A method for establishing a multi-node optical storage direct flexible system energy storage scheduling model

Abstract

The present application relates to the technical field of optical storage direct flexible scheduling, and discloses a method for establishing a multi-node optical storage direct flexible system energy storage scheduling model. The method comprises obtaining real-time operation data of each node of the system, including photovoltaic output fluctuation curves, energy storage state of charge, load demand time sequence, grid interaction power limit value and environmental irradiance variation characteristics, and performing multi-dimensional alignment processing to form a multi-dimensional operation data set. The data set is subjected to operation state mapping topology construction processing to separate target operation data and state indication data. Based on the target operation data, combined with each node scheduling priority configuration information, a multi-node collaborative energy storage scheduling initial model containing energy storage charging and discharging time sequence strategy and power distribution scheme is generated. Based on the state indication data, node operation abnormality characteristics are identified, and the initial model is dynamically reconstructed accordingly to output an updated energy storage scheduling model, so as to ensure stable and efficient operation of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optical storage direct flexible scheduling, in particular to a method for establishing a multi-node optical storage direct flexible system energy storage scheduling model. BACKGROUND

[0002] With the rapid development of new energy generation technology, photovoltaic as an important part of clean and renewable energy continues to increase its share in the energy system. The optical storage direct flexible system integrates photovoltaic power generation, energy storage devices, direct power distribution and flexible control technology, which can effectively improve energy utilization efficiency and reduce dependence on traditional power grids. However, in the multi-node optical storage direct flexible system, the operating states of each node are significantly different and are greatly affected by external environmental factors, which poses many challenges to the stability and reliability of the overall system operation.

[0003] Currently, the energy storage scheduling methods for optical storage direct flexible systems are mostly focused on single-node or simple multi-node scenarios, lacking comprehensive consideration of complex operating environments in multi-node scenarios. In actual operation, the photovoltaic output of each node is significantly volatile due to changes in environmental irradiance, the state of charge of energy storage devices dynamically changes during charging and discharging, and load demand also has temporal differences, while the grid interaction power is also subject to strict limit constraints. These real-time operating data come from various sources and have different dimensions. If they cannot be effectively integrated and processed, it is easy to cause deviations or conflicts between data, which in turn affects the accuracy of energy storage scheduling decisions.

[0004] In the construction process of existing scheduling models, the differences in scheduling priorities of each node are often not fully considered, making it difficult to achieve collaborative optimization scheduling in multi-node scenarios, which may cause insufficient or excessive energy supply in some nodes, reducing the overall energy utilization efficiency of the system. At the same time, when an abnormal operation occurs in a node of the system, the existing model lacks the ability to quickly identify and dynamically adjust, and cannot timely reconstruct the scheduling strategy, which may lead to the spread of abnormalities and affect the stable operation of the entire system. The existence of these problems makes it difficult for multi-node optical storage direct flexible systems to fully exert their advantages in practical applications, limiting their promotion and application in large-scale energy systems. SUMMARY

[0005] The present application aims to provide a method for establishing a multi-node optical storage direct flexible system energy storage scheduling model to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a method for establishing a multi-node optical storage direct flexible system energy storage scheduling model, which comprises:

[0007] Real-time operation data of each node in a photovoltaic energy storage direct flexible system is acquired, multi-dimensional alignment processing is performed on the real-time operation data to form a multi-dimensional operation data set, the real-time operation data including photovoltaic output fluctuation curves, energy storage state of charge, load demand time sequence, grid interaction power limit value and environmental irradiance variation characteristics;

[0008] Operation state mapping topology construction processing is performed on the multi-dimensional operation data set, target operation data used for scheduling model construction and state indication data used for scheduling abnormal response are separated out;

[0009] Based on the target operation data, in combination with scheduling priority configuration information of each node, an energy storage scheduling initial model of multi-node cooperation is generated, the energy storage scheduling initial model including an energy storage charging and discharging time sequence strategy and a power distribution scheme;

[0010] Based on the state indication data, node operation abnormality characteristics in the real-time operation data updating process are identified, and the energy storage scheduling initial model is dynamically reconstructed according to the node operation abnormality characteristics, and an updated energy storage scheduling model is output.

[0011] Preferably, the operation state mapping topology construction processing on the multi-dimensional operation data set includes:

[0012] The multi-dimensional operation data set aligned in the time, node and device dimensions is converted into a scheduling state mapping matrix;

[0013] State characteristic decoupling is performed on the scheduling state mapping matrix to obtain a steady state characteristic matrix and a transient state characteristic matrix, wherein the steady state characteristic matrix represents a normalized scheduling model basic structure, and the transient state characteristic matrix represents abnormal state characteristics corresponding to power over-limit, energy storage attenuation and irradiance mutation;

[0014] The steady state characteristic matrix is taken as the target operation data, and the transient state characteristic matrix is taken as the state indication data.

[0015] Preferably, the generation of the energy storage scheduling initial model of multi-node cooperation includes:

[0016] A photovoltaic output prediction section, an energy storage state of charge interval, load demand peak-valley characteristics and a grid interaction constraint boundary in the steady state characteristic matrix are extracted;

[0017] Optimization weight coefficients are configured for each node based on the scheduling priority configuration information;

[0018] Multi-objective fusion calculation is performed on the photovoltaic output prediction section, the energy storage state of charge interval, the load demand peak-valley characteristics and the grid interaction constraint boundary according to the optimization weight coefficients, and the energy storage charging and discharging time sequence strategy and the power distribution scheme are generated.

[0019] Preferably, the generated multi-node cooperative energy storage scheduling initial model further comprises:

[0020] According to the business attribute of the light-storage direct flexible system, a flexible adjustment target is configured, and the flexible adjustment target comprises a photovoltaic accommodation maximization target, an energy storage life equalization target, a load peak clipping and valley filling target, and a power grid interaction cost optimization target.

[0021] The flexible adjustment target is subjected to hierarchical identification processing to determine a main optimization target, a secondary optimization target, and a boundary constraint target.

[0022] Based on the main optimization target and the secondary optimization target, a key adjustment parameter set is screened, and scheduling influence quantization analysis of the key adjustment parameter set is performed.

[0023] Preferably, the scheduling influence quantization analysis of the key adjustment parameter set comprises:

[0024] An adjustable parameter range is extracted from the steady-state characteristic matrix.

[0025] The main optimization target and the secondary optimization target are subjected to parameter sensitivity mapping to generate a target influence quantization gradient.

