An ai-based energy storage system dynamic risk quantification and predictive control method

Abstract

This invention provides an AI-based method for dynamic risk quantification and predictive control of energy storage systems, relating to the field of energy storage risk control technology. The method collects data on wind power output fluctuations, grid peak-shaving demand, reserve capacity, and equipment status characteristics to construct a dynamic risk topology network. This network enables risk propagation path tracking and conflict intensity accumulation calculation, outputting the grid risk field strength distribution. Based on this risk field strength distribution, a deformable control manifold is constructed within the energy storage operating space. Dynamic deformation of the manifold is achieved through rule constraints, and the optimal cooperative trajectory satisfying the minimum field strength disturbance criterion is searched on its surface. The wind power fluctuation rate, peak-shaving demand mutation rate, and reserve capacity change rate are monitored in real time. When the comprehensive coefficient of variation exceeds the deformation tolerance threshold, the risk field strength distribution is updated, and the control manifold is reshaped. This method accurately quantifies dynamic risks, enabling predictive control of energy storage systems and improving risk response capabilities and control accuracy in complex dynamic environments.

Description

Technical Field

[0001] This invention relates to the field of energy storage risk control technology, specifically to an AI-based method for dynamic risk quantification and predictive control of energy storage systems. Background Technology

[0002] With the large-scale grid connection of wind power, the core value of energy storage systems in mitigating power fluctuations, responding to peak-shaving demands, and providing backup capacity is becoming increasingly prominent. Currently, the field faces a lack of ability to dynamically quantify the risks associated with energy storage after wind power generation. Traditional methods cannot characterize the attenuation patterns during risk transmission, especially the real-time impact of energy transfer efficiency between characteristics on the accumulation of conflict intensity; they also struggle to respond to the synergistic characteristics of wind power fluctuation rates and peak-shaving demand abrupt changes, resulting in updates to risk field strength distribution lagging behind actual dynamic evolution. Therefore, this severely restricts the ability of energy storage systems to predictively avoid high-risk convergence areas and the accuracy of dynamic matching between control actions and risk disturbances. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-based method for dynamic risk quantification and predictive control of energy storage systems, which solves the problems of inaccurate dynamic risk quantification and lack of predictability in control strategies of energy storage systems in wind power integration scenarios.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] S1. By synchronously collecting wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics and equipment status characteristics through multi-source sensing units, a dynamic risk topology network is constructed. In this network, nodes represent the target conflict intensity, edges represent the risk transmission path across targets, and edge weights are determined by the coupling degree between features.

[0006] S2. Perform risk propagation path tracing and conflict intensity accumulation calculations based on the dynamic risk topology network, and output the risk field strength distribution of the power grid. The strength value of the risk field distribution decays exponentially with the risk propagation distance, and the decay rate is dynamically controlled by the current edge weights. The line direction is determined by the edge weight gradient of the dynamic risk topology network.

[0007] S3. Based on the distribution of risk field strength in the power grid, a deformable control manifold is constructed in the energy storage operation space. Its deformation is constrained by the following rules: the extreme field strength region triggers local concavity of the manifold curvature, the field line convergence region forms a saddle point of the manifold, and the field strength gradient direction controls the deflection angle of the manifold normal vector.

[0008] S4. Search for the optimal cooperative trajectory that satisfies the minimum field strength perturbation criterion on the deformable control manifold surface. The curvature change of the optimal cooperative trajectory is negatively correlated with the field strength change, and it automatically avoids the manifold saddle point region.

[0009] S5. Monitor the wind power fluctuation rate, peak demand mutation rate, and reserve capacity change rate in real time, calculate the comprehensive coefficient of variation of the three, and when the comprehensive coefficient of variation exceeds the deformation tolerance threshold of the deformable control manifold, trigger the grid risk field strength distribution update and reshape the deformable control manifold.

