Energy storage power station operation scheduling optimization method and system based on digital twin

Abstract

The present invention provides an energy storage power station operation scheduling optimization method and system based on digital twins, which relates to the technical field of scheduling optimization, including: decomposing the energy storage power station into four standard sub-models of battery, converter, energy management and environment, simplifying, collecting operation data, extracting typical working conditions, searching for optimal parameter combinations, updating standard sub-models and integrating and dynamically decoupling to obtain a digital twin model; establishing a domain knowledge ontology, semantically associating quantitative indicators with ontology concepts, collecting engineering data to generate classic working condition data, converting them into semantic rules and adding them to an initial knowledge base, determining novel association rules, and obtaining a health status assessment knowledge base; fusing to obtain a hybrid intelligent agent and a hybrid reinforcement learning model, establishing an environmental model and an intelligent agent model, interacting, diagnosing online, obtaining initial diagnostic results, optimizing the hybrid reinforcement learning model, generating an optimal scheduling strategy, and fusing to obtain a collaborative optimization scheduling solution.

Description

Technical Field

[0001] The present invention relates to the technical field of scheduling optimization, and in particular to a method and system for optimizing the operation scheduling of an energy storage power station based on digital twins. Background Art

[0002] As the global energy transformation deepens, large-scale energy storage power stations play an increasingly important role in the power system. Energy storage power stations can achieve multiple application scenarios such as peak load shaving, frequency regulation, and power quality improvement through reasonable scheduling, providing strong support for the safe and stable operation of the power grid. However, electrochemical energy storage equipment is easily affected by charging and discharging conditions, resulting in capacity attenuation and life degradation, and the external environment changes are complex and changeable, which brings great challenges to the operation and scheduling of energy storage power stations.

[0003] In the existing technology, the dispatch optimization of energy storage power stations is mostly based on simplified physical models or data-driven models, which make it difficult to accurately describe the dynamic characteristics and nonlinear effects of the equipment. At the same time, the dispatch decision and equipment status monitoring are independent of each other, and there is a lack of coordinated consideration of health management and control optimization.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The present invention provides a method and system for optimizing the operation and scheduling of an energy storage power station based on digital twins, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of the present invention provides a method for optimizing the operation and scheduling of an energy storage power station based on digital twins, comprising:

[0007] Based on the modular modeling concept, the energy storage power station is decomposed into four standard sub-models: battery, converter, energy management and environment. The standard sub-models are simplified, and the operating data of the energy storage power station under actual operating conditions is collected in combination with the parameter self-calibration algorithm. The typical operating conditions are extracted in combination with the mean clustering algorithm. The optimal parameter combination of the standard sub-model is searched through the Bayesian optimization algorithm according to the typical operating conditions. The standard sub-model is updated based on the optimal parameter combination, and the updated standard sub-model is integrated and dynamically decoupled through the collaborative simulation technology to obtain the digital twin model corresponding to the energy storage power station;

[0008] The key concepts, attributes and relationships in the field of health assessment of energy storage power stations are formally defined through an ontology construction method and a domain knowledge ontology is established. The quantitative indicators in the standard sub-model are semantically associated with the ontology concepts based on the ontology mapping rules. The engineering data of the energy storage power station is collected and mapped to the domain knowledge ontology through an ontology matching algorithm. The classic operating condition data is generated by associating and fusion in combination with semantic links. The classic operating condition data is converted into semantic rules through ontology reasoning technology and added to a preset initial knowledge base. The initial knowledge base is adaptively updated according to the operating data and novel association rules are determined through association rule mining technology. The novel association rules are added to the initial knowledge base to obtain a health status assessment knowledge base.

[0009] The digital twin model and the health status assessment knowledge base are integrated to obtain a hybrid agent, a hybrid reinforcement learning model is constructed based on the hybrid agent, an environment model and an agent model are established in combination with model predictive control theory, hybrid simulation is combined to make the environment model and the agent model interact, online diagnosis is performed in combination with a knowledge reasoning algorithm to obtain an initial diagnosis result, extreme working conditions are generated by generating an adversarial network and adversarial samples are added to optimize the hybrid reinforcement learning model to generate an optimal scheduling strategy, the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling solution.

[0010] Preferably,

[0011] Based on the modular modeling concept, the energy storage power station is decomposed into four standard sub-models: battery, converter, energy management and environment. The standard sub-models are simplified, and the operating data of the energy storage power station under actual operating conditions is collected in combination with the parameter self-calibration algorithm. The typical working conditions are extracted in combination with the mean clustering algorithm. According to the typical working conditions, the optimal parameter combination of the standard sub-model is searched through the Bayesian optimization algorithm. The standard sub-model is updated based on the optimal parameter combination, and the updated standard sub-model is integrated and dynamically decoupled through the collaborative simulation technology. The digital twin model corresponding to the energy storage power station includes:

[0012] Based on the modular modeling concept, the energy storage power station is divided into a battery sub-model, a converter sub-model, an energy management sub-model and an environmental sub-model as standard sub-models according to the energy flow. For the standard sub-models, the state variables are removed and the order of the differential equation is reduced by parameter aggregation and variable elimination to obtain a reduced-order simplified model. The standard sub-model in differential form is Laplace transformed to determine the transfer characteristics and establish the transfer function in the frequency domain to obtain a simplified sub-model.

[0013] The simplified sub-model is connected to the monitoring system of the energy storage power station, and the operating data of the energy storage power station is collected to generate a sample set, wherein the operating data includes active power, reactive power and working status, and the samples in the sample set are preprocessed and key characteristic variables including external characteristics and internal characteristics are extracted. Based on the key characteristic variables, the similarity between each sample is calculated in the feature space and a distance matrix is ​​generated, and a plurality of cluster centers are randomly selected to generate a plurality of clusters and each sample is divided into the cluster with the closest distance according to the distance matrix, and the cluster center is updated according to the mean value of the samples in each cluster, and it is repeated until the samples in the cluster are stable. For each cluster, the center, boundary and distribution density of the samples in the cluster are calculated, and the samples are selected as typical samples according to the pre-set screening conditions to obtain the typical working condition;

[0014] An objective function is constructed based on the typical working condition and the output of the standard sub-model, a constrained optimization problem is generated in combination with the physical constraints and prior distribution corresponding to the standard sub-model, a model parameter value range of the standard sub-model is quantized and discretized by binary coding, a continuous parameter space is mapped into a discrete search space, a Gaussian process regression model is constructed based on the initial sampling points and a posterior distribution is obtained, new sampling points are selected in combination with the confidence upper bound and substituted into the objective function of the standard sub-model to obtain a true value and update the Gaussian process regression model based on the true value, the update is repeated until the maximum number of updates is reached, the sampling point with the maximum true value is used as the optimal sampling point, and the optimal parameter combination is obtained;

[0015] The parameters in the standard sub-model are updated based on the optimal parameter combination. For the updated standard sub-model, the corresponding adjoint equation is solved to determine the sensitivity of the performance indicators in the current standard sub-model to the parameters. The standard sub-model is divided into a fast sub-model and a slow sub-model based on the sensitivity. The fast sub-model is solved in the continuous time domain, and the slow sub-model is solved in the discrete time domain. The fast sub-model is quickly solved based on the current boundary conditions and the input of the slow sub-model is updated according to the solution result until convergence. The parallel iteration is performed in combination with the parallel computing method and the relaxation factor to obtain the digital twin model corresponding to the energy storage power station.

[0016] Preferably,

[0017] Solving the corresponding adjoint equation to determine the sensitivity of the performance index to the parameters in the current standard sub-model is shown in the following formula:

[0018] ;

[0019] in, J Represents performance indicators, p i Representation parameters p The vector iQuantity, f j Representation vector f ( x , t , p ) j Quantity, Indicates the state equation j Quantity f j For i Parameters p i The partial derivative of T Indicates the maximum time. λ j ( t ) represents the accompanying variable, which is used to indicate the performance index J Sensitivity to state variables at different time points t The cumulative effect on t represents the time variable, x ( t ) indicates that the system is at time t The state of the moment.

