Multi-target operation optimization method and system for optical storage micro-grid

Abstract

The invention discloses a multi-target operation optimization method and system for an optical storage micro-grid, and relates to the technical field of new energy power system optimization, and the method comprises the steps: carrying out the collection and digital twinning mapping to form a health semantic graph as a state base; calculating an equilibrium priority index on the health semantic graph in a self-adaptive clustering manner to generate a health weight; dynamic weights are fused, charge and discharge power proportions are distributed according to clusters, and health constraint, aging suppression and load response collaboration are achieved; inputting the updated electric quantity vector and the predicted aging trajectory into a multi-target scheduling solver, and synchronously optimizing energy consumption cost, carbon emission estimation and frequency deviation penalty to form a planned power sequence; when the sudden change of the electricity price or the load is detected, the solver agent model is called to quickly recalculate the plan and write the plan back to the health semantic graph to ensure that the control strategy evolves along with the scene. Through mutual feedback and control interlocking of each layer of information, the overall gain of power grid frequency difference convergence and comprehensive operation cost reduction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy power system optimization, in particular to a multi-objective operation optimization method and system for a light-storage micro-grid. BACKGROUND

[0002] In the micro-grid scenario of real-time interaction between photovoltaic power generation, energy storage and load, the system operator hopes to reduce energy cost, reduce carbon emissions and maintain frequency stability at the same time, and all this relies on the response of the energy storage cluster to fluctuating power and price signals within seconds. With the introduction of digital twins, knowledge graphs and self-learning algorithms, micro-grids are gradually moving from static planning to holographic perception-prediction-optimization-control closed loop, but the complexity is rising sharply: the health status, temperature field distribution and aging rate of each battery group are monitored and written into the evolving semantic network to support multi-objective collaborative decision-making. This highly coupled operating background has laid the groundwork for subsequent technical challenges.

[0003] The current light-storage micro-grid is facing an insidious but extremely destructive hidden danger - the state of charge of multi-cluster lithium-ion energy storage is discrete in the long-term charging and discharging process, which is caused by manufacturing differences, temperature rise, load profile and charging strategy. When individual clusters are overcharged or overdischarged, the local internal resistance rises rapidly, thermal-chemical side reactions are amplified, and chain aging is induced, which weakens the overall capacity and efficiency. The result is that the depreciation cost of energy storage soars, the renewable power output absorption space is squeezed, and both economic and emission reduction targets are frustrated, and the health difference of battery clusters also narrows the power margin of energy storage in primary frequency regulation, which deteriorates the frequency stability. If the scheduling algorithm still controls according to the average state, it is difficult to distinguish which clusters should be balanced first, and it is easy to issue mismatch instructions when high-frequency price jumps or load peaks arrive, further amplifying the multi-objective conflict.

[0004] Therefore, the present application provides a multi-objective operation optimization method and system for a light-storage micro-grid. SUMMARY

[0005] (I) Technical problems solved In view of the deficiencies of the prior art, the present application provides a multi-objective operation optimization method and system of a light storage micro-grid, which forms a health semantic graph as a state base through collection and digital twin mapping; generates a health weight by self-adaptive clustering calculation of an equal priority index on the health semantic graph; realizes health constraint, aging suppression and load response coordination by fusing dynamic weight and distributing charge and discharge power proportion according to clusters; inputs the updated power vector and the predicted aging trajectory into a multi-objective scheduling solver to synchronously optimize energy consumption cost, carbon emission estimation and frequency deviation penalty, and forms a planned power sequence; when a price or load mutation is detected, a solver agent model is called to quickly recalculate the plan and write back to the health semantic graph, so as to ensure that the control strategy evolves with the scene. Through mutual feedback of information and mutual locking of control, the technical problems recorded in the background art are solved.

[0006] (II) Technical solutions To achieve the above object, the present application is implemented by the following technical solutions: A multi-objective operation optimization method of a light storage micro-grid, comprising, The main control layer collects the state of charge, the alternating current impedance spectrum and the shell temperature of the energy storage cluster in real time and synchronously maps them to the digital twin; A health semantic graph is constructed and the edge weight is updated by seconds to form a state base; the node attributes are periodically extracted and the health weight is calculated by self-adaptive clustering; Based on the health weight, the state of charge and the load demand, a reinforcement learning equalization controller is used to dynamically allocate charge and discharge power, and the execution result is written back to the health semantic graph; Under the health constraint, a multi-objective genetic algorithm is used to generate a planned power sequence; when a price or load mutation is detected, a neural network agent is called to recalculate by seconds and write back to the health semantic graph.

[0007] Further, the main control layer performs denoising, time synchronization and abnormality rejection on the collected state of charge, alternating current impedance spectrum and shell temperature at the local edge node, and then uploads them to the digital twin.

[0008] Further, the digital twin synchronously simulates the shell temperature rise and internal resistance change through a thermal-electric linkage model, and compares the simulation output with the measured value after each charge and discharge cycle to automatically correct the model parameters, so as to keep the error below a set threshold and improve the prediction accuracy.

