Source-network-load-storage hierarchical optimization method based on multi-objective cma-es improvement and dynamic constraint propagation

Abstract

The application discloses a source network load storage hierarchical optimization method based on multi-target CMA-ES improvement and dynamic constraint propagation, belongs to the field of new energy and energy-saving technology, and comprises the following steps: S1. dynamic constraint propagation network construction; S2. multi-target CMA-ES improved algorithm design; S3. hierarchical optimization architecture design and implementation; S4. constraint processing and safety verification mechanism; S5. multi-source information fusion and state perception; S6. historical high-quality decision trajectory acquisition and learning mechanism. The application solves technical problems in large-scale source network load storage collaborative optimization, such as dimension disaster, low constraint processing efficiency, poor multi-target convergence, difficult multi-time scale coupling and insufficient algorithm self-adaptive capacity, and realizes multi-time scale collaborative optimization scheduling of a power system.

Description

Technical Field

[0001] This invention relates to the field of new energy and energy conservation technology, specifically to a source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation. Background Technology

[0002] With the rapid increase in the penetration rate of new energy sources such as wind power and photovoltaic power in the power system, coordinated and optimized dispatching of power generation, grid, load, and storage has become a key technology to ensure the safe and stable operation of the power system. Traditional power dispatching methods are mainly designed for controllable power sources such as thermal power units, and are difficult to effectively cope with the strong randomness and volatility of new energy output. When the penetration rate of new energy exceeds 30%, the power system faces severe pressure on balance regulation, requiring the coordinated participation of diversified resources such as energy storage systems and demand response.

[0003] Existing source-grid-load-storage optimization methods mainly suffer from the following technical defects: (I) The Curse of Dimensionality is a Prominent Problem. Traditional optimization algorithms face a severe "curse of dimensionality" problem when dealing with large-scale power systems. When considering thousands of generator units, energy storage devices, and controllable loads in a provincial power grid, the dimensionality of decision variables can reach tens of thousands or more, leading to a sharp increase in optimization solution time and making it difficult to meet the real-time requirements of power dispatch. Existing methods usually adopt simplified modeling or decomposition strategies, but these methods often sacrifice the quality of the solution and cannot obtain the globally optimal solution.

[0004] (ii) Inefficient Constraint Processing. Existing technologies have significant shortcomings in constraint processing. Power system operation requires strict adherence to various physical constraints, such as grid security constraints, energy storage state of charge constraints, and unit ramping constraints. Traditional optimization algorithms typically employ penalty functions or death penalty strategies to handle these constraints, resulting in a significant waste of computational resources on evaluating infeasible solutions. Especially in scenarios with a high proportion of renewable energy, system constraints become more complex and diverse, significantly increasing the probability of constraint violations and drastically reducing the efficiency of traditional constraint processing methods.

[0005] (III) Poor Performance of Multi-Objective Optimization. Existing multi-objective optimization methods also have significant limitations in power system applications. Power dispatching needs to consider multiple conflicting objectives such as economy, safety, and environmental protection simultaneously. Traditional multi-objective algorithms, such as NSGA-II and MOEA / D, perform poorly when dealing with the unique objective space distribution of power systems. These algorithms typically assume a uniform distribution of the objective space, but the cost-emissions-safety objective space of a power system often exhibits a non-convex, discontinuous, and complex form, resulting in poor algorithm convergence and difficulty in obtaining high-quality Pareto fronts.

[0006] (iv) Limited ability to handle multi-timescale coupling. Traditional methods have limited effectiveness in handling the multi-timescale coupling characteristics of power generation, grid, load, and storage. Generation planning, grid power flow control, energy storage charging and discharging scheduling, and load demand response each have different time characteristics, ranging from minute-level to day-level. Existing methods often use a single timescale for modeling, which cannot effectively capture the coupling relationships between multiple timescales. This simplistic approach leads to a significant reduction in the effectiveness of optimization results in actual implementation.

[0007] (v) Lack of effective algorithm adaptive mechanisms. Current technologies lack effective algorithm adaptive mechanisms. The power system operating environment is complex and variable; key parameters such as load levels, renewable energy output, and equipment availability exhibit significant time-varying characteristics. Traditional optimization algorithms typically rely on historical experience or offline tuning for parameter configuration, making it difficult to adapt to real-time changes in system conditions. This deficiency is particularly pronounced under special operating conditions such as extreme weather and equipment failures. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation, so as to solve the technical problems in large-scale source-grid-load-storage collaborative optimization such as the curse of dimensionality, low constraint processing efficiency, poor multi-objective convergence, difficulty in multi-time scale coupling and insufficient algorithm adaptive ability, so as to realize multi-time scale collaborative optimization scheduling of power system.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0010] The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation includes the following steps: S1. Construct a dynamic constraint propagation network based on the physical characteristics of the power system, model the power grid power flow security constraints, energy storage state of charge constraints, and unit technical constraints as constraint propagation graphs, realize dynamic constraint reasoning and feasible domain pruning through the Waltz constraint propagation algorithm, and maintain the consistency between system-level and device-level constraints through a hierarchical structure; S2. Design an improved multi-objective CMA-ES algorithm for source-grid-load-storage characteristics, construct an economic-safety-environmental multi-objective function vector, establish an adaptive covariance matrix update mechanism based on Pareto front geometric features, and achieve parallel search through multi-objective decomposition and adaptive weight vector; S3. Establish a hierarchical coordination and optimization architecture for source-grid-load-storage, decompose the scheduling problem into a day-ahead planning layer, an hourly adjustment layer, and a real-time control layer, realize vertical interaction of information between layers through inter-layer constraint transfer functions, and achieve multi-time-scale coordination by each layer using an appropriate optimization algorithm; S4. Develop an intelligent identification and processing mechanism for power system security constraints, construct a constraint importance evaluation system based on power grid topology analysis, integrate N-1 security analysis, and establish a constraint violation risk assessment model to quantify security risks; S5. Integrate a system state perception and prediction mechanism that fuses multi-source information, construct a comprehensive information vector that includes meteorological, load, new energy and equipment status, use attention mechanism to realize dynamic weighting of multi-source information, and improve prediction accuracy through rolling prediction and error correction; S6. Construct a mechanism for acquiring and learning historical high-quality decision-making trajectories, extract the optimal scheduling trajectory of typical historical days, use trajectory clustering to guide the initialization of evolutionary algorithms, and adaptively adjust the search direction based on trajectory similarity.

[0011] Preferably, step S1 specifically includes: S11. Establish the node definitions and topology of the power system constraint propagation graph: Define constraint propagation graph ,in, For a set of nodes, ; is the edge set; the node set Nodes containing decision variables Constraint nodes and auxiliary variable nodes The decision variable node is represented as: Including generator output Energy storage charging and discharging power Controllable load adjustment amount ,in, , , These represent the indices for generator sets, energy storage devices, and controllable loads, respectively. The total number of generator sets. This represents the total number of energy storage devices. The total number of controllable loads; the constraint nodes This represents the physical constraints that the system must satisfy to operate, including power balance constraints. Line power flow constraints Energy storage SOC constraints Unit ramping constraints ,in, For route indexing; S12. Design constraint propagation rules and inference mechanisms: For power balance constraints Establish rules for dissemination:

[0012] in, To meet system load requirements, For network losses; when a generator set outputs power When changes occur, the constraint propagation mechanism automatically adjusts the output of other units or the energy storage power to maintain power balance; For energy storage SOC constraints, a recursive rule is established:

[0013] in, for t Real-time energy storage devices j The state of charge; for t- 1-time energy storage device j The state of charge; for t The energy storage charging and discharging power at any given moment; For time step; For energy storage devices The charge and discharge efficiency; Rated capacity; For the propagation rules of line power flow constraints, based on the DC power flow model:

[0014] in, For the line The trend and The phase angle of the voltage at both ends of the line. For line reactance; S13. Implementing customized improvements to the Waltz constraint propagation algorithm for power systems: Establish a constraint propagation queue A priority scheduling strategy is used to process different types of constraints, and the constraint priority is defined as follows:

[0015] in, Indicate relevant constraints; To constrain Indicators of impact on system security; To constrain Indicators affecting the system's economic performance; and These are the weighting coefficients, and , ; Define variables feasible domain :

[0016] in, and For variables The physical boundary For variables The relevant set of constraints; Indicates division variable External and Constraints A vector consisting of the remaining relevant decision variables; The iterative update formula for the constraint propagation algorithm is:

[0017] in, It represents the feasible region of other variables.