[0026] An interaction influence relationship matrix of all target items in the flexible adjustment target is constructed, and the interaction influence relationship matrix comprises a target cooperative relationship and a target conflict relationship.

[0027] Based on the target influence quantization gradient and the interaction influence relationship matrix, a sensitivity characteristic value of an adjustable parameter set is calculated.

[0028] According to the sensitivity characteristic value, a parameter optimization constraint set is established.

[0029] Preferably, the output of the updated energy storage scheduling model comprises:

[0030] A control boundary of the key adjustment parameter set is set, and an initial solution space is created based on a current running state.

[0031] A fitness evaluation is performed on solutions in the initial solution space, and an optimization direction vector and a gradient optimization step are generated in combination with the parameter optimization constraint set and the fitness evaluation result.

[0032] Based on the optimization direction vector and the gradient optimization step, the initial solution space is iteratively updated.

[0033] According to the iterative update result, the updated energy storage scheduling model is output.

[0034] Preferably, the identification of the node running abnormality feature in the real-time running data update process based on the state indication data comprises:

[0035] monitoring an abnormal marking entry of the transient feature matrix, and associating the node operation deviation, the energy storage decay rate and the irradiation mutation intensity corresponding to the abnormal marking entry;

[0036] generating a cross-node abnormal association feature based on the node operation deviation, the energy storage decay rate and the irradiation mutation intensity;

[0037] when the cross-node abnormal association feature exceeds a preset threshold, triggering a multi-node cooperative rescheduling strategy, and updating the power allocation scheme and the energy storage charging and discharging timing strategy according to the transient feature matrix.

[0038] Preferably, the triggering of the multi-node cooperative rescheduling strategy comprises:

[0039] extracting the node number, the abnormal period and the abnormal type feature vector of the node marked as abnormal in the transient feature matrix;

[0040] configuring a node coupling weight coefficient for the feature vector;

[0041] performing clustering analysis on the scheduling association feature of the abnormal node group according to the node coupling weight coefficient, to generate a first abnormal cluster and a second abnormal cluster;

[0042] wherein the first abnormal cluster represents scheduling abnormality caused by photovoltaic output fluctuation, and the second abnormal cluster represents scheduling abnormality caused by load mutation or energy storage failure;

[0043] the first abnormal cluster and the second abnormal cluster are distinguished by node scheduling deviation.

[0044] Preferably, the iterative updating of the initial solution space based on the optimization direction vector and the gradient optimization step length comprises:

[0045] establishing an iteration path identifier for each solution, and classifying the path state through the fitness value of each iteration;

[0046] identifying the path update state within a preset iteration evaluation window, to generate a gradient optimization mode classification, the gradient optimization mode classification comprising a convergence mode, an exploration mode and a divergence mode;

[0047] performing solution space search optimization management based on the gradient optimization mode classification.

[0048] Preferably, the solution space search optimization management based on the gradient optimization mode classification comprises:

[0049] configuring a gradient agent model for the convergence mode, and predicting the optimization trend direction through the gradient agent model;

[0050] An optimization taboo window is configured for the divergent configuration, and invalid search directions are excluded through the optimization taboo window;

[0051] Based on the optimization trend direction and the optimization taboo window, fine iteration is performed on the convergent configuration solution, mixed direction iteration is performed on the exploration configuration solution, and random restart iteration is performed on the divergent configuration solution.

[0052] Compared with the prior art, the present application has the following advantages:

[0053] Through comprehensive acquisition and multi-dimensional alignment processing of real-time operation data of each node in the optical storage direct flexible system, the data such as photovoltaic output fluctuation curves with different sources and dimensions, energy storage state of charge, load demand time sequence, grid interaction power limit, and environmental irradiance variation characteristics can be effectively integrated to form a unified multi-dimensional operation data set, avoiding the adverse effects of data deviation or conflict on scheduling decisions, and making the subsequent scheduling model construction have a more reliable data foundation.

[0054] In the operation state mapping topology construction processing link, the method can accurately separate the target operation data used for scheduling model construction and the state indication data used for scheduling abnormal response, realizing the classified use of data. This classification processing method, on the one hand, enables the target operation data to focus more on the core needs of the scheduling model, reducing the interference of irrelevant data, and helps to improve the construction efficiency and accuracy of the initial model of energy storage scheduling; on the other hand, the separate separation of state indication data provides clear data support for the subsequent identification of node operation abnormal characteristics, making the abnormal identification process more targeted.

[0055] The multi-node collaborative energy storage scheduling initial model generated based on the target operation data and combined with the scheduling priority configuration information of each node fully considers the importance and operation demand differences of different nodes in the system. By reasonably formulating the energy storage charging and discharging time sequence strategy and power allocation scheme, energy collaborative optimization between nodes can be realized, ensuring the reasonable allocation of energy among nodes and avoiding the imbalance between energy supply and demand in some nodes, thereby improving the energy utilization efficiency of the entire optical storage direct flexible system and promoting the economy of the overall system operation. Based on the state indication data, the method can identify the node operation abnormal characteristics in the real-time operation data updating process in time, and dynamically reconstruct the energy storage scheduling initial model according to the abnormal characteristics. This dynamic adjustment capability enables the scheduling model to quickly adapt to changes in node operation state, and when an abnormality occurs in a node, the model can adjust the charging and discharging time sequence strategy and power allocation scheme in time, avoiding greater impact on the overall system operation caused by the abnormality, and ensuring the stability and reliability of the operation of the multi-node optical storage direct flexible system. At the same time, the dynamic reconstruction mechanism also makes the scheduling model more flexible, which can cope with changes under different environmental conditions and operation scenarios, further expanding the application range of the multi-node optical storage direct flexible system. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working principle of the method for establishing a multi-node photovoltaic-storage-direct-drive-flexible system energy storage scheduling model according to the present invention.

[0057] Figure 2 A flowchart for constructing the runtime state mapping topology;

[0058] Figure 3 A flowchart for generating the initial model of multi-node collaborative energy storage scheduling;

[0059] Figure 4 The flowchart for the updated energy storage scheduling model is output.