[0010] As a further technical solution, the process of establishing the dynamic risk topology network includes: extracting the fluctuation components of synchronously collected wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics, and equipment status characteristics to obtain the fundamental frequency oscillation modes of these four types of characteristics; calculating the real-time cross-correlation delay and energy transfer efficiency between the fundamental frequency of wind power output fluctuation and the fundamental frequencies of the other three types of characteristics, where the energy transfer efficiency is defined as the ratio of the amplitude change of peak-shaving demand, reserve capacity, and equipment status characteristics within the wind power fluctuation cycle to their input fluctuation energy; mapping each characteristic to a network node, with the node value independently calculated by the normalized energy transfer efficiency, characterizing the strength of the characteristic in absorbing wind power fluctuation risk energy; generating directed edges between nodes based on the directionality of the cross-correlation delay, pointing from the feature with the shortest delay to the feature with the longest delay, with the edge weight assigned an exponential transformation of the reciprocal of the delay, the exponential transformation enhancing the weight sensitivity of weakly coupled features. The system establishes an inherent property that the weights decay exponentially with the increase of risk transmission distance, forming a dynamic topology network with dynamic transmission attenuation effect. For fluctuation component extraction, wavelet transform is used for multi-scale decomposition to extract fluctuation components at different time scales, capturing short-term fluctuations and long-term trend changes in wind power output. In cross-correlation delay calculation, the cross-correlation function is used to calculate the correlation coefficient of two feature sequences at different delays, and the delay corresponding to the peak of the correlation coefficient is found as the real-time cross-correlation delay. Normalized energy transfer efficiency uses a linear normalization method to map the energy transfer efficiency of each feature to the [0,1] interval for unified calculation and comparison. The exponential transformation formula is: edge weight = exponential factor × reciprocal of delay. The exponential factor is preset according to the actual system scale and feature importance. The relationship between the reciprocal of delay and edge weight is analyzed. If the weight sensitivity of weakly coupled features is strengthened, a larger value is taken for the exponential factor.

[0011] As a further technical solution, the risk field strength distribution of the power grid is based on a dynamic risk topology network. Taking the wind power output fluctuation characteristic node as the starting point of the risk source, risk propagation path tracing is performed, traversing all nodes connected by directed edges originating from this starting point and recording the propagation path sequence. During the traversal, a conflict intensity accumulation calculation is performed on each node. Starting from the risk source starting point, the cumulative conflict intensity value of the node is calculated sequentially along the propagation path sequence. The output intensity value of the upstream node is multiplied by the edge weight value of the current propagation edge, and this is used as the input intensity value of the downstream node. This edge weight value is processed using an exponential decay function, specifically calculated as follows: Downstream node input intensity value = Upstream node output intensity value * Edge weight * exp(-decay rate parameter * Propagation path length increment), where the decay rate parameter is determined in real-time by the edge weight value of the current propagation edge itself. Dynamic assignment, i.e., the attenuation rate parameter = k * current edge weight value, where k is a preset proportional coefficient, determined in advance based on historical data and system characteristics, with a value range between 0.1 and 0.5; simultaneously, the direction of the risk field line of the power grid is determined by calculating the gradient vector of the edge weights between adjacent nodes in the dynamic risk topology network. The gradient direction is the steepest upward path of risk transmission, and the calculation formula is gradient vector = (downstream edge weight - upstream edge weight) / spatial step size. The spatial step size is set according to the distance between network nodes; if the distance between network nodes is large, the spatial step size takes a larger value; finally, the power grid risk field strength distribution is output, where the field strength value = cumulative conflict intensity value multiplied by the node sensitivity coefficient, and the node sensitivity coefficient is set according to the node type. Different types of nodes, such as wind power nodes and peak-shaving nodes, have different sensitivity coefficients, reflecting their different sensitivities to risk.

