Micro-grid multi-target rolling optimization operation method considering voltage stability and energy storage cooperation

Abstract

The invention discloses a micro-grid multi-target rolling optimization operation method considering voltage stability and energy storage cooperation, and belongs to the technical field of power system operation and control. The invention aims to solve the problems of low precision, poor robustness and high operation cost when an existing micro-grid energy management strategy is used for processing the uncertainty of renewable energy sources. The method comprises the following steps: preprocessing original photovoltaic output and meteorological data, carrying out smoothing processing by adopting an improved Akima interpolation method, screening related input characteristics through a Pearson's correlation coefficient, and decomposing the related input characteristics into different frequency band sub-signals by adopting a wavelet packet; training the LSTM prediction sub-model, and carrying out linear weighted fusion output; establishing a micro-grid system mathematical model, and taking power generation cost and environment cost as optimization targets; solving is carried out based on an improved NSGA2 algorithm, a Pareto optimal solution set is obtained by introducing an adaptive learning mechanism and a constraint satisfaction strategy, and a TOPSIS method is adopted to decide and select an optimal scheduling scheme; and executing rolling horizon optimization. The method is used for micro-grid sustainable energy management.

Description

Technical Field

[0001] This invention relates to a multi-objective rolling optimization operation method for microgrids that considers voltage stability and energy storage synergy, belonging to the field of power system operation and control technology. Background Technology

[0002] With the deepening of the global energy structure transformation, the application of renewable energy sources, represented by photovoltaics and wind power, in modern power systems continues to expand. These energy sources offer advantages such as cleanliness and sustainability, but their intermittent, volatile, and low-inertia characteristics also pose significant challenges to the stable operation and precise dispatch of the power grid. Unlike traditional synchronous generators, most renewable energy sources are connected to the system via power electronic converters, and their output is significantly affected by weather conditions, exhibiting high uncertainty. This makes it difficult for traditional dispatch methods to meet the operational requirements under conditions of high renewable energy integration.

[0003] To address the uncertainties of renewable energy, forecasting technology has become a crucial supporting tool. Existing forecasting methods are mainly divided into two categories: mechanistic models and data-driven models. Mechanistic models rely on high-precision numerical weather prediction, which is often limited by the reliability of meteorological data in practical applications. Data-driven methods, especially deep learning-based time series forecasting models (such as RNN, LSTM, GRU, and TCN), have demonstrated superior performance in extracting nonlinear time-series features. In recent years, physics-data hybrid deep learning models have further improved the robustness of forecasts. However, existing research largely focuses on network structure improvement or physical information embedding, and the in-depth mining and comprehensive utilization of the time-frequency domain characteristics of renewable energy output remains insufficient, limiting further improvements in forecast accuracy and scheduling effectiveness.

[0004] In power system dispatching, existing strategies mainly include rule-based and optimization-based methods. While rule-based strategies are simple and easy to implement, they struggle to achieve economical system operation and multi-objective coordination. Optimization-based methods solve for optimal dispatching schemes by constructing mathematical models. Robust optimization effectively handles uncertainty and has good interpretability, but it is highly conservative, often requiring relaxation of non-standard constraints during modeling, and involves linear matrix inequalities, resulting in a heavy computational burden. On the other hand, deep reinforcement learning (DRL) has attracted widespread attention due to its excellent real-time decision-making capabilities, often employing an Actor-Critic architecture for policy learning. However, DRL has significant limitations in handling hard constraints, adapting to complex power grid environments, and coordinating multi-objective optimization. Existing safe DRL methods typically use Lagrange multipliers or penalty functions to transform hard constraints into soft constraints, failing to completely avoid constraint violations; and in multi-objective processing, weighted summation is often used, ignoring the dimensional differences between objectives and the subjectivity of weight setting.

[0005] Model predictive control (MPC), with its rolling optimization and feedback correction features, has been applied to energy management in distribution networks and microgrids. However, most existing MPC studies are oriented towards a single economic objective, failing to fully consider safety constraints such as voltage quality during operation, and also failing to coordinate the optimization of economic and environmental goals under the "carbon neutrality" objective. Therefore, how to construct a multi-objective optimization scheduling strategy that can simultaneously take into account voltage safety, economic operation, and low-carbon benefits in microgrids with high renewable energy penetration has become a key problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to address the problems of low accuracy, poor robustness, and high operating costs in existing microgrid energy management strategies that include distributed renewable energy sources when dealing with the uncertainty of renewable energy. This invention provides a multi-objective rolling optimization operation method for microgrids that considers voltage stability and energy storage synergy.

[0007] The multi-objective rolling optimization operation method for microgrids that considers voltage stability and energy storage synergy, as described in this invention, includes:

[0008] Preprocessing of raw photovoltaic power output and meteorological data: The improved Akima interpolation method is used to smooth the missing data, the input features related to photovoltaic power output are screened by Pearson correlation coefficient, and wavelet packet decomposition is used to decompose the sequence data of the screened input features into multiple sub-signals of different frequency bands.

[0009] Construct a hybrid prediction model: For each frequency band sub-signal, train a corresponding LSTM prediction sub-model, where each sub-model takes the multivariate sub-signal under a specific frequency band as input and the contribution of that frequency band to photovoltaic power output as output; perform linear weighted fusion of the outputs of all sub-models to obtain the photovoltaic power output prediction result in the prediction time domain;

[0010] Constructing a multi-objective optimization model: Based on the photovoltaic power output prediction results, establish mathematical models of each device in the microgrid system, define a multi-objective function with power generation cost and environmental cost as optimization objectives, introduce key constraints, and form a multi-objective optimization model;

[0011] Multi-objective optimization solution: The multi-objective optimization model is solved based on the improved NSGA2 algorithm. Pareto optimal solution set is obtained by introducing an adaptive learning mechanism and constraint satisfaction strategy. The optimal scheduling scheme at each time step is selected from the solution set by using the TOPSIS method based on objective weights.

[0012] Perform rolling time-domain optimization: Based on the model predictive control mechanism, the optimal scheduling scheme is applied to the microgrid system, the system state is rolled to the next time step, and the above steps are repeated to achieve closed-loop optimized operation of the entire scheduling cycle.

[0013] Preferably, the specific method for smoothing missing data using the improved Akima interpolation method includes:

[0014] Improved Akima piecewise cubic Hermite interpolation is used to smooth the data:

[0015] ;

[0016] in, Represents the interpolation function. This represents the interpolation variable, and the interpolation interval is... , , , and Both represent constants.

[0017] Preferably, the specific method for screening input features related to photovoltaic output using the Pearson correlation coefficient includes:

[0018] The original dataset contains the following features: temperature, relative humidity, total radiation, diffuse radiation, precipitation, global tilt irradiance, and diffuse tilt irradiance.