[0060] Figure 5 This is a flowchart of the multi-node collaborative rescheduling strategy triggering. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 This invention provides a method for establishing an energy storage scheduling model for a multi-node photovoltaic-storage-direct current-flexible system, the method comprising:

[0063] Real-time operation data of each node in the optical storage direct flexible system is acquired, which covers photovoltaic output fluctuation curve, energy storage state of charge, load demand time sequence, grid interaction power limit, and environmental irradiance variation characteristics. The real-time operation data is processed in multiple dimensions to form a multi-dimensional operation data set. The processing includes timestamp alignment, node identification unification, and data sampling frequency normalization, ensuring consistency and comparability of data from different sources in time and space dimensions. Subsequently, the multi-dimensional operation data set is processed to construct a running state mapping topology, which separates the data set into two parts: target operation data for scheduling model construction and state indication data for scheduling abnormal response. Based on the target operation data, and combined with the pre-configured scheduling priority information of each node, an initial energy storage scheduling model of multi-node cooperation is generated, which includes specific energy storage charging and discharging time sequence strategy and power allocation scheme between nodes. At the same time, based on the state indication data, the node operation abnormal characteristics appearing in the real-time operation data updating process are identified, including power limit exceeding, state mutation, etc. According to the identified abnormal characteristics, the generated initial energy storage scheduling model is dynamically reconstructed, and its strategy and scheme are adjusted, and finally the updated energy storage scheduling model is output to adapt to the changes of the system real-time operation state.

[0064] Embodiment 1: see Figure 2 The embodiment relates to a process of converting a multi-dimensional operation data set into a scheduling state mapping matrix, and further decoupling it into steady-state and transient characteristic matrices. The implementation of this process begins with further structural processing of the multi-dimensionally aligned data. Although the multi-dimensional operation data set has been aligned in time, node, and device dimensions, it is still in the form of a discrete data point set and needs to be converted into a matrix form that is more conducive to global analysis and feature extraction. The scheduling state mapping matrix is constructed in a two-dimensional table structure with time as the rows and node-device combinations as the columns. Each row represents a unified timestamp, which comes from the global time sequence determined during the previous alignment process, and the interval is determined by the data acquisition frequency of the system. Each column represents a node and a specific device type on it, such as "Node A photovoltaic component", "Node B energy storage unit", "Node C load", etc. Each element in the matrix fills in the specific operation parameter value of the corresponding time and corresponding node-device combination. These values need to be checked for scale before being filled in to ensure the uniformity of their physical meaning, such as power units being in kilowatts and state of charge being in percentages, but they have not been normalized yet. The construction of this matrix successfully freezes the panoramic running state of the system in a specific time period into a two-dimensional data structure, providing an operating object for subsequent feature separation.

[0065] Decoupling the state features from the scheduling state mapping matrix is the core step, and its purpose is to separate the components representing the long-term stable operation trend of the system from the components representing short-term fluctuations and abnormal events. This process draws on the idea of separating fundamental waves from harmonic waves in signal processing, but applies it to the field of multi-dimensional operation data. In implementation, a sequence analysis method combining sliding window mean filtering and residual calculation is adopted. For each data sequence in the matrix (i.e. each column), a sliding window of appropriate length is defined, which needs to cover most of the normal operation fluctuation period, while being much shorter than the system's main daily or seasonal period changes. The arithmetic mean of the data in each window is calculated, and this mean sequence is used as the steady-state trend line of the data sequence. The difference between the original data sequence and this steady-state trend line constitutes the transient fluctuation component of the sequence. For photovoltaic output, load and other data with strong periodicity, the periodic component will be removed before the above operation to more accurately extract the non-periodic steady-state trend. The steady-state trend lines of all data columns together form the steady-state feature matrix, while the transient fluctuation components of all data columns together form the transient feature matrix.

[0066] The formation of the steady-state feature matrix marks the quantitative expression of the system's basic operation structure, and each value in the matrix represents the "baseline" operation level at the corresponding time and node, excluding short-term disturbances. Before delivering it as target operation data to the subsequent model construction module, it needs to be normalized. Normalization is independent for each node-device type data column, and its purpose is to eliminate the differences brought by different physical dimensions and numerical ranges, so that all data are in a unified, dimensionless numerical interval, usually mapped between zero and one. The maximum and minimum values used for normalization are not the absolute extreme values of the column data, but the upper and lower limits of the reasonable operation range determined based on the device nameplate parameters, historical operation statistics or safe operation procedures. For example, the normalization of the energy storage state of charge is based on its minimum discharge state and maximum charge state, not its physical limits of zero and one hundred. After normalization, the steady-state feature matrix is transformed into a structured data set that purely represents the relative operation state level of each node, and the numerical size directly reflects the position of the node's operation state at that moment relative to its normal range, which greatly facilitates the configuration and calculation of multi-objective weights in subsequent optimization algorithms.

[0067] The processing path of the transient feature matrix is completely different from that of the steady-state matrix, and its core value lies in capturing and quantifying abnormal events deviating from the "baseline" state in system operation. The values in this matrix are the differences between the original data and the steady-state trend, which can contain positive and negative values, and the magnitude of the values directly reflects the amplitude of the fluctuations. In order to more clearly identify the abnormalities, feature extraction and labeling of these transient fluctuation components are needed. In the implementation process, threshold values for abnormality determination are set for each type of operating parameter, and these threshold values are usually adaptive, determined based on the standard deviation multiple of the historical fluctuation statistics of the parameter. When the absolute value of a certain value in the transient matrix exceeds the threshold value set for it, the corresponding data point will be labeled as abnormal. In addition, for some specific types of abnormalities, further feature calculation is needed. For example, for energy storage units, not only the state of charge fluctuation at a single point is concerned, but also the rate of change of the state of charge in a period of time, i.e. the decay or growth rate, which needs to be calculated by analyzing the trend of multiple consecutive points in the transient sequence. For environmental irradiance, the severity of the change, i.e. the mutation strength, is concerned, which is represented by the change amount per unit time. All these abnormal labels, rate calculation values and strength representation values are integrated into the metadata of the transient feature matrix, or in the form of a parallel flag matrix, together constituting the state indication data. The state indication data is like a diagnostic report of the system, clearly indicating where, when, and what type of abnormal condition occurred in the equipment and its severity.

[0068] The separation of steady-state and transient feature matrices constructs a dual description system of "baseline" and "deviation" of system operation state in physical sense. The steady-state matrix depicts the operation trajectory that the system should follow in ideal conditions, which is the target state that the scheduling model strives to achieve and maintain. The transient matrix, on the other hand, monitors and reports all deviations of the actual operation trajectory from this ideal baseline in real time. This separation processing enables the subsequent model construction and abnormal response modules to perform their respective functions efficiently. The model construction module can focus on formulating the optimal scheduling plan based on the clear, noise-free steady-state "blueprint", while the abnormal response module can quickly determine whether and how to intervene and adjust the established plan based on the accurate, quantitative "diagnostic" report. The relative independence and mutual coordination of these two parts of work are an important foundation for the entire method to achieve efficient and robust scheduling function. The entire implementation process relies on a deep understanding of the operation rules of the system and the proper application of data processing techniques, ultimately refining the original, mixed operation data stream into two high-level data products with clear structure and distinct missions, driving the subsequent intelligent scheduling decisions.