[0012] As a further technical solution, the logical implementation process of the deformable control manifold is as follows: Based on the field strength distribution and line direction of the power grid risk field strength distribution, a basic manifold is initialized within the energy storage operating space. The basic manifold is represented by a parametric surface, and the surface equation is: Manifold surface = basic shape function + field strength control term. The basic shape function is set according to the operating range of the energy storage system, i.e., an ellipsoid or a parabola, and the initial value of the field strength control term is 0. Dynamic deformation is applied to the basic manifold according to the field strength distribution, the extreme value region in the field strength distribution is identified, and the values ​​within the region are calculated. The difference between the field strength value of a point and the average field strength of its neighborhood is used to trigger local concavity of the manifold curvature at that point when the difference exceeds a threshold. The convergence region of field lines in the field strength distribution is detected, and the saddle point is located by calculating the convergence point of the direction vectors of adjacent field lines. The location method is that when the divergence between the direction vector of a field line at a point and the direction vectors of the field lines at surrounding points is less than a set threshold, it is determined to be a saddle point. At this location, the manifold is deformed into a saddle-shaped structure with bidirectional concave curvature. The saddle-shaped manifold surface is equal to the basic manifold surface plus a saddle-shaped modulation term, where the saddle-shaped modulation term = saddle depth coefficient * (x 2 -y 2), where x and y are coordinate parameters on the surface of the manifold, used to describe the geometry of the manifold; the saddle depth coefficient is determined according to the field strength distribution and the manifold deformation requirements. If the field strength distribution is complex and a clear saddle structure is needed to guide the trajectory to avoid risks, the saddle depth coefficient takes a larger value; according to the direction of the spatial gradient vector of the field strength distribution, the normal vector at each point of the manifold is adjusted in real time to form a deformable controllable manifold.

[0013] As a further technical solution, the rule constraint is as follows: For the extreme field strength region, the positive difference between the field strength value and the neighborhood mean is input into a piecewise smoothing curve. The piecewise smoothing curve uses cubic spline interpolation. When the difference is lower than the first threshold, a linear concavity depth is output. The linear concavity depth = linear coefficient * difference. The linear coefficient is determined by analyzing the linear relationship between the field strength difference and the concavity depth. If the trajectory curvature changes relatively gently at low field strength differences, a smaller value is taken for the linear coefficient, such as 0.5. Here, 0.5 is a moderate value that can match the trend of concavity depth change with the gradual characteristics of the actual physical field while maintaining a certain response speed. After exceeding the first threshold, an exponential growth segment is activated to accelerate the deformation response in the high field strength region, and the concavity depth increases exponentially. The concavity depth = exponential coefficient * exp(β * (difference - first threshold)). The exponential coefficient is such that if the field strength difference exceeds the first threshold, the concavity depth increases rapidly to enhance the risk response. The exponential coefficient is set to a large value, such as 1.0, because the exponential function itself has the characteristic of rapid growth. When the exponential coefficient is 1.0, the concavity depth exhibits a relatively obvious exponential growth trend after exceeding the first threshold. β is comprehensively evaluated based on the system's risk control requirements and equipment response capabilities. If the system needs to quickly respond to high field strength risks, β is set to a large value; conversely, if the system has relatively low real-time requirements for risk control, β is set to a small value. The concavity depth value is converted into the normal displacement of the manifold surface at the extreme point, forming a local concavity. For the field line convergence region, a local coordinate system is established at the detected saddle point coordinates. Its principal axis direction is determined by the principal direction vector of the converging field lines. Negative curvature constraints are applied along the principal axis, and positive curvature constraints are applied along the secondary axis. By solving the surface partial differential equation with a preset curvature target value, the manifold is forced to transform into a saddle-shaped structure with hyperbolic paraboloid characteristics. The surface partial differential equation is in the form of: Δz / Δx 2 -Δz / Δy 2 = Target value of saddle curvature, which is set according to the field strength distribution and control requirements; the target deflection angle of the normal vector at each point of the manifold is calculated in real time based on the field strength gradient direction vector.