[0018] Preferably, step S2 specifically includes: S21. Constructing a multi-objective function and objective space analysis framework for power systems: Define three optimization objectives: economic objective Safety objectives and environmental protection goals The multi-objective function vector representation is as follows: ; The economic objective function is expressed as:

[0019] in, T This represents the total number of discrete time intervals within the optimized scheduling period. for t The generator set output at any given time, , , For generator sets The cost coefficient, This indicates the unit's start-up and shutdown status. For energy storage devices The cost of degradation; The security objective function is defined as follows:

[0020] in, This indicates the total number of transmission lines participating in the safety verification in the system. For the line Due to the limitations of transport capacity, For the unit Maximum output and This refers to the safety margin factor. The environmental protection objective function is expressed as:

[0021] in, For the unit The carbon emission coefficient per unit of electricity generated; S22. Design a Pareto front adaptive covariance matrix update mechanism: Based on the current non-dominated solution set Based on the geometric characteristics, calculate the distribution density index of the Pareto front:

[0022] in, Represents the non-dominated solution set Middle Desolution Another candidate non-dominated solution, This represents the Euclidean distance between two solutions in the target space; Calculate the covariance learning rate Covariance learning rate The adaptive calculation formula is:

[0023] in, For decision variables; Let be the variance-efficient selection number, where The number of individuals in the parent generation. Individual weights; Stochasticity index The calculation is as follows:

[0024] in, For the first Evolutionary path vector with step size; ,in, For expectation operator, It has a mean of 0 and a covariance matrix that is the identity matrix. The n-dimensional standard normal distribution, It is the Gamma function; Introducing the Pareto front density adjustment factor:

[0025] in, The base learning rate represents the covariance matrix. For the target distribution density, The Pareto front distribution density index corresponding to the current non-dominated solution set; The covariance matrix is ​​updated using the adaptive update formula:

[0026] in, For the current generation, The sorted search vector; S23. Implement a multi-objective decomposition and weight vector adaptive mechanism: Determine the decomposition accuracy parameters based on computing resources and accuracy requirements. :

[0027] in, Population size; Using a uniform weight vector generation strategy, for a three-objective optimization problem, the set of weight vectors... Generated using the simplex lattice method:

[0028] in, ; The ideal point is determined by the minimum value of each objective. :

[0029] in, For the current population, To solve In the The objective function value on the optimization objective; The function is scalarized using the Tchebycheff method. :

[0030] The multi-objective problem is decomposed into single-objective sub-problems, where... Weight vector The i One component; Calculate local solution density :

[0031] in, Weight vector The neighborhood, This is the optimal solution corresponding to the current weight vector. Small positive number, Weight vector Within the neighborhood The current candidate solutions corresponding to each adjacent subproblem; Calculate the adaptive learning rate :

[0032] in, The maximum number of generations; The weight vector is adaptively adjusted based on the current solution distribution.

[0033] in, After adaptive adjustment, the first i New values ​​for each weighted component; For the first i The values ​​of each weight vector component before adaptive adjustment.

[0034] Preferably, step S3 specifically includes: S31. Establish a multi-timescale hierarchical decision-making framework: Design a three-tier optimized architecture: The current planning layer has a time resolution of 1 hour and an optimization cycle of 24 hours. Hourly adjustment layer: time resolution 15 minutes, optimization cycle 4 hours; Real-time control layer: time resolution of 1 minute, optimization cycle of 15 minutes; The optimization variables at the day-ahead planning level include unit combination variables. and continuous output variables ,in, ; The optimization variable for the hourly adjustment layer is the output adjustment amount. and energy storage power adjustment ,in, Indexed for 15-minute time intervals; The control variable for the real-time control layer is the energy storage fast response power. and demand response adjustment amount ,in, Indexed by minute-level time; S32. Design inter-layer information transmission and coordination mechanisms: The optimization results from the upper layer are transmitted to the lower layer through the command issuance interface, and the execution results from the lower layer are reported back to the upper layer through the status feedback interface; the unit combination scheme output by the day-ahead planning layer. and output reference value As a constraint boundary for the hourly adjustment layer; The inter-layer constraint transfer function is defined as:

[0035] in, For hourly adjustment of the middle unit During the period The minimum allowable output boundary; For hourly adjustment of the middle unit During the period Maximum permissible output boundary; For the unit The minimum output limit of the technology; For the unit The maximum output limit of the technology; For the corresponding hourly time period, For the units in the current planning stage In the Planned baseline output for an hourly time period; and For the unit The downward and upward adjustment rates; Hours, for adjusting time intervals; S33. Implement customized design of optimization algorithms for each layer: The current planning layer employs a hybrid coding multi-objective CMA-ES algorithm, in which discrete variables... Using binary encoding, continuous variables Use real number encoding; Mutation probability Adaptive calculation is as follows:

[0036] in, This represents the total number of binary variables. The mutation operation involves bit flipping the binary portion:

[0037] in, For the first i The generator set is t Start-stop status before time period variation. For the first i The generator set is t Start-stop status after time period variation. The probability of a single binary gene bit being flipped; Gaussian transformation is applied to the real part:

[0038] in, ; The hourly adjustment layer employs a fast non-dominated sorting and elite retention strategy, defining the constraint violation degree. for:

[0039] in, For the first A quantity that violates an inequality constraint. For the first A violation of an equality constraint. and These are the number of inequality and equality constraints, respectively. The ranking criteria are based on the degree of constraint violation and the objective function value. The specific ranking criteria are as follows:

[0040] in, This is a non-dominant ranking level.

[0041] Preferably, step S4 specifically includes: S41. Establish a dynamic identification system for the N-1 security constraints of the power grid: Define the set of critical transmission sections It is determined based on power grid topology analysis and historical fault data; For each line Calculate the system vulnerability index based on the failure conditions:

[0042] in, For the line Cross-section after fault The trend As the importance weight of the cross section, cross-section The permissible stability limits under safe operating conditions; The mathematical expression for the N-1 safety constraint is:

[0043] in, For all transmission lines, The maximum allowable vulnerability index; Safety constraint verification using fast DC power flow:

[0044] in, Inject power vectors into nodes. Here is the nodal susceptance matrix. This is the voltage phase angle vector; S42. Design a dynamic update mechanism for the safe operation boundary of the energy storage system: Establish an energy storage health status perception model to assess battery health. The update formula is:

[0045] in, For energy storage devices j exist t Battery health status at any time; For energy storage devicesj exist t Battery health at time -1; This is due to calendar aging loss; This is due to cyclic aging losses; The formula for calculating the cyclic aging loss is as follows:

[0046] in, Power index, J / mol is the activation energy. J / (mol·K) is the gas constant. Battery temperature; The formula for calculating the calendar aging loss is as follows:

[0047] in, Calendar aging factor, J / mol is the activation energy for calendar aging. The SOC impact index; Update security boundaries based on health status:

[0048] in, and For energy storage devices The baseline SOC boundary; S43. Construct a constraint violation risk quantification and early warning system: Define a constraint violation probability prediction model, considering the uncertainties in renewable energy output and load forecasting. For power balance constraints, the violation probability is calculated as follows:

[0049] in, For power balance tolerance, This is the system's rated load; Assuming load forecasting error New energy prediction error ,in , , If the total output of new energy sources is considered, then the distribution of power imbalance is as follows:

[0050] The risk level of constraint violation is classified as follows:

[0051] in, To constrain the probability of violation; Computational system safety margin index :

[0052] in, For thermal stability margin, where, For the line n The actual transmission active power flow, For the line n Thermal stability limit; This is the voltage stability margin, where, For nodes i The actual voltage amplitude, For nodes i The reference voltage, This represents the maximum allowable voltage deviation of the system. For frequency stability margin, where, This is the current actual frequency of the system. For system reference frequency, This represents the maximum allowable frequency deviation of the system.