[0009] Further, the health semantic graph automatically performs node integrity and time consistency check after each edge weight update; if missing or abnormality is detected, it triggers rollback to the nearest valid snapshot and re-synchronizes the data to ensure the reliability of the state base.

[0010] Further, the self-adaptive clustering algorithm adjusts the neighborhood radius by real-time calculation of the contour coefficient; When the average profile value is lower than the safety threshold, the radius is automatically expanded or the minimum sample points are increased to maintain the balance between the compactness within the cluster and the separation between the clusters.

[0011] Further, the health weight is obtained by nonlinear mapping of the impedance growth trend and the temperature gradient; when the weight of a certain energy storage cluster is higher than a threshold value, the system raises its balance priority and shortens the power distribution period to suppress chain aging.

[0012] Further, the reinforcement learning balance controller sets a variable exploration rate, which is automatically increased when the frequency deviation increases or the aging indicator deteriorates, to jump out of the local optimum and ensure the stability and reliability of the charging and discharging power distribution.

[0013] Further, the multi-objective genetic algorithm combines non-dominated sorting and congestion distance to maintain population diversity. After the cost, carbon emission and frequency three targets meet the convergence threshold at the same time, the non-dominated solution set is output for the scheduling system to select the optimal planned power sequence.

[0014] Further, when the price or load mutation is monitored, the neural network agent outputs the adjusted planned power sequence within seconds and writes back the health semantic graph after passing the aging constraint and frequency safety check, realizing fast closed-loop response to disturbance.

[0015] A multi-objective operation optimization system of a light storage micro-grid, comprising, The acquisition mapping module acquires the state of charge, the alternating current impedance spectrum and the shell temperature of the energy storage cluster in real time and synchronously maps them to the digital twin. The semantic base module constructs a health semantic graph and updates the edge weight to form a state base every second; the node attribute is extracted periodically and adaptively clustered to calculate the health weight. The balance control module uses a reinforcement learning balance controller to dynamically distribute the charging and discharging power based on the health weight, the state of charge and the load demand, and writes the execution result back to the health semantic graph. The scheduling optimization module generates a planned power sequence under health constraints using a multi-objective genetic algorithm; when the price or load mutation is detected, a neural network agent is called to recalculate every second and write back the health semantic graph.

[0016] (Three) beneficial effects The application provides a multi-objective operation optimization method and system of a light storage micro-grid, which has the following beneficial effects: The master control layer continuously collects the state of charge, the internal resistance spectrum and the shell temperature of each cluster, and maps them into a health semantic graph by means of a digital twin, forming a real-time evolving state base, realizing the same scale coupling and traceability of multi-source data, avoiding information silos under traditional static planning, and improving the accuracy and real-time performance of health perception of the energy storage cluster.

[0017] Adaptive clustering is performed on the health semantic graph, and a balanced priority index is calculated and updated, so that each cluster obtains a comparable health weight for subsequent scheduling reference, this step innovatively introduces a spatial clustering method based on noise application density, cooperatively fuses impedance growth rate and temperature field gradient, generates a dynamic health weight, and significantly reduces the aging risk induced by the health difference between clusters.

[0018] The reinforcement learning balanced controller synthesizes the dynamic weight reflected by the voltage gradient change and the temperature rise rate to allocate the charge-discharge power proportion to each cluster, so that the health constraint, aging suppression and load response are cooperatively satisfied, this self-learning mechanism creatively balances the electro-thermal dynamics and health weight, ensures that the power distribution is adaptively reconstructed according to the decay risk, avoids mismatched instructions caused by average control, and enhances the overall efficiency of the energy storage cluster.

[0019] The updated power vector and the predicted aging trajectory are input into a multi-objective scheduling solver, and the energy cost, carbon emission estimation and frequency deviation penalty are simultaneously optimized to form a planned power sequence for the next pricing period, this optimization process innovatively uses a non-dominated sorting genetic algorithm to process a Pareto optimal solution set, coordinates multiple objective constraints, realizes the synergistic gain of economic, environmental and stable targets, and reduces the energy storage depreciation cost and carbon emission.

[0020] When a price or load mutation is detected, a solver agent model is called to quickly recalculate the plan and write back to the health semantic graph, ensuring that the control strategy continuously evolves with the scene, this second-level response mechanism creatively uses a neural network to approximate the Pareto frontier, reduces the strategy lag, and supports frequency stability maintenance and renewable power absorption under high-frequency disturbance. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 A flowchart of the multi-objective operation optimization method of the optical storage micro-grid of the present application is shown. Figure 2 A structural diagram of the multi-objective operation optimization system of the optical storage micro-grid of the present application is shown. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0023] Please refer to Figure 1 The present application provides a multi-objective operation optimization method for an optical storage micro-grid, comprising, Step one, the master layer obtains the SoC, internal resistance spectrum and shell temperature of each cluster through the sensor network in real time, maps them to the digital twin and obtains the normalized parameter vector through finite element-electrochemical coupling; in the health semantic graph, the capacity, impedance, thermal node and their relationship are created by ontology, the edge weight is updated according to the second-level event, the SoC deviation is fused and the impedance growth rate is calculated, the mapping space-time consistency is maintained through integrity verification, and the real-time evolving state foundation is formed; In the photovoltaic power generation, energy storage and real-time interaction of load, the state of charge of multi-cluster lithium ion energy storage is differentiated due to manufacturing differences, temperature rise, load profile and charging and discharging strategy, which leads to local internal resistance rise, thermal-chemical side reaction amplification, and induces chain aging, thereby weakening the overall capacity and efficiency, and further weakening the multi-objective optimization of energy cost reduction, carbon emission reduction and frequency stability maintenance.