[0019] The correlation between these characteristics and photovoltaic output was evaluated and screened using the Pearson correlation coefficient:

[0020] ;

[0021] in, This represents the correlation coefficient between each feature and photovoltaic output. Indicates the number of features. Indicates photovoltaic output data, Representing feature data, Indicates the first One data point, This indicates the total number of data points.

[0022] Preferably, the specific method for decomposing the sequence data from which the input features have been selected into multiple sub-signals of different frequency bands using wavelet packet decomposition includes:

[0023] The wavelet basis function is "db3", and the decomposition level is three.

[0024] Discrete wavelet packet decomposition is expressed as:

[0025] ;

[0026] in, Indicates the first wavelet level coefficient, Indicates a low-pass filter. Indicates a high-pass filter. and Both represent discrete-time indices;

[0027] The refactoring process is described as follows:

[0028] ;

[0029] The dataset was normalized to:

[0030] ;

[0031] in, This represents the normalization result. This represents the maximum value of the target normalization range, which is 1. The minimum value of the target normalization range is -1; Represents the original value. Indicates the maximum original value. This represents the minimum original value.

[0032] Preferably, the LSTM prediction sub-model specifically includes:

[0033] The loss function of LSTM is:

[0034] ;

[0035] in, , Indicates the predicted segment length. Represents the actual value. This represents the predicted value.

[0036] Preferably, the specific method for defining the multi-objective function with power generation cost and environmental cost as optimization objectives includes:

[0037] Electricity generation cost objective function for:

[0038] ;

[0039] in, Indicates time, Indicates optimization of the time domain, This represents the cost of electricity generation from distributed photovoltaic (PV) power plants. Indicates time Active power of distributed photovoltaic (PV) at that time This represents the power generation cost (WT) of a distributed wind turbine. Indicates time The active power of the distributed wind turbine (WT) at that time. and These represent the active power generation cost and reactive power generation cost of dispatchable generators (DG), respectively. Indicates time The active power exchanged with the main power grid at that time. Indicates time The reactive power exchanged with the main power grid at that time and These represent the costs of purchasing active power and reactive power from the main power grid, respectively. for , for , and These refer to the prices for active power and reactive power sold to the main power grid. for , for , and These refer to the operating costs of the energy storage device ES during charging and discharging. Indicates time The charging power of the energy storage device ES. Indicates time The discharge power of the energy storage device ES;

[0040] Environmental cost target for:

[0041] ;

[0042] in, This represents the equivalent environmental cost coefficient.

[0043] Preferably, the introduced key constraints include: power flow constraints, voltage and power balance constraints, specifically:

[0044]

[0045]

[0046] in, and They represent time respectively Time-injection node Active power and reactive power, ;

[0047] It is with nodes The set of adjacent nodes;

[0048] It is time Time node The voltage amplitude satisfies , and These are the lower and upper limits of the voltage amplitude. It represents the voltage amplitude at all nodes;

[0049] It is time Time node The voltage amplitude;

[0050] and They represent the nodes respectively To the node Conductivity and susceptance;

[0051] It is time Time from node To the node The phase difference;

[0052] Indicates the number of dispatchable generators. Indicates the amount of energy stored. Indicates the number of photovoltaic power plants. Indicates the number of wind farms. Indicates the quantity that meets the requirements. Indicates the serial number of the power generation equipment / energy storage / load. Indicates network loss. This represents the load power of each node;

[0053] Indicates the sequence number is The active power exchanged between the dispatchable generator (DG) and the main power grid. Indicates the sequence number is The discharge power of the energy storage device ES. Indicates the sequence number is The active power of distributed photovoltaic (PV) Indicates the sequence number is The active power of the distributed wind turbine WT Indicates the sequence number is The charging power of the energy storage device ES.

[0054] Preferably, the multi-objective optimization model is specifically as follows:

[0055] ;

[0056] in, It is an optimization variable. It is a feasible region;

[0057] The optimization problem is constrained by equipment operation constraints, power flow constraints, and power balance constraints.

[0058] Preferably, the specific method for solving the multi-objective optimization model based on the improved NSGA2 algorithm includes:

[0059] The two operators in the NSGA2 algorithm are fast non-dominated sorting and crowd distance calculation;

[0060] The specific process of quick nondominated sort is as follows:

[0061] Determine the set of individuals in the first ranking order:

[0062] set up For population groups, In order to be in Individuals ( The set of all individuals under control. for medium dominance The total number of individuals;

[0063] for Each individual in ,implement:

[0064] For each individual Perform initialization operations. , ;

[0065] for Each individual in ,implement:

[0066] if Then will be subject to Dominant Individual deposit middle, ;

[0067] if So, for each dominant entity discovered individuals, Increase by 1, ;

[0068] if So, individual sorting level And these individuals are stored in the first-level sort. In the set;

[0069] Determine the set of individuals for the remaining ranking levels:

[0070] set up It is the first Let a set of individuals ordered by level be... ;

[0071] when At that time, execute:

[0072] ,in, This represents a temporary variable used to store information about the next level of individuals.

[0073] for Each individual in ,implement:

[0074] for Each individual in ,implement:

[0075] ;if ,So , , Represents an individual The sorting order;

[0076] Update the sorting level number: ;

[0077] Obtain the A set of individuals ordered by hierarchy ;

[0078] All individuals are assigned to different levels of sorted sets;

[0079] The specific process for calculating congestion distance is as follows:

[0080] For each sorted set ,implement:

[0081] ,in, Represents a sorted set The number of individuals in;

[0082] For each ,make , Indicates storage A temporary variable representing the crowding distance for each individual;

[0083] For each objective function ,implement:

[0084] Will Individuals are stored in ascending order according to their corresponding objective function. middle: ;

[0085] ;

[0086] for arrive :

[0087] No. The crowding distance of an individual is equivalent to that of the first individual. The sum of the distances between two individuals in each objective function direction, i.e.:

[0088] ;

[0089] Indicates the first The first individual The objective function value, It is a divisor used to unify dimensions;

[0090] get Crowding distance of all individuals in the group:

[0091] .