[0053] Preferably, step S5 specifically includes: S51. Construct a power system state awareness framework based on multimodal data fusion: Establish a comprehensive information vector Includes meteorological data Load characteristics New energy power output and device status :

[0054] Meteorological data vector ,in, For ambient temperature, Relative humidity, For wind speed, Solar irradiance; Load characteristic vector Includes the load power of each node, where, For nodes i The load power, The total number of nodes; New energy power output vector Including photovoltaic and wind power output, among which, For photovoltaic power station i contribute, For wind farm j contribute, This represents the total number of new energy power stations; Device state vector It includes the energy storage's state of charge and health status, among which, For energy storage j The state of charge, In good health; S52. Design a multi-source information weight allocation algorithm based on attention mechanism: An attention weighting calculation model is established to dynamically evaluate the contribution of different information sources in system state prediction; the attention score calculation formula is as follows:

[0055] in, and The first The and the first The hidden state vector of each information source , and The parameter matrix of the attention mechanism, This is the bias vector for the attention mechanism; Normalize the attention weights using the softmax function:

[0056] in, The total number of information sources. For the first The information source for the first Attention scores calculated from each information source; Generate the fused state feature vector:

[0057] S53. Establish a rolling forecast and online error correction mechanism: Using the sliding window prediction method, the window length is... Hourly, predicted step size Hour; The prediction model uses a long short-term memory network with a hidden layer dimension of [missing information]. The state update equation is:

[0058] in, , , These are the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, Output in hidden state For the sigmoid function, Element-wise product; weight matrix Recursive weight matrix Bias vector ; for t Input at any time, for t The hidden state output of the LSTM at time -1 This represents the cell state at the previous moment. Online error correction is performed using Kalman filtering:

[0059] in, For prior state estimation, For posterior state estimation, This is the state transition matrix; To estimate the error covariance matrix a priori, The posterior estimation error covariance matrix; Here is the Kalman gain matrix. The covariance of the new information; To combine t The posterior state estimate after correction of the observations at time step. for t The actual observed state vector at time t; for t The posterior estimation error covariance matrix at time step; The observation matrix and process noise covariance are given. Observation noise covariance ,in , .

[0060] Preferably, step S6 specifically includes: S61. Establish a high-quality decision-making trajectory extraction method based on historical data: A mixed-integer linear programming (MILP) model is used to solve for the optimal scheduling scheme on typical historical days. This solution is then used as training data to generate high-quality decision trajectories. The historical data-driven MILP model is expressed as follows:

[0061] The constraints include power balance, unit technical constraints, energy storage operation constraints, and network security constraints. For the unit i The power generation cost coefficient, For start-up and shutdown costs; Principal component analysis is used to extract features from high-dimensional decision vectors. Projected onto a low-dimensional feature space:

[0062] in, Principal component weight matrix, The number of principal components to be retained; The evaluation indicators for a high-quality trajectory are defined by comprehensively considering economic efficiency, safety, and environmental friendliness:

[0063] in, Total operating cost, Total carbon emissions and As the baseline value; S62. Initialization strategy for evolutionary algorithms guided by decision-making trajectories: The extracted high-quality decision trajectories are used to guide the initial population generation and search direction setting of the CMA-ES algorithm: Population initialization employs trajectory clustering and sampling methods:

[0064] in, For the first The decision trajectory corresponding to each cluster center To initialize the noise standard deviation; The initialization of the covariance matrix is ​​based on the covariance structure of the historical trajectories:

[0065] in, The average of the historical trajectories. The number of trajectories; S63. Constructing a trajectory similarity evaluation and adaptive learning weight mechanism: Considering load patterns, renewable energy output patterns, and equipment availability factors, a similarity metric between the current operating status and historical trajectories is defined: System state vector Includes standardized load levels Proportion of new energy power output and the proportion of available energy storage capacity :

[0066] in:

[0067] Similarity is calculated as follows:

[0068] in, For similarity bandwidth parameters, This is the current system operating state vector. For the first k The state vector corresponding to each historical high-quality decision trajectory; Learning weights are assigned based on similarity:

[0069] Dynamically adjust the trajectory learning rate based on trajectory quality. :

[0070] in, For the first Quality evaluation of historical trajectories; The formula for adjusting the search direction guided by the trajectory is:

[0071] in, For the first The search center of the era For the first k The decision vector corresponding to a historical high-quality decision trajectory.

[0072] The technological advancements achieved by this invention are as follows, thanks to the adoption of the above technical solutions.

[0073] This invention significantly improves constraint handling efficiency: Through the innovative design of a dynamic constraint propagation network, this invention models the physical constraints of the power system as a constraint propagation graph. The Waltz constraint propagation algorithm maintains the feasibility of solutions in real time during the evolutionary search process, transforming constraint handling from passive punishment to active prevention. This avoids the invalid evaluation of a large number of infeasible solutions in traditional penalty function methods, significantly improving the algorithm's search efficiency. The network is designed based on the physical constraint characteristics of the power system, employing a hierarchical structure to maintain consistency between system-level and device-level constraints. It accurately reflects complex physical limitations such as power flow constraints and energy storage technology constraints, ensuring the engineering feasibility of the optimization results.

[0074] This invention exhibits superior multi-objective optimization performance: Addressing the unique distribution characteristics of the power system's objective space, this invention designs an adaptive covariance matrix update mechanism based on the geometric features of the Pareto front, enabling the CMA-ES algorithm to effectively handle the non-convex and discontinuous distribution of economic, safety, and environmental objectives. By dynamically sensing the Pareto front distribution density index, the covariance learning rate and search step size are adaptively adjusted to ensure the population remains uniformly distributed across the entire Pareto front. This design not only enhances the diversity of knowledge but also strengthens the decision-maker's flexibility in choosing between different objective trade-offs.

[0075] This invention offers effective multi-timescale coordination: The hierarchical optimization architecture of this invention effectively solves the technical challenge of coupled optimization across multiple timescales of power generation, grid, load, and storage. By decomposing the complex power system scheduling problem into three sub-problems at different timescales—day-ahead planning (24h / 1h), hourly adjustment (4h / 15min), and real-time control (15min / 1min)—the optimization algorithms at each level can focus on decision optimization within a specific time range, avoiding the computational complexity explosion caused by mixed optimization of variables at different timescales in traditional methods. Inter-layer constraint transfer functions ensure the consistency and executability of the optimization scheme across different timescales.

[0076] This invention significantly improves computational efficiency: the combination of the constraint propagation mechanism and the improved CMA-ES algorithm significantly reduces the computational complexity of large-scale power system optimization. The constraint propagation network reduces the search space through feasible region pruning, and the improved covariance matrix update strategy improves the convergence speed of the algorithm. The synergistic effect of these two mechanisms enables the algorithm to solve large-scale optimization problems with thousands of decision variables within a practical timeframe. The initialization strategy guided by historical high-quality decision trajectories further accelerates the algorithm's convergence.

[0077] This invention enhances security capabilities: Through an intelligent constraint identification and processing mechanism, it can dynamically identify and prioritize critical constraints with the greatest impact on system security. A constraint importance evaluation system based on power grid topology analysis ensures that the optimization algorithm, while pursuing economic efficiency and environmental friendliness, always prioritizes system security. The integrated N-1 security analysis function enables the optimized scheme to maintain the safe and stable operation of the system even in the event of equipment failure. The dynamic update mechanism for the energy storage safety operation boundary effectively extends the lifespan of energy storage devices.