[0014] As a further technical solution, the optimal cooperative trajectory is as follows: On the deformable controllable manifold surface, a dynamic trajectory search is established with the local field strength value of the manifold as the potential energy. An initial energy storage control command sequence is used as candidate trajectories. A genetic algorithm is employed to optimize the candidate trajectories, with a population size of 50 to 100 and an iteration count of 100 to 300. The manifold field strength value at the current point of the candidate trajectory is calculated and defined as the upper limit of the trajectory curvature constraint. It is then detected whether the trajectory point of the candidate trajectory enters the vicinity of the saddle point. If it does, a repulsion mechanism is generated with the curvature center of the saddle point as the origin. Force field; with the minimum field strength disturbance as the optimization objective, under the conditions of satisfying the upper limit of curvature constraint and repulsive force field avoidance, the Lagrange multiplier method is used to solve the negative correlation between the rate of change of trajectory curvature and the rate of change of manifold field strength. This allows the trajectory to automatically increase the radius of curvature in the region where the manifold field strength increases to reduce the control action amplitude, and decrease the radius of curvature in the region where the field strength decreases to accelerate the response. Finally, the optimal cooperative trajectory is output, which dynamically matches the curvature change with the spatial distribution of field strength and avoids the saddle point region. After the trajectory is output, it is smoothed by using cubic spline interpolation to make the trajectory smooth and continuous.

[0015] As a further technical solution, the real-time parallel acquisition of time-series data on wind power fluctuation rate, peak demand mutation rate, and reserve capacity change rate is performed, and their coefficients of variation are calculated separately within a sliding time window. The length of the sliding time window is set according to the dynamic characteristics of the system. The three coefficients of variation are fused through dynamic weighting to calculate the comprehensive coefficient of variation. Based on the current principal curvature spatial distribution of the deformable control manifold surface, the average value of the principal curvature change rate at its sampling point is calculated. When the fused comprehensive coefficient of variation exceeds the real-time deformation tolerance threshold, a global update of the power grid risk field strength distribution is triggered and the deformable control manifold is reshaped. The update cycle is set to 1 to 5 minutes to ensure the timeliness and effectiveness of the control strategy.

[0016] This invention provides an AI-based method for dynamic risk quantification and predictive control of energy storage systems, which has the following advantages:

[0017] 1. This invention constructs a dynamic risk topology network by synchronously collecting wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics, and equipment status characteristics through multi-source sensing units. This enables multi-objective coupled quantification of wind power fluctuation risks and improves the predictive perception capability of energy storage systems for hidden transmission paths.

[0018] 2. This invention drives the real-time deformation of the deformable control manifold through field strength distribution, realizing the geometric mapping between the energy storage control surface and the risk situation, thereby improving the accuracy of active avoidance in high-risk areas.

[0019] 3. This invention establishes an optimal collaborative trajectory search mechanism with a negative correlation between curvature and field strength to achieve dynamic matching between control actions and risk disturbances, thereby improving the responsiveness of energy storage in fluctuating scenarios. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0022] refer to Figure 1 This invention provides an artificial intelligence-based method for dynamic risk quantification and predictive control of energy storage systems, aiming to solve the problems of inaccurate dynamic risk quantification and lack of predictability in control strategies for energy storage systems in wind power integration scenarios. The method constructs a dynamic risk topology network by synchronously collecting wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics, and equipment status characteristics through multi-source sensing units. Based on this network, it performs operations such as risk propagation path tracking, grid risk field strength distribution calculation, deformable control manifold construction, optimal cooperative trajectory search, and real-time monitoring and updating to ensure the efficient operation of the energy storage system in complex dynamic environments.