[0024] The multi-objective operation optimization of micro-grid depends on the accurate perception of the health state of energy storage cluster to support subsequent adaptive clustering, reinforcement learning equilibrium and multi-objective scheduling decision; however, the heterogeneity and real-time evolution characteristics of multi-source data require a unified mapping mechanism to integrate the state of charge, internal resistance spectrum and shell temperature, thereby providing a traceable state foundation for the health semantic graph, ensuring that the subsequent steps can reference consistent health information.

[0025] Continuous collection of multi-source data: The master layer collects the state of charge, internal resistance spectrum and shell temperature of each lithium ion energy storage cluster in real time through the sensor network, wherein the state of charge is represented as the percentage of remaining capacity, the internal resistance spectrum is obtained by the complex impedance value sequence obtained by the alternating current impedance spectrum method, and the shell temperature is the surface thermal distribution measurement value; these parameters directly reflect the electrochemical and thermodynamic behavior of the energy storage cluster, avoiding the limitations of single indicators, ensuring that the data covers the influence of manufacturing differences, temperature rise and load profile, forming an original data set, reducing the amplification of aging risk caused by data lag, and promoting the real-time construction of the health semantic graph.

[0026] Mapping of digital twin parameters: The collected state of charge, internal resistance spectrum and shell temperature are mapped to the virtual entity using the digital twin model, and the thermal conduction process and electrochemical model parameters are calculated through finite element simulation. Specifically, the shell temperature distribution is mapped to the temperature field gradient, and the thermal conductivity is calculated by multiplying the temperature gradient, and the divergence of the heat source term caused by the internal resistance spectrum is equal to the density multiplied by the specific heat capacity multiplied by the temperature derivative with respect to time, to simulate the temperature field distribution; the internal resistance spectrum is mapped to the equivalent circuit model parameters, and the state of charge is directly synchronized to the virtual battery state. Digital twin allows the mirror simulation of physical entities, captures the coupled dynamics of temperature rise rate and impedance change, rather than static average, thereby accurately reproducing the actual behavior of the energy storage cluster, generating a standardized parameter vector, improving the comparability and traceability of the data, and promoting the accurate association of the health semantic graph nodes.

[0027] Health semantic graph node construction: Based on the mapped parameters, nodes are created in the semantic network, where each node corresponds to a cluster of energy storage and is associated with attributes: state of charge as a capacity node, internal resistance spectrum as an impedance node, and shell temperature as a thermal node. The relationship is defined by ontology, such as the internal resistance spectrum affecting the increase of shell temperature. The semantic graph supports the structured representation of multi-source information, allowing the query of the causal chain of aging rate and charge differentiation, avoiding the non-associated storage of traditional databases, and thus realizing the system integration of energy storage cluster health information. An evolving graph structure is formed, which facilitates adaptive clustering to reference health weights, ensures that the differences between clusters are explicitly encoded, and promotes the accurate calculation of subsequent balance priority indices.

[0028] Real-time evolution mechanism integration: Through an event-triggered mechanism, the edge weights of the health semantic graph are updated every second, such as adjusting the connection strength between the impedance node and the thermal node according to the change in the internal resistance spectrum. The impedance growth rate is calculated by subtracting the initial internal resistance spectrum value from the current internal resistance spectrum value and dividing by the running time. At the same time, the charge state deviation is integrated to dynamically refresh the graph. Real-time evolution captures transient changes during charging and discharging, preventing mismatched instructions under static planning, and thus maintaining the dynamic consistency of the health semantic graph. A state base is established to support the input of multi-objective scheduling solvers, which can ensure the accuracy of aging trajectory prediction and promote the continuous satisfaction of energy storage cluster health constraints.

[0029] Data consistency verification: After updating, the integrity of the health semantic graph is checked to verify whether the node attributes cover all clusters and to calibrate the spatiotemporal consistency of the mapping parameters. This avoids semantic breaks caused by acquisition noise or mapping errors, maintains the reliability of the graph, and thus guarantees the integrity of the health semantic graph as a state base. The robustness of the state base is strengthened, which facilitates adaptive clustering to provide reliable health semantic graph input and promotes the overall stability of multi-objective operation optimization.