[0092] Preferably, the specific method for obtaining the Pareto optimal solution set by introducing an adaptive learning mechanism and a constraint satisfaction strategy includes:

[0093] For cost-based metrics, the solution in the Pareto front should be standardized as follows:

[0094] ;

[0095] in, It is the first The solution is the first The standardized value of the objective function. It is the first The solution is the first The original values ​​of the objective function. and They represent the first Minimum and maximum values ​​of the objective function;

[0096] The standard deviation is calculated as follows:

[0097] ;

[0098] in, It is the first The standard deviation of each objective It is the size of the Pareto front. It is the first The average of the targets;

[0099] Calculation target and The correlation coefficient is:

[0100] ;

[0101] in, The goal and The correlation coefficient;

[0102] No. The solution of the first The information content of each target is calculated as follows:

[0103] ;

[0104] in, It's about information content. The number of targets;

[0105] Weight of each objective The calculation is as follows:

[0106] ;

[0107] in, For the target sequence number, Indicates the target sequence number is The information content of the target;

[0108] Then, the weighted standardized decision matrix is ​​constructed as follows:

[0109] ;

[0110] in, It is the first The solution is the first The weighted standardized value of each indicator;

[0111] Then the positive ideal solution and the negative ideal solution and yes:

[0112] ;

[0113] in, This represents the ideal solution for the last objective. , ;

[0114] The distance between the solution and the ideal solution is calculated as follows:

[0115] ;

[0116] in, It is the distance between the solution and the ideal solution. It is the distance between the solution and the negative ideal solution;

[0117] Relative closeness of each solution The calculation is as follows:

[0118] ;

[0119] The larger, the solution The better.

[0120] The multi-objective rolling optimization operation method for microgrids that considers voltage stability and energy storage synergy proposed in this invention has the following advantages:

[0121] 1. A novel hybrid prediction model was constructed, which significantly improved prediction accuracy:

[0122] A novel hybrid prediction model is constructed by combining wavelet packet decomposition with a long short-term memory network. This framework can effectively extract multi-scale time-frequency features from photovoltaic output sequences, overcoming the deficiency of insufficient feature extraction in traditional LSTM models under complex weather conditions. Experimental results show that, under highly volatile cloudy conditions, the root mean square error of the proposed method is reduced by 32.4% compared to the traditional LSTM model, significantly improving the accuracy and reliability of renewable energy prediction.

[0123] 2. It achieves objective decision-making for multi-objective optimization, eliminating the subjectivity of weight setting:

[0124] By embedding an online Pareto front assessment module within the model predictive control framework, this invention enables multidimensional real-time analysis of power generation costs and environmental costs. This module can dynamically derive the weight distribution of each objective based on the system state within the prediction range, completely avoiding the subjectivity and arbitrariness of manually setting weights in the traditional weighted summation method, thus making dispatch decisions more scientific and objective.

[0125] 3. An efficient optimization algorithm was designed, which improved the solution speed and the quality of the solution set:

[0126] At the algorithmic level, an improved NSGA2 algorithm with a fast non-dominated sorting mechanism is proposed. By optimizing the population initialization strategy and introducing adaptive learning coefficients, the initial convergence process of the algorithm is effectively accelerated, while avoiding premature convergence. In addition, the innovative Pareto optimal solution construction method can automatically generate a high-quality optimal trade-off solution set without pre-setting weight vectors, ensuring a balance between the economy and environmental friendliness of the scheduling scheme.

[0127] 4. A complete closed-loop optimization system has been established, enhancing the system's practicality:

[0128] By organically integrating advanced forecasting, multi-objective optimization, and rolling scheduling, a complete closed-loop optimization scheme for microgrid energy management has been formed. This scheme can not only effectively cope with the random fluctuation characteristics of renewable energy, but also ensure the safe and stable operation of the system through a hard constraint satisfaction mechanism, providing a practical and feasible technical path for the optimized scheduling of microgrids with a high proportion of renewable energy access. Attached Figure Description

[0129] Figure 1 This is a schematic diagram of the microgrid system structure described in this invention;

[0130] Figure 2 This is an overall flowchart of the energy management strategy proposed in this invention;

[0131] Figure 3 This is a structural diagram of a photovoltaic power prediction model based on hybrid LSTM;

[0132] Figure 4 This is a schematic diagram of the rolling optimization mechanism of Model Predictive Control (MPC).

[0133] Figure 5 This is an overall flowchart of the microgrid multi-objective rolling optimization method described in this invention;

[0134] Figure 6 The graph shows the prediction results of a photovoltaic power prediction model based on hybrid LSTM under two typical weather conditions.

[0135] Figure 7 This is a voltage distribution diagram of each node in the microgrid after scheduling using the method of this invention; Figure 8 This is a microgrid dispatch plan diagram for 24 hours. Detailed Implementation

[0136] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0137] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0138] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0139] Example 1:

[0140] The following is combined Figures 1-5This embodiment describes a multi-objective rolling optimization operation method for microgrids that considers voltage stability and energy storage synergy. It includes:

[0141] Preprocessing of raw photovoltaic power output and meteorological data: The improved Akima interpolation method is used to smooth the missing data, the input features related to photovoltaic power output are screened by Pearson correlation coefficient, and wavelet packet decomposition is used to decompose the sequence data of the screened input features into multiple sub-signals of different frequency bands.

[0142] Construct a hybrid prediction model: For each frequency band sub-signal, train a corresponding LSTM prediction sub-model, where each sub-model takes the multivariate sub-signal under a specific frequency band as input and the contribution of that frequency band to photovoltaic power output as output; perform linear weighted fusion of the outputs of all sub-models to obtain the photovoltaic power output prediction result in the prediction time domain;

[0143] Constructing a multi-objective optimization model: Based on the photovoltaic power output prediction results, establish mathematical models of each device in the microgrid system, define a multi-objective function with power generation cost and environmental cost as optimization objectives, introduce key constraints, and form a multi-objective optimization model;

[0144] Multi-objective optimization solution: The multi-objective optimization model is solved based on the improved NSGA2 algorithm. Pareto optimal solution set is obtained by introducing an adaptive learning mechanism and constraint satisfaction strategy. The optimal scheduling scheme at each time step is selected from the solution set by using the TOPSIS method based on objective weights.

[0145] Perform rolling time-domain optimization: Based on the model predictive control mechanism, the optimal scheduling scheme is applied to the microgrid system, the system state is rolled to the next time step, and the above steps are repeated to achieve closed-loop optimized operation of the entire scheduling cycle.

[0146] Furthermore, the specific method for smoothing missing data using the improved Akima interpolation method includes:

[0147] Improved Akima piecewise cubic Hermite interpolation is used to smooth the data:

[0148] ;

[0149] in, Represents the interpolation function. This represents the interpolation variable, and the interpolation interval is... , , , and Both represent constants.

[0150] Furthermore, the specific method for screening input features related to photovoltaic output using the Pearson correlation coefficient includes:

[0151] The original dataset contains the following features: temperature, relative humidity, total radiation, diffuse radiation, precipitation, global tilt irradiance, and diffuse tilt irradiance.

[0152] The correlation between these characteristics and photovoltaic output was evaluated and screened using the Pearson correlation coefficient:

[0153] ;

[0154] in, This represents the correlation coefficient between each feature and photovoltaic output. Indicates the number of features. Indicates photovoltaic output data, Representing feature data, Indicates the first One data point, This indicates the total number of data points.