[0078] This invention exhibits strong adaptability and robustness: its multi-source information fusion mechanism and adaptive parameter adjustment strategy enable the optimization algorithm to effectively cope with complex changes in the power system operating environment. The attention-based dynamic weighting of multi-source information automatically adjusts the contribution of each information source according to the actual operating state. The deep learning-driven state awareness system accurately captures multi-dimensional information such as load changes, new energy output fluctuations, and equipment status changes. Rolling prediction and Kalman filtering error correction mechanisms ensure the system's rapid response to changes in the operating environment. Historical trajectory similarity evaluation and adaptive learning weight mechanisms enable the algorithm to continuously learn from historical experience. Attached Figure Description

[0079] Figure 1 This is a flowchart of the present invention; Figure 2 This is a framework diagram of the present invention. Detailed Implementation

[0080] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0081] A hierarchical optimization method for power system, grid, load, and storage based on improved multi-objective CMA-ES and dynamic constraint propagation is proposed. Implemented using a linear logic architecture, it first constructs a dynamic constraint propagation network to intelligently handle the physical constraints of the power system. Second, it develops an improved multi-objective CMA-ES algorithm to adapt to the objective space characteristics of the power system. Third, it establishes a hierarchical optimization framework to coordinate multi-timescale decisions. Next, it designs constraint handling and security verification mechanisms to ensure the feasibility of the solution. Then, it integrates a state awareness and prediction system to provide accurate input information. Finally, it constructs a historical decision trajectory learning mechanism to accelerate algorithm convergence, achieving complete optimized scheduling functionality. Figures 1 to 2 As shown, it includes the following steps: S1. Construction of dynamic constraint propagation network.

[0082] A dynamic constraint propagation network based on the physical characteristics of the power system is constructed. Power flow security constraints, energy storage state of charge constraints, and unit technical constraints are modeled as constraint propagation graphs. The Waltz constraint propagation algorithm is used to achieve dynamic constraint reasoning and feasible region pruning. A hierarchical structure maintains the consistency of system-level and device-level constraints. This network can identify and eliminate infeasible solutions in real time during the evolution process, transforming constraint handling from passive punishment to active prevention, significantly reducing the ineffective search space. The constraint propagation network adopts a hierarchical structure design. The upper layer handles system-level constraints such as power balance and reserve capacity, while the lower layer handles device-level constraints such as unit output limits and energy storage state of charge (SOC) boundaries. Consistency maintenance of constraints across the entire system is achieved through inter-layer information transmission.

[0083] This step specifically includes: S11. Establish the node definitions and topology of the power system constraint propagation graph: Define constraint propagation graph ,in, For a set of nodes, ; is the set of edges; is the set of nodes Nodes containing decision variables Constraint nodes and auxiliary variable nodes ; Decision variable nodes are represented as Including generator output Energy storage charging and discharging power Controllable load adjustment amount ,in, , , These represent the indices for generator sets, energy storage devices, and controllable loads, respectively. The total number of generator sets. This represents the total number of energy storage devices. Total number of controllable loads; constraint nodes This represents the physical constraints that the system must satisfy to operate, including power balance constraints. Line power flow constraints Energy storage SOC constraints Unit ramping constraints ,in, This is a route index.

[0084] S12. Design constraint propagation rules and inference mechanisms: For power balance constraints Establish rules for dissemination:

[0085] in, To meet system load requirements, For network losses; when a generator set outputs power When changes occur, the constraint propagation mechanism automatically adjusts the output of other units or the energy storage power to maintain power balance; For energy storage SOC constraints, a recursive rule is established:

[0086] in, for t Real-time energy storage devices j The state of charge; for t- 1-time energy storage device j The state of charge; for t The energy storage charging and discharging power at any given moment; For time step; For energy storage devices The charge and discharge efficiency; This is the rated capacity.

[0087] For the propagation rules of line power flow constraints, based on the DC power flow model:

[0088] in, For the line The trend and The phase angle of the voltage at both ends of the line. This refers to the line reactance.

[0089] S13. Implementing customized improvements to the Waltz constraint propagation algorithm for power systems: The constraint propagation network adopts a hierarchical structure design. The upper layer processes system-level constraints, including power balance constraints and reserve capacity constraints, while the lower layer processes equipment-level constraints, including unit output limit constraints and energy storage SOC boundary constraints. Consistent maintenance of constraints across the entire system is achieved through inter-layer information transmission.

[0090] Establish a constraint propagation queue A priority scheduling strategy is used to process different types of constraints, and the constraint priority is defined as follows:

[0091] in, Indicate relevant constraints; To constrain Indicators of impact on system security; To constrain Indicators affecting the system's economic performance; and These are the weighting coefficients, and , ; Define variables feasible domain :

[0092] in, and For variables The physical boundary For variables The relevant set of constraints; Indicates division variable External and Constraints A vector consisting of the remaining relevant decision variables; The iterative update formula for the constraint propagation algorithm is:

[0093] in, It represents the feasible region of other variables.

[0094] S2. Design of an improved multi-objective CMA-ES algorithm.

[0095] An improved multi-objective CMA-ES algorithm targeting the characteristics of power system generation, grid, load, and storage is designed. Through analysis of the target space characteristics of the power system, a multi-objective function vector of economy, safety, and environmental protection is constructed. An adaptive covariance matrix update mechanism based on the geometric features of the Pareto front is established, enabling the evolutionary search to effectively adapt to the non-convex distribution characteristics of the economic, safety, and environmental protection objectives. The improved algorithm introduces a covariance matrix adjustment strategy based on the geometric features of the Pareto front, dynamically adjusting the search direction and step size according to the distribution density and convergence degree of the current non-dominated solution set, ensuring that the population remains uniformly distributed across the entire Pareto front. Simultaneously, a multi-objective decomposition and aggregation mechanism is established, decomposing the complex multi-objective problem into multiple single-objective sub-problems, and achieving parallel search of different target regions through adaptive adjustment of the weight vector.

[0096] This step specifically includes: S21. Constructing a multi-objective function and objective space analysis framework for power systems: Define three optimization objectives: economic objective Safety objectives and environmental protection goals The multi-objective function vector representation is as follows: ; The economic objective function is expressed as:

[0097] in, T This represents the total number of discrete time intervals within the optimized scheduling period. for t The generator set output at any given time, , , For generator sets The cost coefficient, This indicates the unit's start-up and shutdown status. For energy storage devices The cost of degradation; The security objective function is defined as follows:

[0098] in, This indicates the total number of transmission lines participating in the safety verification in the system. For the line Due to the limitations of transport capacity, For the unit Maximum output and This refers to the safety margin factor. The environmental protection objective function is expressed as:

[0099] in, For the unit The carbon emission coefficient per unit of electricity generated.

[0100] S22. Design a Pareto front adaptive covariance matrix update mechanism: Based on the current non-dominated solution set Based on the geometric characteristics, calculate the distribution density index of the Pareto front:

[0101] in, Represents the non-dominated solution set Middle Desolution Another candidate non-dominated solution, This represents the Euclidean distance between two solutions in the target space; Calculate the covariance learning rate Covariance learning rate The adaptive calculation formula is:

[0102] in, For decision variables; Let be the variance-efficient selection number, where The number of individuals in the parent generation. Individual weights; Stochasticity index The calculation is as follows:

[0103] in, For the first Evolutionary path vector with step size; ,in, For expectation operator, It has a mean of 0 and a covariance matrix that is the identity matrix. The n-dimensional standard normal distribution, It is the Gamma function; Introducing the Pareto front density adjustment factor:

[0104] in, The base learning rate represents the covariance matrix. For the target distribution density, The Pareto front distribution density index corresponding to the current non-dominated solution set; The covariance matrix is ​​updated using the adaptive update formula:

[0105] in, For the current generation, This is the sorted search vector.

