Dynamic management and control method for lithium battery energy storage system

Abstract

The invention relates to the technical field of dynamic management and control of lithium battery energy storage, and discloses a dynamic management and control method of a lithium battery energy storage system. The method comprises a battery state report text, environment monitoring waveform data and an operation record data set of a load demand configuration file; the method comprises the following steps: performing semantic analysis and enhancement on a text to obtain a state semantic vector, performing decomposition and feature synthesis on waveform data to obtain an environment waveform vector, and performing multi-level fusion on the state semantic vector and the environment waveform vector to form a fusion feature map; mining an association relationship in the atlas by using an analysis architecture, and constructing an association network between a load demand and a battery state; cooperative analysis is carried out based on the network, a dynamic control command of identification parameter modification and strategy optimization is generated, and an adjustment plan is constructed according to the dynamic control command to be issued and executed. According to the method, deep analysis and fusion of multi-source heterogeneous data are realized, and the accuracy and response adaptivity of dynamic management and control of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of dynamic management and control technology for lithium battery energy storage, specifically a dynamic management and control method for a lithium battery energy storage system. Background Technology

[0002] Current dynamic management methods for lithium battery energy storage systems often employ a separate processing approach at the data processing level. For textual data such as battery status reports, existing technologies mostly rely on keyword matching or simple classification for status identification, lacking deep semantic understanding. For environmental monitoring waveform data, conventional methods generally use direct statistical feature extraction or waveform comparison with fixed thresholds, simplifying complex continuous waveforms into a few statistical quantities. In the feature fusion stage, existing solutions typically perform simple vector concatenation or shallow model-based fusion of processed text features and waveform features, failing to fully explore the deep, non-linear interactions between heterogeneous data.

[0003] The aforementioned fragmented and coarse-grained data processing methods have shortcomings. The lack of textual semantic understanding makes it impossible to accurately capture the complex operating condition descriptions implicit in status reports; simple statistical processing of waveforms loses their dynamic details and frequency domain characteristics that change over time, resulting in insufficient ability to identify sudden environmental disturbances or gradual failure modes. Furthermore, shallow feature fusion struggles to establish accurate and interpretable correlation models between battery status, environmental fluctuations, and load demands, causing system control strategies to rely on experience and coarse mapping, failing to achieve precise dynamic control based on multi-dimensional deep correlations. How to perform more refined analysis of environmental waveforms to retain key dynamic information, and how to achieve deep, structured fusion between text and waveform heterogeneous features, are key issues for improving the accuracy of dynamic control of energy storage systems. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic management and control method for lithium battery energy storage systems to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a dynamic management and control method for a lithium battery energy storage system, the method comprising: Collect a dataset of operation records of the target energy storage system over a continuous time interval. The dataset of operation records consists of battery status report text, environmental monitoring waveform data, and load demand configuration file. Semantic parsing and enhancement operations are performed on the battery status report text to output a status semantic vector. Waveform decomposition and feature synthesis operations are performed on the environmental monitoring waveform data to output an environmental waveform vector. The status semantic vector and the environmental waveform vector are fused at multiple levels to generate a fused feature map. A pre-configured analysis architecture is used to perform correlation mining operations on the fused feature map to obtain the correlation network between load demand configuration and battery status report; Based on the relationship network, collaborative analysis is performed on the load demand configuration file and battery status report text to generate dynamic management commands. These dynamic management commands are used to identify modifications to battery operating parameters and optimizations to load scheduling strategies. Based on the dynamic control command, construct the parameter adjustment plan for the energy storage system and transmit the parameter adjustment plan to the control interface of the energy storage system to initiate parameter changes.

[0006] Preferably, the process of performing semantic parsing and enhancement operations on the battery status report text to output a status semantic vector involves the following steps: The battery status report text is split into multiple semantic paragraph units; For each semantic paragraph unit, perform semantic transformation processing with domain knowledge injection to generate a basic semantic vector; Access the lithium battery-specific terminology database, extract term groups related to semantic paragraph units, perform term encoding processing on the term groups, and generate term vector groups; The basic semantic vector and the term vector group are aligned and merged in vector space to obtain the enhanced semantic vector. The enhanced semantic vectors are concatenated into state semantic vectors according to the timestamp order.

[0007] Preferably, the operation stage of performing waveform decomposition and feature synthesis on environmental monitoring waveform data and outputting environmental waveform vectors involves the following processing: Abnormal waveform segment localization processing is performed on environmental monitoring waveform data to identify abnormal waveform intervals in temperature waveform, humidity waveform and voltage waveform; Perform multi-resolution waveform slicing processing within the abnormal waveform range to obtain multiple waveform slices; Perform local waveform feature extraction processing on each waveform slice to generate a slice feature vector; All slice feature vectors are input into the waveform feature aggregator. The matching coefficient between each slice feature vector and the reference waveform template is calculated through the waveform pattern matching mechanism. Based on the matching coefficient, the slice feature vectors are weighted and combined to output the environment waveform vector.

[0008] Preferably, the operation stage of performing multi-level fusion of state semantic vectors and environmental waveform vectors to generate a fused feature map involves the following processing: Perform time axis synchronization processing on the state semantic vector and the environment waveform vector to make them have a consistent time scale; Establish a cross-attention mechanism between semantic vectors and waveform vectors, and calculate the interaction intensity map between the semantic components of each time scale in the state semantic vector and the waveform components of the corresponding time scale in the environment waveform vector. The interaction intensity map guides the bidirectional feature tuning process of the state semantic vector and the environment waveform vector to generate tuned semantic vector and tuned waveform vector. The tuned semantic vector and the tuned waveform vector are input into the feature map generator, and the fused feature map is output through feature interleaving and dimensionality reduction operations.