[0155] Furthermore, specific methods for using wavelet packet decomposition to decompose the sequence data from which input features have been selected into multiple sub-signals of different frequency bands include:

[0156] The wavelet basis function is "db3", and the decomposition level is three.

[0157] Discrete wavelet packet decomposition is expressed as:

[0158] ;

[0159] in, Indicates the first wavelet level coefficient, Indicates a low-pass filter. Indicates a high-pass filter. and Both represent discrete-time indices;

[0160] The refactoring process is described as follows:

[0161] ;

[0162] The dataset was normalized to:

[0163] ;

[0164] in, This represents the normalization result. This represents the maximum value of the target normalization range, which is 1. The minimum value of the target normalization range is -1; Represents the original value. Indicates the maximum original value. This represents the minimum original value.

[0165] Furthermore, the LSTM prediction sub-model specifically includes:

[0166] The loss function of LSTM is:

[0167] ;

[0168] in, , Indicates the predicted segment length. Represents the actual value. This represents the predicted value.

[0169] Furthermore, the specific methods for defining the multi-objective function with power generation cost and environmental cost as optimization objectives include:

[0170] Electricity generation cost objective function for:

[0171] ;

[0172] in, Indicates time, Indicates optimization of the time domain, This represents the cost of electricity generation from distributed photovoltaic (PV) power plants. Indicates time Active power of distributed photovoltaic (PV) at that time This represents the power generation cost (WT) of a distributed wind turbine. Indicates time The active power of the distributed wind turbine (WT) at that time. and These represent the active power generation cost and reactive power generation cost of dispatchable generators (DG), respectively. Indicates time The active power exchanged with the main power grid at that time. Indicates time The reactive power exchanged with the main power grid at that time and These represent the costs of purchasing active power and reactive power from the main power grid, respectively. for , for , and These refer to the prices for active power and reactive power sold to the main power grid. for , for , and These refer to the operating costs of the energy storage device ES during charging and discharging. Indicates time The charging power of the energy storage device ES. Indicates time The discharge power of the energy storage device ES;

[0173] Environmental cost target for:

[0174] ;

[0175] in, This represents the equivalent environmental cost coefficient.

[0176] Furthermore, the key constraints introduced include: power flow constraints, voltage and power balance constraints, specifically:

[0177]

[0178]

[0179] in, and They represent time respectively Time-injection node Active power and reactive power, ;

[0180] It is with nodes The set of adjacent nodes;

[0181] It is time Time node The voltage amplitude satisfies , and These are the lower and upper limits of the voltage amplitude. It represents the voltage amplitude at all nodes;

[0182] It is time Time node The voltage amplitude;

[0183] and They represent the nodes respectively To the node Conductivity and susceptance;

[0184] It is time Time from node To the node The phase difference;

[0185] Indicates the number of dispatchable generators. Indicates the amount of energy stored. Indicates the number of photovoltaic power plants. Indicates the number of wind farms. Indicates the quantity that meets the requirements. Indicates the serial number of the power generation equipment / energy storage / load. Indicates network loss. This represents the load power of each node;

[0186] Indicates the sequence number is The active power exchanged between the dispatchable generator (DG) and the main power grid. Indicates the sequence number is The discharge power of the energy storage device ES. Indicates the sequence number is The active power of distributed photovoltaic (PV) Indicates the sequence number is The active power of the distributed wind turbine WT Indicates the sequence number is The charging power of the energy storage device ES.

[0187] Furthermore, the multi-objective optimization model is specifically as follows:

[0188] ;

[0189] in, It is an optimization variable. It is a feasible region;

[0190] The optimization problem is constrained by equipment operation constraints, power flow constraints, and power balance constraints.

[0191] Furthermore, the specific method for solving the multi-objective optimization model based on the improved NSGA2 algorithm includes:

[0192] The two operators in the NSGA2 algorithm are fast non-dominated sorting and crowd distance calculation;

[0193] The specific process of quick nondominated sort is as follows:

[0194] Determine the set of individuals in the first ranking order:

[0195] set up For population groups, In order to be in Individuals ( The set of all individuals under control. for medium dominance The total number of individuals;

[0196] for Each individual in ,implement:

[0197] For each individual Perform initialization operations. , ;

[0198] for Each individual in ,implement:

[0199] if Then will be subject to Dominant Individual deposit middle, ;

[0200] if So, for each dominant entity discovered individuals, Increase by 1, ;

[0201] if So, individual sorting level And these individuals are stored in the first-level sort. In the set;

[0202] Determine the set of individuals for the remaining ranking levels:

[0203] set up It is the first Let a set of individuals ordered by level be... ;

[0204] when At that time, execute:

[0205] ,in, This represents a temporary variable used to store information about the next level of individuals.

[0206] for Each individual in ,implement:

[0207] for Each individual in ,implement:

[0208] ;if ,So , , Represents an individual The sorting order;

[0209] Update the sorting level number: ;

[0210] Obtain the A set of individuals ordered by hierarchy ;

[0211] All individuals are assigned to different levels of sorted sets;

[0212] The specific process for calculating congestion distance is as follows:

[0213] For each sorted set ,implement:

[0214] ,in, Represents a sorted set The number of individuals in;

[0215] For each ,make , Indicates storage A temporary variable representing the crowding distance for each individual;

[0216] For each objective function ,implement:

[0217] Will Individuals are stored in ascending order according to their corresponding objective function. middle: ;

[0218] ;

[0219] for arrive :

[0220] No. The crowding distance of an individual is equivalent to that of the first individual. The sum of the distances between two individuals in each objective function direction, i.e.:

[0221] ;

[0222] Indicates the first The first individual The objective function value, It is a divisor used to unify dimensions;

[0223] get Crowding distance of all individuals in the group:

[0224] .

[0225] Furthermore, the specific method for obtaining the Pareto optimal solution set by introducing an adaptive learning mechanism and a constraint satisfaction strategy includes:

[0226] For cost-based metrics, the solution in the Pareto front should be standardized as follows:

[0227] ;

[0228] in, It is the first The solution is the first The standardized value of the objective function. It is the first The solution is the first The original values ​​of the objective function. and They represent the first Minimum and maximum values ​​of the objective function;

[0229] The standard deviation is calculated as follows:

[0230] ;

[0231] in, It is the first The standard deviation of each objective It is the size of the Pareto front. It is the first The average of the targets;

[0232] Calculation target and The correlation coefficient is:

[0233] ;

[0234] in, The goal and The correlation coefficient;

[0235] No. The solution of the first The information content of each target is calculated as follows:

[0236] ;

[0237] in, It's about information content. The number of targets;

[0238] Weight of each objective The calculation is as follows:

[0239] ;

[0240] in, For the target sequence number, Indicates the target sequence number is The information content of the target;

[0241] Then, the weighted standardized decision matrix is ​​constructed as follows:

[0242] ;

[0243] in, It is the first The solution is the first The weighted standardized value of each indicator;

[0244] Then the positive ideal solution and the negative ideal solution and yes:

[0245] ;

[0246] in, This represents the ideal solution for the last objective. , ;

[0247] The distance between the solution and the ideal solution is calculated as follows:

[0248] ;

[0249] in, It is the distance between the solution and the ideal solution. It is the distance between the solution and the negative ideal solution;

[0250] Relative closeness of each solution The calculation is as follows:

[0251] ;

[0252] The larger, the solution The better.