[0020] Preferably,

[0021] The key concepts, attributes and relationships in the field of health assessment of energy storage power stations are formally defined through an ontology construction method, and a domain knowledge ontology is established. The quantitative indicators in the standard sub-model are semantically associated with the ontology concepts based on the ontology mapping rules. The engineering data of the energy storage power station is collected and mapped to the domain knowledge ontology through an ontology matching algorithm. The classic operating condition data is generated by associating and fusion in combination with semantic links. The classic operating condition data is converted into semantic rules through ontology reasoning technology and added to a preset initial knowledge base. The initial knowledge base is adaptively updated according to the operating data and novel association rules are determined through association rule mining technology. The novel association rules are added to the initial knowledge base, and the health status assessment knowledge base obtained includes:

[0022] The key concepts in the health assessment field corresponding to the energy storage power station are formally defined by ontology construction methods. Knowledge elements are extracted from the three dimensions of system, component and fault through top-down and bottom-up methods and organized into a hierarchical structure. For each key concept, data attributes and object attributes are defined by analyzing the intrinsic attribute characteristics, and a concept attribute set is generated. The vertical semantic relationship and the horizontal semantic relationship are determined based on the attributes between different concepts to obtain a concept relationship network. Based on the key concepts, the concept attribute set and the concept relationship network, combined with constraint rules and ontology description language, the concepts, attributes and relationships are converted into a standardized form and the domain knowledge ontology is established;

[0023] Based on the domain knowledge ontology, data resources of the monitoring system in the energy storage power station are collected, the grammatical features and semantic features of the data resources are analyzed, ontology mapping rules are defined, domain terms in the data resources are identified through natural language analysis technology, and the quantitative indicators in the standard sub-model are semantically associated with the ontology concepts corresponding to the domain knowledge ontology, and the engineering data of the energy storage power station from design, manufacturing to operation and maintenance throughout its life cycle are collected, wherein the engineering data refers to multi-source heterogeneous data consisting of design parameters, manufacturing processes, operating conditions and fault repairs, and the engineering data is mapped to the domain knowledge ontology through an ontology matching algorithm, and association fusion is achieved through semantic linking to obtain the classic operating condition data;

[0024] For the classic operating condition data, the ontology reasoning technology is combined with forward linking to realize the automatic conversion of the classic operating condition data into knowledge rules, generate semantic rules and add them to the preset initial knowledge base, and the initial knowledge base is adaptively updated based on the pre-acquired operating data. For the updated initial knowledge base, the association rule mining algorithm is used in combination with a pre-set difference threshold to identify novel association rules that are different from the original association rules, and the novel association rules are added to the initial knowledge base to obtain the health status assessment knowledge base.

[0025] Preferably,

[0026] By using the association rule mining algorithm and combining the preset difference threshold, novel association rules that are different from the original association rules are identified as shown in the following formula:

[0027] ;

[0028] in, D Represents association rules r 1 and association rules r 2 The difference between w 1 ( i ) indicates an item i In the association rules r 1 The weight in w 2 ( i ) indicates an item i In the association rules r 2 The weight in I ( r 1 ) represents association rules r 1 The set of all items contained in I (r 2 ) represents association rules r 2 The set of all items contained in μ represents the balance factor, sup ( r 1 ) represents association rules r 1 The support of sup ( r 2 ) represents association rules r 2 support.

[0029] Preferably,

[0030] The digital twin model and the health status assessment knowledge base are integrated to obtain a hybrid agent, a hybrid reinforcement learning model is constructed based on the hybrid agent, an environment model and an agent model are established in combination with model predictive control theory, the environment model and the agent model are interacted with each other in combination with hybrid simulation, online diagnosis is performed in combination with a knowledge reasoning algorithm to obtain an initial diagnosis result, extreme working conditions are generated by generating adversarial networks and adversarial samples are added to optimize the hybrid reinforcement learning model, an optimal scheduling strategy is generated, and the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling solution including:

[0031] Based on the digital twin model, the standard sub-models in the digital twin model are uniformly described to obtain a hybrid simulation environment, the association rules in the health status assessment knowledge base are encapsulated as a knowledge reasoning engine, and the hybrid simulation environment and the knowledge reasoning engine are coupled to obtain the hybrid intelligent agent;

[0032] Based on the hybrid agent, the state data corresponding to the energy storage power station is used as the state space, the control variables corresponding to the energy storage power station is used as the action space, the optimization target is determined in combination with the task requirements and the reward function is set, and the policy network of the hybrid agent is trained through a deep reinforcement learning algorithm to obtain the hybrid reinforcement learning model;

[0033] In the hybrid simulation environment, based on the model predictive control theory and combined with the rolling horizon optimization mechanism, the environmental evolution within multiple time steps is abstracted as a time-varying linear system, and a hierarchical intelligent agent framework is constructed in combination with the hybrid reinforcement learning model, wherein the execution layer of the hybrid intelligent agent corresponds to the control system of the energy storage power station, the planning layer corresponds to the scheduling decision system of the energy storage power station, and the perception layer corresponds to the online diagnosis system of the energy storage power station, and the environmental model and the intelligent agent model are obtained. The environmental model receives the scheduling action and deduces the evolution trajectory of the system state in the energy storage power station in combination with the external input. The intelligent agent model obtains the environmental state and evaluates the rationality of the current decision. If the rationality is less than the preset rationality threshold, the optimization scheduling curve is regenerated, and the optimization is repeated to generate the optimal control sequence.

[0034] The energy storage power station is dispatched and optimized based on the optimal control sequence. The knowledge reasoning engine receives status data based on the knowledge reasoning algorithm and performs multi-time scale diagnosis in combination with the health status assessment knowledge base. It performs hierarchical diagnosis in combination with the ontology reasoning algorithm and graph embedding technology, and generates equipment health status reports and key parameter warnings as initial diagnosis results to be fed back to the dispatch decision system.

[0035] Based on the conditional generative adversarial network, the data distribution of the real working condition is learned through random noise vectors, the extreme abnormal working condition is constructed by the generator and judged by the discriminator to obtain adversarial samples, the adversarial samples are added to the training set, the hybrid reinforcement learning model is optimized based on the training set, the optimal scheduling strategy is generated based on the optimized hybrid reinforcement learning model and the real working condition, the optimal scheduling strategy is integrated with the initial diagnosis result to obtain a collaborative optimization scheduling solution.

[0036] Preferably,

[0037] The optimization objective function corresponding to the hybrid reinforcement learning model is shown in the following formula:

[0038] ;

[0039] in, J ( θ ) represents the expected value of the expected cumulative reward, E Express expectations, θ represents the parameters of the agent’s policy network, Indicates execution strategy π θ Then we get the marginal distribution of states, Represents the state extracted from the state marginal distribution s , γ represents the discount factor, k represents the offset of the time step, αrepresents the income weight factor, Indicates at time t + k Direct operating income of the power station, β represents the life weight factor, Indicates at time t + k The contribution of equipment health status to lifespan, s Indicates the current status.

[0040] A second aspect of the present invention provides an energy storage power station operation scheduling optimization system based on digital twins, comprising:

[0041] The first unit is used to decompose the energy storage power station into four standard sub-models of battery, converter, energy management and environment based on the modular modeling concept, simplify the standard sub-models, collect the operating data of the energy storage power station under the actual operating conditions in combination with the parameter self-calibration algorithm, extract the typical working conditions in combination with the mean clustering algorithm, search for the optimal parameter combination of the standard sub-model according to the typical working conditions through the Bayesian optimization algorithm, update the standard sub-model based on the optimal parameter combination, and integrate and dynamically decouple the updated standard sub-models through the collaborative simulation technology to obtain the digital twin model corresponding to the energy storage power station;

[0042] The second unit is used to formally define the key concepts, attributes and relationships in the field of health assessment of energy storage power stations through an ontology construction method and establish a domain knowledge ontology, semantically associate the quantitative indicators in the standard sub-model with the ontology concepts based on the ontology mapping rules, collect the engineering data of the energy storage power station and map it to the domain knowledge ontology through an ontology matching algorithm, associate and fuse the data in combination with semantic links to generate classic operating condition data, convert the classic operating condition data into semantic rules through ontology reasoning technology and add them to a preset initial knowledge base, adaptively update the initial knowledge base according to the operating data and determine novel association rules through association rule mining technology, add the novel association rules to the initial knowledge base, and obtain a health status assessment knowledge base;

[0043] The third unit is used to fuse the digital twin model and the health status assessment knowledge base to obtain a hybrid agent, construct a hybrid reinforcement learning model based on the hybrid agent, establish an environment model and an agent model in combination with model predictive control theory, combine hybrid simulation to make the environment model and the agent model interact, combine knowledge reasoning algorithm to perform online diagnosis to obtain initial diagnosis results, generate extreme working conditions through a generative adversarial network and add adversarial samples to optimize the hybrid reinforcement learning model, generate an optimal scheduling strategy, fuse the optimal scheduling strategy with the initial diagnosis result, and obtain a collaborative optimization scheduling plan.