[0030] In use, through continuous collection and digital twin mapping, a health semantic graph integrating state of charge, internal resistance spectrum, and shell temperature is constructed, serving as a state base for health differentiation perception of energy storage clusters in a photovoltaic energy storage microgrid. This ensures that the subsequent balance priority index calculation can directly reference uniform semantic nodes, thereby supporting aging suppression and multi-objective coordination.

[0031] Step two, real-time extraction of semantic graph node attributes forms a high-dimensional vector, and DBSCAN adaptive clustering is used to automatically determine the cluster center. The balance priority index is obtained by exponentially fusing the impedance growth rate and temperature gradient. A time window is set to cycle the clustering and write back the health weight to the node. The compactness and separation are evaluated by the silhouette coefficient, and the neighborhood radius is adjusted automatically, so that each energy storage cluster obtains accurate health weights.

[0032] In the photovoltaic power generation, energy storage and real-time interaction of load, the state of charge of multi-cluster lithium ion energy storage is differentiated by manufacturing differences, temperature rise, load profile and charging and discharging strategy, which leads to the increase of local resistance, the amplification of thermal-chemical side reactions, and the induction of chain aging, thereby weakening the overall capacity and efficiency, and further weakening the multi-objective optimization of energy cost reduction, carbon emission reduction and frequency stability maintenance.

[0033] The multi-objective operation optimization of the photovoltaic energy storage microgrid depends on the cluster analysis of the health state of the energy storage cluster to generate comparable health weight to support balanced scheduling and aging suppression; however, the heterogeneity and dynamic evolution of the nodes in the health semantic graph require an adaptive mechanism to divide the similarity between clusters, thereby calculating the balanced priority index to ensure that the reinforcement learning balanced controller can reference unified weight information.

[0034] Node extraction from health semantic graph: The node attributes of each energy storage cluster are extracted from the health semantic graph, including the capacity node, the impedance node associated with the internal resistance spectrum, and the thermal node associated with the shell temperature, forming a multi-dimensional feature vector. These node attributes are directly derived from digital twin mapping, capturing the coupling relationship between state of charge, internal resistance spectrum and shell temperature, avoiding analysis bias of isolated parameters, and ensuring that the cluster input covers the influence of aging rate and temperature rise. A standardized feature set can be generated to promote accurate division of health differences between clusters by adaptive clustering, supporting subsequent calculation of balanced priority index.

[0035] Adaptive cluster initialization: The extracted multi-dimensional feature vector is initially divided by using a density-based spatial clustering method based on noise application, which automatically determines the number of cluster centers and adapts to the node distribution of the health semantic graph through the minimum sample point and neighborhood radius parameters. The density-based spatial clustering method based on noise application can handle non-convex clusters and noise points, and is suitable for non-uniform distribution caused by manufacturing differences in the health semantic graph, rather than a fixed number of clusters, thereby accurately identifying subgroups with discrete health states. Health cluster groups can be initially formed, reducing misclassification caused by fixed parameters and improving the robustness of balanced priority index calculation.

[0036] Balanced priority index calculation: Based on the clustering results, the equilibrium priority index of each energy storage cluster is calculated, and the impedance growth rate and temperature field gradient are fused using an exponential function, where the impedance growth rate is obtained by subtracting the initial internal resistance spectrum value from the current internal resistance spectrum value and dividing by the running time, and the temperature field gradient is represented by the spatial derivative of the shell temperature. The exponential term of the exponential function is the decay coefficient of the negative impedance growth rate multiplied by the impedance growth rate plus the amplification coefficient of the temperature field gradient multiplied by the spatial derivative of the shell temperature. The exponential function captures the nonlinear effects of aging coupling, avoids the amplification effect of linear weighting that ignores the thermal-chemical side reactions, and ensures that the index reflects the priority balancing needs of the energy storage cluster, thereby accurately quantifying the differences in cluster health weights and promoting the allocation of power by the reinforcement learning balancing controller. This avoids mismatched instructions under multi-objective conflict and improves the coordination efficiency of energy storage clusters during charging and discharging.

[0037] Health weight update mechanism: The clustering is re-executed and the equilibrium priority index is updated every fixed time interval, and the newly calculated health weight is written into the node attribute of the health semantic graph. The real-time evolution of the health semantic graph requires dynamic updating to respond to changes in voltage gradient and temperature rise rate during charging and discharging, avoiding mismatched instructions caused by static weights, thereby maintaining the timeliness of the health weight and supporting the multi-objective scheduling solver to input the predicted aging trajectory, ensuring the coordinated response of the energy storage cluster when the price jumps.

[0038] Clustering result verification: The clustering results are evaluated by the silhouette coefficient to verify the compactness within the cluster and the separation between clusters, and the neighborhood radius is adjusted to optimize the distribution of the equilibrium priority index. The silhouette coefficient provides an objective measure to prevent overfitting or underfitting of the clustering, ensuring that the health weight accurately represents the differentiation of the state of charge, thereby improving the overall accuracy of the clustering division, enhancing the reliability of adaptive clustering, and facilitating seamless data transfer from the health semantic graph to the reinforcement learning balancing controller.