[0009] Preferably, the operation stage of performing correlation mining on the fused feature map using a pre-configured analysis architecture to obtain the correlation network between load demand configuration and battery status report involves the following processing: The feature cleanup module of the fused feature map input analysis architecture removes irrelevant features and retains the core feature map. In the relationship building module of the analysis architecture, the implicit relationship between each feature dimension in the core feature map and the load demand configuration file and battery status report text is analyzed. The influence weight of each feature dimension on the correlation between load demand configuration and battery status report is calculated based on the implicit relationship, forming a list of influence weights; Based on the influence weight list, feature recombination and mapping processing is performed on the core feature map to generate an association network. The association network is represented by a graph structure to show the association paths between various parameters in the load demand configuration and various indicators in the battery status report.

[0010] Preferably, the operation phase of performing collaborative analysis on load demand configuration files and battery status report texts based on the relationship network to generate dynamic control commands involves the following processing: Traverse the network of relationships, filter out the paths with a relationship strength higher than a set value, and form a set of key relationships; For each associated path in the key associated set, decode the corresponding load demand deviation information and battery status anomaly information; The load demand deviation information and battery status anomaly information are fed into the collaborative analyzer for simulation, generating multiple candidate control sequences; For each candidate control sequence, sequence evaluation and screening are performed, and the most suitable candidate control sequence is selected as the dynamic control command.

[0011] Preferably, the operational phase of sending load demand deviation information and battery status anomaly information into a collaborative analyzer for deduction to generate multiple candidate control sequences involves the following processing: The load demand deviation information is processed by deviation encoding and converted into a demand deviation feature representation. Perform anomaly encoding processing on abnormal battery status information and convert it into anomaly feature representation; Construct a reasoning graph between demand deviation feature representations and state anomaly feature representations, and perform iterative reasoning propagation on the reasoning graph; The inference chain that connects the demand deviation feature representation and the state anomaly feature representation is extracted from the inference graph, and each inference chain corresponds to a candidate control sequence.

[0012] Preferably, the operation phase of performing sequence evaluation and screening for each candidate control sequence, and selecting the most suitable candidate control sequence as the dynamic control command, involves the following processes: Analyze the chain of operation nodes in each candidate control sequence; The sequence evaluation model is invoked to perform logical compliance checks and resource requirement calculations on the operation node chain. Based on the combined results of the logical compliance check and the resource requirement calculation, the fit value of each candidate control sequence is calculated. The candidate control sequence with the highest fitness value is selected as the best-fit candidate control sequence.

[0013] Preferably, the operational phase of constructing a parameter adjustment plan for the energy storage system based on dynamic control commands involves the following processes: Analyze the battery operating parameter modification items and load scheduling strategy optimization items in the dynamic management and control commands; Search the historical management log database to find historical execution records that match battery operating parameter modifications and load scheduling strategy optimizations. Valid execution segments are filtered out from historical execution records and integrated into an initial set of adjustment plans; Perform multi-condition optimization on the initial set of adjustment schemes and output a parameter adjustment plan. The multi-condition optimization includes the coordinated optimization of system stability constraints, energy consumption constraints and response time constraints.

[0014] Preferably, the method further includes: Based on the system response data after the parameter adjustment plan is executed, the pre-configured analysis architecture is updated. The update operations include: collecting operational monitoring data after parameter changes, comparing actual operational characteristics with expected operational characteristics, and adjusting the internal parameter configurations of the feature purification module and relationship building module in the analysis architecture based on the comparison differences.

[0015] Compared with the prior art, the beneficial effects of the present invention are: Performing waveform decomposition and feature synthesis on environmental monitoring waveform data changes the conventional approach of directly extracting statistical features. By decomposing continuous waveform data into different components and then synthesizing them in a targeted manner based on their correlation with specific physicochemical processes such as battery state decay and thermal runaway risk, feature vectors that better characterize the nature of environmental stress and the details of dynamic changes can be extracted from the original waveforms. This enables the system to more accurately identify subtle anomalies and gradual trends caused by environmental factors, improving the accuracy and timeliness of analyzing the impact of the battery operating environment, and providing richer and more targeted inputs for subsequent correlation analysis.

[0016] By performing multi-level fusion operations on state semantic vectors and environmental waveform vectors to generate a fused feature map, this approach breaks through the limitations of simple splicing or weighted averaging. Through sequential interaction and integration, multi-level fusion gradually couples textual semantic information with waveform detail information at the vector, feature cluster, and relational levels, ultimately forming a structured map that can represent the complex topological relationships between multiple factors. This graph-based representation makes the hidden relationships between load demand, battery status, and environmental disturbances explicit and structured. The mining of this relational network can therefore uncover deeper and more stable causal relationships and collaborative patterns. The generation of dynamic control commands is thus built upon a more comprehensive and accurate system state cognition model, improving the targeting and overall synergy of control strategies. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the working principle of the dynamic control method for the lithium battery energy storage system described in this invention. Figure 2 This is a flowchart of the semantic parsing and enhancement operations; Figure 3 A flowchart for multi-level fusion operations; Figure 4 A comparison chart of multidimensional resource demand calculations for candidate control sequences; Figure 5 Adjust the trend chart to analyze the architecture parameters. Detailed Implementation

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

[0019] Please see Figure 1This invention provides a dynamic management and control method for a lithium battery energy storage system. The method includes: collecting a dataset of operation records of the target energy storage system over a continuous time interval, wherein the dataset consists of battery status report text, environmental monitoring waveform data, and load demand configuration file; performing semantic parsing and enhancement operations on the battery status report text to output a status semantic vector; performing waveform decomposition and feature synthesis operations on the environmental monitoring waveform data to output an environmental waveform vector; performing multi-level fusion operations on the status semantic vector and the environmental waveform vector to generate a fused feature map; performing correlation mining operations on the fused feature map using a pre-configured analysis architecture to obtain a correlation network between the load demand configuration and the battery status report; performing collaborative analysis operations on the load demand configuration file and the battery status report text based on the correlation network to generate dynamic management and control commands, wherein the dynamic management and control commands are used to identify modification items of battery operating parameters and optimization items of load scheduling strategies; constructing a parameter adjustment plan for the energy storage system based on the dynamic management and control commands, and transmitting the parameter adjustment plan to the control interface of the energy storage system to initiate parameter changes.