[0044] The third aspect of the present invention,

[0045] An electronic device is provided, comprising:

[0046] processor;

[0047] a memory for storing processor-executable instructions;

[0048] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0049] A fourth aspect of the present invention,

[0050] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0051] In the present invention, collaborative simulation technology is used to integrate and dynamically decouple each sub-model, which not only retains the flexibility of modularization, but also takes into account the overall synergy, so that the digital twin model can accurately map the real-time operating status of the energy storage power station, and provide a high-fidelity digital space for subsequent diagnosis and optimization. The health status assessment knowledge base breaks through the boundaries of data, models, and knowledge, realizes cross-domain semantic layer fusion, makes up for the limitations of a single knowledge source, and provides a global analysis perspective. The diagnostic results generated by knowledge reasoning and the optimal strategy generated by reinforcement learning are mutually verified to form a closed-loop feedback, realizing the collaborative optimization of scheduling and diagnosis. The hybrid reinforcement learning paradigm inherits the interpretability advantage of knowledge-driven, and absorbs the learning evolution ability of data-driven, realizing the collaborative enhancement of knowledge and data. In summary, the present invention realizes the system integration of multiple fields, multiple targets, and multiple spatiotemporal scales, so that energy storage scheduling bids farewell to the static, single-point, and solidified traditional mode, and forms a dynamic collaborative and continuously evolving intelligent paradigm, which is of great significance for improving the operating efficiency of energy storage power stations, reducing operation and maintenance costs, and extending equipment life. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a method for optimizing the operation and scheduling of an energy storage power station based on digital twins according to an embodiment of the present invention;

[0053] Figure 2 It is a structural schematic diagram of an energy storage power station operation scheduling optimization system based on digital twins according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0056] Figure 1 Schematic diagram of the process of optimizing the operation and scheduling of energy storage power stations based on digital twins according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] S1. Based on the modular modeling concept, the energy storage power station is decomposed into four standard sub-models: battery, converter, energy management and environment. The standard sub-models are simplified, and the operating data of the energy storage power station under actual operating conditions is collected in combination with the parameter self-calibration algorithm. The typical operating conditions are extracted in combination with the mean clustering algorithm. According to the typical operating conditions, the optimal parameter combination of the standard sub-model is searched through the Bayesian optimization algorithm. Based on the optimal parameter combination, the standard sub-model is updated, and the updated standard sub-model is integrated and dynamically decoupled through the collaborative simulation technology to obtain the digital twin model corresponding to the energy storage power station;

[0058] The parameter self-calibration algorithm refers to an algorithm that automatically adjusts model parameters in a simulation or model to minimize the difference between the model's output and the actual observed data. The mean clustering algorithm is a common unsupervised learning algorithm used to divide data sets into different clusters. The typical operating condition refers to a specific state or operating condition that often occurs during the operation of a complex system or equipment. The Bayesian optimization is a global optimization method for optimizing black box functions. The standard submodel refers to a standardized model used to describe a specific part in a complex system. The optimal parameter combination refers to a set of parameter values ​​found in a multi-parameter optimization problem so that the objective function reaches the optimal value. The collaborative simulation refers to multiple simulation models working together on the same simulation platform to simulate different parts of a complex system. The dynamic decoupling refers to separating the dynamic characteristics of each part of the system in a multivariable control system so that each part can be independently controlled.

[0059] In an optional embodiment,

[0060] Based on the modular modeling concept, the energy storage power station is decomposed into four standard sub-models: battery, converter, energy management and environment. The standard sub-models are simplified, and the operating data of the energy storage power station under actual operating conditions is collected in combination with the parameter self-calibration algorithm. The typical working conditions are extracted in combination with the mean clustering algorithm. According to the typical working conditions, the optimal parameter combination of the standard sub-model is searched through the Bayesian optimization algorithm. The standard sub-model is updated based on the optimal parameter combination, and the updated standard sub-model is integrated and dynamically decoupled through the collaborative simulation technology. The digital twin model corresponding to the energy storage power station includes:

[0061] Based on the modular modeling concept, the energy storage power station is divided into a battery sub-model, a converter sub-model, an energy management sub-model and an environmental sub-model as standard sub-models according to the energy flow. For the standard sub-models, the state variables are removed and the order of the differential equation is reduced by parameter aggregation and variable elimination to obtain a reduced-order simplified model. The standard sub-model in differential form is Laplace transformed to determine the transfer characteristics and establish the transfer function in the frequency domain to obtain a simplified sub-model.

[0062] The simplified sub-model is connected to the monitoring system of the energy storage power station, and the operating data of the energy storage power station is collected to generate a sample set, wherein the operating data includes active power, reactive power and working status, and the samples in the sample set are preprocessed and key characteristic variables including external characteristics and internal characteristics are extracted. Based on the key characteristic variables, the similarity between each sample is calculated in the feature space and a distance matrix is ​​generated, and a plurality of cluster centers are randomly selected to generate a plurality of clusters and each sample is divided into the cluster with the closest distance according to the distance matrix, and the cluster center is updated according to the mean value of the samples in each cluster, and it is repeated until the samples in the cluster are stable. For each cluster, the center, boundary and distribution density of the samples in the cluster are calculated, and the samples are selected as typical samples according to the pre-set screening conditions to obtain the typical working condition;

[0063] An objective function is constructed based on the typical working condition and the output of the standard sub-model, a constrained optimization problem is generated in combination with the physical constraints and prior distribution corresponding to the standard sub-model, a model parameter value range of the standard sub-model is quantized and discretized by binary coding, a continuous parameter space is mapped into a discrete search space, a Gaussian process regression model is constructed based on the initial sampling points and a posterior distribution is obtained, new sampling points are selected in combination with the confidence upper bound and substituted into the objective function of the standard sub-model to obtain a true value and update the Gaussian process regression model based on the true value, the update is repeated until the maximum number of updates is reached, the sampling point with the maximum true value is used as the optimal sampling point, and the optimal parameter combination is obtained;

[0064] The parameters in the standard sub-model are updated based on the optimal parameter combination. For the updated standard sub-model, the corresponding adjoint equation is solved to determine the sensitivity of the performance indicators in the current standard sub-model to the parameters. The standard sub-model is divided into a fast sub-model and a slow sub-model based on the sensitivity. The fast sub-model is solved in the continuous time domain, and the slow sub-model is solved in the discrete time domain. The fast sub-model is quickly solved based on the current boundary conditions and the input of the slow sub-model is updated according to the solution result until convergence. The parallel iteration is performed in combination with the parallel computing method and the relaxation factor to obtain the digital twin model corresponding to the energy storage power station.

[0065] The modular modeling concept is a method of decomposing a complex system into smaller, self-contained modules that can be developed and analyzed independently. The energy flow refers to the energy transfer within a system or between different systems. The order of the differential equation refers to the highest-order derivative of the dependent variable with respect to the independent variable in the equation. The distance matrix is ​​a table showing the distance between each pair of points in a given space. The cluster center is the center point of a cluster in a data set, indicating the average position of all cluster points and serving as a reference for assigning new data points to the nearest cluster. The discrete search space refers to a search space in which variables can only take discrete, independent values, rather than a continuous range. The Gaussian process regression is a nonparametric, Bayesian-based regression. Method, provides a distribution for a possible function that fits a set of data points, the confidence bound is a strategy for making decisions under uncertainty, balancing exploration and exploitation by choosing the option with the highest potential reward based on current knowledge and uncertainty, the adjoint equation is a class of differential equations used to calculate sensitivity in optimization problems, the fast submodel is a component of a larger system model that operates on a faster time scale, the slow submodel is the counterpart of the fast submodel that operates on a slower time scale, the relaxation factor is a parameter used in iterative methods to control the convergence rate, the parallel computing method involves using multiple computing resources simultaneously to solve a problem, and the computing time can be significantly reduced by breaking the problem into smaller subproblems that can be solved in parallel.

[0066] According to the energy flow of the energy storage power station, it is divided into four subsystems: battery, converter, energy management and environment. A physical model is established for each subsystem to obtain a standard submodel containing differential equations. The standard submodel is simplified by parameter aggregation and variable elimination to reduce the model order. The simplified differential equation is Laplace transformed to obtain a simplified submodel in the form of a transfer function. The simplified submodel is connected to the energy storage power station monitoring system to collect operating data such as active power, reactive power and working status. The collected data is preprocessed to extract key characteristic variables that reflect the external and internal characteristics of the system. The similarity between samples is calculated to generate a distance matrix. The samples are divided using clustering algorithms such as k-means to obtain multiple typical operating condition clusters. The center, boundary and distribution density of each operating condition cluster are analyzed to screen out typical samples.

[0067] Taking the typical working condition as the optimization target and the output of the standard sub-model as the optimization variable, the objective function is constructed. The physical constraints and prior distribution of the standard sub-model are used as optimization constraints to form a constrained optimization problem. The value range of the model parameters is binary-encoded, and the continuous parameter space is discretized. The Gaussian process regression model is trained based on the initial sampling points to obtain the posterior distribution. New sampling points are selected in combination with the confidence upper bound, and the true values ​​are substituted into the objective function. The Gaussian process regression model is updated with the true values. It is iterated repeatedly until convergence, and the optimal sampling points are decoded into model parameters to obtain the optimal parameter combination.