[0069] Example 2: see Figure 3The embodiments focus on extracting key operating features from the steady-state feature matrix and generating an initial energy storage scheduling model based on multi-objective fusion calculation, while involving hierarchical adjustment of flexible adjustment targets and screening of key parameter sets. The starting point of the process is to deeply analyze the steady-state feature matrix that has been constructed. The steady-state feature matrix is a structured data set that has been cleaned, aligned and normalized, and its internal contains the core information of the system's future operating situation for a period of time. The extraction operation is not a simple data reading, but a purposeful identification and interval division for specific physical meaning fields. The extraction of photovoltaic output prediction section depends on the analysis of photovoltaic node data columns in the matrix, combined with recent historical data and weather forecast information, and through time series prediction algorithm to deduce the expected power generation curve of several time sections in the future, which reflects the expected change of photovoltaic power generation capacity. The determination of the state of charge interval of the energy storage requires reading the normalized state value of each energy storage node, and according to the maximum and minimum allowable charge and discharge thresholds provided by the battery management system, the safe state of charge range of each energy storage unit at the current time is calculated. This range is a dynamic window. The identification of load demand peak and valley characteristics is completed by statistical analysis of load data columns in the matrix, and the typical power level of the peak load period and the valley load period is found out by using clustering algorithm or rule-based judgment, so as to predict the form of future load curve. The grid interaction constraint boundary is a set of relatively fixed external conditions, but the key constraint values such as time-of-use price, maximum purchasable power and maximum reversible power can be analyzed from the related fields of the matrix. These boundary conditions define the rule framework of energy exchange between the system and the external grid.

[0070] After the above feature extraction is completed, scheduling priority configuration information is introduced to guide the subsequent optimization direction. The priority information is a set of strategic parameters predefined by the system operator, which quantifies the importance of different nodes in global optimization. The setting of priority may be based on various factors, such as a node connected to a critical load, which requires higher reliability of power supply; or the unit cost of energy storage of a node is higher, which needs more precise control of its cycle times; or the light resources in the area where some nodes are located are more abundant, and the priority of photovoltaic consumption is higher. These qualitative strategies are converted into specific optimization weight coefficients and assigned to each node. The weight coefficient is a positive real number, and its absolute value reflects the contribution degree of the node in the objective function, while the relative proportion between the weight coefficients of different nodes determines the allocation tendency of optimization resources among different nodes. A node with a higher weight coefficient has a higher cost when its state deviates from the expected value in the optimization process, so the algorithm will tend to prioritize meeting the operating requirements of the node or more fully utilizing its resources.

[0071] Multi-objective fusion computation is the core optimization process launched after the above preparation is completed. This process needs to coordinate multiple, possibly conflicting system operation objectives extracted from the steady-state feature matrix. These objectives usually contain multiple aspects such as maximizing local consumption of photovoltaic power generation, maintaining and balancing the state of energy storage, smoothing the load curve, and minimizing the cost of grid interaction. Fusion computation uses mathematical programming methods to build an optimization problem model containing multiple objective functions. Optimization weight coefficients are used to construct a weighted sum form of the objective function, converting the multi-objective problem into a single-objective problem for solving, or to define the optimization order based on priority. The calculation process also needs to strictly respect various constraints extracted from the matrix, including the dynamic upper and lower limits of the state of charge of the energy storage, the hard constraints of grid power interaction, and the rigid demand that the load must be met. The choice of optimization algorithm depends on the size and complexity of the problem, and may use linear programming, quadratic programming or more complex heuristic algorithms. Through iterative solving, the calculation process finally outputs a detailed energy storage charging and discharging timing strategy. The strategy is a time series plan that clearly specifies the operation that each energy storage unit is instructed to perform in each scheduling period in the future: charging to absorb excess power, discharging to supplement power shortages, or idle standby. At the same time, the power allocation scheme, as a quantitative supplement to the timing strategy, precisely specifies how the total power demand or surplus is allocated among multiple energy storage units when they need to act simultaneously, and the allocation is usually based on the remaining capacity, rated power and priority weight of each unit.

[0072] In addition to generating the main framework of the initial scheduling model, configuring the flexible adjustment target according to the specific application scenario of the photovoltaic energy storage flexible system is an important work. The flexible adjustment target is a macro effect defined by the user or operator, which gives the mathematical model a business soul. The photovoltaic consumption maximization target requires the system to prioritize scheduling energy storage and other flexible resources to absorb photovoltaic power generation, reduce the "abandoned light" phenomenon, and improve the use of renewable energy. The energy storage life equalization target focuses on avoiding the overuse of some energy storage units while others are idle through control strategies, so that the aging rate of the entire energy storage system tends to be consistent, extending the overall service life. The load peak clipping and valley filling target aims to use the energy storage system to discharge during the peak load period and charge during the valley period, thereby smoothing the overall load curve and reducing the pressure on the grid. The grid interaction cost optimization target focuses on electricity costs by intelligently adjusting the purchase, sale and self-use of electricity during different price periods to achieve optimal operational economy. These targets may have synergies or internal conflicts, such as the need for energy storage to discharge at the same time to achieve peak clipping and reduce electricity costs during peak pricing periods, but frequent charging and discharging cycles can affect energy storage life.

[0073] In the face of multiple possible competing flexible targets, hierarchical identification processing is a key prerequisite for effective optimization. This process sorts and classifies the importance of multiple established flexible targets according to the current system operating conditions, external market environment and operation strategy. The sorting result will determine a core main optimization target in the current scheduling period, for example, in the period of large photovoltaic generation and low electricity price, photovoltaic consumption maximization may be determined as the main target; while in the period of power grid peak load and high electricity price, reducing the cost of power grid interaction may become the main target. The slightly less important targets are listed as secondary optimization targets, which need to be optimized on the premise that the main target is satisfied as much as possible. In addition, some hard conditions that must be met, such as ensuring the power supply of important loads and not violating the power grid power constraints, are explicitly defined as boundary constraint targets, which usually appear in the form of constraint conditions in the optimization model, rather than the objective function. This hierarchical identification converts the complex multi-objective decision-making problem into a structured problem with clear priority order.