[0106] S23. Implement a multi-objective decomposition and weight vector adaptive mechanism: A multi-objective decomposition and aggregation mechanism is established to decompose complex multi-objective problems into multiple single-objective sub-problems, and parallel search of different target regions is achieved through adaptive adjustment of weight vectors.

[0107] Determine the decomposition accuracy parameters based on computing resources and accuracy requirements. :

[0108] in, Population size; Using a uniform weight vector generation strategy, for a three-objective optimization problem, the set of weight vectors... Generated using the simplex lattice method:

[0109] in, ; The ideal point is determined by the minimum value of each objective. :

[0110] in, For the current population, To solve In the The objective function value on the optimization objective; The function is scalarized using the Tchebycheff method. :

[0111] The multi-objective problem is decomposed into single-objective sub-problems, where... Weight vector The i One component; Calculate local solution density :

[0112] in, Weight vector The neighborhood, This is the optimal solution corresponding to the current weight vector. Small positive number, Weight vector Within the neighborhood The current candidate solutions corresponding to each adjacent subproblem; Calculate the adaptive learning rate :

[0113] in, The maximum number of generations; The weight vector is adaptively adjusted based on the current solution distribution.

[0114] in, After adaptive adjustment, the first i New values ​​for each weighted component; For the first i The values ​​of each weight vector component before adaptive adjustment.

[0115] S3. Layered optimization architecture design and implementation.

[0116] A hierarchical coordinated optimization architecture for source-grid-load-storage is established, decomposing the scheduling problem into a day-ahead planning layer, an hourly adjustment layer, and a real-time control layer, with each layer employing different time resolutions and optimization algorithms. Vertical information exchange between layers is achieved through inter-layer constraint transfer functions, and each layer uses an adapted optimization algorithm to achieve multi-timescale coordination.

[0117] This step specifically includes: S31. Establish a multi-timescale hierarchical decision-making framework: Design a three-tier optimized architecture: The current planning layer has a time resolution of 1 hour and an optimization cycle of 24 hours. Hourly adjustment layer: time resolution 15 minutes, optimization cycle 4 hours; Real-time control layer: time resolution of 1 minute, optimization cycle of 15 minutes; The optimization variables at the current planning level include unit combination variables. and continuous output variables ,in, ; The optimization variable for the hourly adjustment layer is the output adjustment amount. and energy storage power adjustment ,in, Indexed for 15-minute time intervals; The control variable for the real-time control layer is the energy storage fast response power. and demand response adjustment amount ,in, It is a minute-level time index.

[0118] S32. Design inter-layer information transmission and coordination mechanisms: The optimization results from the upper layer are transmitted to the lower layer through the command issuance interface, and the execution results from the lower layer are reported back to the upper layer through the status feedback interface; the unit combination scheme output by the day-ahead planning layer. and output reference value As a constraint boundary for the hourly adjustment layer; The inter-layer constraint transfer function is defined as:

[0119] in, For hourly adjustment of the middle unit During the period The minimum allowable output boundary; For hourly adjustment of the middle unit During the period Maximum permissible output boundary; For the unit The minimum output limit of the technology; For the unit The maximum output limit of the technology; For the corresponding hourly time period, For the units in the current planning stage In the Planned baseline output for an hourly time period; and For the unit The downward and upward adjustment rates; The hour is used to adjust the time interval.

[0120] S33. Implement customized design of optimization algorithms for each layer: The current planning layer employs a hybrid coding multi-objective CMA-ES algorithm, in which discrete variables... Using binary encoding, continuous variables Use real number encoding; Mutation probability Adaptive calculation is as follows:

[0121] in, This represents the total number of binary variables. The mutation operation involves bit flipping the binary portion:

[0122] in, For the first i The generator set is t Start-stop status before time period variation. For the first i The generator set is t Start-stop status after time period variation. The probability of a single binary gene bit being flipped; Gaussian transformation is applied to the real part:

[0123] in, ; The hourly adjustment layer employs a fast non-dominated sorting and elite retention strategy, defining the constraint violation degree. for:

[0124] in, For the first A quantity that violates an inequality constraint. For the first A violation of an equality constraint. and These are the number of inequality and equality constraints, respectively. The ranking criteria are based on the degree of constraint violation and the objective function value. The specific ranking criteria are as follows:

[0125] in, This is a non-dominant ranking level.

[0126] S4. Constraint handling and security verification mechanism.

[0127] This paper develops an intelligent identification and processing mechanism for power system security constraints, establishes a constraint importance evaluation system based on power grid topology analysis and power flow calculation, dynamically identifies the key constraints with the greatest impact on system security, and prioritizes high-risk constraints to improve algorithm efficiency. The mechanism integrates N-1 power grid security analysis, considers the impact of line and transformer faults on system operation, and establishes constraint propagation rules under fault scenarios to ensure that the optimization scheme maintains system security even in the event of equipment failure. Simultaneously, a constraint violation risk assessment model is established to quantify the impact of different constraint violations on system security and economy, providing priority guidance for constraint processing in the optimization algorithm.

[0128] This step specifically includes: S41. Establish a dynamic identification system for the N-1 security constraints of the power grid: Define the set of critical transmission sections It is determined based on power grid topology analysis and historical fault data; For each line Calculate the system vulnerability index based on the failure conditions:

[0129] in, For the line Cross-section after fault The trend As the importance weight of the cross section, cross-section The permissible stability limits under safe operating conditions; The mathematical expression for the N-1 safety constraint is:

[0130] in, For all transmission lines, The maximum allowable vulnerability index; Safety constraint verification using fast DC power flow:

[0131] in, Inject power vectors into nodes. Here is the nodal susceptance matrix. This is the voltage phase angle vector.

[0132] S42. Design a dynamic update mechanism for the safe operation boundary of the energy storage system: Establish an energy storage health status perception model to assess battery health. The update formula is:

[0133] in, For energy storage devices j exist t Battery health status at any time; For energy storage devices j exist t Battery health at time -1; This is due to calendar aging loss; This is due to cyclic aging losses; The formula for calculating cyclic aging loss is:

[0134] in, Power index, J / mol is the activation energy. J / (mol·K) is the gas constant. Battery temperature; The formula for calculating calendar aging loss is:

[0135] in, Calendar aging factor, J / mol is the activation energy for calendar aging. The SOC impact index; Update security boundaries based on health status:

[0136] in, and For energy storage devices The baseline SOC boundary.

[0137] S43. Construct a constraint violation risk quantification and early warning system: Establish a constraint violation risk assessment model to quantify the impact of different constraint violations on system security and economy, and provide priority guidance for constraint handling in optimization algorithms.

[0138] Define a constraint violation probability prediction model, considering the uncertainties in renewable energy output and load forecasting. For power balance constraints, the violation probability is calculated as follows:

[0139] in, For power balance tolerance, This is the system's rated load; Assuming load forecasting error New energy prediction error ,in , , If the total output of new energy sources is considered, then the distribution of power imbalance is as follows:

[0140] The risk level of constraint violation is classified as follows:

[0141] in, To constrain the probability of violation; Computational system safety margin index :

[0142] in, For thermal stability margin, where, For the line n The actual transmission active power flow, For the line n Thermal stability limit; This is the voltage stability margin, where, For nodes i The actual voltage amplitude, For nodes i The reference voltage, This represents the maximum allowable voltage deviation of the system. For frequency stability margin, where, This is the current actual frequency of the system. For system reference frequency, This represents the maximum allowable frequency deviation of the system.

[0143] S5. Multi-source information fusion and state perception.

[0144] A system state perception and prediction mechanism integrating multi-source information fusion is proposed. This mechanism combines meteorological data, historical load patterns, and renewable energy output characteristics to construct a comprehensive prediction model for the power system's operating state. Deep learning is used to process multimodal data fusion, and an attention mechanism is employed to identify the contribution of different information sources to system state prediction, dynamically adjusting information weights to improve prediction accuracy. A rolling prediction and error correction mechanism is established to continuously update the prediction model parameters based on real-time observation data, ensuring the accuracy and timeliness of system state perception and providing reliable input information for the optimization algorithm.