[0068] Substitute the optimal parameter combination into the standard sub-model, update the model parameters, calculate the sensitivity of the performance indicators to the parameters by solving the adjoint equation of the model, divide the standard sub-model into a fast sub-model and a slow sub-model according to the sensitivity analysis results, adopt different time domain integration methods, use continuous time domain solution for the fast sub-model, and use discrete time domain solution for the slow sub-model. At the current time step, first fix the slow sub-model, quickly iterate and solve the fast sub-model, use the result of the fast sub-model as the input of the slow sub-model, update the slow sub-model, repeat and alternately solve the fast and slow sub-models until the whole system converges, introduce relaxation factors to accelerate convergence, and combine parallel computing methods to improve computing efficiency, so as to obtain the digital twin model corresponding to the energy storage power station;

[0069] For example, assuming that there is an electrochemical energy storage power station, it is necessary to build a corresponding digital twin model, and the battery pack, PCS converter, EMS energy management, and temperature and humidity environment are modeled as four independent sub-models. The battery sub-model considers SOC, temperature, voltage and other states, PCS considers IGBT switching characteristics, EMS considers scheduling strategies and prediction algorithms, and the environment sub-model considers dynamic temperature and humidity and heat dissipation conditions. The battery sub-model is simplified by Thevenin equivalent circuit, and the PCS sub-model is reduced by the average value modeling method. The simplified differential equation is Laplace transformed to obtain the battery OCV-SOC curve, PCS efficiency surface and other transfer characteristics, and the power station SCADA system and BMS battery management system are connected to collect one month of operating data, clean and feature engineer the collected data, extract key features reflecting revenue, life, and efficiency, use dynamic time warping to calculate the similarity of working condition samples, generate 64 typical working condition clusters through k-means clustering, calculate the center, boundary, density and other indicators for each cluster, and obtain 8 most representative typical samples after screening;

[0070] Taking battery life and power station revenue under typical working conditions as optimization goals, a weighted objective function is constructed, battery capacity, PCS ratio, and EMS parameters are used as optimization variables, and working condition constraints and parameter prior distribution are used as constraints. For battery capacity, 16 discrete levels are selected in the range of [100, 200], and for PCS ratio, 8 levels are selected in the range of [0.5, 1.5]. Latin hypercube sampling is used to generate initial sample points, which are substituted into the physical model to solve the revenue and life. The confidence upper bound criterion is used to select new sampling points and solve the objective function. The Gaussian process model is continuously updated. After repeated sampling 50 times, the optimal battery capacity of 175Ah, the optimal PCS ratio of 1.2, and the optimal EMS parameters of (0.2, 0.8) are obtained.

[0071] The standard sub-model is updated using the optimal parameter combination, and the battery model is divided into a SOC fast sub-model and a capacity slow sub-model. PCS and EMS are used as intermediate sub-models. The fast sub-model is solved using the explicit Euler method, and the slow sub-model is optimized using the differential evolution algorithm. The SOC sub-model is quickly iterated on the millisecond scale, and the results are fed back to the PCS and EMS sub-models. After optimizing the intermediate sub-model, the capacity attenuation sub-model is updated on the day scale, and a relaxation factor of 0.8 is introduced to accelerate the convergence of multi-scale coupling. 32 typical operating conditions are solved in parallel, and the converged standard sub-models are assembled into a digital twin model of the energy storage power station.

[0072] In this embodiment, the multi-time scale analysis method can separate the fast and slow time scales according to the dynamic characteristics of the sub-model, and solve them separately using appropriate numerical methods. Multi-physical field coupling is achieved through parallel iteration, avoiding the rigidity problem caused by time scale differences, and significantly improving the simulation speed of digital twins. Machine learning methods such as Gaussian process regression are used to adaptively sample the unknown parameter space, and Bayesian optimization is performed in combination with the prior constraints of the physical model. The accuracy of parameter identification is improved while ensuring the optimization efficiency. The optimized model parameters have a wider range of applicability and enhance the adaptability of the digital twin model to changes in operating conditions. In summary, this embodiment achieves a balanced improvement in accuracy, efficiency, and reliability, and provides a new solution for the safe and economical operation of new power systems.

[0073] In an optional embodiment,

[0074] Solving the corresponding adjoint equation to determine the sensitivity of the performance index to the parameters in the current standard sub-model is shown in the following formula:

[0075] ;

[0076] in, J Represents performance indicators, p i Representation parameters p The vector i Quantity, f j Representation vector f ( x , t , p ) j Quantity, Indicates the state equation j Quantity f j For i Parameters p i The partial derivative of T Indicates the maximum time. λ j ( t ) represents the accompanying variable, which is used to indicate the performance index J Sensitivity to state variables at different time points t The cumulative effect on t represents the time variable, x ( t ) indicates that the system is at time t The state of the moment.

[0077] In this embodiment, sensitivity analysis can quantitatively evaluate the influence of each parameter on the system performance, identify key parameters with large contributions to performance indicators and non-key parameters with small contributions, provide a theoretical basis for model reduction and simplification, and avoid model distortion caused by empiricism and blind simplification. By analyzing the time scale of the state variable response caused by different parameters, the system can be divided into fast sub-models and slow sub-models according to the sensitivity, avoiding the computational redundancy caused by the use of a unified time step, thereby greatly improving the efficiency of numerical solution. By accumulating the sensitivity at different times, the transmission and amplification effects of parameter disturbances in the dynamic evolution of the system can be analyzed, revealing the inherent mechanism between parameters and performance, providing theoretical guidance for parameter optimization and control, and providing a quantitative basis for system design and improvement. In summary, this embodiment can guide model simplification, accelerate solution, reveal mechanism, integrate multiple disciplines, and achieve self-adaptation, significantly improving the application effect of digital twins in engineering practice, and has important theoretical significance and practical value.

[0078] S2. Formalize the key concepts, attributes and relationships in the field of health assessment of energy storage power stations through ontology construction methods and establish a domain knowledge ontology. Based on the ontology mapping rules, semantically associate the quantitative indicators in the standard sub-model with the ontology concepts. Collect the engineering data of the energy storage power station and map it to the domain knowledge ontology through the ontology matching algorithm. Combine semantic links to generate classic operating condition data through association fusion. Convert the classic operating condition data into semantic rules through ontology reasoning technology and add them to the preset initial knowledge base. Adaptively update the initial knowledge base according to the operating data and determine novel association rules through association rule mining technology. Add the novel association rules to the initial knowledge base to obtain a health status assessment knowledge base.

[0079] The ontology construction method refers to the process of creating and defining an ontology. Ontology is a data model used to represent concepts and their relationships within a field. The ontology mapping rules refer to the rules for mapping concepts, attributes and relationships in one ontology to corresponding elements in another ontology. The semantic association refers to the association between data or information based on their meaning. The ontology matching algorithm is an algorithm for identifying and aligning similar or related concepts between two or more ontologies. The classic operating condition data refers to operating conditions or data sets that are widely used in specific fields or applications and are considered standard or typical. The ontology reasoning technology is a technology for automatic reasoning based on the logical structure of the ontology. The association rule mining technology is a data mining technology used to discover the relationship between different items in large data sets. The novel association rules refer to association rules that are significantly different from existing knowledge or association rules.

[0080] In an optional embodiment,

[0081] The key concepts, attributes and relationships in the field of health assessment of energy storage power stations are formally defined through an ontology construction method, and a domain knowledge ontology is established. The quantitative indicators in the standard sub-model are semantically associated with the ontology concepts based on the ontology mapping rules. The engineering data of the energy storage power station is collected and mapped to the domain knowledge ontology through an ontology matching algorithm. The classic operating condition data is generated by associating and fusion in combination with semantic links. The classic operating condition data is converted into semantic rules through ontology reasoning technology and added to a preset initial knowledge base. The initial knowledge base is adaptively updated according to the operating data and novel association rules are determined through association rule mining technology. The novel association rules are added to the initial knowledge base, and the health status assessment knowledge base obtained includes:

[0082] The key concepts in the health assessment field corresponding to the energy storage power station are formally defined by ontology construction methods. Knowledge elements are extracted from the three dimensions of system, component and fault through top-down and bottom-up methods and organized into a hierarchical structure. For each key concept, data attributes and object attributes are defined by analyzing the intrinsic attribute characteristics, and a concept attribute set is generated. The vertical semantic relationship and the horizontal semantic relationship are determined based on the attributes between different concepts to obtain a concept relationship network. Based on the key concepts, the concept attribute set and the concept relationship network, combined with constraint rules and ontology description language, the concepts, attributes and relationships are converted into a standardized form and the domain knowledge ontology is established;

[0083] Based on the domain knowledge ontology, data resources of the monitoring system in the energy storage power station are collected, the grammatical features and semantic features of the data resources are analyzed, ontology mapping rules are defined, domain terms in the data resources are identified through natural language analysis technology, and the quantitative indicators in the standard sub-model are semantically associated with the ontology concepts corresponding to the domain knowledge ontology, and the engineering data of the energy storage power station from design, manufacturing to operation and maintenance throughout its life cycle are collected, wherein the engineering data refers to multi-source heterogeneous data consisting of design parameters, manufacturing processes, operating conditions and fault repairs, and the engineering data is mapped to the domain knowledge ontology through an ontology matching algorithm, and association fusion is achieved through semantic linking to obtain the classic operating condition data;

[0084] For the classic operating condition data, the ontology reasoning technology is combined with forward linking to realize the automatic conversion of the classic operating condition data into knowledge rules, generate semantic rules and add them to the preset initial knowledge base, and the initial knowledge base is adaptively updated based on the pre-acquired operating data. For the updated initial knowledge base, the association rule mining algorithm is used in combination with a pre-set difference threshold to identify novel association rules that are different from the original association rules, and the novel association rules are added to the initial knowledge base to obtain the health status assessment knowledge base.