[0074] Based on the explicit main and secondary optimization targets, the system can screen out the key adjustment parameter set from the numerous adjustable operating parameters, which are most sensitive and have the most significant impact on achieving the current core target. For example, when the main target is photovoltaic consumption maximization, parameters such as the maximum charge power limit of the energy storage unit and the starting threshold of the adjustable load become key adjustment parameters; while when the main target is power grid cost optimization, parameters such as the threshold for triggering energy storage action based on time-of-use electricity price and the planned value of power exchange with the power grid become crucial. The screening process relies on a deep understanding of the system operation mechanism, and sometimes needs the assistance of tools such as sensitivity analysis described in the foregoing for auxiliary judgment. After determining the key adjustment parameter set, a scheduling impact quantification analysis needs to be performed. This analysis aims to evaluate the direction and extent of the impact on the main and secondary optimization targets when these key parameters vary within their allowed range. For example, the analysis will determine how much the photovoltaic consumption will increase if the maximum charge power of the energy storage unit is increased by one unit, and how much negative impact it will have on the energy storage life decay indicator. This quantification analysis provides a decision basis for how to adjust these parameters in the subsequent optimization process, making the optimization no longer a blind search, but a directional exploration with certain prior knowledge guidance. The entire implementation process embodies the layer-by-layer progression and transformation from data to features, from features to constraints, from constraints to models, and from models to strategies, finally generating an initial scheduling scheme that conforms to the physical operation law of the system and is close to the actual business demand.

[0075] Embodiment 3: see Figure 4Embodiments relate to in-depth scheduling impact quantification analysis on the set of key regulation parameters, and based on the analysis results, output an updated energy storage scheduling model through an iterative optimization process. The starting point of the process is to accurately extract the allowable regulation range of each parameter in the set of key regulation parameters from the constructed steady-state characteristic matrix. This range is not simply a physical limit, but takes into account the safe operating interval specified by the equipment manufacturer, the current actual operating condition of the system, and the reliable operation boundary exhibited in the history of operation. For example, for a battery energy storage unit, the adjustable range of the maximum allowable charging power parameter, the lower limit may be set as the minimum charging power allowed in the current state (to avoid potential damage to the battery caused by small current charging), and the upper limit is strictly limited to the maximum safe power received calculated by the battery management system in real time, which may be dynamically adjusted with the changes in the state of charge, temperature and internal resistance of the battery. Each extracted adjustable range is represented in the form of an interval, which defines the explicit solution space boundary for subsequent optimization search.

[0076] Parameter sensitivity mapping of the main optimization objective and the secondary optimization objective is the core link of quantification analysis, which aims to reveal how the small changes of each key regulation parameter will affect the value of each optimization objective function. In the implementation process, a local sensitivity analysis method based on partial derivative is used. For each objective function and each key parameter, a small perturbation step is defined near the current operating point of the parameter, and the response ratio of the objective function value to the change of the parameter is calculated. This response ratio constitutes the elements of the objective impact quantification gradient matrix, the rows of which correspond to different optimization objectives and the columns of which correspond to different key regulation parameters. The size of the gradient value indicates the strength of the parameter's impact on the objective, and its positive or negative sign indicates the direction of the impact. A positive gradient value means that increasing the value of the parameter will lead to the corresponding objective function value improving in the better direction (for the minimization objective, it is deteriorating), and vice versa.

[0077] Building the interaction influence relationship matrix among all the flexible regulation objectives is another key work, which aims to depict the complex mutual relationship among multiple objectives. Each element of the matrix qualitatively or semi-quantitatively describes the collateral effect on another objective when one of the objectives is optimized. This relationship is mainly determined in two ways: one is based on the derivation of the system physical model to analyze the internal mathematical association between the objective functions; the other is through a large number of simulation calculations of design scenarios to statistically analyze the correlation between the values of each objective in different scenarios. The interaction influence relationship matrix explicitly identifies which pairs of objectives have a synergistic relationship, i.e. optimizing one objective will also promote the improvement of another objective; which pairs of objectives have a conflict relationship, i.e. improving one objective will be at the expense of another objective. For example, the photovoltaic consumption maximization objective and the grid interaction cost optimization objective may exhibit synergistic behavior during the valley period of the daytime electricity price, and may exhibit conflict behavior during the peak period of the electricity price.

[0078] Based on the obtained objective influence quantitative gradient and the interaction influence relationship matrix, an index that comprehensively reflects the sensitivity of the key regulation parameters, i.e. the sensitivity eigenvalue, can be calculated. The calculation of this value aims to comprehensively consider the direct influence of a parameter on multiple objectives and the indirect influence caused by the conflict or synergy between objectives. A feasible comprehensive calculation method is as follows:

[0079]

[0080] Among them: represents the sensitivity eigenvalue of the kth key regulation parameter, and the larger the value, the greater the influence of the parameter on the overall objective set, which needs to be paid more attention in the optimization process. represents the total number of flexible regulation objectives. represents the weight coefficient allocated to the kth optimization objective after hierarchical identification, and the weight of the primary optimization objective is usually higher than that of the secondary optimization objective. represents the kth optimization objective after hierarchical identification. represents the weight coefficient allocated to the kth optimization objective after hierarchical identification, and the weight of the primary optimization objective is usually higher than that of the secondary optimization objective. represents the kth objective function represents the kth objective function with respect to the kth key parameter , i.e. the corresponding element in the objective influence quantitative gradient matrix. is a reference factor, which is used to represent the net interaction relationship strength between the kth objective and other objectives. Its calculation will refer to the interaction influence relationship matrix. If the objective has more synergistic relationships with other main objectives, this factor may be less than 1 to weaken its influence; if it has more conflict relationships, this factor may be greater than 1 to enhance its influence, so as to reflect the amplification effect brought by the objective conflict in the comprehensive sensitivity.

[0081] ​​According to the sensitivity characteristic values of the calculated key parameters, a parameter optimization constraint set can be established, which not only contains the previously extracted physical allowable range, but also can introduce dynamic constraints based on sensitivity analysis. For example, for parameters with extremely high sensitivity characteristic values, their optimization step can be limited to be smaller to avoid excessive optimization process and cause system oscillation; or a more conservative operating boundary can be set for it to reduce the risk of operation. Conversely, for parameters with lower sensitivity characteristic values, more aggressive search can be allowed within their range. The parameter optimization constraint set provides important rule guidance for the subsequent iterative optimization process.