[0145] This step specifically includes: S51. Construct a power system state awareness framework based on multimodal data fusion: Establish a comprehensive information vector Includes meteorological data Load characteristics New energy power output and device status :

[0146] Meteorological data vector ,in, For ambient temperature, Relative humidity, For wind speed, Solar irradiance; Load characteristic vector Includes the load power of each node, where, For nodes i The load power, The total number of nodes; New energy power output vector Including photovoltaic and wind power output, among which, For photovoltaic power station i contribute, For wind farm j contribute, This represents the total number of new energy power stations; Device state vector It includes the energy storage's state of charge and health status, among which, For energy storage j The state of charge, It indicates a healthy state.

[0147] S52. Design a multi-source information weight allocation algorithm based on attention mechanism: An attention weighting calculation model is established to dynamically evaluate the contribution of different information sources in system state prediction; the attention score calculation formula is as follows:

[0148] in, and The first The and the first The hidden state vector of each information source , and The parameter matrix of the attention mechanism, This is the bias vector for the attention mechanism; Normalize the attention weights using the softmax function:

[0149] in, The total number of information sources. For the first The information source for the first Attention scores calculated from each information source; Generate the fused state feature vector:

[0150] This attention mechanism gives higher weight to the most valuable information sources for the current forecast under different operating conditions. For example, the attention weight of solar irradiance is significantly increased under clear weather conditions, and the attention weight of meteorological data is significantly increased under extreme weather conditions.

[0151] S53. Establish a rolling forecast and online error correction mechanism: Using the sliding window prediction method, the window length is... Hourly, predicted step size Hour; The prediction model uses a long short-term memory network with a hidden layer dimension of [missing information]. The state update equation is:

[0152] in, , , These are the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, Output in hidden state For the sigmoid function, Element-wise product; weight matrix Recursive weight matrix Bias vector ; for t Input at any time, for t The hidden state output of the LSTM at time -1 This represents the cell state at the previous moment. Online error correction is performed using Kalman filtering:

[0153] in, For prior state estimation, For posterior state estimation, This is the state transition matrix; To estimate the error covariance matrix a priori, The posterior estimation error covariance matrix; Here is the Kalman gain matrix. The covariance of the new information; To combine t The posterior state estimate after correction of the observations at time step. for t The actual observed state vector at time t; for t The posterior estimation error covariance matrix at time step; The observation matrix and process noise covariance are given. Observation noise covariance ,in , .

[0154] S6. Mechanism for acquiring and learning historical high-quality decision-making trajectories.

[0155] A mechanism for acquiring and learning historical high-quality decision-making trajectories is constructed. The optimal scheduling trajectory of typical historical days is extracted, trajectory clustering is used to guide the initialization of the evolutionary algorithm, and the search direction is adaptively adjusted based on trajectory similarity.

[0156] S61. Establish a high-quality decision-making trajectory extraction method based on historical data: A mixed-integer linear programming (MILP) model is used to solve for the optimal scheduling scheme on typical historical days. This solution is then used as training data to generate high-quality decision trajectories. The historical data-driven MILP model is expressed as follows:

[0157] The constraints include power balance, unit technical constraints, energy storage operation constraints, and network security constraints. For the unit i The power generation cost coefficient, For start-up and shutdown costs; Principal component analysis is used to extract features from high-dimensional decision vectors. Projected onto a low-dimensional feature space:

[0158] in, Principal component weight matrix, The number of principal components to be retained; The evaluation indicators for a high-quality trajectory are defined by comprehensively considering economic efficiency, safety, and environmental friendliness:

[0159] in, Total operating cost, Total carbon emissions and This is the baseline value.

[0160] S62. Initialization strategy for evolutionary algorithms guided by decision-making trajectories: The extracted high-quality decision trajectories are used to guide the initial population generation and search direction setting of the CMA-ES algorithm: Population initialization employs trajectory clustering and sampling methods:

[0161] in, For the first The decision trajectory corresponding to each cluster center To initialize the noise standard deviation; The initialization of the covariance matrix is ​​based on the covariance structure of the historical trajectories:

[0162] in, The average of the historical trajectories. This represents the number of trajectories.

[0163] S63. Constructing a trajectory similarity evaluation and adaptive learning weight mechanism: Considering load patterns, renewable energy output patterns, and equipment availability factors, a similarity metric between the current operating status and historical trajectories is defined: System state vector Includes standardized load levels Proportion of new energy power output and the proportion of available energy storage capacity :

[0164] in:

[0165] Similarity is calculated as follows:

[0166] in, For similarity bandwidth parameters, This is the current system operating state vector. For the first k The state vector corresponding to each historical high-quality decision trajectory; Learning weights are assigned based on similarity:

[0167] Dynamically adjust the trajectory learning rate based on trajectory quality. :

[0168] in, For the first Quality evaluation of historical trajectories; The formula for adjusting the search direction guided by the trajectory is:

[0169] in, For the first The search center of the era For the first k The decision vector corresponding to a historical high-quality decision trajectory.

[0170] This step specifically includes: The working principle of this invention is as follows: 1. The Active Constraint Processing Principle of Constraint Propagation Networks: Traditional optimization methods passively process constraints using penalty functions, wasting significant computational resources on infeasible solutions. This invention models the physical constraints of the power system as a constraint propagation graph and dynamically infers the logical relationships between constraints during the evolutionary search process using the Waltz constraint propagation algorithm. When a decision variable changes, the constraint propagation mechanism automatically adjusts relevant variables along the constraint propagation graph, actively removing infeasible regions from the search space. This ensures that the evolutionary algorithm always searches within the feasible region, fundamentally solving the efficiency problem of constraint processing.

[0171] 2. Multi-objective search principle for adaptive covariance matrix update: The economic, safety, and environmental protection objective space of the power system presents a complex, non-convex, and discontinuous form. This invention dynamically senses the geometric characteristics of the objective space by calculating the distribution density index of the Pareto front in real time. When the distribution density index is small, the covariance learning rate is automatically increased to expand the search range; when the distribution density index is large, the learning rate is decreased for a more refined search. The density adjustment factor guides the search direction to always be towards the sparse region of the Pareto front, ensuring a uniform distribution of the population in the objective space.

[0172] 3. Multi-timescale hierarchical coordination principle: Different components of a power system have vastly different time constants: power generation planning is measured in hours, energy storage regulation in minutes, and frequency response in seconds. This invention decomposes the multi-timescale coupling problem into three levels: the day-ahead planning level handles 24-hour unit combination and economic dispatch, the hourly adjustment level handles 15-minute rolling corrections, and the real-time control level handles second-level power balance. Inter-level constraint transfer functions ensure that upper-level decisions serve as constraint boundaries for lower-level decisions, and lower-level execution results are fed back to the upper level for rolling corrections, forming a closed-loop coordination mechanism of "planning-correction-execution".

[0173] 4. Attention-Driven Multi-Source Information Fusion Principle: The operating state of a power system is affected by multiple factors such as weather, load, new energy sources, and equipment status. The importance of different factors varies under different operating conditions. This invention employs an attention mechanism to weightedly fuse the hidden states of various information sources. The attention weights are dynamically calculated based on the input data, enabling the system to adaptively focus on the most relevant information sources under different operating conditions, thereby improving the accuracy and robustness of state perception.

[0174] 5. Evolutionary Learning Principle Guided by Historical Experience: Power system operation exhibits periodic patterns, and scheduling schemes for typical historical days contain rich prior knowledge. This invention extracts historical optimal scheduling trajectories through MILP and utilizes PCA dimensionality reduction and cluster analysis to extract common features, which guide the population initialization and search direction setting of the CMA-ES algorithm. During online operation, the learning weights are adaptively adjusted by calculating the similarity between the current state and historical trajectories, enabling the algorithm to continuously learn from historical experience and achieve an intelligent optimization closed loop of "offline learning - online guidance - continuous evolution."