[0085] The data resources refer to a collection of data that is stored or available for use, which usually includes structured data (such as databases), unstructured data (such as text, images), and semi-structured data. The natural language analysis technology refers to a method of using computer algorithms and models to analyze and understand human language (natural language). The forward link refers to a link from the current step or state to the subsequent step or state in the information flow or process flow. The difference threshold refers to a threshold used to determine whether the difference between the two is significant during data analysis or pattern recognition.

[0086] Using top-down and bottom-up methods, knowledge elements are extracted from the three dimensions of system, component and fault, and the extracted knowledge elements are organized into a hierarchical structure. For each key concept, data attributes and object attributes are defined by analyzing the intrinsic attribute characteristics, and a concept attribute set is generated. Based on the attributes between different concepts, vertical semantic relationships and horizontal semantic relationships are determined to obtain a concept relationship network. Based on key concepts, concept attribute sets and concept relationship networks, combined with constraint rules and ontology description languages, concepts, attributes and relationships are converted into standardized forms to establish a domain knowledge ontology. For example, for the lithium battery system of the energy storage power station, key concepts such as battery packs and battery management systems can be defined. The battery pack has data attributes such as voltage, current, and temperature, and object attributes such as series connection and parallel connection. There is a semantic relationship of monitoring between the battery pack and the battery management system. Concepts, attributes and relationships are formally represented using an ontology language such as OWL to construct a domain ontology for lithium battery health assessment.

[0087] Analyze the grammatical and semantic features of data resources, define ontology mapping rules, identify domain terms in data resources through natural language analysis technology, semantically associate the quantitative indicators in the standard sub-model with the ontology concepts corresponding to the domain knowledge ontology, collect engineering data from the design, manufacturing to operation and maintenance of energy storage power stations (including multi-source heterogeneous data such as design parameters, manufacturing processes, operating conditions and fault repair), map the engineering data to the domain knowledge ontology through the ontology matching algorithm, realize association fusion through semantic links, and obtain classic operating condition data. For example, collect monitoring data such as voltage, current, and temperature of lithium battery packs, identify domain terms such as overcharge and over-discharge, and associate them with the fault type concept in the ontology, collect engineering data such as design parameters (such as capacity, material), production date, and number of operating cycles of battery packs, map them to the attributes of the battery pack concept through ontology matching, and fuse them to form typical data of battery packs under different operating conditions;

[0088] Combined with ontology reasoning technology, the automatic conversion of classic operating condition data to knowledge rules is realized through forward linking, and the generated semantic rules are added to the preset initial knowledge base. The initial knowledge base is adaptively updated based on the pre-acquired operating data. For the updated initial knowledge base, the association rule mining algorithm is used in combination with the pre-set difference threshold to identify novel association rules that are different from the original association rules. The novel association rules are added to the initial knowledge base to obtain a health status assessment knowledge base. For example, according to the typical data of the battery pack under different operating conditions, the semantic rules are converted and generated: IF voltage>4.2V AND temperature>45℃ THEN the battery pack is overcharged, and the rule is added to the initial knowledge base. When the operating data of the battery pack is continuously collected, the knowledge base is updated with the new data, and a new association rule is mined: IF the number of overcharges>30 THEN the battery pack health decreases by 20%, which is added to the knowledge base to form a complete health assessment knowledge base for diagnosing and predicting the health status of the battery pack.

[0089] In this embodiment, the ontology-based knowledge representation is easy for machines to understand and process, which is conducive to knowledge sharing and reuse. Standardized semantic knowledge representation lays the foundation for automatic knowledge reasoning and decision-making. Through ontology mapping, multi-source heterogeneous engineering data is mapped to the domain ontology, and semantic associations are established with ontology concepts to match engineering data with ontology concepts. On the one hand, the intrinsic semantics of the data are explained, and on the other hand, the ontology knowledge is enriched with data. The use of data-driven knowledge discovery can supplement and improve the original knowledge, break through the limitations of manually constructed knowledge bases, expand the coverage of knowledge, realize the automatic generation and update of knowledge bases, and improve the efficiency of knowledge acquisition. Knowledge reasoning integrates expert experience and has better interpretability and robustness, providing strong support for the health management of energy storage power stations. In summary, this embodiment realizes the automatic generation and update of the knowledge base of energy storage power station health assessment, greatly improves the efficiency of knowledge acquisition and application, and has important guiding significance for ensuring the safe, reliable and economical operation of energy storage power stations.

[0090] In an optional embodiment,

[0091] By using the association rule mining algorithm and combining the preset difference threshold, novel association rules that are different from the original association rules are identified as shown in the following formula:

[0092] ;

[0093] in, D Represents association rules r 1 and association rules r 2 The difference between w 1 ( i) indicates an item i In the association rules r 1 The weight in w 2 ( i ) indicates an item i In the association rules r 2 The weight in I ( r 1 ) represents association rules r 1 The set of all items contained in I ( r 2 ) represents association rules r 2 The set of all items contained in μ represents the balance factor, sup ( r 1 ) represents association rules r 1 The support of sup ( r 2 ) represents association rules r 2 support.

[0094] In this embodiment, association rule mining can be used to automatically extract association rules between items from data, revealing the influencing factors and propagation paths of performance degradation and failure occurrence. The mined novel association rules can be used as new knowledge to supplement the original knowledge base, expand the coverage of the knowledge base, and make up for the blind spots of expert experience. By setting a difference threshold, the novelty of the rules added to the knowledge base can be controlled to avoid duplication and redundancy of knowledge. The overlap of the item sets contained in the two rules is considered to characterize the semantic similarity of the rules. By setting a suitable balance factor, a balance can be achieved between the interpretability of knowledge representation and the universality of knowledge models, ensuring the accuracy of rule semantics while taking into account the data support of the rules, thereby improving the quality of knowledge. In summary, this embodiment improves the timeliness of knowledge reasoning and decision-making, balances the semantic similarity and data support of knowledge, takes into account the interpretability and universality of knowledge, and provides strong knowledge support and intelligent decision-making for tasks such as health status assessment, fault diagnosis, and life prediction of energy storage power stations.

[0095] S3. The digital twin model and the health status assessment knowledge base are integrated to obtain a hybrid agent, a hybrid reinforcement learning model is constructed based on the hybrid agent, an environmental model and an agent model are established in combination with model predictive control theory, hybrid simulation is combined to enable the environmental model and the agent model to interact, online diagnosis is performed in combination with a knowledge reasoning algorithm to obtain an initial diagnosis result, extreme working conditions are generated by a generative adversarial network and adversarial samples are added to optimize the hybrid reinforcement learning model, an optimal scheduling strategy is generated, the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling solution.

[0096] The hybrid intelligent agent refers to a complex formed by combining multiple intelligent technologies or models. The hybrid reinforcement learning model refers to a model that combines different types of reinforcement learning algorithms or combines reinforcement learning with other optimization technologies. The model predictive control theory is an optimization method for controlling dynamic systems. Based on the current state of the system and the prediction model, the control input in the future is calculated to make the system reach the optimal state. The knowledge reasoning algorithm refers to an algorithm that performs automatic reasoning based on existing knowledge and logical rules. The generative adversarial network is a deep learning model consisting of a generator and a discriminator. The collaborative optimization scheduling scheme refers to a resource scheduling strategy generated by a collaborative optimization algorithm under multi-objective and multi-constraint conditions.

[0097] In an optional embodiment,

[0098] The digital twin model and the health status assessment knowledge base are integrated to obtain a hybrid agent, a hybrid reinforcement learning model is constructed based on the hybrid agent, an environment model and an agent model are established in combination with model predictive control theory, the environment model and the agent model are interacted with each other in combination with hybrid simulation, online diagnosis is performed in combination with a knowledge reasoning algorithm to obtain an initial diagnosis result, extreme working conditions are generated by generating adversarial networks and adversarial samples are added to optimize the hybrid reinforcement learning model, an optimal scheduling strategy is generated, and the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling solution including:

[0099] Based on the digital twin model, the standard sub-models in the digital twin model are uniformly described to obtain a hybrid simulation environment, the association rules in the health status assessment knowledge base are encapsulated as a knowledge reasoning engine, and the hybrid simulation environment and the knowledge reasoning engine are coupled to obtain the hybrid intelligent agent;

[0100] Based on the hybrid agent, the state data corresponding to the energy storage power station is used as the state space, the control variables corresponding to the energy storage power station is used as the action space, the optimization target is determined in combination with the task requirements and the reward function is set, and the policy network of the hybrid agent is trained through a deep reinforcement learning algorithm to obtain the hybrid reinforcement learning model;

[0101] In the hybrid simulation environment, based on the model predictive control theory and combined with the rolling horizon optimization mechanism, the environmental evolution within multiple time steps is abstracted as a time-varying linear system, and a hierarchical intelligent agent framework is constructed in combination with the hybrid reinforcement learning model, wherein the execution layer of the hybrid intelligent agent corresponds to the control system of the energy storage power station, the planning layer corresponds to the scheduling decision system of the energy storage power station, and the perception layer corresponds to the online diagnosis system of the energy storage power station, and the environmental model and the intelligent agent model are obtained. The environmental model receives the scheduling action and deduces the evolution trajectory of the system state in the energy storage power station in combination with the external input. The intelligent agent model obtains the environmental state and evaluates the rationality of the current decision. If the rationality is less than the preset rationality threshold, the optimization scheduling curve is regenerated, and the optimization is repeated to generate the optimal control sequence.