[0082] The output of the updated energy storage scheduling model is a dynamic iterative optimization process. According to the parameter optimization constraint set, clear control boundaries are set for the key parameter set. All subsequent searches will be conducted within the feasible region defined by these boundaries. Based on the current real-time operating state of the system, an initial solution space is created, which contains multiple candidate solutions generated randomly or based on the current operating point. Each solution represents a specific assignment scheme of a set of key parameters. Fitness evaluation is performed on each candidate solution in the initial solution space. The fitness function is usually a scalar function that considers all hierarchical objectives (and their weights) and constraint violation penalty terms. The value of the fitness function directly reflects the pros and cons of the parameter scheme. Based on the parameter optimization constraint set and the fitness evaluation results, the optimization algorithm generates an optimization direction vector and a gradient optimization step. The vector indicates which parameter adjustment direction is most likely to improve the fitness within the current solution space, and the step size suggests the magnitude of parameter adjustment.

[0083] Based on the optimization direction vector and the gradient optimization step, the optimization algorithm iteratively updates the initial solution space. This process usually uses gradient descent algorithms, evolutionary algorithms or other meta-heuristic algorithms. In each iteration, the algorithm generates a new generation of candidate solution population based on the fitness information and direction guidance of the current population. The new generation of solutions may include existing solutions adjusted along the gradient direction, new solutions generated by random combination, and random exploration solutions introduced to maintain diversity. The iteration process continues until the preset termination condition is met, such as reaching the maximum number of iterations, the fitness improvement is not significant, or a solution that meets all requirements is found. The final iteration update result outputs one or more candidate solutions with the best performance. The solution with the highest fitness is selected and its corresponding key parameter combination scheme is embedded into the scheduling strategy, thereby forming and outputting the final updated energy storage scheduling model for guiding the actual operation of the system.

[0084] Example 4: see Figure 5, the core of the embodiment is to identify operation abnormalities in real time based on state indication data and trigger a multi-node collaborative rescheduling strategy. The operation of this process relies on continuous monitoring and analysis of the transient characteristic matrix. The transient characteristic matrix, as a quantitative record of the deviation of system operation state from the "baseline", has dynamic internal data, including power limit indicators, state mutation rates, and other abnormal indication signals. The system operation monitoring module scans the matrix at a very high frequency, focusing on data entries marked as abnormal by preset rules. Each abnormality marker is like an alarm signal that needs to be immediately associated with its source. The association operation first locates the specific node identifier and device type corresponding to the abnormal entry, and then extracts the deep feature data related to the abnormal event from the matrix or parallel database. These feature data include but are not limited to: node operation deviation, which is the absolute or relative deviation between the current actual operation parameter value of the node and the predicted baseline value in the steady-state characteristic matrix, this value directly reflects the degree of abnormality; energy storage decay rate, for abnormal energy storage devices, calculate the abnormal change trend of state of charge or health index in the recent time window, the rate value indicates the emergency degree of energy storage device performance degradation or failure; irradiance mutation intensity, for photovoltaic device related abnormalities, quantify the change amplitude of environmental irradiance in a short time, the intensity value reflects the intensity of the change in light conditions, which is directly related to the sudden drop or rise of photovoltaic output.

[0085] After successfully associating and extracting the abnormal characteristics of a single node, the analysis perspective needs to be elevated from a single node to the entire system, aiming to discover abnormal association patterns across nodes. The generation of cross-node abnormal association features is achieved by analyzing the time proximity, spatial proximity, and abnormal type similarity of abnormal events occurring on different nodes. The system calculates the time difference between abnormal events to determine whether they occur almost simultaneously; checks the electrical distance of nodes in the network topology to determine whether they are physically adjacent; and compares the type feature vectors of each abnormal event to determine whether their sources are likely to be the same. For example, if multiple photovoltaic nodes in close geographical proximity report output fluctuation abnormalities caused by irradiance mutation almost at the same time, and the fluctuation direction is consistent (both sudden drops), the system will identify a strong cross-node association feature, suggesting that the area may be experiencing a cloud cover. The strength of this association feature is quantified by a comprehensive index that considers the number of common abnormal nodes, the weighted average of abnormality degree, and the tightness of time and space coupling.

[0086] When the calculated cross-node abnormal correlation feature strength value exceeds the preset threshold value set according to historical experience and system safety regulations, the system determines that the current abnormality is no longer a local isolated event, but a correlated failure or disturbance that may affect the global operation state of the system. At this time, the multi-node cooperative rescheduling strategy is automatically triggered. The core task of this strategy is to dynamically adjust the current or just formulated initial model of energy storage scheduling according to the latest transient feature matrix containing abnormal information. The rescheduling process first evaluates the impact of abnormal events on the current power balance and energy storage plan, and then quickly calculates and outputs the updated power distribution scheme and energy storage charging and discharging timing strategy. For example, in the face of the above regional photovoltaic output sudden drop, the rescheduling strategy may instruct the energy storage units in the unaffected area to discharge in advance or increase the discharge power to make up for the sudden drop in power generation. At the same time, it may appropriately increase the upper limit of power purchase from the grid and recalculate the power support allocation ratio between nodes.

[0087] After triggering the rescheduling strategy, in order to more accurately perform cooperative adjustment, the detailed information of all marked abnormalities needs to be systematically extracted from the transient feature matrix. These information is structured into three core dimensions: the unique number of abnormal nodes, which indicates which nodes deviate from the normal operation track; the specific time period of abnormal occurrence, which accurately defines the time window of disturbance start and duration; and the feature vector of abnormal type, which is a data group containing multiple attributes, used to describe the nature of the abnormality, such as abnormal code, deviation value, change trend, etc. The construction of the feature vector makes the abnormality no longer a simple Boolean flag, but has rich information that can be quantified and classified. Configuring node coupling weight coefficients for these abnormal feature vectors is a key step to achieve effective cooperative rescheduling. The coupling weight coefficient aims to quantify the mutual influence between any two nodes in operation. The assignment of this coefficient is not simply based on physical distance, but takes into account the electrical connection impedance, historical power interaction data, and the relevance of control logic. Two nodes that are closely connected and often have power interaction will have a higher coupling weight coefficient; conversely, nodes that are loosely connected or have little direct interaction will have a lower coupling weight coefficient. The weight coefficient constitutes a matrix that describes the coupling relationship between all nodes.