[0020] Example 1: See Figure 2 The process of performing semantic parsing and enhancement operations on the battery status report text to output a status semantic vector involves the following steps: splitting the battery status report text into multiple semantic paragraph units; performing semantic transformation processing with domain knowledge injection on each semantic paragraph unit to generate a basic semantic vector; accessing a lithium battery-specific terminology library to extract terminology groups related to the semantic paragraph units, performing terminology encoding processing on the terminology groups to generate a terminology vector group; performing vector space alignment and merging processing on the basic semantic vector and the terminology vector group to obtain an enhanced semantic vector; and concatenating the enhanced semantic vectors into a status semantic vector according to the timestamp order. The process of performing waveform decomposition and feature synthesis on environmental monitoring waveform data to output environmental waveform vectors involves the following steps: Abnormal waveform segment localization processing is performed on the environmental monitoring waveform data to identify abnormal waveform intervals in temperature, humidity, and voltage waveforms; multi-resolution waveform slicing is performed within the abnormal waveform intervals to obtain multiple waveform slices; local waveform feature extraction processing is performed on each waveform slice to generate slice feature vectors; all slice feature vectors are input into a waveform feature aggregator, and the matching coefficient between each slice feature vector and the reference waveform template is calculated through a waveform pattern matching mechanism; based on the matching coefficients, a weighted combination processing is performed on the slice feature vectors to output the environmental waveform vector.

[0021] In practice, the process of performing semantic parsing and enhancement operations on the battery status report text to output state semantic vectors involves a series of steps. The battery status report text is divided into multiple semantic paragraph units, each corresponding to a text block with independent semantics within the battery status report text. Semantic transformation processing with domain knowledge injection is performed on each semantic paragraph unit. This domain knowledge injection semantic transformation is achieved by loading a lithium battery domain knowledge graph into the parameters of the transformation model, generating a basic semantic vector. A lithium battery-specific terminology library is accessed to extract terminology groups related to the semantic paragraph units. This library stores standard terms and their definitions in the lithium battery technology field, and each terminology group contains multiple related terms. Terminology encoding is performed on the terminology groups. This terminology encoding process uses an embedding layer to map each term into a numerical vector, generating a terminology vector group. Vector space alignment and merging are performed on the basic semantic vectors and terminology vector groups. Vector space alignment projects the basic semantic vectors and terminology vector groups into a unified dimensional space using a linear transformation matrix. Merging is achieved by concatenating vectors and then fusing them through a fully connected layer to obtain enhanced semantic vectors. Finally, the enhanced semantic vectors are concatenated into state semantic vectors according to timestamp order, forming a time-series feature representation. In some embodiments, the segmentation of semantic paragraph units is based on punctuation marks and line breaks, and the semantic transformation processing with domain knowledge injection uses a Transformer-based encoder model. Optionally, the embedding layer parameters in the terminology encoding process are obtained through pre-training with lithium battery domain text, and the linear transformation matrix in the vector space alignment and merging process is optimized through supervised learning.

[0022] In practical implementation, the processing of environmental monitoring waveform data into an environmental waveform vector involves a series of steps, including waveform decomposition and feature synthesis. First, abnormal waveform segments are located using a Z-score-based statistical method to identify waveform ranges deviating from normal ranges in temperature, humidity, and voltage waveforms. Second, multi-resolution waveform slicing is performed within these abnormal waveform segments, dividing them into multiple segments using windows of varying time lengths. Third, local waveform feature extraction is performed on each segment, using a one-dimensional convolutional neural network to extract local pattern features and generate a segment feature vector. Finally, all segment feature vectors are input into a waveform feature aggregator, which calculates the matching coefficient between each segment feature vector and a baseline waveform template using a waveform pattern matching mechanism. The cosine similarity between the segment feature vector and the baseline waveform template is used as the matching coefficient. Based on the matching coefficient, a weighted combination process is performed on the segment feature vectors, summing them according to the matching coefficient to output the environmental waveform vector. It can be understood that the baseline waveform template in the waveform pattern matching mechanism is a representative feature vector obtained by clustering from historical normal waveform data. In some embodiments, the Z-score threshold for abnormal waveform segment localization is dynamically adjusted according to the statistical distribution of historical waveform data, and the time window length for multi-resolution waveform slice processing is set to multiple values ​​at different scales. Optionally, the one-dimensional convolutional neural network for local waveform feature extraction processing contains multiple convolutional layers and pooling layers, and the weighted combination processing of the waveform feature aggregator calculates the environmental waveform vector using the following formula: in: Represents an environmental waveform vector. Indicates the first Each slice feature vector Indicates the first Matching coefficients of the feature vectors of each slice. This represents the total number of feature vectors in the slice. Matching coefficient. The reference waveform template is calculated through a waveform pattern matching mechanism and is a predefined waveform feature vector.

[0023] Example 2: See Figure 3The operation stage of multi-level fusion of state semantic vectors and environment waveform vectors to generate a fused feature map involves the following processes: performing time axis synchronization processing on the state semantic vectors and environment waveform vectors to ensure they have a consistent time scale; establishing a cross-attention mechanism between the semantic vectors and waveform vectors, calculating the interaction intensity map between the semantic components of each time scale in the state semantic vector and the waveform components of the corresponding time scale in the environment waveform vector; using the interaction intensity map to guide the state semantic vectors and environment waveform vectors to perform bidirectional feature tuning processing, generating tuned semantic vectors and tuned waveform vectors; inputting the tuned semantic vectors and tuned waveform vectors into the feature map generator, and outputting the fused feature map through feature interleaving and dimensionality reduction operations.