[0102] The energy storage power station is dispatched and optimized based on the optimal control sequence. The knowledge reasoning engine receives status data based on the knowledge reasoning algorithm and performs multi-time scale diagnosis in combination with the health status assessment knowledge base. It performs hierarchical diagnosis in combination with the ontology reasoning algorithm and graph embedding technology, and generates equipment health status reports and key parameter warnings as initial diagnosis results to be fed back to the dispatch decision system.

[0103] Based on the conditional generative adversarial network, the data distribution of the real working condition is learned through random noise vectors, the extreme abnormal working condition is constructed by the generator and judged by the discriminator to obtain adversarial samples, the adversarial samples are added to the training set, the hybrid reinforcement learning model is optimized based on the training set, the optimal scheduling strategy is generated based on the optimized hybrid reinforcement learning model and the real working condition, the optimal scheduling strategy is integrated with the initial diagnosis result to obtain a collaborative optimization scheduling solution.

[0104] The knowledge reasoning engine is a computing system or software used to execute the knowledge reasoning process. The rolling time domain optimization mechanism is a dynamic optimization method, in which the optimization process is carried out rollingly within a fixed time window. The time-varying linear system refers to a linear system in which system parameters change over time. The planning layer refers to the layer in the system or architecture responsible for formulating long-term strategies or macro decisions. The perception layer is the layer responsible for obtaining data from the external environment or internal state. The execution layer is the layer in the system or architecture responsible for performing actual operations or realizing specific functions. The optimization scheduling curve is a time series curve generated by an optimization algorithm, which is used to guide the operating status or resource allocation of the system or equipment in different time periods. The multi-time scale diagnosis refers to the status monitoring and fault diagnosis of the system or equipment at different time scales. The graph embedding technology is a technology for converting graph-structured data into low-dimensional vector space representation. The random noise vector refers to a vector with random distribution characteristics, which is usually used in generative adversarial networks as the input of the generator.

[0105] The standard sub-models in the digital twin model are described uniformly to obtain a hybrid simulation environment, the association rules in the health status assessment knowledge base are encapsulated into a knowledge reasoning engine, and the hybrid simulation environment and the knowledge reasoning engine are coupled to obtain a hybrid intelligent agent. For example, for the digital twin model of the energy storage power station, the sub-models such as battery, PCS, and EMS are described uniformly to build a hybrid simulation environment, and the association rules in the battery health assessment knowledge base: IF temperature>45℃ AND number of overcharges>30 THEN SOH drops by 20%, etc. are encapsulated into a knowledge reasoning engine, and the two are coupled to form an energy storage hybrid intelligent agent, which can perform system simulation and health diagnosis at the same time;

[0106] The state data corresponding to the energy storage power station (such as SOC, temperature) is used as the state space, and the control variables (such as charging and discharging power) are used as the action space. The optimization goal is determined in combination with the task requirements (such as minimizing the operating cost), and the reward function (such as the profit per kilowatt-hour) is set. The policy network of the hybrid agent is trained through a deep reinforcement learning algorithm (such as DDPG) to obtain a hybrid reinforcement learning model. Based on the model predictive control theory, the environmental evolution within multiple time steps is abstracted into a time-varying linear system in combination with the rolling time domain optimization mechanism. A hierarchical agent framework is constructed in combination with the hybrid reinforcement learning model. Among them, the execution layer of the hybrid agent corresponds to the control system of the energy storage power station, the planning layer corresponds to the scheduling decision system, and the perception layer corresponds to the online diagnosis system. The environmental model and the agent model are obtained. The environmental model receives the scheduling action and is combined with external inputs (such as grid demand) to deduce the system state in the energy storage power station. Evolution trajectory, the intelligent agent model obtains the environmental state and evaluates the rationality of the current decision. If the rationality is less than the preset threshold, the optimized scheduling curve is regenerated, and the optimization is repeated to generate the optimal control sequence. For example, based on the MPC theory, each scheduling decision considers the demand forecast for the next 24 hours, and the 24 hours are discretized into several time steps. The energy storage power station can be regarded as a linear system in each time step. In the hierarchical intelligent agent, the execution layer controls the real-time charging and discharging of the battery, the planning layer makes a scheduling decision every 15 minutes, and the perception layer continuously monitors the battery health status. The environmental model receives the scheduling curve, combines the load forecast, and calculates the system's SOC, temperature and other state changes. The intelligent agent model evaluates the impact of the scheduling curve on the operating cost and battery life. If the cost is too high or the life decays too fast, the weight of the optimization target is adjusted, and a new scheduling curve is generated again. This is repeated until the optimal control sequence is found;

[0107] The energy storage power station is dispatched and optimized based on the optimal control sequence. The knowledge reasoning engine receives status data based on the knowledge reasoning algorithm, performs multi-time scale diagnosis in combination with the health status assessment knowledge base, and performs hierarchical diagnosis in combination with the ontology reasoning algorithm and graph embedding technology. It generates equipment health status reports and key parameter warnings, which are fed back to the dispatch decision system as initial diagnosis results. For example, the energy storage power station performs charging and discharging scheduling according to the optimal control sequence. The knowledge reasoning engine receives BMS data in real time, monitors key parameters such as the voltage and temperature of single cells at the second time scale, diagnoses the balanced state of the battery pack at the minute time scale, and evaluates the health state of the battery system at the hour time scale. Ontology reasoning is used to perform comprehensive diagnosis from different levels of single cells, battery clusters, battery packs, and battery systems, and generates: a health report that the battery system SOH is 90% and the estimated remaining life is 5 years, and a warning that the temperature of the No. 3 battery cluster is too high and heat dissipation needs to be strengthened, which are fed back to the dispatch decision system;

[0108] The data distribution of the real working condition is learned through the random noise vector, the extreme abnormal working condition is constructed through the generator, and the adversarial sample is obtained through the discriminator. The adversarial sample is added to the training set, and the hybrid reinforcement learning model is optimized based on the training set. Based on the optimized hybrid reinforcement learning model and the real working condition, the optimal scheduling strategy is generated, and the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling solution. For example,

[0109] Based on the conditional generative adversarial network, abnormal operating condition data such as extreme high temperature and SOC saturation are generated and added to the training set as adversarial samples. The hybrid reinforcement learning model is retrained on richer data. Under real operating conditions, the hybrid reinforcement learning model gives the optimal scheduling strategy. For example, the current SOC is 90%, and the load forecast for the next 4 hours is 100kW. The optimal strategy is to discharge 20kW for 4 hours. Combined with the health status given by the knowledge reasoning engine: the current battery SOH is 90%, and the maximum continuous discharge power is 30kW. The optimal strategy is constrained and corrected to obtain a collaborative optimization scheduling plan: the current SOC is 90%, and the power will be continuously discharged at 20kW for the next 4 hours. At the same time, the discharge is suspended for 5 minutes every half an hour to avoid over-discharge of the battery.

[0110] In this embodiment, a high-precision virtual model of the energy storage power station is constructed through digital twin technology, the operating status and environmental parameters of the physical equipment are synchronized in real time, and the hybrid intelligent model of knowledge-driven and data-driven is integrated to provide an explainable basis for diagnostic decision-making. Data-driven provides generalized learning capabilities in unknown scenarios. The integrated hybrid intelligent agent has both explainability and adaptability. By generating an adversarial network to construct extreme abnormal working conditions, the robustness of the scheduling strategy to deal with black swan events is enhanced. Through target decomposition and behavior coordination, the distributed optimization of multiple intelligent agents is achieved under the premise of meeting system-level constraints. The entire scheduling system exhibits emergent intelligence, continuous learning and evolution in a dynamic environment. In summary, this embodiment realizes systematic innovation of models, methods, and architectures, making energy storage scheduling move from static to dynamic, from single point to coordination, and from specific to multiple, comprehensively improving the intelligence level of energy storage power stations, simplifying manual operations, improving equipment utilization efficiency, and extending battery life, with significant technical benefits and economic value.

[0111] In an optional embodiment,

[0112] The optimization objective function corresponding to the hybrid reinforcement learning model is shown in the following formula:

[0113] ;

[0114] in, J ( θ ) represents the expected value of the expected cumulative reward,E Express expectations, θ represents the parameters of the agent’s policy network, Indicates execution strategy π θ Then we get the marginal distribution of states, Represents the state extracted from the state marginal distribution s , γ represents the discount factor, k represents the offset of the time step, α represents the income weight factor, Indicates at time t + k Direct operating income of the power station, β represents the life weight factor, Indicates at time t + k The contribution of equipment health status to lifespan, s Indicates the current status.

