanism and a behavior trajectory transformation network is used to determine the differences in the site response behavior when the scheduling strategy is ambiguous. S2 specifically includes the following steps: S201. After identifying the ambiguous situation in the scheduling strategy, extract the behavior records of multiple energy storage sites in the historical scheduling tasks for the ambiguous scheduling strategy. Construct a response mode vector based on the response time offset, power adjustment trajectory, load reference type and feedback adjustment cycle before and after the strategy is triggered. Each response mode vector includes the behavior start time difference, power response function sequence and strategy field weight factor group. S202. Input all response mode vectors into the response prediction model that integrates the multi-head attention mechanism and the behavior trajectory transformation network. Calculate the attention weight distribution of each policy field to the response behavior through the multi-head attention mechanism. Use the behavior trajectory transformation network to perform time-series remapping on the power response function sequence and output the predicted response behavior trajectory of each energy storage site under the current ambiguous policy conditions. S203. Based on the predicted response behavior trajectory of each energy storage site, perform time dimension alignment and numerical normalization processing, calculate the difference indicators between trajectories in response direction, response rhythm and output intensity, and construct a site response behavior difference matrix to represent the degree and type of difference in response behavior between different sites when there is ambiguity in the scheduling strategy. S3. Based on the differences in site response behavior, generate policy parsing prompts through behavior deviation value mapping, and combine the policy parsing prompts with the original scheduling policy to construct a structured scheduling policy with consistent parsing paths, thereby achieving unified parsing and constraint control of the scheduling policy execution logic. S4. Construct a consistent annotation set based on the structured scheduling strategy, embed the consistent annotation set into the semantic expression of the scheduling strategy, and explicitly replace and reorganize ambiguous fields to form a scheduling strategy structure with a clear parsing path. S5. During the execution of the structured scheduling strategy, a dynamic correction loop is established based on the offset function between the actual response trajectory and the expected response trajectory of the station. The explicitness of the annotation and the tolerance range of the behavior are adjusted in real time according to the trajectory offset to achieve consistent closed-loop control of the scheduling strategy execution logic.

2. The multi-site collaborative energy storage centralized control and scheduling method according to claim 1, characterized in that, S1 specifically includes the following steps: S101. Extract the response baseline field, timing parameter field, and behavior intensity field from the scheduling strategy. Based on the causal and constraint relationships of the fields in the logical semantics, establish a multi-level nested field structure and construct a strategy field dependency tree containing the response baseline field, timing parameter field, and behavior intensity field. S102. Traverse the policy field dependency tree to generate multiple policy semantic paths. Each policy semantic path contains a response baseline field, a timing parameter field, and a behavior intensity field, and records the field parsing direction and the logical pointer between fields. S103. Perform field-level alignment and cross-path reorganization operations on all policy semantic paths, calculate the field cross-rate and semantic overlap between semantic paths. If the field cross-rate of multiple paths is lower than the preset threshold and the semantic overlap is inconsistent, it is determined that there are multiple interpretation paths in the scheduling policy and that there is ambiguity.

3. The multi-site collaborative energy storage centralized control and scheduling method according to claim 2, characterized in that, S103 specifically refers to: The position index sequence of the response baseline field, the timing parameter field and the behavior intensity field in each policy semantic path is extracted sequentially, and a field-level alignment operation is performed based on the order of field appearance to construct a path field comparison matrix. Based on the path field comparison matrix, a cross-path recombination operation is performed to extract the set of deviation paths composed of all inconsistent field combinations. The field cross rate and semantic overlap between each pair of paths are calculated, where the field cross rate is the ratio of the number of co-occurring fields to the total number of fields, and the semantic overlap is the weighted average of the semantic label similarity of the fields. Compare the field crossover rate and semantic overlap of all path pairs. If there are path pairs with a field crossover rate lower than a preset threshold and inconsistent semantic overlap, these paths are identified as multiple interpretation paths, and the scheduling strategy is judged to be ambiguous.

4. The multi-site collaborative energy storage centralized control and scheduling method according to claim 1, characterized in that, S202 specifically refers to: For each response pattern vector, expand the field dimensions, extract the initial weight values ​​corresponding to the response baseline field, time series parameter field and behavior intensity field, and input the initial weight values ​​and power response function sequence into the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network to complete the input vector format standardization and dimension alignment. In the response prediction model that integrates multi-head attention mechanism and behavior trajectory transformation network, each attention head calculates the attention weight of each policy field to the behavior at different time periods in the power response function sequence, forming multiple weight distribution vectors. All weight distribution vectors are then weighted and synthesized to generate a unified field attention weight distribution matrix. The field attention weight distribution matrix is ​​input into the behavior trajectory transformation network, and a time-series remapping operation is performed on the power response function sequence. Through time-domain interpolation and feature smoothing, the predicted response behavior trajectory of each energy storage site under the current ambiguous strategy condition is obtained, which serves as a high-order representation of the site response mode.