[0024] In practical implementation, the process of generating a fused feature map by performing multi-level fusion operations on state semantic vectors and environmental waveform vectors involves a series of steps. First, time-axis synchronization processing is performed on the state semantic vectors and environmental waveform vectors. This time-axis synchronization processing, based on the timestamp information carried by each vector, resamples vector sequences with different sampling frequencies to a unified time scale sequence, ensuring that the state semantic vectors and environmental waveform vectors have consistent time scales. Second, a cross-attention mechanism is established between the semantic vectors and waveform vectors. This mechanism uses the state semantic vectors as the query vector sequence and the environmental waveform vectors as the key and value vector sequences. It calculates the interaction intensity map between the semantic component of each time scale in the state semantic vector and the corresponding waveform component of the environmental waveform vector. The interaction intensity map is a two-dimensional matrix, where each element represents the association weight between a semantic component at a specific time scale and a waveform component at another specific time scale. A bidirectional feature tuning process is performed on the state semantic vector and the environment waveform vector using an interaction intensity map. This process involves weighting the state semantic vector based on the environment waveform vector information using the interaction intensity map, and simultaneously weighting the environment waveform vector based on the state semantic vector information using the transpose of the interaction intensity map, generating tuned semantic and waveform vectors. These are then input into a feature map generator. The generator outputs a fused feature map through feature interleaving and dimensionality reduction operations. The feature interleaving operation concatenates the tuned semantic and waveform vectors along the feature dimension, while the dimensionality reduction operation uses a multilayer perceptron to perform nonlinear transformations and compression on the concatenated high-dimensional features.

[0025] In some embodiments, the unified timescale for timeline synchronization is chosen as the union of the timestamp sets of the state semantic vector and the environment waveform vector, and the missing vector values ​​are filled in using a linear interpolation method. In the cross-attention mechanism, the query vector, key vector, and value vector are obtained by linearly projecting the original vectors. Optionally, the calculation of each association weight in the interaction intensity graph is implemented using the following formula: in: In the interaction intensity map, time... State semantic vector query components and time The correlation weights between the key components of the environmental waveform vector. Indicates time The query vector is obtained by linear projection of the state semantic vector. Indicates time The key vector is obtained by linearly projecting the environmental waveform vector. This represents the total number of times the time scale is unified. It can be understood that this formula implements attention weight calculation based on dot products, and the interaction intensity map is ultimately a combination of all time points. Composition. In the bidirectional feature tuning process, the tuned semantic vector is obtained by adding the weighted sum of the environmental waveform vector value sequence based on the interaction intensity map to the original state semantic vector, and the tuned waveform vector is obtained by adding the weighted sum of the state semantic vector value sequence based on the transpose of the interaction intensity map to the original environmental waveform vector.

[0026] In some embodiments, the multilayer perceptron in the feature map generator comprises two fully connected layers, with ReLU activation functions used between the layers. Optionally, a dimensionality reduction operation reduces the dimension of the concatenated high-dimensional feature vector to a fixed-length low-dimensional feature vector, which is the fused feature map.

[0027] Example 3: The fused feature map is input into the feature purification module of the analysis architecture to remove irrelevant features and retain the core feature map. In the relationship construction module of the analysis architecture, the implicit relationship between each feature dimension in the core feature map and the load requirement configuration file and battery status report text is parsed. The influence weight of each feature dimension on the correlation between the load requirement configuration and the battery status report is calculated based on the implicit relationship, forming an influence weight list. Based on the influence weight list, feature recombination and mapping processing is performed on the core feature map to generate a correlation network. The correlation network represents the correlation path between each parameter in the load requirement configuration and each indicator in the battery status report in a graph structure.

[0028] In practical implementation, the pre-configured analysis architecture performs correlation mining on the fused feature map to obtain the correlation network. This involves a series of steps. The fused feature map is input into the feature purification module of the analysis architecture. The feature purification module evaluates the information content of the features by calculating the variance-to-mean ratio of each feature dimension in the fused feature map, removing feature dimensions with information content below a set threshold, and retaining the core feature map. In the relationship construction module of the analysis architecture, the implicit relationships between each feature dimension in the core feature map and the load demand configuration file and battery status report text are parsed. The parsing of implicit relationships is achieved by training a multi-task neural network. The multi-task neural network takes the core feature map as input and outputs the prediction weights of each feature dimension for load demand configuration parameters and battery status report indicators. Based on the implicit relationships, the influence weight of each feature dimension on the correlation between load demand configuration and battery status report is calculated. The influence weight is calculated based on the product of the weights of the feature dimension when predicting load demand configuration parameters and battery status report indicators, forming an influence weight list. Based on the influence weight list, feature recombination and mapping are performed on the core feature map. The feature recombination and mapping process sorts the feature dimensions in the core feature map from high to low according to their influence weights, and selects the top k feature dimensions for linear combination to generate a correlation network. The correlation network is represented by a graph structure to show the correlation paths between various parameters in the load demand configuration and various indicators in the battery status report. The nodes in the graph represent load demand configuration parameters or battery status report indicators, and the edges represent the correlation strength.

[0029] In some embodiments, the threshold setting in the feature cleansing module is determined based on the percentile of the information content of the feature dimension in the historical fused feature map data, where the information content is measured by the absolute value of the ratio of variance to mean. The multi-task neural network in the relationship building module is trained using a historical dataset labeled with load demand configuration parameters and battery status report indicators. The multi-task neural network contains a shared encoding layer and two independent task output layers. Optionally, the calculation of the influencing weights is implemented using the following formula: in: Indicates the first The influence weights of each feature dimension Indicates the first The prediction weights of each feature dimension for load requirement configuration parameters Indicates the first The predictive weights of each feature dimension for the battery status report metric.