[0115] In this embodiment, by adjusting the size of the income weight factor and the life weight factor, the relative importance of income and life can be flexibly balanced. Through cumulative reward optimization, the scheduling strategy is made more reasonable in the time dimension, myopic behavior is avoided, and long-term benefits are maximized. The weighted sum of the health status reward and the operating income reward integrates the equipment life and the operating income into the same optimization framework, so that the scheduling decision can actively adapt to the health status of the equipment, while improving the income and delaying the battery aging as much as possible. The distribution-based method comprehensively considers all possible state distributions, reduces sampling bias, and makes the training more stable and efficient. The policy gradient theorem reveals the zero bias and consistency of the objective function gradient, ensures the convergence of the policy optimization process, and avoids the problems of gradient vanishing and explosion. In summary, this embodiment comprehensively describes the technical and economic characteristics of energy storage scheduling, provides a scientific and reasonable optimization direction for subsequent reinforcement learning training, can give full play to the comprehensive benefits of energy storage power stations, and promote the formation of a sustainable scheduling model that takes into account current and long-term interests, local benefits and global benefits.

[0116] Figure 2 Schematic diagram of the structure of the energy storage power station operation scheduling optimization system based on digital twins according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0117] The first unit is used to decompose the energy storage power station into four standard sub-models of battery, converter, energy management and environment based on the modular modeling concept, simplify the standard sub-models, collect the operating data of the energy storage power station under the actual operating conditions in combination with the parameter self-calibration algorithm, extract the typical working conditions in combination with the mean clustering algorithm, search for the optimal parameter combination of the standard sub-model according to the typical working conditions through the Bayesian optimization algorithm, update the standard sub-model based on the optimal parameter combination, and integrate and dynamically decouple the updated standard sub-models through the collaborative simulation technology to obtain the digital twin model corresponding to the energy storage power station;

[0118] The second unit is used to formally define the key concepts, attributes and relationships in the field of health assessment of energy storage power stations through an ontology construction method and establish a domain knowledge ontology, semantically associate the quantitative indicators in the standard sub-model with the ontology concepts based on the ontology mapping rules, collect the engineering data of the energy storage power station and map it to the domain knowledge ontology through an ontology matching algorithm, associate and fuse the data in combination with semantic links to generate classic operating condition data, convert the classic operating condition data into semantic rules through ontology reasoning technology and add them to a preset initial knowledge base, adaptively update the initial knowledge base according to the operating data and determine novel association rules through association rule mining technology, add the novel association rules to the initial knowledge base, and obtain a health status assessment knowledge base;

[0119] The third unit is used to fuse the digital twin model and the health status assessment knowledge base to obtain a hybrid agent, construct a hybrid reinforcement learning model based on the hybrid agent, establish an environment model and an agent model in combination with model predictive control theory, combine hybrid simulation to make the environment model and the agent model interact, combine knowledge reasoning algorithm to perform online diagnosis to obtain initial diagnosis results, generate extreme working conditions through a generative adversarial network and add adversarial samples to optimize the hybrid reinforcement learning model, generate an optimal scheduling strategy, fuse the optimal scheduling strategy with the initial diagnosis result, and obtain a collaborative optimization scheduling plan.

[0120] According to a third aspect of the embodiments of the present invention,

[0121] An electronic device is provided, comprising:

[0122] processor;

[0123] a memory for storing processor-executable instructions;

[0124] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0125] A fourth aspect of the embodiments of the present invention is:

[0126] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0127] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The energy storage power station operation scheduling optimization method based on digital twin is characterized by: include: Based on the modular modeling concept, the energy storage power station is decomposed into four standard sub-models: battery, converter, energy management and environment. The standard sub-models are simplified, and the operating data of the energy storage power station under actual operating conditions is collected in combination with the parameter self-calibration algorithm. The typical working conditions are extracted in combination with the mean clustering algorithm. According to the typical working conditions, the optimal parameter combination of the standard sub-model is searched through the Bayesian optimization algorithm. The standard sub-model is updated based on the optimal parameter combination, and the updated standard sub-model is integrated and dynamically decoupled through the collaborative simulation technology to obtain the digital twin model corresponding to the energy storage power station, including: Based on the optimal parameter combination, the parameters in the standard sub-model are updated. For the updated standard sub-model, the corresponding adjoint equation is solved to determine the sensitivity of the performance index in the current standard sub-model to the parameters. Based on the sensitivity, the standard sub-model is divided into a fast sub-model and a slow sub-model. The fast sub-model is solved in the continuous time domain, and the slow sub-model is solved in the discrete time domain. The fast sub-model is quickly solved based on the current boundary conditions and the input of the slow sub-model is updated according to the solution result until convergence. The parallel computing method and the relaxation factor are combined to perform parallel iteration to obtain the digital twin model corresponding to the energy storage power station. The fast sub-model is a component of a larger system model and operates on a faster time scale. The slow sub-model is the corresponding part of the fast sub-model and operates on a slower time scale. The key concepts, attributes and relationships in the field of health assessment of energy storage power stations are formally defined through an ontology construction method and a domain knowledge ontology is established. The quantitative indicators in the standard sub-model are semantically associated with the ontology concepts based on the ontology mapping rules. The engineering data of the energy storage power station is collected and mapped to the domain knowledge ontology through an ontology matching algorithm. The classic operating condition data is generated by associating and fusion in combination with semantic links. The classic operating condition data is converted into semantic rules through ontology reasoning technology and added to a preset initial knowledge base. The initial knowledge base is adaptively updated according to the operating data and novel association rules are determined through association rule mining technology. The novel association rules are added to the initial knowledge base to obtain a health status assessment knowledge base. The digital twin model and the health status assessment knowledge base are integrated to obtain a hybrid agent, a hybrid reinforcement learning model is constructed based on the hybrid agent, an environment model and an agent model are established in combination with model predictive control theory, hybrid simulation is combined to make the environment model and the agent model interact, online diagnosis is performed in combination with a knowledge reasoning algorithm to obtain an initial diagnosis result, extreme working conditions are generated by generating an adversarial network and adversarial samples are added to optimize the hybrid reinforcement learning model, an optimal scheduling strategy is generated, the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling plan, and the hybrid agent refers to a complex formed by combining multiple intelligent technologies or models.

2. The method according to claim 1, characterized in that Based on the modular modeling concept, the energy storage power station is decomposed into four standard sub-models: battery, converter, energy management and environment. The standard sub-models are simplified, and the operating data of the energy storage power station under actual operating conditions is collected in combination with the parameter self-calibration algorithm. The typical operating conditions are extracted in combination with the mean clustering algorithm. According to the typical operating conditions, the optimal parameter combination of the standard sub-model is searched through the Bayesian optimization algorithm, including: Based on the modular modeling concept, the energy storage power station is divided into a battery sub-model, a converter sub-model, an energy management sub-model and an environmental sub-model as standard sub-models according to the energy flow. For the standard sub-models, the state variables are removed and the order of the differential equation is reduced by parameter aggregation and variable elimination to obtain a reduced-order simplified model. The standard sub-model in differential form is Laplace transformed to determine the transfer characteristics and establish the transfer function in the frequency domain to obtain a simplified sub-model. The simplified sub-model is connected to the monitoring system of the energy storage power station, and the operating data of the energy storage power station is collected to generate a sample set, wherein the operating data includes active power, reactive power and working status, and the samples in the sample set are preprocessed and key characteristic variables including external characteristics and internal characteristics are extracted. Based on the key characteristic variables, the similarity between each sample is calculated in the feature space and a distance matrix is ​​generated, and a plurality of cluster centers are randomly selected to generate a plurality of clusters and each sample is divided into the cluster with the closest distance according to the distance matrix, and the cluster center is updated according to the mean value of the samples in each cluster, and it is repeated until the samples in the cluster are stable. For each cluster, the center, boundary and distribution density of the samples in the cluster are calculated, and the samples are selected as typical samples according to the pre-set screening conditions to obtain the typical working condition; An objective function is constructed based on the typical working condition and the output of the standard sub-model, a constrained optimization problem is generated in combination with the physical constraints and prior distribution corresponding to the standard sub-model, the model parameter value range of the standard sub-model is quantized and discretized through binary coding, the continuous parameter space is mapped into a discrete search space, a Gaussian process regression model is constructed based on the initial sampling points and the posterior distribution is obtained, new sampling points are selected in combination with the confidence upper bound and substituted into the objective function of the standard sub-model to obtain the true value and update the Gaussian process regression model based on the true value, the update is repeated until the maximum number of updates is reached, the sampling point with the maximum true value is used as the optimal sampling point, and the optimal parameter combination is obtained.

3. The method according to claim 1, characterized in that Solving the corresponding adjoint equation to determine the sensitivity of the performance index to the parameters in the current standard sub-model is shown in the following formula: ； in, J Represents performance indicators, p i Representation parameters p The vector i Quantity, f j Representation vector f ( x , t , p ) j Quantity, Indicates the state equation j Quantity f j right p i The partial derivative of T Indicates the maximum time. λ j ( t ) represents the accompanying variable, which is used to indicate the performance index J Sensitivity to state variables at different time variables t The cumulative effect on t represents the time variable, x ( t ) indicates that the system is in the state where the time variable is t The state of the moment.