[0088] Based on the coupling weight coefficients of nodes, the system performs clustering analysis on all groups of nodes with abnormality. The clustering algorithm (such as hierarchical clustering or K-means variant) groups abnormal nodes according to the similarity of their abnormalities and the strength of their coupling relationships. The analysis results usually generate two abnormal clusters with typical significance. The nodes in the first abnormal cluster have abnormal characteristics that are highly consistent with photovoltaic output fluctuations caused by external environmental changes, such as sudden drops or rises in irradiance. The nodes in the second abnormal cluster have abnormal characteristics mainly caused by unexpected changes in load demand (such as the unexpected start or stop of large equipment) or fault signs of the energy storage system itself (such as abnormal increase in internal resistance or sudden decrease in available capacity). The distinction between these two abnormal clusters is not based on subjective judgment, but on a quantitative indicator calculated by the system - the node scheduling deviation. This deviation calculates the comprehensive deviation between the actual operating point of the node and the expected operating point of the original scheduling plan when the abnormality occurs. Abnormalities caused by photovoltaic fluctuations usually show unanticipated power shortages or surpluses in the scheduling deviation, while abnormalities caused by load fluctuations or energy storage failures usually show unanticipated power demand or loss of supply capacity. This fundamental difference allows the clustering algorithm to effectively distinguish between them, as shown in Table 1.

[0089] Table 1: Abnormal events and node coupling weights.

[0090]

[0091] Referring to Table 1, which shows the abnormal events and node coupling weights, the data in the table indicates that at around 10:05, photovoltaic node PV-Node-05 experienced a sudden drop in output, and it had a high coupling weight with energy storage node ESS-Node-02, meaning that this energy storage node was an important source of power support for it. Another abnormality occurred at a later time, an abnormal increase in the internal resistance of energy storage node ESS-Node-08 almost simultaneously with a sudden abnormality in load node Load-Node-03, and they had a very high coupling weight between them, strongly suggesting that they may belong to the same fault event cluster and need to be handled together. The entire implementation process embodies a complete closed loop from abnormality perception, feature correlation, global evaluation to coordinated response, ensuring that the system can quickly and coordinately respond when facing local disturbances, maintaining the stability and efficiency of overall operation.

[0092] Embodiment 5: The implementation focuses on the management and adaptive adjustment of the iterative process, which is centered on the fine classification and guidance of the solution space search behavior. The implementation of this process begins with the establishment of an independent iteration path identifier for each candidate solution generated in the iterative optimization process. This identifier is not a simple serial number, but a structured record that continuously tracks and records the update history of the solution in each iteration. The recorded information includes the fitness value of the solution in each generation, the change direction and amplitude of its parameter combination compared to the previous generation, and the way it is generated (for example, through crossover, mutation or gradient guidance). Through these historical trajectories, the state of the search path represented by the solution can be dynamically classified. The main basis for classification is to observe the trend and stability of the fitness value in multiple iterations. If the fitness value of a path shows a stable and significant trend of approaching better values in consecutive iterations, its state is classified as convergent. If the fitness value fluctuates within a certain range without forming a clear and continuous improvement direction, but also does not show significant performance degradation, its state is classified as exploratory. If the fitness value continues to deteriorate in multiple iterations, or its parameter values gradually deviate from the boundary of the feasible region, its state is classified as divergent.

[0093] Within the preset iteration evaluation window, the path update state of all active solutions is systematically identified and statistically analyzed. This evaluation window is usually set as a fixed iteration number interval, for example, every ten iterations, a comprehensive evaluation of the path state of all solutions in the current solution population is performed. The purpose of evaluation is not only to label the state of each path, but more importantly, to generate an overall understanding of the current search situation of the entire solution space, i.e., the gradient optimization form classification. This classification describes the stage and characteristics of the current optimization process from a macro perspective. Dominance of convergent form may indicate that the algorithm is performing local fine search in a promising region. Dominance of exploratory form may indicate that the algorithm is in the global exploration stage, evaluating the potential of different regions extensively. Frequent occurrence of divergent form may be a warning signal, indicating that the current search strategy may have problems, or the algorithm is trapped in an unfavorable region.

[0094] Based on the gradient optimization morphology classification, the system performs solution space search optimization management, which is a highly adaptive strategy allocation mechanism. For search paths identified as being in convergent morphology, the system configures a gradient proxy model for them. This proxy model is a lightweight, computationally efficient mathematical model that learns and fits the response characteristics of the fitness function in the vicinity of the path to predict the effects of further parameter fine-tuning. Its essence is to use a local model to predict the optimization trend direction, thereby avoiding the time-consuming complete fitness evaluation of each fine-tuning, and achieving rapid and refined iteration in promising areas. For search paths identified as being in divergent morphology, the system adopts a more cautious strategy and configures an optimization taboo window for them. This window records the parameter movement directions that lead to the continuous deterioration of the path's fitness and the related parameter regions. In the subsequent several iterations, the path's updates will be prohibited from exploring these directions and regions that have been marked as invalid or even harmful, thereby forcing it to turn to other search directions or guiding it out of the current unfavorable region.

[0095] For the most numerous exploration morphology paths, a hybrid strategy is adopted. These paths have not yet shown a clear convergent or divergent trend and contain the potential to explore new regions. Their iterative updates will combine multiple sources of information, including the macroscopic optimization trend direction extracted from the gradient proxy model of convergent paths (as heuristic guidance), the invalid directions excluded from the taboo window of divergent paths (to avoid repeating mistakes), and a certain proportion of random exploration components to maintain the diversity of the population and global exploration ability.

[0096] This classification management mechanism is ultimately translated into differentiated iteration update strategies for different forms of solutions. For solutions in the convergent form, fine iteration is performed. This means that a smaller optimization step is adopted, and small-scale and accurate parameter adjustment is strictly carried out along the direction of the local optimization trend predicted by the gradient proxy model, aiming to fully tap the potential of the current local optimal solution and seek the ultimate performance improvement. For solutions in the exploratory form, mixed direction iteration is performed. The parameter update vector is a synthesis of multiple directions, which may be partially based on the attraction of the global optimal solution, partially based on random disturbance, and partially based on the experience of other successful paths, aiming to balance the exploration and utilization capabilities of the algorithm. For solutions in the divergent form, random restart iteration is performed. This is a more aggressive strategy. When a path is confirmed to be continuously divergent and the taboo strategy is ineffective, the solution will not be fine-tuned in the current region, but its parameters will be randomly reset to a new region in the solution space that has not been fully explored and has not been marked as taboo, and the evaluation will be completely restarted, which helps the algorithm to jump out of the local optimal or ineffective region and reinvigorate the global search. Through this fine management of different strategies for different search states, the efficiency and effectiveness of the entire optimization process are significantly improved, which can quickly approach high-quality solutions and effectively avoid premature convergence or ineffective search, thereby providing a guarantee for outputting high-performance updated energy storage scheduling models.