5. The multi-site collaborative energy storage centralized control and scheduling method according to claim 1, characterized in that, S203 specifically refers to: For the predicted response behavior trajectory of each energy storage site, a time dimension alignment operation is performed to uniformly map the time axis in the trajectory to a standard time reference frame, and the power value sequence is subjected to maximum and minimum normalization processing to ensure that all trajectories are comparable under a unified time scale and power range. Based on the normalized predicted response behavior trajectory, the trajectories between any two energy storage sites are compared sequentially. By calculating the consistency of power change direction, the synchronization error of response rhythm, and the mean deviation of output intensity, the difference values ​​of response direction, response rhythm, and output intensity between each pair of trajectories are obtained. A three-dimensional response behavior difference matrix is ​​constructed by combining the three types of difference values ​​between all pairs of energy storage sites. Each element in the matrix represents the degree and type of difference in the behavioral response of two sites when there is ambiguity in the scheduling strategy, which is used for the adjustment and control of subsequent scheduling parsing logic.

6. The multi-site collaborative energy storage centralized control and scheduling method according to claim 1, characterized in that, S3 specifically refers to: Based on the response direction difference value, response rhythm difference value, and output intensity difference value in the site response behavior difference matrix, a behavior deviation value mapping rule set is constructed. The difference values ​​of different dimensions are mapped to field parsing offset weights. Field groups with significant response deviations are extracted through field parsing offset weights, and strategy parsing prompts containing field identifiers, recommended parsing methods, and adjustment levels are generated. Perform field-level structure mapping operations on the policy parsing prompts and the original scheduling policy. Based on the field identifiers in the prompts, perform structural anchoring and semantic association on the corresponding fields in the original scheduling policy, and rewrite the field parsing path according to the recommended parsing method to construct a structured scheduling policy with consistent parsing paths. Based on the structured scheduling strategy, adjustment level parameters are introduced from the strategy parsing prompt instructions to construct field-level parsing boundary constraints and behavioral response strength limits. By unifying the field parsing logic and behavioral response specifications through boundary constraints and strength limits, the unified parsing and constraint control of the scheduling strategy execution logic is achieved.

7. The multi-site collaborative energy storage centralized control and scheduling method according to claim 1, characterized in that, S4 specifically refers to: Based on the semantic parsing path, parsing offset weight and field boundary constraint parameters of each field in the structured scheduling strategy, the semantic combination relationship of the response benchmark field, the timing parameter field and the behavior intensity field is extracted, a field interpretation logic index table is constructed, and annotation information entries containing field labels, interpretation scope, behavior response mode and parsing prompts are generated according to the semantic dependency and logical order between fields, forming a consistent annotation set. The consistent annotation set is embedded into the semantic expression of the scheduling strategy. The field position index in the structured scheduling strategy is identified through the field label matching mechanism. The corresponding annotation information entries are inserted into the scheduling strategy according to the logical hierarchy using the semantic embedding operation. The semantic structure tree is reconstructed so that each field has a clear parsing boundary and response expectation, realizing the semantic fusion of annotation content and strategy structure. Based on the semantic structure tree of the scheduling strategy, ambiguous field nodes are identified. By comparing the interpretation range and response mode of the fields in the consistent annotation set, the field rewriting operation is performed to replace the original ambiguous fields with explicit semantic fields and adjust the semantic order and nesting level to form a scheduling strategy structure with a unique parsing path and a clear structural hierarchy. This structure is used for unified parsing of the scheduling strategy execution logic and behavioral response specification control.

8. The multi-site collaborative energy storage centralized control and scheduling method according to claim 1, characterized in that, S5 specifically refers to: During the execution of the structured scheduling strategy, the actual response trajectory of each energy storage site is collected and aligned with the expected response trajectory in terms of time and power dimensions. A multi-dimensional offset function is constructed based on the response start time difference, power offset value and response rhythm error. The offset function is used to reflect the dynamic difference between the actual response trajectory and the expected response trajectory. Based on the time offset component, power offset component and rhythm difference component of the multidimensional offset function, the annotation explicitness adjustment parameter and the behavior tolerance range adjustment parameter are calculated respectively. The adjustment parameters are mapped to the field parsing control factor and the behavior response threshold control factor to construct the annotation control model and the tolerance control model. The field parsing control factor is applied to the annotation display level of each field in the structured scheduling strategy to achieve real-time adjustment of the annotation explicitness. At the same time, the behavior response threshold control factor is used to dynamically adjust the behavior tolerance range parameter, establishing a dynamic correction loop that coordinates the parsing logic adjustment and behavior specification correction, and completing the consistent closed-loop control of the scheduling strategy execution logic.

Images