[0023] First, wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics, and equipment status characteristics are collected synchronously through multi-source sensing units. These characteristics reflect the fluctuation of wind power, the grid's peak-shaving demand for energy storage, the availability of reserve capacity, and the health status of energy storage equipment, respectively. For example, wind power output fluctuation characteristics are obtained through power sensors installed in wind farms, grid peak-shaving demand characteristics are obtained from the grid dispatch center, reserve capacity characteristics are obtained through the energy management system of the energy storage system, and equipment status characteristics are obtained through equipment health monitoring sensors. After preprocessing, the collected data is used for risk quantification and control strategy formulation.

[0024] When constructing a dynamic risk topology network, fluctuation components are extracted from the collected features to obtain the fundamental frequency oscillation modes of four types of features. For example, the fundamental frequency oscillation mode of wind power output fluctuation features is decomposed into multiple scales through wavelet transform to extract fluctuation components at different time scales and capture short-term fluctuations and long-term trend changes in wind power output. The real-time cross-correlation delay and energy transfer efficiency of the fundamental frequency of wind power output fluctuations and the fundamental frequencies of the other three types of features are calculated. Among them, the energy transfer efficiency is defined as the ratio of the amplitude change of peak-shaving demand, reserve capacity, and equipment status features within the wind power fluctuation cycle to its input fluctuation energy. By calculating the directionality of the cross-correlation delay, directed edges between nodes are generated, and an exponential transformation is performed based on the reciprocal of the cross-correlation delay to assign edge weight values. The exponential transformation strengthens the weight sensitivity of weakly coupled features and establishes the inherent property that the weight decays exponentially with the increase of risk transmission distance, thus forming a dynamic topology network with dynamic transmission attenuation effect.

[0025] Based on a dynamic risk topology network, this method uses wind power output fluctuation characteristic nodes as the starting points of risk sources and performs risk propagation path tracing. It traverses all nodes connected by directed edges originating from this starting point and records the propagation path sequence. During the traversal, it performs conflict intensity accumulation calculations on each node. Starting from the risk source starting point, it sequentially calculates the cumulative conflict intensity value of nodes along the propagation path sequence. The output intensity value of the upstream node is multiplied by the edge weight value of the current propagation edge, and this is used as the input intensity value of the downstream node. This edge weight value is subjected to exponential decay processing, so that the cumulative conflict intensity value of each node exhibits an exponential decay characteristic as the propagation path length between it and the risk source starting point increases. Simultaneously, it determines the direction of the power grid risk field line by calculating the gradient vector of the edge weights between adjacent nodes in the dynamic risk topology network. This gradient direction indicates the steepest ascending path of risk propagation. Finally, it outputs the power grid risk field strength distribution, whose field strength value decays exponentially with the risk propagation distance, and the decay rate is dynamically controlled by the current edge weight.

[0026] Within the energy storage operation space, a deformable control manifold is constructed based on the grid risk field strength distribution. The deformation of this manifold is constrained by the following rules: extreme field strength regions trigger local concavity of the manifold curvature, convergence regions of field lines form saddle points of the manifold, and the field strength gradient direction controls the deflection angle of the manifold normal vector. Specifically, extreme regions in the field strength distribution are identified, and the difference between the field strength value at each point in the region and the average field strength of its neighborhood is calculated. When the difference exceeds a preset threshold, for example, a threshold set to 0.5, local concavity of the manifold curvature at that point is triggered, and the depth of concavity is positively correlated with the magnitude of the difference. Convergence regions of field lines in the field strength distribution are detected, and the saddle point is located by calculating the convergence point of the direction vectors of adjacent field lines. At this location, the manifold is deformed into a saddle-shaped structure with bidirectional concave curvature. Based on the spatial gradient vector direction of the field strength distribution, the normal vector at each point of the manifold is adjusted in real time so that the deflection angle of the normal vector is in the same direction as the gradient vector direction, and the deflection amplitude is proportional to the gradient magnitude, thereby obtaining a deformable control manifold.