[0030] In some embodiments, the number of k feature dimensions selected in the feature recombination and mapping process is determined by a preset retention ratio, which is dynamically calculated based on the total number of dimensions of the core feature map. Linear combination maps the k feature dimensions to node embedding vectors in the association network through a fully connected layer; the dimension of the node embedding vector is equal to the total number of nodes in the graph. The edge association strength of the association network is obtained by calculating the cosine similarity between corresponding node embedding vectors. Optionally, the graph structure of the association network is represented using an adjacency matrix, where each element in the adjacency matrix represents the cosine similarity between corresponding node pairs.

[0031] Example 4: Traverse the relationship network, filter out the association paths with association strength higher than a set value to form a key association set; for each association path in the key association set, decode the corresponding load demand deviation information and battery status anomaly information; send the load demand deviation information and battery status anomaly information into the collaborative analyzer for deduction to generate multiple candidate control sequences; perform sequence evaluation and screening processing on each candidate control sequence, and select the most suitable candidate control sequence as the dynamic control command. The operation stage of sending the load demand deviation information and battery status anomaly information into the collaborative analyzer for deduction to generate multiple candidate control sequences involves the following processing: perform deviation encoding processing on the load demand deviation information to convert it into a demand deviation feature representation; perform anomaly encoding processing on the battery status anomaly information to convert it into a status anomaly feature representation; construct a reasoning graph between the demand deviation feature representation and the status anomaly feature representation, and perform iterative reasoning propagation on the reasoning graph; extract the reasoning chain connecting the demand deviation feature representation and the status anomaly feature representation from the reasoning graph, with each reasoning chain corresponding to a candidate control sequence. The process of evaluating and screening each candidate control sequence to select the most suitable candidate control sequence as the operational phase of the dynamic control command involves the following steps: parsing the operation node chain in each candidate control sequence; calling the sequence evaluation model to perform logical compliance checks and resource requirement calculations on the operation node chain; and a rule-based judgment system accessing a predefined energy storage system operation specification library to perform sequential logical verification on each adjacent operation node pair in the operation node chain, checking whether it conforms to safe operation procedures and policy dependencies, and outputting a binary logical compliance check result. The regression prediction model is trained using historical operation sequences and their resource consumption records, and analyzes the characteristics of the operation node chain through a neural network structure to predict the computational resource consumption, time delay, and physical equipment load required to execute the chain, outputting a multi-dimensional vector form of resource requirement calculation result. Combining the logical compliance check result and the resource requirement calculation result, the fit value of each candidate control sequence is calculated; the candidate control sequence with the highest fit value is selected as the most suitable candidate control sequence.

[0032] In practical implementation, the process of generating dynamic control commands by performing collaborative analysis on load demand configuration files and battery status report texts based on the relational network involves a series of steps. First, the relational network is traversed to filter out relational paths with a correlation strength higher than a set value. The correlation strength is represented by the weight of the edges in the relational network, and the set value is a predefined threshold, forming a set of key correlations. For each relational path in the key correlation set, the corresponding load demand deviation information and battery status anomaly information are decoded. The decoding operation, based on the load demand configuration parameter nodes and battery status report indicator nodes connected by the relational path, compares the actual values ​​of the corresponding parameters in the load demand configuration file with the standard values ​​to generate load demand deviation information, and compares the current values ​​of the corresponding indicators in the battery status report text with the normal range to generate battery status anomaly information. The load demand deviation information and battery status anomaly information are then fed into a collaborative analyzer for deduction. The collaborative analyzer is a model based on causal reasoning, generating multiple candidate control sequences. Sequence evaluation and screening are performed on each candidate control sequence. The sequence evaluation and screening process calls an independent evaluation model to score the candidate control sequences from multiple dimensions, selecting the most suitable candidate control sequence as the dynamic control command. In practical implementation, the process of feeding load demand deviation information and battery status anomaly information into a collaborative analyzer to generate multiple candidate control sequences involves a series of steps. First, deviation encoding is performed on the load demand deviation information, converting the deviation type, magnitude, and parameter name into a fixed-dimensional numerical vector, thus transforming it into a demand deviation feature representation. Second, anomaly encoding is performed on the battery status anomaly information, converting the anomaly type, level, and indicator name into a fixed-dimensional numerical vector, thus transforming it into a status anomaly feature representation. Third, a reasoning graph is constructed between the demand deviation feature representation and the status anomaly feature representation. The nodes in the reasoning graph include demand deviation feature representation nodes, status anomaly feature representation nodes, and multiple intermediate reasoning concept nodes, with edges representing causal or correlational relationships between nodes. Fourth, iterative reasoning propagation is performed on the reasoning graph, starting from the demand deviation feature representation nodes and the status anomaly feature representation nodes, activating adjacent intermediate reasoning concept nodes along the edges. After multiple rounds of propagation, a reasoning graph with activation weights is formed. The reasoning graph extracts reasoning chains representing nodes with connection requirements deviation features and nodes with state anomaly features. Each reasoning chain consists of a series of nodes and edges, and each reasoning chain corresponds to a candidate control sequence. The operation nodes in the candidate control sequence are mapped from the intermediate reasoning concept nodes passed through in the reasoning chain.