Claims

1. A source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation, characterized in that: Includes the following steps: S1. Construct a dynamic constraint propagation network based on the physical characteristics of the power system, model the power grid power flow security constraints, energy storage state of charge constraints, and unit technical constraints as constraint propagation graphs, realize dynamic constraint reasoning and feasible domain pruning through the Waltz constraint propagation algorithm, and maintain the consistency between system-level and device-level constraints through a hierarchical structure; S2. Design an improved multi-objective CMA-ES algorithm for source-grid-load-storage characteristics, construct an economic-safety-environmental multi-objective function vector, establish an adaptive covariance matrix update mechanism based on Pareto front geometric features, and achieve parallel search through multi-objective decomposition and adaptive weight vector; S3. Establish a hierarchical coordination and optimization architecture for source-grid-load-storage, decompose the scheduling problem into a day-ahead planning layer, an hourly adjustment layer, and a real-time control layer, realize vertical interaction of information between layers through inter-layer constraint transfer functions, and achieve multi-time-scale coordination by each layer using an appropriate optimization algorithm; S4. Develop an intelligent identification and processing mechanism for power system security constraints, construct a constraint importance evaluation system based on power grid topology analysis, integrate N-1 security analysis, and establish a constraint violation risk assessment model to quantify security risks; S5. Integrate a system state perception and prediction mechanism that fuses multi-source information, construct a comprehensive information vector that includes meteorological, load, new energy and equipment status, use attention mechanism to realize dynamic weighting of multi-source information, and improve prediction accuracy through rolling prediction and error correction; S6. Construct a mechanism for acquiring and learning historical high-quality decision-making trajectories, extract the optimal scheduling trajectory of typical historical days, use trajectory clustering to guide the initialization of evolutionary algorithms, and adaptively adjust the search direction based on trajectory similarity.

2. The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation as described in claim 1, characterized in that: Step S1 specifically includes: S11. Establish the node definitions and topology of the power system constraint propagation graph: Define constraint propagation graph ,in, For a set of nodes, ; is the edge set; the node set Nodes containing decision variables Constraint nodes and auxiliary variable nodes The decision variable node is represented as: Including generator output Energy storage charging and discharging power Controllable load adjustment amount ,in, , , These represent the indices for generator sets, energy storage devices, and controllable loads, respectively. The total number of generator sets. This represents the total number of energy storage devices. The total number of controllable loads; the constraint nodes This represents the physical constraints that the system must satisfy to operate, including power balance constraints. Line power flow constraints Energy storage SOC constraints Unit ramping constraints ,in, For route indexing; S12. Design constraint propagation rules and inference mechanisms: For power balance constraints Establish rules for dissemination: in, To meet system load requirements, For network losses; when a generator set outputs power When changes occur, the constraint propagation mechanism automatically adjusts the output of other units or the energy storage power to maintain power balance; For energy storage SOC constraints, a recursive rule is established: in, for t Real-time energy storage devices j The state of charge; for t- 1-time energy storage device j The state of charge; for t The energy storage charging and discharging power at any given moment; For time step; For energy storage devices The charge and discharge efficiency; Rated capacity; For the propagation rules of line power flow constraints, based on the DC power flow model: in, For the line The trend and The phase angle of the voltage at both ends of the line. For line reactance; S13. Implementing customized improvements to the Waltz constraint propagation algorithm for power systems: Establish a constraint propagation queue A priority scheduling strategy is used to process different types of constraints, and the constraint priority is defined as follows: in, Indicate relevant constraints; To constrain Indicators of impact on system security; To constrain Indicators affecting the system's economic performance; and These are the weighting coefficients, and , ; Define variables feasible domain : in, and For variables The physical boundary For variables The relevant set of constraints; Indicates division variable External and Constraints A vector consisting of the remaining relevant decision variables; The iterative update formula for the constraint propagation algorithm is: in, It represents the feasible region of other variables.

3. The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation as described in claim 2, characterized in that: Step S2 specifically includes: S21. Constructing a multi-objective function and objective space analysis framework for power systems: Define three optimization objectives: economic objective Safety objectives and environmental protection goals The multi-objective function vector representation is as follows: ; The economic objective function is expressed as: in, T This represents the total number of discrete time intervals within the optimized scheduling period. for t The generator set output at any given time, , , For generator sets The cost coefficient, This indicates the unit's start-up and shutdown status. For energy storage devices The cost of degradation; The security objective function is defined as follows: in, This indicates the total number of transmission lines participating in the safety verification in the system. For the line Due to the limitations of transport capacity, For the unit Maximum output and This refers to the safety margin factor. The environmental protection objective function is expressed as: in, For the unit The carbon emission coefficient per unit of electricity generated; S22. Design a Pareto front adaptive covariance matrix update mechanism: Based on the current non-dominated solution set Based on the geometric characteristics, calculate the distribution density index of the Pareto front: in, Represents the non-dominated solution set Middle Desolution Another candidate non-dominated solution, This represents the Euclidean distance between two solutions in the target space; Calculate the covariance learning rate Covariance learning rate The adaptive calculation formula is: in, For decision variables; Let be the variance-efficient selection number, where The number of individuals in the parent generation. Individual weights; Stochasticity index The calculation is as follows: in, For the first Evolutionary path vector with step size; ,in, For expectation operator, It has a mean of 0 and a covariance matrix that is an identity matrix. The n-dimensional standard normal distribution, It is the Gamma function; Introducing the Pareto front density adjustment factor: in, The base learning rate represents the covariance matrix. For the target distribution density, The Pareto front distribution density index corresponding to the current non-dominated solution set; The covariance matrix is ​​updated using the adaptive update formula: in, For the current generation, The sorted search vector; S23. Implement a multi-objective decomposition and weight vector adaptive mechanism: Determine the decomposition accuracy parameters based on computing resources and accuracy requirements. : in, Population size; Using a uniform weight vector generation strategy, for a three-objective optimization problem, the set of weight vectors... Generated using the simplex lattice method: in, ; The ideal point is determined by the minimum value of each objective. : in, For the current population, To solve In the The objective function value on the optimization objective; The function is scalarized using the Tchebycheff method. : The multi-objective problem is decomposed into single-objective sub-problems, where... Weight vector The i One component; Calculate local solution density : in, Weight vector The neighborhood, This is the optimal solution corresponding to the current weight vector. Small positive number, Weight vector Within the neighborhood The current candidate solutions corresponding to each adjacent subproblem; Calculate the adaptive learning rate : in, The maximum number of generations; The weight vector is adaptively adjusted based on the current solution distribution. in, After adaptive adjustment, the first i New values ​​for each weighted component; For the first i The values ​​of each weight vector component before adaptive adjustment.