4. The method according to claim 1, characterized in that The key concepts, attributes and relationships in the field of health assessment of energy storage power stations are formally defined through an ontology construction method, and a domain knowledge ontology is established. The quantitative indicators in the standard sub-model are semantically associated with the ontology concepts based on the ontology mapping rules. The engineering data of the energy storage power station is collected and mapped to the domain knowledge ontology through an ontology matching algorithm. The classic operating condition data is generated by associating and fusion in combination with semantic links. The classic operating condition data is converted into semantic rules through ontology reasoning technology and added to a preset initial knowledge base. The initial knowledge base is adaptively updated according to the operating data and novel association rules are determined through association rule mining technology. The novel association rules are added to the initial knowledge base, and the health status assessment knowledge base obtained includes: The key concepts in the health assessment field corresponding to the energy storage power station are formally defined by ontology construction methods. Knowledge elements are extracted from the three dimensions of system, component and fault through top-down and bottom-up methods and organized into a hierarchical structure. For each key concept, data attributes and object attributes are defined by analyzing the intrinsic attribute characteristics, and a concept attribute set is generated. The vertical semantic relationship and the horizontal semantic relationship are determined based on the attributes between different concepts to obtain a concept relationship network. Based on the key concepts, the concept attribute set and the concept relationship network, combined with constraint rules and ontology description language, the concepts, attributes and relationships are converted into a standardized form and the domain knowledge ontology is established; Based on the domain knowledge ontology, data resources of the monitoring system in the energy storage power station are collected, the grammatical features and semantic features of the data resources are analyzed, ontology mapping rules are defined, domain terms in the data resources are identified through natural language analysis technology, and the quantitative indicators in the standard sub-model are semantically associated with the ontology concepts corresponding to the domain knowledge ontology, and the engineering data of the energy storage power station from design, manufacturing to operation and maintenance throughout its life cycle are collected, wherein the engineering data refers to multi-source heterogeneous data consisting of design parameters, manufacturing processes, operating conditions and fault repairs, and the engineering data is mapped to the domain knowledge ontology through an ontology matching algorithm, and association fusion is achieved through semantic linking to obtain the classic operating condition data; For the classic operating condition data, the ontology reasoning technology is combined with forward linking to realize the automatic conversion of the classic operating condition data into knowledge rules, generate semantic rules and add them to the preset initial knowledge base, and the initial knowledge base is adaptively updated based on the pre-acquired operating data. For the updated initial knowledge base, the association rule mining algorithm is used in combination with a pre-set difference threshold to identify novel association rules that are different from the original association rules, and the novel association rules are added to the initial knowledge base to obtain the health status assessment knowledge base.

5. The method according to claim 4, characterized in that By using the association rule mining algorithm and combining the preset difference threshold, novel association rules that are different from the original association rules are identified as shown in the following formula: ； in, D Represents association rules r 1 and association rules r 2 The difference between w 1 ( i ) indicates an item i In the association rules r 1 The weight in w 2 ( i ) indicates an item i In the association rules r 2 The weight in I ( r 1 ) represents association rules r 1 The set of all items contained in I ( r 2 ) represents association rules r 2 The set of all items contained in μ represents the balance factor, sup ( r 1 ) represents association rules r 1 The support of sup ( r 2 ) represents association rules r 2 support.

6. The method according to claim 1, characterized in that The digital twin model and the health status assessment knowledge base are integrated to obtain a hybrid agent, a hybrid reinforcement learning model is constructed based on the hybrid agent, an environment model and an agent model are established in combination with model predictive control theory, the environment model and the agent model are interacted with each other in combination with hybrid simulation, online diagnosis is performed in combination with a knowledge reasoning algorithm to obtain an initial diagnosis result, extreme working conditions are generated by generating adversarial networks and adversarial samples are added to optimize the hybrid reinforcement learning model, an optimal scheduling strategy is generated, and the optimal scheduling strategy and the initial diagnosis result are integrated to obtain a collaborative optimization scheduling solution including: Based on the digital twin model, the standard sub-models in the digital twin model are uniformly described to obtain a hybrid simulation environment, the association rules in the health status assessment knowledge base are encapsulated as a knowledge reasoning engine, and the hybrid simulation environment and the knowledge reasoning engine are coupled to obtain the hybrid intelligent agent, which refers to a complex formed by combining multiple intelligent technologies or models; Based on the hybrid agent, the state data corresponding to the energy storage power station is used as the state space, the control variables corresponding to the energy storage power station is used as the action space, the optimization target is determined in combination with the task requirements and the reward function is set, and the policy network of the hybrid agent is trained through a deep reinforcement learning algorithm to obtain the hybrid reinforcement learning model; In the hybrid simulation environment, based on the model predictive control theory and combined with the rolling horizon optimization mechanism, the environmental evolution within multiple time steps is abstracted as a time-varying linear system, and a hierarchical intelligent agent framework is constructed in combination with the hybrid reinforcement learning model, wherein the execution layer of the hybrid intelligent agent corresponds to the control system of the energy storage power station, the planning layer corresponds to the scheduling decision system of the energy storage power station, and the perception layer corresponds to the online diagnosis system of the energy storage power station, and the environmental model and the intelligent agent model are obtained. The environmental model receives the scheduling action and deduces the evolution trajectory of the system state in the energy storage power station in combination with the external input. The intelligent agent model obtains the environmental state and evaluates the rationality of the current decision. If the rationality is less than the preset rationality threshold, the optimization scheduling curve is regenerated, and the optimization is repeated to generate the optimal control sequence. The time-varying linear system refers to a linear system whose system parameters change with time. The energy storage power station is dispatched and optimized based on the optimal control sequence. The knowledge reasoning engine receives status data based on the knowledge reasoning algorithm and performs multi-time scale diagnosis in combination with the health status assessment knowledge base. It performs hierarchical diagnosis in combination with the ontology reasoning algorithm and graph embedding technology, and generates equipment health status reports and key parameter warnings as initial diagnosis results to be fed back to the dispatch decision system. Based on the conditional generative adversarial network, the data distribution of the real working condition is learned through random noise vectors, the extreme abnormal working condition is constructed by the generator and judged by the discriminator to obtain adversarial samples, the adversarial samples are added to the training set, the hybrid reinforcement learning model is optimized based on the training set, the optimal scheduling strategy is generated based on the optimized hybrid reinforcement learning model and the real working condition, the optimal scheduling strategy is integrated with the initial diagnosis result to obtain a collaborative optimization scheduling solution.

7. The method according to claim 6, characterized in that The optimization objective function corresponding to the hybrid reinforcement learning model is shown in the following formula: ； in, J ( θ ) represents the expected value of the expected cumulative reward, E Express expectations, θ represents the parameters of the agent’s policy network, Indicates execution strategy π θ Then we get the marginal distribution of states, Indicates that the current state is extracted from the state marginal distribution s , γ represents the discount factor, k represents the offset of the time step, α represents the income weight factor, Indicates at time t + k Direct operating income of the power station, β represents the life weight factor, Indicates at time t + k The contribution of equipment health status to lifespan, s Indicates the current status.

8. A digital twin-based energy storage power station operation scheduling optimization system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to decompose the energy storage power station into four standard sub-models of battery, converter, energy management and environment based on the modular modeling concept, simplify the standard sub-models, collect the operating data of the energy storage power station under the actual operating conditions in combination with the parameter self-calibration algorithm, extract the typical working conditions in combination with the mean clustering algorithm, search for the optimal parameter combination of the standard sub-model according to the typical working conditions through the Bayesian optimization algorithm, update the standard sub-model based on the optimal parameter combination, and integrate and dynamically decouple the updated standard sub-models through the collaborative simulation technology to obtain the digital twin model corresponding to the energy storage power station; The second unit is used to formally define the key concepts, attributes and relationships in the field of health assessment of energy storage power stations through an ontology construction method and establish a domain knowledge ontology, semantically associate the quantitative indicators in the standard sub-model with the ontology concepts based on the ontology mapping rules, collect the engineering data of the energy storage power station and map it to the domain knowledge ontology through an ontology matching algorithm, associate and fuse the data in combination with semantic links to generate classic operating condition data, convert the classic operating condition data into semantic rules through ontology reasoning technology and add them to a preset initial knowledge base, adaptively update the initial knowledge base according to the operating data and determine novel association rules through association rule mining technology, add the novel association rules to the initial knowledge base, and obtain a health status assessment knowledge base; The third unit is used to fuse the digital twin model and the health status assessment knowledge base to obtain a hybrid agent, construct a hybrid reinforcement learning model based on the hybrid agent, establish an environment model and an agent model in combination with model predictive control theory, combine hybrid simulation to make the environment model and the agent model interact, combine knowledge reasoning algorithm to perform online diagnosis to obtain initial diagnosis results, generate extreme working conditions through a generative adversarial network and add adversarial samples to optimize the hybrid reinforcement learning model, generate an optimal scheduling strategy, fuse the optimal scheduling strategy with the initial diagnosis result, and obtain a collaborative optimization scheduling plan.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images