[0033] In some embodiments, the causal reasoning model of the collaborative analyzer includes a predefined knowledge graph that stores causal relationship rules between load demand deviation, battery state anomalies, and potential control operations. The reasoning graph is constructed by extracting subgraphs from the predefined knowledge graph that are related to the current demand deviation feature representation and state anomaly feature representation. Optionally, iterative reasoning propagation employs a random walk and weight accumulation algorithm, with node activation weights updated during propagation. When extracting reasoning chains from the reasoning graph, the top N paths with the highest sum of node activation weights are extracted first. In specific implementations, the operation of performing sequence evaluation and screening for each candidate control sequence to select the most suitable candidate control sequence as the dynamic control command involves a series of steps. The operation node chain in each candidate control sequence is parsed; the operation node chain is the control operation identifier arranged sequentially in the candidate control sequence. The sequence evaluation model is invoked to perform logical compliance checks and resource requirement calculations on the operation node chain. The logical compliance check verifies whether the logical order between operations in the operation node chain conforms to the energy storage system operation specifications, and the resource requirement calculation calculates the computational, time, and physical resource consumption required to execute the operation node chain. Based on the combined results of logical compliance checks and resource requirement calculations, a fit score is calculated for each candidate control sequence. The candidate control sequence with the highest fit score is selected as the most suitable candidate control sequence, and this most suitable candidate control sequence is output as a dynamic control command. The calculation of the fit score integrates the logical compliance score and the resource requirement score. The logical compliance score is a Boolean-converted value, and the resource requirement score is a function of normalized resource consumption values.

[0034] In some embodiments, the sequence evaluation model is a combination of a rule-based judgment system and a regression prediction model. The rule-based judgment system outputs a logical compliance check result, and the regression prediction model outputs a resource requirement calculation result. The logical compliance check result is represented by 0 or 1, where 0 indicates a logical violation and 1 indicates logical compliance. The resource requirement calculation result is a multi-dimensional vector containing quantitative estimates of computational, time, and physical resource consumption. Optionally, the fitness value is calculated using the following formula: in: This represents the fitness value of the candidate control sequence. This indicates the results of the logical compliance check. Indicates the logical compliance score. This represents a vector of resource demand calculation results. Indicates the score for logical compliance. The normalization function, Represents a vector of resource demand measurement results. The comprehensive evaluation function, and These are preset weighting coefficients, and It's understandable that when the logic compliance check results... When it is 0, the fitness value is... Forced to 0, ensuring that logically non-compliant sequences are not selected; when When the fitness value is 1, the fitness value is... The result is determined by a weighted sum of logical compliance score and resource requirement score. Selecting the candidate control sequence with the highest fit score means choosing the optimal sequence based on comprehensive evaluation under the premise of logical compliance. Resource requirement calculation result vector. It may include resource estimates from multiple dimensions, and a comprehensive evaluation function. Weighted summation and normalization were performed on each dimension. See Table 1 for an evaluation result containing three candidate control sequences.

[0035] Table 1: Evaluation Results of Candidate Control Sequences

[0036] As can be understood, based on the examples in Table 1, the logical compliance check results for candidate control sequences SEQ-001 and SEQ-002 are both 1, while that for SEQ-003 is 0. Although the resource requirement calculation result for SEQ-003 is higher (0.90), its fit value is 0 due to logical non-compliance. Among the compliant sequences, the fit value of SEQ-001 (0.81) is higher than that of SEQ-002 (0.80), therefore SEQ-001 is selected as the most suitable candidate control sequence. In some embodiments, the weighting coefficient... and Adjustments should be made based on the actual operating strategy of the energy storage system; if the focus is on operating efficiency, then improve... The weighting should be increased if the focus is on security and stability. Weighting. Optional, logical compliance score. The resource requirement calculation result vector is obtained by querying the compliance rule base for each pair of adjacent operation nodes in the operation node chain and accumulating the scores. The prediction model is obtained by inputting the chain of operation nodes into a trained neural network. The neural network structure contains multiple fully connected hidden layers, and the ReLU activation function is used between the hidden layers to introduce nonlinear transformation capabilities. The final output layer dimension is consistent with the dimension of the resource demand measurement result vector, directly generating a multi-dimensional vector containing estimates of computational resources, time resources, and physical resource consumption.

[0037] See Figure 4In the resource requirement calculation phase of the candidate control sequences, the quantitative results of five candidate sequences (SEQ-001 to SEQ-005) in three dimensions—computational resource consumption, time resource consumption (ms), and physical resource load (%)—are presented intuitively. Specifically, the resource consumption of each sequence exhibits differentiated characteristics: the time resource consumption (ms) of SEQ-002 is significantly higher than other sequences, and its physical resource load (%) is also at a high level; the time resource consumption of SEQ-001, SEQ-004, and SEQ-005 is in the medium range, while their physical resource load is relatively stable; the time resource consumption and physical resource load of SEQ-003 are both at a low level, but considering the sequence evaluation logic, its suitability value needs to be comprehensively determined in conjunction with the logical compliance results. This multi-dimensional comparison method provides visual support for the resource dimension evaluation of candidate control sequences and is an intuitive presentation of the resource requirement calculation results in the process of selecting the most suitable sequence.

[0038] Example 5: The operational phase of constructing a parameter adjustment plan for an energy storage system based on dynamic control commands involves the following processes: parsing battery operating parameter modification items and load scheduling strategy optimization items in the dynamic control commands; searching the historical control log database to find historical execution records that match the battery operating parameter modification items and load scheduling strategy optimization items; filtering out valid execution segments from the historical execution records and integrating them into an initial adjustment scheme set; performing multi-condition optimization processing on the initial adjustment scheme set to output the parameter adjustment plan. The multi-condition optimization processing includes the coordinated optimization of system stability constraints, energy consumption constraints, and response time constraints. This method also includes updating the pre-configured analysis architecture based on the system response data after the parameter adjustment plan is executed. The update operation includes: collecting operational monitoring data after parameter changes, comparing actual operational characteristics with expected operational characteristics, and adjusting the internal parameter configurations of the feature purification module and relationship construction module in the analysis architecture based on the comparison differences.