[0039] In use, the health weight of each energy storage cluster is generated through adaptive clustering and equilibrium priority index calculation, serving as a quantitative basis for aging coupling suppression in the photovoltaic energy storage microgrid, ensuring that the reinforcement learning balancing controller can directly reference these weights, thereby supporting adaptive reconstruction of power distribution and multi-objective coordination.

[0040] Step three, the reinforcement learning balancing controller reads the health weight and monitors the voltage gradient and temperature rise rate, and fuses them into a dynamic weight vector. Taking the SoC vector, health weight, and load demand as the state, and the energy storage cluster charging and discharging power ratio as the action, the MDP is constructed based on historical data, and the reward function is composed of health satisfaction, load matching, and aging penalty. The action value is updated through Q-learning iteration and the optimal charging and discharging distribution is output, and the result is fed back to the health semantic graph in real time. In photovoltaic power generation, energy storage and real-time interaction of load, the state of charge of multi-cluster lithium ion energy storage is differentiated by manufacturing differences, temperature rise, load profile and charging and discharging strategy, which leads to the increase of local resistance, the amplification of thermal-chemical side reactions, and the induction of chain aging, thereby weakening the capacity and efficiency of the whole group, and further affecting the multi-objective optimization of energy cost reduction, carbon emission reduction and frequency stability maintenance.

[0041] The multi-objective operation optimization of the photovoltaic energy storage microgrid relies on dynamic power distribution to integrate health weights to achieve aging inhibition and load response; however, the instantaneous fluctuation of voltage gradient change and temperature rise rate requires a self-learning mechanism to adjust the dynamic weight, so as to allocate the charging and discharging power ratio according to the cluster, and ensure that the multi-objective scheduling solver can refer to the balanced charge vector.

[0042] Dynamic weight extraction and integration: The reinforcement learning balancing controller reads the health weight from the health semantic graph, and monitors the voltage gradient change and temperature rise rate in real time, wherein the voltage gradient change is calculated by the time derivative of the voltage at the end of the energy storage cluster, and the temperature rise rate is obtained by the partial derivative of the shell temperature with respect to time. These are integrated into a comprehensive dynamic weight vector with the health weight. The health weight is derived from adaptive clustering, and the voltage gradient change and temperature rise rate reflect the real-time electrical-thermal dynamics, avoiding the risk of overcharging individual clusters by average allocation, and ensuring that the input covers the influence of aging coupling. A unified weight vector can be formed to promote the targeted allocation of charging and discharging power ratio, reduce chain aging induction, and support subsequent optimization of the multi-objective scheduling solver.

[0043] State-action space definition: The state space is defined as the combination of the current state of charge vector, health weight and load demand, and the action space is the charging and discharging power ratio of each energy storage cluster. Markov decision process is used for modeling, wherein the transition probability is estimated based on historical charging and discharging data; Markov decision process supports sequence decision-making in uncertain environment, which is suitable for photovoltaic fluctuation and price signal response, rather than deterministic planning, so as to capture the synergy of health constraints and load response; a learnable decision-making framework is established to facilitate the design of reward function, so as to ensure that the power distribution adapts to the risk of attenuation, and promotes the realization of aging inhibition.

[0044] Reward function design and optimization: The reward function is designed as a health constraint satisfaction coefficient multiplied by one minus overcharge and over-discharge deviation plus a load response benefit coefficient multiplied by a load matching rate minus an aging penalty coefficient multiplied by a predicted aging increment, wherein the health satisfaction, load benefit and aging penalty are balanced by weighted summation to avoid single item dominating the decision, and the trade-off between multiple objectives is quantified to guide the learning to converge in a collaborative direction. The reward function balances multiple objectives, avoids single optimization ignoring frequency deviation penalty, ensures that dynamic weights suppress thermal-chemical side reactions in learning, thereby accurately reflecting the impact of aging coupling on power allocation. The guide controller converges to the optimal strategy, promotes the maintenance of energy storage cluster capacity, supports the input of multi-objective scheduling solver, and improves overall operation efficiency.

[0045] Q-learning iterative update: The Q-learning algorithm is used to iteratively update the action value function, and the charge and discharge power ratio is selected by the exploration-exploitation strategy. The update formula is calculated as the current action value function plus the learning rate multiplied by the instantaneous reward plus the discount factor multiplied by the maximum action value function in the next state minus the current action value function. Based on the approximate Bellman equation, the optimal value function is gradually approached, and the knowledge is accumulated from the experience samples to adapt to environmental uncertainty. Q-learning does not require an environment model, and directly learns from experience, which is suitable for responding to price jumps on a second scale, avoids the computational complexity of gradient descent, and thus realizes real-time adjustment of dynamic weights. The realization of adaptive allocation of power ratio can ensure the collaboration of aging suppression and load response, reduce the amplification of multi-objective conflicts, and improve the stability of energy storage clusters in fluctuating scenarios.