Claims

1. A method for establishing a multi-node optical energy storage direct flexible system energy storage scheduling model, characterized in that, The method comprises the following steps: acquiring real-time operation data of each node in a photovoltaic-storage-direct-flexible system, performing multi-dimensional alignment processing on the real-time operation data to form a multi-dimensional operation data set, the real-time operation data including photovoltaic output fluctuation curves, energy storage state of charge, load demand time sequence, grid interaction power limit value and environmental irradiance variation characteristics; performing operation state mapping topology construction processing on the multi-dimensional operation data set to separate out target operation data for scheduling model construction and state indication data for scheduling abnormal response; based on the target operation data, combining scheduling priority configuration information of each node, generating an energy storage scheduling initial model for multi-node collaboration, the energy storage scheduling initial model including energy storage charging and discharging time sequence strategy and power distribution scheme; based on the state indication data, identifying node operation abnormality characteristics in the real-time operation data updating process, dynamically reconstructing the energy storage scheduling initial model according to the node operation abnormality characteristics, and outputting an updated energy storage scheduling model; the method of generating the energy storage scheduling initial model for multi-node collaboration further comprises: configuring a flexible adjustment target according to the business attributes of the photovoltaic-storage-direct-flexible system, the flexible adjustment target including a photovoltaic consumption maximization target, an energy storage life equalization target, a load peak clipping and valley filling target and a grid interaction cost optimization target; performing hierarchical identification processing on the flexible adjustment target to determine a main optimization target, a secondary optimization target and a boundary constraint target; based on the main optimization target and the secondary optimization target, filtering a key adjustment parameter set and performing scheduling influence quantization analysis on the key adjustment parameter set; the scheduling influence quantization analysis on the key adjustment parameter set comprises: extracting an adjustable parameter range from a steady state characteristic matrix; performing parameter sensitivity mapping on the main optimization target and the secondary optimization target to generate a target influence quantization gradient; constructing an interaction influence relationship matrix of all target items in the flexible adjustment target, the interaction influence relationship matrix including target synergy relationship and target conflict relationship; based on the target influence quantization gradient and the interaction influence relationship matrix, calculating the sensitivity eigenvalue of the adjustable parameter set; establishing a parameter optimization constraint set according to the sensitivity eigenvalue; the output of the updated energy storage scheduling model comprises: setting control boundaries of the key adjustment parameter set and creating an initial solution space based on the current operation state; performing fitness evaluation on solutions in the initial solution space, generating an optimization direction vector and a gradient optimization step length based on the parameter optimization constraint set and the fitness evaluation result; iteratively updating the initial solution space based on the optimization direction vector and the gradient optimization step length; outputting the updated energy storage scheduling model according to the iterative update result; the operation state mapping topology construction processing on the multi-dimensional operation data set comprises: converting the multi-dimensional operation data set after time, node and device dimension alignment into a scheduling state mapping matrix; The state characteristic decoupling is performed on the scheduling state mapping matrix to obtain a steady-state characteristic matrix and a transient-state characteristic matrix, wherein the steady-state characteristic matrix represents a normalized scheduling model infrastructure, and the transient-state characteristic matrix represents abnormal state characteristics corresponding to power overrun, energy storage attenuation, and irradiance mutation. The steady-state characteristic matrix is taken as the target operation data, and the transient-state characteristic matrix is taken as the state indication data. 2.The method of claim 1, wherein, The generation of the multi-node cooperative energy storage scheduling initial model includes: extracting photovoltaic output prediction segments, energy storage state of charge intervals, load demand peak-valley characteristics, and power grid interaction constraint boundaries from the steady-state characteristic matrix; configuring optimization weight coefficients for each node based on the scheduling priority configuration information; performing multi-objective fusion calculation on the photovoltaic output prediction segments, energy storage state of charge intervals, load demand peak-valley characteristics, and power grid interaction constraint boundaries according to the optimization weight coefficients to generate the energy storage charging and discharging timing strategy and the power distribution scheme. 3.The method of claim 2, wherein, The identification of node operation abnormal characteristics in the real-time operation data updating process based on the state indication data includes: monitoring abnormal marker entries of the transient-state characteristic matrix, and associating node operation deviation degrees, energy storage attenuation rates, and irradiance mutation intensities corresponding to the abnormal marker entries; generating cross-node abnormal association characteristics based on the node operation deviation degrees, energy storage attenuation rates, and irradiance mutation intensities; when the cross-node abnormal association characteristics exceed a preset threshold, triggering a multi-node cooperative rescheduling strategy, and updating the power distribution scheme and the energy storage charging and discharging timing strategy according to the transient-state characteristic matrix.

4. The method of claim 3, wherein the method further comprises: The triggering of the multi-node cooperative rescheduling strategy includes: extracting node numbers, abnormal time periods, and abnormal type characteristic vectors of marked abnormal nodes in the transient-state characteristic matrix; configuring node coupling weight coefficients for the characteristic vectors; performing clustering analysis on scheduling association characteristics of abnormal node groups according to the node coupling weight coefficients to generate a first abnormal cluster and a second abnormal cluster; wherein the first abnormal cluster represents scheduling abnormalities caused by photovoltaic output fluctuations, and the second abnormal cluster represents scheduling abnormalities caused by load mutations or energy storage failures; the first abnormal cluster and the second abnormal cluster are distinguished by node scheduling deviation degrees.

5. The method of claim 4, wherein the method further comprises: The iterative updating of the initial solution space based on the optimization direction vector and the gradient optimization step length includes: establishing an iteration path identifier for each solution, and classifying path states through fitness values of each iteration round; identifying path update states within a preset iteration evaluation window to generate gradient optimization mode classifications, the gradient optimization mode classifications including a convergence mode, an exploration mode, and a divergence mode; performing solution space search optimization management based on the gradient optimization mode classifications. 6.The method of claim 5, wherein, The solution space search optimization management based on the gradient optimization mode classifications includes: configuring a gradient proxy model for the convergence mode to predict an optimization trend direction through the gradient proxy model; configuring an optimization taboo window for the divergence mode to exclude invalid search directions through the optimization taboo window; Based on the optimization trend direction and the optimization taboo window, fine iteration is performed on the convergent shape solution, mixed direction iteration is performed on the exploratory shape solution, and random restart iteration is performed on the divergent shape solution.

Images