[0039] In practical implementation, the operational phase of constructing a parameter adjustment plan for the energy storage system based on dynamic control commands involves a series of processing steps. This includes parsing battery operating parameter modification items and load scheduling strategy optimization items from the dynamic control commands. Battery operating parameter modification items include the target parameter name, target modified value, and effective time range. Load scheduling strategy optimization items include the strategy adjustment type, adjustment magnitude, and execution priority. The historical control log database is searched for historical execution records that match the battery operating parameter modification items and load scheduling strategy optimization items. The historical control log database stores each dynamic control command and its corresponding complete execution record. The matching operation calculates the similarity between the parameter name in the battery operating parameter modification items and the strategy type in the load scheduling strategy optimization items, based on historical records. Historical execution records with similarity exceeding a threshold are extracted. Valid execution segments are filtered from the historical execution records. Filtering conditions include successful execution result marking, system stability indicators within allowable ranges, and response time meeting requirements. These are integrated into an initial adjustment scheme set. Each scheme in the initial adjustment scheme set contains a specific set of parameter adjustment steps and strategy execution action sequences. A multi-condition optimization process is performed on the initial set of adjustment schemes. Taking this set as input, and satisfying system stability, energy consumption, and response time constraints, the optimization algorithm finds the scheme with the minimum overall cost, outputting a parameter adjustment plan, which is a detailed sequence of executable instructions. In practice, this method also includes updating the pre-configured analysis architecture based on system response data after the parameter adjustment plan is executed, collecting operational monitoring data after parameter changes (including battery status report text, environmental monitoring waveform data, and new instances of load demand configuration files), comparing actual operational characteristics with expected operational characteristics (predefined based on the average performance of the parameter adjustment plan in similar historical scenarios), and adjusting the internal parameter configurations of the feature purification and relationship construction modules within the analysis architecture based on the differences.

[0040] In some embodiments, the similarity calculation when retrieving the historical management log database employs a cosine similarity method based on word vectors. Battery operating parameter modifications and load scheduling strategy optimizations are converted into feature vectors and compared with the feature vectors of historical records. When filtering valid execution segments from historical execution records, the allowable range of system stability indicators and response time requirements are read from the energy storage system configuration file. Optionally, multi-condition optimization is achieved by solving a constrained optimization problem, with the goal of minimizing the comprehensive cost of the execution parameter adjustment plan. The comprehensive cost is measured using the following formula: in: This represents the overall cost of the parameter adjustment plan. This represents the total number of solutions in the initial set of adjustment solutions. During the optimization process, a value will be calculated for each solution. value, Indicates execution of the first The predicted system stability deviation for each scheme Indicates execution of the first The predicted energy cost for each option. Indicates execution of the first The predicted total response time for each option , and These are weighting coefficients for system stability deviation, energy consumption cost, and total response time, respectively. It can be understood that the goal of multi-condition optimization is to select from the initial set of adjustment schemes that minimizes the overall cost... The solution with the minimum value is used as the output parameter adjustment plan. System stability constraints, energy consumption constraints, and response time constraints are expressed as inequalities. , , The upper limit is limited.

[0041] In some embodiments, when updating the pre-configured analysis architecture, a comparison difference is generated between the actual running features and the expected running features. This difference is the vector difference between the actual running feature vector and the expected running feature vector in each dimension. Based on the comparison difference, the internal parameter configurations of the feature cleansing module and the relationship building module in the analysis architecture are adjusted. The adjustment operation employs a gradient descent algorithm, aiming to minimize the norm of the comparison difference. This updates the threshold parameters used to evaluate the feature information content in the feature cleansing module and the network weight parameters of the multi-task neural network in the relationship building module. Optionally, new instances of running monitoring data undergo the same preprocessing steps as the original running record dataset before being input into the analysis architecture to ensure consistency in the feature space.

[0042] See Figure 5During the architecture update phase, the dynamic trends of feature purification threshold adjustment and network weight adjustment were observed with each adjustment round. Specifically, the feature purification threshold adjustment (blue curve) and network weight adjustment (red curve) exhibited phased fluctuations: during adjustment rounds 1 and 2, the network weight adjustment rapidly climbed to its peak, while the feature purification threshold adjustment remained in the negative range; in adjustment round 5, the feature purification threshold adjustment reached a positive peak, simultaneously accompanied by a negative trough in the network weight adjustment; in subsequent rounds, the fluctuation trends of the two showed a certain inverse correlation. This parameter adjustment trend corresponds to the implementation logic in the architecture update operation of "adjusting the threshold of the feature purification module and the network weight of the relationship construction module through the gradient descent algorithm based on the comparison between actual and expected operating characteristics," reflecting the dynamic adaptation process of the parameters of the feature purification module and the relationship construction module during iterative optimization.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A dynamic control method for a lithium battery energy storage system, characterized in that, The method achieves dynamic control through a series of sequentially executed operation stages, the specific operation stages of which are as follows: Collect a dataset of operation records of the target energy storage system over a continuous time interval. The dataset of operation records consists of battery status report text, environmental monitoring waveform data, and load demand configuration file. Semantic parsing and enhancement operations are performed on the battery status report text to output a status semantic vector. Waveform decomposition and feature synthesis operations are performed on the environmental monitoring waveform data to output an environmental waveform vector. The status semantic vector and the environmental waveform vector are fused at multiple levels to generate a fused feature map. A pre-configured analysis architecture is used to perform correlation mining operations on the fused feature map to obtain the correlation network between load demand configuration and battery status report; Based on the relationship network, collaborative analysis is performed on the load demand configuration file and battery status report text to generate dynamic management commands. These dynamic management commands are used to identify modifications to battery operating parameters and optimizations to load scheduling strategies. Based on the dynamic control command, construct the parameter adjustment plan for the energy storage system and transmit the parameter adjustment plan to the control interface of the energy storage system to initiate parameter changes.

2. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The process of performing semantic parsing and enhancement on the battery status report text to output a status semantic vector involves the following steps: The battery status report text is split into multiple semantic paragraph units; For each semantic paragraph unit, perform semantic transformation processing with domain knowledge injection to generate a basic semantic vector; Access the lithium battery-specific terminology database, extract term groups related to semantic paragraph units, perform term encoding processing on the term groups, and generate term vector groups; The basic semantic vector and the term vector group are aligned and merged in vector space to obtain the enhanced semantic vector. The enhanced semantic vectors are concatenated into state semantic vectors according to the timestamp order.

3. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The process of performing waveform decomposition and feature synthesis operations on environmental monitoring waveform data to output environmental waveform vectors involves the following steps: Abnormal waveform segment localization processing is performed on environmental monitoring waveform data to identify abnormal waveform intervals in temperature waveform, humidity waveform and voltage waveform; Perform multi-resolution waveform slicing processing within the abnormal waveform range to obtain multiple waveform slices; Perform local waveform feature extraction processing on each waveform slice to generate a slice feature vector; All slice feature vectors are input into the waveform feature aggregator. The matching coefficient between each slice feature vector and the reference waveform template is calculated through the waveform pattern matching mechanism. Based on the matching coefficient, the slice feature vectors are weighted and combined to output the environment waveform vector.

4. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The operation of fusing the state semantic vector and the environment waveform vector at multiple levels to generate a fused feature map involves the following processing: Perform time axis synchronization processing on the state semantic vector and the environment waveform vector to make them have a consistent time scale; Establish a cross-attention mechanism between semantic vectors and waveform vectors, and calculate the interaction intensity map between the semantic components of each time scale in the state semantic vector and the waveform components of the corresponding time scale in the environment waveform vector. The interaction intensity map guides the bidirectional feature tuning process of the state semantic vector and the environment waveform vector to generate tuned semantic vector and tuned waveform vector. The tuned semantic vector and the tuned waveform vector are input into the feature map generator, and the fused feature map is output through feature interleaving and dimensionality reduction operations.

5. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The operation of performing correlation mining on the fused feature map using a pre-configured analysis architecture to obtain the correlation network between load demand configuration and battery status report involves the following processing: The feature cleanup module of the fused feature map input analysis architecture removes irrelevant features and retains the core feature map. In the relationship building module of the analysis architecture, the implicit relationship between each feature dimension in the core feature map and the load demand configuration file and battery status report text is analyzed. The influence weight of each feature dimension on the correlation between load demand configuration and battery status report is calculated based on the implicit relationship, forming a list of influence weights; Based on the influence weight list, feature recombination and mapping processing is performed on the core feature map to generate an association network. The association network is represented by a graph structure to show the association paths between various parameters in the load demand configuration and various indicators in the battery status report.

6. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The process of performing collaborative analysis on load demand configuration files and battery status report texts based on the relational network to generate dynamic control commands involves the following steps: Traverse the network of relationships, filter out the paths with a relationship strength higher than a set value, and form a set of key relationships; For each associated path in the key associated set, decode the corresponding load demand deviation information and battery status anomaly information; The load demand deviation information and battery status anomaly information are fed into the collaborative analyzer for simulation, generating multiple candidate control sequences; For each candidate control sequence, sequence evaluation and screening are performed, and the most suitable candidate control sequence is selected as the dynamic control command.

7. The dynamic control method for a lithium battery energy storage system according to claim 6, characterized in that, The operation phase, which involves feeding load demand deviation information and battery status anomaly information into a collaborative analyzer for simulation and generating multiple candidate control sequences, involves the following processing: The load demand deviation information is processed by deviation encoding and converted into a demand deviation feature representation. Perform anomaly encoding processing on abnormal battery status information and convert it into anomaly feature representation; Construct a reasoning graph between demand deviation feature representations and state anomaly feature representations, and perform iterative reasoning propagation on the reasoning graph; The inference chain that connects the demand deviation feature representation and the state anomaly feature representation is extracted from the inference graph, and each inference chain corresponds to a candidate control sequence.

8. The dynamic control method for a lithium battery energy storage system according to claim 6, characterized in that, For each candidate control sequence, sequence evaluation and screening are performed. The operation phase of selecting the most suitable candidate control sequence as the dynamic control command involves the following processes: Analyze the chain of operation nodes in each candidate control sequence; The sequence evaluation model is invoked to perform logical compliance checks and resource requirement calculations on the operation node chain. Based on the combined results of the logical compliance check and the resource requirement calculation, the fit value of each candidate control sequence is calculated. The candidate control sequence with the highest fitness value is selected as the best-fit candidate control sequence.

9. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The operational phase of constructing a parameter adjustment plan for an energy storage system based on dynamic control commands involves the following processes: Analyze the battery operating parameter modification items and load scheduling strategy optimization items in the dynamic management and control commands; Search the historical management log database to find historical execution records that match battery operating parameter modifications and load scheduling strategy optimizations. Valid execution segments are filtered out from historical execution records and integrated into an initial set of adjustment plans; Perform multi-condition optimization on the initial set of adjustment schemes and output a parameter adjustment plan. The multi-condition optimization includes the coordinated optimization of system stability constraints, energy consumption constraints and response time constraints.

10. The dynamic control method for a lithium battery energy storage system according to claim 1, characterized in that, The method further includes: Based on the system response data after the parameter adjustment plan is executed, the pre-configured analysis architecture is updated. The update operations include: collecting operational monitoring data after parameter changes, comparing actual operational characteristics with expected operational characteristics, and adjusting the internal parameter configurations of the feature purification module and relationship building module in the analysis architecture based on the comparison differences.

Images