[0027] On a deformable controllable manifold surface, the optimal cooperative trajectory that satisfies the minimum field strength disturbance criterion is searched. The curvature change of this optimal cooperative trajectory is negatively correlated with the field strength change, and it automatically avoids the saddle point region of the manifold. In specific implementation, the energy storage control command sequence is initialized as a candidate trajectory, the manifold field strength value of the current point of the candidate trajectory is calculated, and it is defined as the upper limit of the trajectory curvature constraint, which is negatively correlated with the field strength value. It is detected whether the trajectory point enters the vicinity of the saddle point. If it does, a repulsive field function is generated with the curvature center of the saddle point as the origin. The magnitude of the repulsive force increases exponentially with the distance, forcing the trajectory to deviate from the saddle point. With the minimum field strength disturbance as the optimization objective, under the conditions of satisfying the upper limit of the curvature constraint and the repulsive field avoidance, the negative correlation function between the trajectory curvature change rate and the manifold field strength change rate is solved. This allows the trajectory to automatically increase the curvature radius in the region where the manifold field strength increases to reduce the control action amplitude, and decrease the curvature radius in the region where the field strength decreases to accelerate the response. Finally, the optimal cooperative trajectory that dynamically matches the curvature change with the spatial distribution of the field strength and avoids the saddle point region is output.

[0028] Real-time parallel acquisition of time-series data on wind power fluctuation rate, peak demand mutation rate, and reserve capacity change rate; calculation of their coefficients of variation within a sliding time window, the length of which is set according to the system's dynamic characteristics, for example, 5 minutes; fusion of the three coefficients of variation using dynamic weights to calculate a comprehensive coefficient of variation, where the weight of wind power fluctuation rate is fixed at a baseline value, for example, 0.4; the weight of peak demand mutation rate is modulated by the ratio of the normalized energy transfer efficiency of the wind power output fluctuation characteristic node to the normalized energy transfer efficiency of the peak demand characteristic node; and the weight of reserve capacity change rate is directly bound to the normalized energy transfer efficiency of the reserve capacity characteristic node; based on the current principal curvature spatial distribution of the deformable control manifold surface, the average value of the principal curvature change rate at its sampling points is calculated, and the average value is defined as the deformation tolerance threshold; when the fused comprehensive coefficient of variation exceeds the real-time deformation tolerance threshold, a global update of the grid risk field strength distribution is triggered, and the deformable control manifold is reshaped, with an update cycle set to 3 minutes.