[0253] This invention elucidates its principles in three parts: modeling of microgrids and distributed energy resources; a multi-objective energy management strategy based on learning-driven model predictive control; and simulation results.

[0254] I. Modeling of Microgrids and Distributed Energy Resources

[0255] The basic structure of a microgrid is as follows Figure 1 As shown, the network comprises various distributed energy sources, including energy storage (ES), dispatchable generators (DG), distributed photovoltaic (PV), distributed wind turbines (WT), and demand. The distribution network can exchange active and reactive power with the main grid. Dispatchable energy sources in the network include DGs and ES, while PV and wind turbines are uncontrollable resources that can only be reduced. To optimize network operation from both economic and environmental perspectives, an effective energy management strategy needs to be designed. The energy management strategy proposed in this invention is centralized, periodically scheduling all resources in the network.

[0256] In designing an energy management strategy, the operating conditions of adjustable energy resources should be analyzed first. This section focuses primarily on the steady-state characteristics of these resources. The operating conditions of distributed generators can be expressed as the following constraints:

[0257] (1)

[0258] (2)

[0259] in, It is the active power of DGs in each scheduling cycle. This is the minimum active power of DGs. This is the highest active power of DGs. This is the maximum gradeability of DGs. It is the current moment. This is an optimization of the time domain. These two constraints mean that the scheduling reference value of DGs cannot exceed the total operating range, and the ramp rate of DGs cannot exceed the limit. DGs are usually set as "PV nodes" without reactive power constraints, so the reactive power of DGs is uncontrollable.

[0260] The operating conditions of an ES device can be described as follows:

[0261] (3)

[0262] in It is the charging power of ES in each scheduling cycle. This is the lowest charging power of ES. This is the highest charging power of the ES. It is the discharge power of the ES in each scheduling cycle. This is the lowest discharge power of the ES. This is the maximum discharge power of the ES. This constraint means that the charging and discharging rates of the ES should not exceed the limit. Simultaneously, the state of charge of the ES also satisfies:

[0263] (4)

[0264] in It is the SOC of the ES in each scheduling cycle. It is the lower limit of SOC. This is the upper limit of the SoC. This constraint means that the total charging and discharging energy cannot exceed the capacity limit.

[0265] The operating conditions for renewable energy sources, including PV and WT, can be described as follows:

[0266] (5)

[0267] in It is the active power of PV in each scheduling cycle. This is the actual output of PV. It is the active power of WT in each scheduling cycle. This is the actual output of WT. This constraint means that PV and WT power can only be reduced.

[0268] II. Design of Energy Management Strategies

[0269] like Figure 2 The diagram shown is an overall flowchart of the proposed energy management strategy. Figure 2 As shown, the strategy comprises three steps: Step 1 involves photovoltaic power prediction, Step 2 solves a multi-objective optimization problem, and Step 3 performs rolling time-domain optimization. In Step 1, an LSTM-based network is proposed, utilizing wavelet packet decomposition. This network predicts PV output power in the prediction domain based on historical PV data and numerical weather forecasts. This step also contributes to the prediction model used in model predictive control. In Step 2, an open-loop optimization problem is solved. An improved multi-objective particle swarm optimization algorithm is proposed to improve the quality of the Pareto front. The initialization phase of the algorithm is optimized using Latin hypercube sampling. Adaptive learning coefficients facilitate exploration in early iterations and help find more refined local optima in later iterations. The solution set in the decision space forms the Pareto front through non-dominated sorting and selection strategies. Then, an objective weighting method based on the Critic algorithm is proposed. The solutions are evaluated using a sequential preference technique based on the similarity of ideal solutions, and the optimal solution is selected. Finally, the first term of the control sequence is applied, and rolling time-domain optimization is performed.

[0270] PV prediction based on hybrid LSTM:

[0271] Photovoltaic power prediction models based on hybrid LSTM, such as Figure 3As shown, the original data is first preprocessed. Then, wavelet packet decomposition is used to decompose the PV dataset into different frequency bands, and the subsequences are projected into the subspace through signal reconstruction. For each sub-signal, four independent LSTMs are trained, and the results of all LSTMs are summed to obtain the final prediction.

[0272] First, the raw PV data needs to be preprocessed. Due to sensor malfunction or other reasons, the raw data may contain some missing data. Therefore, an improved Akima piecewise cubic Hermite interpolation is used to smooth the data.

[0273] (6)

[0274] in It is an interpolation function. It is an interpolation variable, and the interpolation interval is... , It is a constant.

[0275] The next step was to identify the features: the original dataset contained temperature, relative humidity, total radiation, diffuse radiation, precipitation, global tilt irradiance, and diffuse tilt irradiance. The correlation between these features and PV output was evaluated using the Pearson correlation coefficient method.

[0276] (7)

[0277] in It is the correlation coefficient between each feature and the PV output. It is PV output data. It is feature data. It is the total number of data. It is the number of features.

[0278] The results are shown in Table 1, which shows the correlation coefficients between the characteristic quantities and photovoltaic output.

[0279] Table 1

[0280] temperature relative humidity Total radiation direct radiation Precipitation Global tilt irradiance Direct tilt irradiance 0.7341 -0.7038 0.9727 0.6066 -0.0978 0.9690 0.6757

[0281] As can be seen from Table 1, the correlation between PRCP and PV output is very small, therefore the feature set is [GR, GTI, TEMP, RH, DTI, DR].

[0282] The PV output dataset will then be decomposed using wavelet packet decomposition.

[0283] The wavelet basis function is "db3", and the decomposition level is three. Discrete wavelet packet decomposition can be expressed as:

[0284] (8)

[0285] in It is the first wavelet level coefficient, and These are low-pass and high-pass filters, respectively.

[0286] The refactoring process can be described as follows:

[0287] (9)

[0288] The dataset was normalized to:

[0289] (10)

[0290] in It is the result of normalization. and These are the minimum and maximum values ​​of the target normalization range, which are -1 and 1, respectively. It is the original value. and These are the minimum and maximum original values.