[0046] Power allocation execution and feedback: According to the optimal action value function, the charge and discharge power ratio of each energy storage cluster is allocated, and the execution result is fed back to the health semantic graph to update the node attributes for closed-loop adjustment of dynamic weights. The feedback mechanism maintains the continuity of learning, responds to high-frequency load spikes, and avoids mismatch instructions that induce frequency deterioration, thereby strengthening the continuous satisfaction of health constraints. The completion of collaborative satisfaction can promote the accurate input of the predicted aging trajectory and support overall operation optimization.

[0047] In use, the reinforcement learning balancing controller fuses dynamic weights, allocates charge and discharge power ratios by cluster, and serves as an execution mechanism for health constraint satisfaction in the photovoltaic energy storage micro-grid, ensuring that the multi-objective scheduling solver can directly reference the updated power vector, thereby supporting cost optimization and carbon emission reduction.

[0048] Step four, construct constraints with the updated power vector and the predicted aging trajectory, integrate photovoltaic output and energy storage charge and discharge price as cost, estimate fossil reserve proportion as carbon emission, and integrate frequency deviation as penalty; within the health constraints, use NSGA-II evolution to find the Pareto frontier, select the balanced solution that minimizes cost + carbon weight + frequency weight to generate the next pricing period power sequence, and write it into the health semantic graph after simulation verification.

[0049] In photovoltaic power generation, energy storage and real-time interaction of load, the state of charge of multi-cluster lithium ion energy storage is differentiated by manufacturing differences, temperature rise, load profile and charging and discharging strategy, which leads to the increase of local internal resistance, the amplification of thermal-chemical side reactions, and the induction of chain aging, thereby weakening the overall capacity and efficiency, and further weakening the multi-objective optimization of energy cost reduction, carbon emission reduction and frequency stability maintenance.

[0050] The multi-objective operation optimization of photovoltaic energy storage micro-grid depends on the balanced state input to solve the economic, environmental and stable targets simultaneously; however, the coupling complexity of energy cost, carbon emission estimation and frequency deviation penalty requires a Pareto frontier search mechanism to form a planned power sequence, ensuring that the solver agent model can quickly recalculate when mutations occur.

[0051] Input data integration: The updated state of charge vector obtained from the reinforcement learning equilibrium controller and the aging trajectory predicted based on the health semantic graph are taken as inputs, where the updated state of charge vector represents the current remaining capacity percentage of each energy storage cluster, and the aging trajectory is obtained by the time projection of impedance growth rate and temperature field gradient. The updated state of charge vector is derived from power ratio allocation, and the aging trajectory is derived from health weight fusion, ensuring that the input reflects the dynamic state after aging suppression and avoiding ignoring the impact of chain aging on capacity weakening. A complete constraint set is constructed to promote the accurate evaluation of cost and carbon by the multi-objective scheduling solver, and to support the inclusion of frequency deviation penalty.

[0052] Multi-objective function construction: Define the multi-objective function, including the energy cost function as the price integral of photovoltaic output and energy storage charging and discharging, the carbon emission estimation function as the product of the emission factor of fossil fuel backup ratio, and the frequency deviation penalty function as the integral square error of actual frequency and nominal value, and use the non-dominated sorting genetic algorithm as the solving framework.

[0053] The non-dominated sorting genetic algorithm handles the Pareto optimal solution set, which is suitable for simultaneously optimizing conflicting targets, rather than single-target gradient methods, thereby capturing the trade-off between economic targets and emission reduction targets. The target vector is generated to facilitate the constraint integration of the predicted aging trajectory, which can ensure that the planned power sequence covers the overall demand of the next billing period.

[0054] Optimization constraints and solution: Incorporating health constraints as inequality boundaries, where aging trajectories limit the upper bound of energy storage cluster power, a weighted Pareto search of the cost-minimization subject to carbon emission estimation and frequency deviation penalty iteratively evolves the population until convergence; the Pareto search avoids target compromise deviation, sets boundaries based on updated state of charge vectors, and ensures that aging suppression does not undermine frequency stability maintenance. A non-inferior solution set is generated, supporting the selection of planned power sequences, which can promote the expansion of renewable power consumption space.

[0055] Planned power sequence formation: Selecting a balanced solution from the Pareto frontier, output the planned power sequence for the next pricing period, including the timing charge-discharge instructions for each energy storage cluster and the proportion of photovoltaic consumption, where the planned power sequence is determined by finding the minimum value of the evaluation of the energy cost function on power plus the carbon emission weight coefficient multiplied by the carbon emission estimation function plus the frequency deviation weight coefficient multiplied by the frequency deviation penalty function, the minimization of the compound target balances the coupling between multi-objectives, ensuring that the sequence responds to the limitations of the predicted aging trajectory within the feasible region; the minimization of the compound target selects the balanced characteristics of the Pareto frontier, avoiding the amplification of multi-objective conflicts caused by single-target dominance, thereby accurately generating sequences that adapt to price jumps.

[0056] The executable sequence formed facilitates the invocation of the solver agent model, ensuring scenario adaptation of the control strategy and improving the overall performance of energy storage depreciation cost control and frequency stability.