[0029] Through the above steps, this invention realizes a dynamic risk quantification and predictive control method for energy storage systems based on artificial intelligence. This method can accurately quantify the dynamic risks of energy storage systems in wind power access scenarios, and achieve predictive control of energy storage systems by constructing deformable control manifolds and searching for optimal cooperative trajectories. At the same time, through real-time monitoring and updating, it ensures that the control strategy can respond to changes in system state in a timely manner, thereby improving the risk response capability and control accuracy of energy storage systems in complex dynamic environments.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based energy storage system dynamic risk quantification and predictive control method, characterized in that, Includes the following steps: S1. By synchronously collecting wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics, and equipment status characteristics through multi-source sensing units, a dynamic risk topology network is constructed. Nodes represent the target conflict intensity, edges represent the risk transmission path across targets, and edge weights are determined by the coupling degree between features. Each feature is mapped to a network node, and the node value is independently calculated by the normalized energy transfer efficiency, characterizing the strength of the feature in absorbing wind power fluctuation risk energy. Directed edges between nodes are generated based on the directionality of cross-correlation delay: from the feature with the shortest delay to the feature with the longest delay, and the edge weight is assigned as an exponential transformation of the reciprocal of the delay. S2. Based on the dynamic risk topology network, perform risk propagation path tracing and conflict intensity accumulation calculations to output the risk field strength distribution of the power grid. The strength value of the risk field distribution decays exponentially with the risk propagation distance, and the decay rate is dynamically controlled by the current edge weight. Specifically, starting from the risk source origin, calculate the cumulative conflict intensity value of each node sequentially along the propagation path sequence. The output intensity value of the upstream node is multiplied by the edge weight value of the current propagation edge and used as the input intensity value of the downstream node. This edge weight value is subjected to exponential decay processing, so that the cumulative conflict intensity value of each node exhibits an exponential decay characteristic as the propagation path length between it and the risk source origin increases. Determine the direction of the power grid risk field line by calculating the gradient direction of the edge weights between adjacent nodes in the dynamic risk topology network. This gradient direction indicates the risk propagation direction. The steepest ascending path of the guide; the direction of the line is determined by the edge weight gradient of the dynamic risk topology network; the process of applying exponential decay to the edge weight value is as follows: when calculating the input strength value of the downstream node, the output strength value of the upstream node is multiplied by the edge weight value of the current propagation edge, and the result of the multiplication is exponentially decayed; the exponential decay is defined as: output strength = input value * exp(-decay rate parameter * propagation path length increment); where the decay rate parameter is dynamically assigned in real time by the edge weight value of the current propagation edge itself, that is, decay rate parameter = k * current edge weight value, k is a preset proportional coefficient; for propagation edges with small edge weight values, the corresponding decay rate parameter increases, so that the output strength value undergoes exponential decay on that edge; conversely, for propagation edges with large edge weight values, the decay rate parameter decreases; S3. Based on the distribution of risk field strength in the power grid, a deformable control manifold is constructed in the energy storage operation space. Its deformation is constrained by the following rules: the extreme field strength region triggers local concavity of the manifold curvature, the field line convergence region forms a saddle point of the manifold, and the field strength gradient direction controls the deflection angle of the manifold normal vector. S4. Search for the optimal cooperative trajectory that satisfies the minimum field strength perturbation criterion on the deformable control manifold surface. The curvature change of the optimal cooperative trajectory is negatively correlated with the field strength change, and it automatically avoids the manifold saddle point region. S5. Monitor the wind power fluctuation rate, peak demand mutation rate, and reserve capacity change rate in real time, calculate the comprehensive coefficient of variation of the three, and when the comprehensive coefficient of variation exceeds the deformation tolerance threshold of the deformable control manifold, trigger the grid risk field strength distribution update and reshape the deformable control manifold.

2. The method for dynamic risk quantification and predictive control of an AI-based energy storage system according to claim 1, characterized in that, The process of establishing the dynamic risk topology network includes: extracting the fluctuation components of synchronously collected wind power output fluctuation characteristics, grid peak-shaving demand characteristics, reserve capacity characteristics, and equipment status characteristics to obtain the fundamental frequency oscillation modes of the four types of characteristics; calculating the real-time cross-correlation delay and energy transfer efficiency between the fundamental frequency of wind power output fluctuation and the fundamental frequencies of the other three types of characteristics, wherein the energy transfer efficiency is defined as the ratio of the amplitude change of peak-shaving demand, reserve capacity, and equipment status characteristics within the wind power fluctuation cycle to their input fluctuation energy; the exponential transformation enhances the weight sensitivity of weakly coupled characteristics and establishes the inherent property that the weight decays exponentially with the increase of risk transmission distance, forming a dynamic topology network with dynamic transmission attenuation effect.

3. The method for dynamic risk quantification and predictive control of an AI-based energy storage system according to claim 1, characterized in that, The process of realizing the deformable control manifold is as follows: Based on the field strength distribution and line direction of the risk field strength distribution, a basic manifold is initialized in the energy storage operation space; the basic manifold is dynamically deformed according to the field strength distribution, the extreme value region in the field strength distribution is identified, and the difference between the field strength value of each point in the region and the average field strength of its neighborhood is calculated. When the difference exceeds a threshold, the local concavity of the manifold curvature at that point is triggered, and the concavity depth is positively correlated with the magnitude of the difference; the field line convergence region in the field strength distribution is detected, and the saddle point is located by calculating the convergence point of the direction vectors of adjacent field lines, and the manifold is deformed into a saddle-shaped structure with bidirectional concave curvature at that position; according to the spatial gradient vector direction of the field strength distribution, the normal vector at each point of the manifold is adjusted in real time so that the deflection angle of the normal vector is in the same direction as the gradient vector direction, and the deflection amplitude is proportional to the gradient magnitude, thus obtaining the deformable control manifold.