[0291] The loss function of LSTM is designed as follows:

[0292] (11)

[0293] in It is the length of the predicted segment. It is the actual value. This is a predicted value.

[0294] The loss function is designed based on mean squared error.

[0295] The problem statement for rolling time-domain optimization:

[0296] Model predictive control MPC rolling optimization mechanism, such as Figure 4 As shown, MPC comprises two main steps: predicting the system state and solving the open-loop optimization problem. Because MPC corrects the predicted information before solving the optimization problem at each time step, it helps reduce the impact of uncertainties on the system. Figure 4 As shown, the prediction domain is defined as the time range within which the controller predicts the system state starting from the current time. The controller then solves an open-loop optimization problem within the prediction domain and obtains a sequence of solutions. Note that the length of the solution sequence matches the length of the prediction domain, but only the first term of the sequence is implemented in the system. For example, in... Figure 4 In the middle, the current time is The prediction domain is four steps long, therefore the system state changes from... Predicted After solving the optimization problem, we obtain a control sequence. Only when the first item is implemented in the system does the system transition to time. .

[0297] Although the proposed energy management strategy is not a control problem and does not involve reference signals or closed-loop systems, the MPC mechanism still has reference value. MPC-based energy management strategies have several advantages. First, they are easier to implement in real time. Since the accuracy of long-term renewable energy forecasts is generally lower than that of short-term forecasts, MPC-based methods can achieve more accurate forecasts through rolling time-domain adjustments. Furthermore, compared to other methods, MPC-based methods consider the impact of prior information when optimizing the system, thus achieving better cost optimization.

[0298] The optimization objectives include power generation costs and environmental management costs. The objective function for power generation costs is defined as follows:

[0299] (12)

[0300] in It is the cost of generating electricity from PV. It is the electricity generation cost of WT. and These are the operating costs of ES charging and discharging, respectively. and These are the active and reactive power generation costs of DG, respectively. and These are the costs of purchasing active and reactive power from the main power grid, respectively. and These are the prices for active and reactive power sold to the main power grid, respectively. express , express , It is time The active power exchanged with the main power grid at that time. express , express , It is time The reactive power exchanged with the main power grid at that time.

[0301] The environmental cost objective function can be expressed as:

[0302] (13)

[0303] in It is the equivalent environmental cost coefficient.

[0304] When solving optimization problems, power flow constraints, voltage and power balance constraints should be considered:

[0305] (14)

[0306] (15)

[0307] in and It is time Time-injection node Active and reactive power, , It is with nodes The set of adjacent nodes, It is time Time node The voltage amplitude satisfies , and It is a limitation on the voltage amplitude. and From node The side road The conductivity and susceptance, It is time Time from node arrive phase difference, Indicates the quantity of distributed energy resources or loads. Indicates network loss. This represents the load power of each node.

[0308] Therefore, the multi-objective optimization problem can be expressed as:

[0309] (16)

[0310] in, (17)

[0311] These are optimization variables. The feasible region is represented by Equations (1)-(5) and (15)-(16).

[0312] Power balance constraints and power flow constraints are two essential constraints in the energy management of distribution networks and should not be violated.

[0313] This invention uses two methods to handle these two constraints. First, during the iteration of the NSGA2 algorithm, power flow calculations are performed on each solution set in the population to evaluate power balance and power flow convergence conditions. If a particular solution exhibits power flow non-convergence or imbalance, a large penalty factor is applied when calculating the fitness value of that solution. This mechanism guides the particles to evolve towards regions that satisfy the power flow constraints and power balance conditions. Then, to prevent infeasible solutions from entering the Pareto optimal solution set in special cases, the power flow calculation results of each solution in the population are checked after non-dominated sorting, and solutions exhibiting power flow non-convergence or power imbalance are subsequently excluded. Existing literature has demonstrated that when using evolutionary algorithms to solve similar optimization problems, applying penalty terms can effectively handle constraints and ensure that constraint violations do not occur. From a theoretical analysis perspective, it is feasible to use penalty terms for constraint handling in multi-objective particle swarm optimization algorithms. Theoretically, as long as the penalty factor is large enough, any infeasible solution will be systematically excluded from the dominance sorting or congestion calculation, thereby guiding the algorithm to converge to the feasible region. Simultaneously, this invention incorporates a feasibility test into the dominance mechanism to ensure that any solution that violates the constraints will be dominated by feasible solutions. This mechanism guarantees that the Pareto optimal solution set does not contain solutions that violate the constraints. With these two guarantees, the method proposed in this invention, similar to the hard constraint method, ensures that the constraints are satisfied.

[0314] Improved NSGA2 algorithm and Pareto frontier evaluation:

[0315] Fast non-dominated sorting and crowding distance calculation are two particularly important operators in NSGA2. Fast non-dominated sorting uses the dominance relationship between individuals to sort them hierarchically, which allows better solutions to be placed in earlier positions (the first being level 1). Crowding distance is introduced to make the distribution of the obtained solutions more uniform in space.

[0316] Quick Non-Dominant Sort:

[0317] set up For population groups, In order to be in Individuals ( The set of all individuals under control. for medium dominance The total number of individuals.

[0318] The entire fast nondominated sorting process can be divided into two main parts: the first part is determining... Each individual Corresponding and The first part provides the set of individuals ordered at the first level; the second part uses the results obtained above to determine the sets of individuals ordered at the second, third, ... levels. A hierarchically ordered set of individuals.

[0319] Below is the specific programming logic and explanation for Quick Non-Dominant Sort:

[0320] Part 1: Determining the set of individuals in the first ranking order .

[0321] for Each individual in Perform the following operations:

[0322] (1) Let , ; (for each individual) and (Perform initialization operations)

[0323] (2) For Each individual in Perform the following operations:

[0324] if ,So ;(will be subject to) Dominant Individual deposit middle)

[0325] if ,So ;(every time a dominant force is discovered) individual , Add 1)

[0326] (3) If ,So ; ; ( A value of 0 indicates that... There is no one who can control Individuals, therefore these individuals sorting level And these individuals are stored in the first-level sort. (in the set)

[0327] Part Two: Determine the set of individuals for the remaining sorting ranks.

[0328] set up It is the first Let a set of individuals ordered by level be... ;

[0329] when At that time, perform the following operations:

[0330] (1) ( It is a temporary variable used to store the next level of individuals; it must be initialized before use.

[0331] (2) For Each individual in Perform the following operations:

[0332] for Each individual in Perform the following operations:

[0333] ; (the first one on the right side of the equals sign) The value is equal to the value already determined in the first part. (value) if ,So , (Through the entire loop, all individuals in the next level of sorting can be selected and temporarily stored.) This loop is also the most ingenious part of Quick Non-Dominant Sort's ability to perform multi-level sorting.