[0057] Output verification and storage: Verify the constraint satisfaction of the planned power sequence, check the aging trajectory deviation by simulating execution, and write the sequence into the health semantic graph to update the node attributes. Verification prevents mismatch instructions caused by solving errors, and based on the updated state of charge vectors, it ensures that frequency stability does not deteriorate. It can strengthen optimization reliability, promote rapid recalculation when electricity prices mutate, and support the realization of overall technical gains.

[0058] In use, the multi-objective scheduling solver synchronously optimizes input data to form a planned power sequence as the decision output for resolving multi-objective conflicts in a photovoltaic storage microgrid, ensuring that the solver agent model can directly reference these sequences, thereby supporting the reduction of policy lag and the reduction of operating costs.

[0059] Step five, real-time monitoring of electricity prices and load curves to capture mutations and trigger the solver agent; using a neural network to approximate the Pareto frontier, input the updated state of charge vector, aging trajectory, and mutation parameters, recalculate the cost, carbon emission, and frequency deviation compound target, and output the adjusted planned power sequence; compare the deviation of the original sequence and the constraint satisfaction, and write back to the health semantic graph and synchronously update the capacity and hot node weight after verification.

[0060] In photovoltaic power generation, energy storage and real-time interaction of load, the state of charge of multi-cluster lithium ion energy storage is differentiated by manufacturing differences, temperature rise, load profile and charging and discharging strategy, which leads to the increase of local resistance, the amplification of thermal-chemical side reactions, and the induction of chain aging, thereby weakening the overall capacity and efficiency, and further weakening the multi-objective optimization of energy cost reduction, carbon emission reduction and frequency stability maintenance.

[0061] The multi-objective operation optimization of the light storage micro-grid depends on the dynamic adjustment of the planned power sequence to cope with real-time disturbances and maintain coordinated decision-making; however, the instantaneous nature of price or load mutation requires a proxy mechanism to approximate the solution, thereby quickly recalculating the plan and updating the health semantic map to ensure that the scene evolution of the control strategy avoids strategy lag.

[0062] Mutation detection mechanism: By monitoring the price signal and load curve in real time, the price or load mutation is detected, where the price mutation is defined as the difference between the current price and the last period planned price exceeding the threshold, and the load mutation is defined as the deviation of the load demand from the predicted value exceeding the preset percentage. The planned power sequence is derived from the multi-objective scheduling solver, and the mutation detection is based on the input parameters of these sequences to ensure that the high-frequency price jump or load peak is responded to, rather than delayed perception, thereby capturing potential hidden dangers of multi-objective conflict amplification. Triggering a quick response can promote the call of the solver agent model and reduce the frequency deterioration caused by mismatched instructions.

[0063] Solver agent model call: The solver agent model is called as an approximate version of the multi-objective scheduling solver, using a neural network to approximate the Pareto frontier, where the input is the updated charge vector, the predicted aging trajectory and the current mutation parameter, and the output is the adjusted planned power sequence. The neural network agent accelerates the calculation, is trained on historical optimization data, avoids the time-consuming of complete non-dominated sorting genetic algorithm, and ensures the response to fluctuating power signals within seconds. An approximate optimal sequence can be generated to support the maintenance of health constraints and promote the persistence of aging suppression in mutation scenarios.

[0064] Fast recalculation process: In the solver agent model, the composite target of recalculating the energy cost, carbon emission estimation and frequency deviation penalty is calculated, the adjusted planning power sequence is determined by the mapping of the updated power vector, the predicted aging trajectory, the price mutation difference and the load mutation deviation through the neural network function, the pre-trained network layer is used to capture the nonlinear relationship between the input variables to approximate the Pareto frontier, and the iteration burden of complete optimization is avoided. The neural network function fuses the mutation parameters, avoids re-iteration of population evolution, ensures that the recalculation captures the dynamic balance of economic targets and emission reduction targets, so as to accurately respond to the instantaneous changes of the price signal. The output efficient sequence promotes the instant expansion of the renewable power consumption space, can reduce the soaring depreciation cost of energy storage, and improves the overall robustness of the microgrid under disturbance.

[0065] Recalculation result verification: The adjusted planning power sequence is verified by comparing the deviation with the original sequence and the constraint satisfaction degree, and the aging trajectory is checked whether it exceeds the boundary. Based on the predicted aging trajectory, it is ensured that the agent model does not introduce error amplification multi-objective imbalance, instead of directly executing the unchecked sequence, so as to maintain the frequency stability maintenance. The sequence reliability can be confirmed, the accuracy of the write-back health semantic graph is supported, and the closed-loop evolution of the control strategy is promoted.

[0066] Write-back health semantic graph: The adjusted planning power sequence that passes the verification is written back to the health semantic graph, and the relationship weight of the capacity node and the thermal node is updated. The health semantic graph serves as a state base, and the write-back ensures information feedback, avoids chain aging induced by strategy lag, and thus supports the continuous collection and subsequent circulation of the master layer. The scene evolution is completed, which can promote the realization of overall technical gains such as aging rate reduction and operating cost reduction.