4. The method for dynamic risk quantification and predictive control of an AI-based energy storage system according to claim 3, characterized in that, The rules and constraints are as follows: For the extreme field strength region, the positive difference between the field strength value and the neighborhood mean is input into the piecewise smoothing curve. When the difference is lower than the first threshold, a linear concavity depth is output. When it exceeds the first threshold, an exponential growth segment is activated to accelerate the deformation response of the high field strength region. The concavity depth value is directly converted into the normal displacement of the manifold surface at the extreme point, forming a local concavity. For the field line convergence region, a local coordinate system is established at the detected saddle point coordinates. Its principal axis direction is determined by the principal direction vector of the converging field lines. Negative curvature constraints are applied along the principal axis and positive curvature constraints are applied along the secondary axis. By solving the surface partial differential equation with a preset curvature target value, the manifold is forced to deform into a saddle-shaped structure with hyperbolic paraboloid characteristics. The target deflection angle of the normal vector at each point of the manifold is calculated in real time according to the field strength gradient direction vector.

5. The method for dynamic risk quantification and predictive control of an AI-based energy storage system according to claim 1, characterized in that, The optimal cooperative trajectory is as follows: On a deformable control manifold surface, a dynamic trajectory search is established with the local field strength of the manifold as the potential energy; the energy storage control command sequence is initialized as a candidate trajectory, and the manifold field strength value at the current point of the candidate trajectory is calculated and defined as the upper limit of the trajectory curvature constraint, which is negatively correlated with the field strength value; it is detected whether the trajectory point enters the vicinity of the saddle point. If it does, a repulsive field function is generated with the curvature center of the saddle point as the origin. The magnitude of the repulsive force increases exponentially with decreasing distance, forcing the trajectory to deviate from the saddle point; with the minimum field strength disturbance as the optimization objective, under the conditions of satisfying the upper limit of the curvature constraint and the repulsive field avoidance, the negative correlation function between the trajectory curvature change rate and the manifold field strength change rate is solved, so that the trajectory automatically increases the curvature radius in the region where the manifold field strength increases to reduce the control action amplitude, and decreases the curvature radius in the region where the field strength decreases to accelerate the response. Finally, the optimal cooperative trajectory that dynamically matches the curvature change and the spatial distribution of the field strength and avoids the saddle point region is output.

6. The method for dynamic risk quantification and predictive control of an AI-based energy storage system according to claim 1, characterized in that: Real-time parallel acquisition of time-series data on wind power fluctuation rate, peak demand mutation rate, and reserve capacity change rate; calculation of their coefficients of variation within a sliding time window; fusion of the three coefficients of variation through dynamic weighting to calculate a comprehensive coefficient of variation, wherein the weight of wind power fluctuation rate is fixed at the baseline value, the weight of peak demand mutation rate is modulated by the ratio of the normalized energy transfer efficiency of the wind power output fluctuation characteristic node to the normalized energy transfer efficiency of the peak demand characteristic node, and the weight of reserve capacity change rate is directly bound to the normalized energy transfer efficiency of the reserve capacity characteristic node. Based on the current spatial distribution of principal curvature of the deformable control manifold surface, calculate the average value of the rate of change of principal curvature at its sampling points; The average value is defined as the deformation tolerance threshold; When the combined coefficient of variation after fusion exceeds the real-time deformation tolerance threshold, a global update of the power grid risk field strength distribution is triggered and the deformable control manifold is reshaped.

Images