[0334] (3) (Update the sorting rank number)

[0335] (4) ; (obtained the first A set of individuals ordered by hierarchy )

[0336] After the above sorting operations, All individuals are assigned to different levels of sorted sets. In the middle, the total number of levels It is not a fixed value; its value is determined by... It depends on the specific circumstances of each individual.

[0337] Crowded distance calculation:

[0338] After completing the fast non-dominated sort, the order of individuals at different levels becomes easy to determine, but the order relationship between individuals at the same level remains uncertain. Therefore, the concept of crowding distance solves this problem. Crowding distance is used to further sort individuals at the same level, but it is ineffective for individuals at different levels. The physical meaning of crowding distance is that the crowding distance of the i-th individual is equivalent to the sum of the distances between its two preceding and following individuals in each objective function direction, which is half the perimeter of the dashed rectangle in the diagram.

[0339] The specific programming approach and explanation for crowding distance are as follows:

[0340] For each sorted set ( Perform the following operations:

[0341] (1) ; ( It is a sorted set (Number of individuals in)

[0342] (2) For each ,make ;

[0343] (initialization It is storage (A temporary variable representing the crowding distance for each individual)

[0344] (3) For each objective function Perform the following operations:

[0345] ;(Will Individuals are stored in the database in ascending order of their corresponding objective function values. middle)

[0346] (In order to preserve the boundary points later, the crowding distance between the first and last individuals is set to infinity.)

[0347] for arrive :

[0348] (This step is the core formula for calculating congestion distance, the first...) The crowding distance of an individual is equivalent to the sum of the distances between the two individuals immediately before and after it in each objective function direction. To ensure that each objective function has an equivalent influence on the crowding distance, a normalization-like operation (dividing by) is used in the calculation process. To unify the dimensions. Indicates the first The first individual (The objective function value.)

[0349] (4) ;(get (crowding distance of all individuals)

[0350] In the subsequent optimization selection process, the ranking level and crowding distance of each individual will serve as important reference indicators.

[0351] The assessment and decision-making process of the Pareto front:

[0352] For cost-based metrics, the solution in the Pareto front should be standardized as follows:

[0353] (18)

[0354] in It is the first The solution is the first The standardized value of the objective function. It is the first The solution is the first The original values ​​of the objective function. and It is the first The minimum and maximum values ​​of the objective function.

[0355] Then the standard deviation can be calculated as follows:

[0356] (19)

[0357] in No. The standard deviation of each objective It is the size of the Pareto front. It is the first The average of the targets.

[0358] Then the target can be calculated. and The correlation coefficient is:

[0359] (20)

[0360] in The goal and The correlation coefficient.

[0361] No. The solution of the first The information content of an objective can be calculated as follows:

[0362] (twenty one)

[0363] in It's about information content. The weight of each objective. It can be calculated as follows:

[0364] (twenty two)

[0365] Then, the weighted standardized decision matrix is ​​constructed as follows:

[0366] (twenty three)

[0367] in It is the first The solution is the first The weighted standardized value of each indicator.

[0368] Then the positive ideal solution and the negative ideal solution and yes:

[0369] (twenty four)

[0370] in , .

[0371] The distance between the solution and the ideal solution can be calculated as follows:

[0372] (25)

[0373] in and It is the distance between the solution and the positive & negative ideal solutions.

[0374] Relative closeness of each solution It can be calculated as follows:

[0375] (26)

[0376] The larger, the solution The better.

[0377] Note that although the proposed energy management strategy uses the MPC mechanism, the optimization problem studied is an open-loop problem. The overall flowchart of the proposed energy management strategy is as follows: Figure 2 As shown. The relationship between the prediction, optimization, and decision-making components is as follows. Figure 5 As shown. The detailed process of the prediction part can be found in... Figure 3 Found it.

[0378] III. Simulation Results

[0379] The effectiveness of the proposed energy management strategy was verified through simulation. For the prediction part, the proposed hybrid LSTM framework was validated under two typical daytime conditions: sunny and cloudy. The results are as follows: Figure 6 As shown, the proposed hybrid LSTM prediction model can accurately predict photovoltaic output and track fluctuation trends under both weather conditions. For the scheduling part, the proposed energy management strategy is evaluated on a 33-node test system. The node voltage calculation results are shown below. Figure 7 As shown, the voltage amplitude of all nodes is within the constraints. The scheduling plan for the next 24 hours is as follows: Figure 8 As shown in Table 2, the costs of the two algorithms are compared with those of the NSGA3 algorithm.

[0380] Table 2

[0381] algorithm <![CDATA[Power generation cost (10 4 $)]]> <![CDATA[Environmental cost (10 4 $)]]> <![CDATA[Total cost (10 4 $)]]> NSGA2 4.086 0.698 4.784 NSGA3 4.309 0.645 4.954

[0382] Compared with traditional methods, the proposed strategy reduces power generation costs by 8.6% and total costs by 5.9%. The proposed algorithm demonstrates better diversity and uniformity of non-dominated solutions than baseline algorithms. Simulation results illustrate the effectiveness of the proposed energy management strategy.

[0383] Compared with existing inventions:

[0384] This invention deeply integrates wavelet packet decomposition with LSTM architecture to design a novel hybrid prediction framework. This framework explicitly extracts multi-scale time-frequency features of photovoltaic output, significantly improving prediction performance under cloudy conditions compared to using LSTM alone. Compared to other prediction models, the proposed method reduces the root mean square error of prediction results by 32.4% on cloudy days.

[0385] This invention innovatively embeds an online Pareto front assessment module into the MPC rolling scheduling framework. This module performs multi-dimensional real-time analysis of power generation and environmental costs within the prediction range and dynamically derives the target weight distribution, thereby eliminating the subjective influence in scheduling decisions.

[0386] At the optimization algorithm level, this invention proposes an improved NSGA2 algorithm with fast non-dominated sorting to accelerate initial convergence and avoid local optima, and proposes a Pareto optimal solution construction method that can naturally generate the optimal trade-off solution without pre-setting weight vectors.