[0067] In use, by detecting mutations and calling the solver agent model, the planning power sequence is recalculated and written back, which serves as an adaptive mechanism for disturbance response in the light storage microgrid, ensures that the control strategy evolves with the change of price or load, and thus supports real-time resolution of multi-objective conflicts and convergence of grid frequency deviation.

[0068] Please refer to Figure 2 The present application provides a multi-objective operation optimization system of a light storage microgrid, comprising, The acquisition mapping module collects the state of charge, AC impedance spectrum and shell temperature of the energy storage cluster in real time and synchronously maps them to the digital twin; The semantic base module constructs a health semantic graph and updates the edge weight to form a state base every second; the node attributes are periodically extracted for adaptive clustering to calculate the health weight; The balanced control module dynamically allocates charge and discharge power based on the health weight, state of charge and load demand using a reinforcement learning balanced controller, and writes the execution result back to the health semantic graph; The scheduling optimization module generates a planned power sequence under health constraints by using a multi-objective genetic algorithm; when a price or load mutation is detected, a neural network agent is called to perform a second-level recalculation and write back a health semantic graph.

[0069] Those skilled in the art can appreciate that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. A skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0070] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be repeated here.

[0071] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0072] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0073] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any skilled person in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A multi-objective operation optimization method for a photovoltaic and energy storage microgrid, characterized by: include, The main control layer collects the energy storage cluster's state of charge, AC impedance spectrum, and shell temperature in real time and maps them to the digital twin simultaneously; Construct a health semantic graph and update edge weights in seconds to form a state base; periodically extract node attributes and adaptively cluster to calculate health weights; A reinforcement learning balancing controller is used to dynamically allocate charging and discharging power based on health weight, state of charge, and load demand, and the execution results are written back to the health semantic graph. A multi-objective genetic algorithm is used to generate a planned power sequence under health constraints. When a sudden change in electricity price or load is detected, a neural network agent is called to recalculate the data in seconds and write it back to the health semantic graph.

2. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 1 is characterized by: The main control layer performs denoising, time synchronization and anomaly elimination on the collected state of charge, AC impedance spectrum and shell temperature at the local edge node, and then uploads them to the digital twin.

3. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 2 is characterized in that: The digital twin uses a thermal-electric linkage model to synchronously simulate the shell temperature rise and internal resistance changes, and compares the simulation output with the measured value after each charge and discharge cycle, automatically correcting the model parameters to keep the error below the set threshold and improve prediction accuracy.

4. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 3 is characterized by: The healthy semantic graph automatically performs node integrity and time consistency checks after each edge weight update; if a missing or abnormality is detected, it triggers a rollback to the most recent valid snapshot and resynchronizes the data to ensure a reliable state base.

5. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 4 is characterized in that: The adaptive clustering algorithm adjusts the neighborhood radius by calculating the silhouette coefficient in real time; When the average contour value is lower than the safety threshold, the radius is automatically expanded or the minimum sample points are increased to maintain a balance between intra-cluster compactness and inter-cluster separation.

6. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 5, characterized in that: The health weight is obtained by nonlinear mapping the impedance growth trend and temperature gradient. When the weight of a certain energy storage cluster is higher than the threshold, the system increases its balancing priority and shortens the power allocation cycle to suppress chain aging.

7. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 6 is characterized by: The reinforcement learning balancing controller sets a variable exploration rate. When it detects an increase in frequency deviation or a deterioration in aging indicators, it automatically increases the exploration rate, allowing the strategy to escape the local optimum and ensure stable and reliable charging and discharging power distribution.

8. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 7, characterized in that: Multi-objective genetic algorithm combines non-dominated sorting and crowding distance to maintain population diversity; After the three objectives of cost, carbon emissions and frequency meet the convergence threshold at the same time, a non-inferior solution set is output for the dispatching system to select the optimal planned power sequence.

9. The multi-objective operation optimization method of a photovoltaic energy storage microgrid according to claim 8, characterized in that: When the electricity price or load mutation is monitored, the neural network agent outputs the adjusted planned power sequence within seconds and writes back the health semantic graph after passing the aging constraint and frequency safety check, realizing fast closed-loop response to disturbance. 10.A multi-objective operation optimization system of a light storage micro-grid, characterized in that: Comprising, The acquisition mapping module collects the state of charge, AC impedance spectrum and shell temperature of the energy storage cluster in real time and synchronously maps them to the digital twin; The semantic base module constructs the health semantic graph and updates the edge weight to form the state base by seconds; periodically extracts node attributes and adaptively clusters to calculate health weights; The balanced control module dynamically allocates charge and discharge power based on health weights, state of charge and load demand using a reinforcement learning balanced controller, and writes the execution results back to the health semantic graph; The scheduling optimization module generates a planned power sequence under health constraints using a multi-objective genetic algorithm; when detecting electricity price or load mutation, a neural network agent is called to recalculate within seconds and write back the health semantic graph.

Images