4. The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation as described in claim 3, characterized in that: Step S3 specifically includes: S31. Establish a multi-timescale hierarchical decision-making framework: Design a three-tier optimized architecture: The current planning layer has a time resolution of 1 hour and an optimization cycle of 24 hours. Hourly adjustment layer: time resolution 15 minutes, optimization cycle 4 hours; Real-time control layer: time resolution of 1 minute, optimization cycle of 15 minutes; The optimization variables at the day-ahead planning level include unit combination variables. and continuous output variables ,in, ; The optimization variable for the hourly adjustment layer is the output adjustment amount. and energy storage power adjustment ,in, Indexed for 15-minute time intervals; The control variable for the real-time control layer is the energy storage fast response power. and demand response adjustment amount ,in, Indexed by minute-level time; S32. Design inter-layer information transmission and coordination mechanisms: The optimization results from the upper layer are transmitted to the lower layer through the command issuance interface, and the execution results from the lower layer are reported back to the upper layer through the status feedback interface; the unit combination scheme output by the day-ahead planning layer. and output reference value As a constraint boundary for the hourly adjustment layer; The inter-layer constraint transfer function is defined as follows: in, For hourly adjustment of the middle unit During the period The minimum allowable output boundary; For hourly adjustment of the middle unit During the period Maximum permissible output boundary; For the unit The minimum output limit of the technology; For the unit The maximum output limit of the technology; For the corresponding hourly time period, For the units in the current planning stage In the Planned baseline output for an hourly time period; and For the unit The downward and upward adjustment rates; Hours, for adjusting time intervals; S33. Implement customized design of optimization algorithms for each layer: The current planning layer employs a hybrid coding multi-objective CMA-ES algorithm, in which discrete variables... Using binary encoding, continuous variables Use real number encoding; Mutation probability Adaptive calculation is as follows: in, This represents the total number of binary variables. The mutation operation involves bit flipping the binary portion: in, For the first i The generator set is t Start-stop status before time period variation. For the first i The generator set is t Start-stop status after time period variation. The probability of a single binary gene bit being flipped; Gaussian transformation is applied to the real part: in, ; The hourly adjustment layer employs a fast non-dominated sorting and elite retention strategy, defining the constraint violation degree. for: in, For the first A quantity that violates an inequality constraint. For the first A violation of an equality constraint. and These are the number of inequality and equality constraints, respectively. The ranking criteria are based on the degree of constraint violation and the objective function value. The specific ranking criteria are as follows: in, This is a non-dominant ranking level.

5. The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation as described in claim 4, characterized in that: Step S4 specifically includes: S41. Establish a dynamic identification system for the N-1 security constraints of the power grid: Define the set of critical transmission sections It is determined based on power grid topology analysis and historical fault data; For each line Calculate the system vulnerability index based on the failure conditions: in, For the line Cross-section after fault The trend As the importance weight of the cross section, cross-section The permissible stability limits under safe operating conditions; The mathematical expression for the N-1 safety constraint is: in, For all transmission lines, The maximum allowable vulnerability index; Safety constraint verification using fast DC power flow: in, Inject power vectors into nodes. Here is the nodal susceptance matrix. This is the voltage phase angle vector; S42. Design a dynamic update mechanism for the safe operation boundary of the energy storage system: Establish an energy storage health status perception model to assess battery health. The update formula is: in, For energy storage devices j exist t Battery health status at any time; For energy storage devices j exist t Battery health at time -1; This is due to calendar aging loss; Losses due to cyclic aging; The formula for calculating the cyclic aging loss is as follows: in, Power index, J / mol is the activation energy. J / (mol·K) is the gas constant. Battery temperature; The formula for calculating the calendar aging loss is as follows: in, Calendar aging factor, J / mol is the activation energy for calendar aging. The SOC impact index; Update security boundaries based on health status: in, and For energy storage devices The baseline SOC boundary; S43. Construct a constraint violation risk quantification and early warning system: Define a constraint violation probability prediction model, considering the uncertainties in renewable energy output and load forecasting. For power balance constraints, the violation probability is calculated as follows: in, For power balance tolerance, This is the system's rated load; Assuming load forecasting error New energy prediction error ,in , , If the total output of new energy sources is considered, then the distribution of power imbalance is as follows: The risk level of constraint violation is classified as follows: in, To constrain the probability of violation; Computational system safety margin index : in, For thermal stability margin, where, For the line n The actual transmission active power flow, For the line n Thermal stability limit; This is the voltage stability margin, where, For nodes i The actual voltage amplitude, For nodes i The reference voltage, This represents the maximum allowable voltage deviation of the system. For frequency stability margin, where, This is the current actual frequency of the system. For system reference frequency, This represents the maximum allowable frequency deviation of the system.

6. The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation as described in claim 5, characterized in that: Step S5 specifically includes: S51. Construct a power system state awareness framework based on multimodal data fusion: Establish a comprehensive information vector Includes meteorological data Load characteristics New energy power output and device status : Meteorological data vector ,in, For ambient temperature, Relative humidity, For wind speed, Solar irradiance; Load characteristic vector Includes the load power of each node, where, For nodes i The load power, The total number of nodes; New energy power output vector Including photovoltaic and wind power output, among which, For photovoltaic power station i contribute, For wind farm j contribute, This represents the total number of new energy power stations; Device state vector It includes the energy storage's state of charge and health status, among which, For energy storage j The state of charge, In good health; S52. Design a multi-source information weight allocation algorithm based on attention mechanism: An attention weight calculation model is established to dynamically evaluate the contribution of different information sources in system state prediction; the attention score calculation formula is as follows: in, and The first The and the first The hidden state vector of each information source , and The parameter matrix of the attention mechanism, This is the bias vector for the attention mechanism; Normalize the attention weights using the softmax function: in, The total number of information sources. For the first The information source for the first Attention scores calculated from each information source; Generate the fused state feature vector: S53. Establish a rolling forecast and online error correction mechanism: Using the sliding window prediction method, the window length is... Hourly, predicted step size Hour; The prediction model uses a long short-term memory network with a hidden layer dimension of [missing information]. The state update equation is: in, , , These are the forget gate, input gate, and output gate, respectively. Candidate cell state, In cellular state, Output in hidden state For the sigmoid function, Element-wise product; weight matrix Recursive weight matrix Bias vector ; for t Input at any time, for t The hidden state output of the LSTM at time -1 This represents the cell state at the previous moment. Online error correction is performed using Kalman filtering: in, For prior state estimation, For posterior state estimation, This is the state transition matrix; To estimate the error covariance matrix a priori, The posterior estimation error covariance matrix; Here is the Kalman gain matrix. The covariance of the new information; To combine t The posterior state estimate after correction of the observations at time step. for t The actual observed state vector at time t; for t The posterior estimation error covariance matrix at time step; The observation matrix and process noise covariance are given. Observation noise covariance ,in , .

7. The source-grid-load-storage hierarchical optimization method based on multi-objective CMA-ES improvement and dynamic constraint propagation as described in claim 6, characterized in that: Step S6 specifically includes: S61. Establish a high-quality decision-making trajectory extraction method based on historical data: A mixed-integer linear programming (MILP) model is used to solve for the optimal scheduling scheme on typical historical days. This solution is then used as training data to generate high-quality decision trajectories. The historical data-driven MILP model is expressed as follows: The constraints include power balance, unit technical constraints, energy storage operation constraints, and network security constraints. For the unit i The power generation cost coefficient, For start-up and shutdown costs; Principal component analysis is used to extract features from high-dimensional decision vectors. Projected onto a low-dimensional feature space: in, Principal component weight matrix, The number of principal components to be retained; The evaluation indicators for a high-quality trajectory are defined by comprehensively considering economic efficiency, safety, and environmental friendliness: in, Total operating cost, Total carbon emissions, and As the baseline value; S62. Initialization strategy for evolutionary algorithms guided by decision-making trajectories: The extracted high-quality decision trajectories are used to guide the initial population generation and search direction setting of the CMA-ES algorithm: Population initialization employs trajectory clustering and sampling methods: in, For the first The decision trajectory corresponding to each cluster center To initialize the noise standard deviation; The initialization of the covariance matrix is ​​based on the covariance structure of the historical trajectories: in, The average of the historical trajectories. The number of trajectories; S63. Constructing a trajectory similarity evaluation and adaptive learning weight mechanism: Considering load patterns, renewable energy output patterns, and equipment availability factors, a similarity metric between the current operating status and historical trajectories is defined: System state vector Includes standardized load levels Proportion of new energy power output and the proportion of available energy storage capacity : in: Similarity is calculated as follows: in, For similarity bandwidth parameters, This is the current system operating state vector. For the first k The state vector corresponding to each historical high-quality decision trajectory; Learning weights are assigned based on similarity: Dynamically adjust the trajectory learning rate based on trajectory quality. : in, For the first Quality evaluation of historical trajectories; The formula for adjusting the search direction guided by the trajectory is: in, For the first The search center of the era For the first k The decision vector corresponding to a historical high-quality decision trajectory.

Images