[0387] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy, characterized in that, It includes: Preprocessing of raw photovoltaic power output and meteorological data: The improved Akima interpolation method is used to smooth the missing data, the input features related to photovoltaic power output are screened by Pearson correlation coefficient, and wavelet packet decomposition is used to decompose the sequence data of the screened input features into multiple sub-signals of different frequency bands. Construct a hybrid prediction model: For each frequency band sub-signal, train a corresponding LSTM prediction sub-model, where each sub-model takes the multivariate sub-signal under a specific frequency band as input and the contribution of that frequency band to photovoltaic power output as output; perform linear weighted fusion of the outputs of all sub-models to obtain the photovoltaic power output prediction result in the prediction time domain; Constructing a multi-objective optimization model: Based on the photovoltaic output prediction results, a mathematical model of each device in the microgrid system is established, a multi-objective function with power generation cost and environmental cost as optimization objectives is defined, key constraints are introduced, and a multi-objective optimization model is formed; Solving the multi-objective optimization: The multi-objective optimization model is solved based on the improved NSGA2 algorithm. By introducing an adaptive learning mechanism and constraint satisfaction strategy, a Pareto optimal solution set is obtained, and the TOPSIS method based on objective weights is used to select the optimal scheduling scheme at each time step from the solution set; Perform rolling time-domain optimization: Based on the model predictive control mechanism, the optimal scheduling scheme is applied to the microgrid system, the system state is rolled to the next time step, and the above steps are repeated to achieve closed-loop optimized operation of the entire scheduling cycle.

2. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The specific methods for smoothing missing data using the improved Akima interpolation method include: Improved Akima piecewise cubic Hermite interpolation is used to smooth the data: ； in, Represents the interpolation function. This represents the interpolation variable, and the interpolation interval is... , , , and Both represent constants.

3. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The specific method for screening input features related to photovoltaic output using the Pearson correlation coefficient includes: The original dataset contains the following features: temperature, relative humidity, total radiation, diffuse radiation, precipitation, global tilt irradiance, and diffuse tilt irradiance. The correlation between these characteristics and photovoltaic output was evaluated and screened using the Pearson correlation coefficient: ； in, This represents the correlation coefficient between each feature and photovoltaic output. Indicates the number of features. Indicates photovoltaic output data, Representing feature data, Indicates the first One data point, This indicates the total number of data points.

4. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, Specific methods for decomposing sequence data with selected input features into multiple sub-signals of different frequency bands using wavelet packet decomposition include: The wavelet basis function is "db3", and the decomposition level is three. Discrete wavelet packet decomposition is expressed as: ； in, Indicates the first wavelet level coefficient, Indicates a low-pass filter. Indicates a high-pass filter. and Both represent discrete-time indices; The refactoring process is described as follows: ； The dataset was normalized to: ； in, This represents the normalization result. This represents the maximum value of the target normalization range, which is 1. The minimum value of the target normalization range is -1; Represents the original value. Indicates the maximum original value. This represents the minimum original value.

5. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The LSTM prediction sub-model specifically includes: The loss function of LSTM is: ； in, , Indicates the predicted segment length. Represents the actual value. This represents the predicted value.

6. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The specific methods for defining a multi-objective function with power generation cost and environmental cost as optimization objectives include: Electricity generation cost objective function for: ； in, Indicates time, Indicates optimization of the time domain, This represents the cost of electricity generation from distributed photovoltaic (PV) power plants. Indicates time Active power of distributed photovoltaic (PV) at that time This represents the power generation cost (WT) of a distributed wind turbine. Indicates time The active power of the distributed wind turbine (WT) at that time. and These represent the active power generation cost and reactive power generation cost of dispatchable generators (DG), respectively. Indicates time The active power exchanged with the main power grid at that time. Indicates time The reactive power exchanged with the main power grid at that time and These represent the costs of purchasing active power and reactive power from the main power grid, respectively. for , for , and These refer to the prices for active power and reactive power sold to the main power grid. for , for , and These refer to the operating costs of the energy storage device ES during charging and discharging. Indicates time The charging power of the energy storage device ES. Indicates time The discharge power of the energy storage device ES; Environmental cost target for: ； in, This represents the equivalent environmental cost coefficient.

7. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The key constraints introduced include: power flow constraints, voltage and power balance constraints, specifically: in, and They represent time respectively Time-injection node Active power and reactive power, ; It is with nodes The set of adjacent nodes; It is time Time node The voltage amplitude satisfies , and These are the lower and upper limits of the voltage amplitude. It represents the voltage amplitude at all nodes; It is time Time node The voltage amplitude; and They represent the nodes respectively To the node Conductivity and susceptance; It is time Time from node To the node The phase difference; Indicates the number of dispatchable generators. Indicates the amount of energy stored. Indicates the number of photovoltaic power plants. Indicates the number of wind farms. Indicates the quantity that meets the requirements. Indicates the serial number of the power generation equipment / energy storage / load. Indicates network loss. This represents the load power of each node; Indicates the sequence number is The active power exchanged between the dispatchable generator (DG) and the main power grid. Indicates the sequence number is The discharge power of the energy storage device ES. Indicates the sequence number is The active power of distributed photovoltaic (PV) Indicates the sequence number is The active power of the distributed wind turbine WT Indicates the sequence number is The charging power of the energy storage device ES.

8. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The multi-objective optimization model is specifically as follows: ； in, It is an optimization variable. It is a feasible region; The optimization problem is constrained by equipment operation constraints, power flow constraints, and power balance constraints.

9. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The specific method for solving the multi-objective optimization model based on the improved NSGA2 algorithm includes: The two operators in the NSGA2 algorithm are fast non-dominated sorting and crowd distance calculation; The specific process of quick nondominated sort is as follows: Determine the set of individuals in the first ranking order: set up For population groups, In order to be in Individuals ( The set of all individuals under control. for medium dominance The total number of individuals; for Each individual in ,implement: For each individual Perform initialization operations. , ; for Each individual in ,implement: if Then will be subject to Dominant Individual deposit middle, ; if So, for each dominant entity discovered individuals, Increase by 1, ; if So, individual sorting level And these individuals are stored in the first-level sort. In the set; Determine the set of individuals for the remaining ranking levels: set up It is the first Let a set of individuals ordered by level be... ; when At that time, execute: ,in, This represents a temporary variable used to store information about the next level of individuals. for Each individual in ,implement: for Each individual in ,implement: ;if ,So , , Represents an individual The sorting order; Update the sorting level number: ; Obtain the A set of individuals ordered by hierarchy ; All individuals are assigned to different levels of sorted sets; The specific process for calculating congestion distance is as follows: For each sorted set ,implement: ,in, Represents a sorted set The number of individuals in; For each ,make , Indicates storage A temporary variable representing the crowding distance for each individual; For each objective function ,implement: Will Individuals are stored in ascending order according to their corresponding objective function. middle: ; ； for arrive : No. The crowding distance of an individual is equivalent to that of the first individual. The sum of the distances between two individuals in each objective function direction, i.e.: ； Indicates the first The first individual The objective function value, It is a divisor used to unify dimensions; get Crowding distance of all individuals in the group: 。 10. The multi-objective rolling optimization operation method for microgrids considering voltage stability and energy storage synergy as described in claim 1, characterized in that, The specific methods for obtaining the Pareto optimal solution set by introducing an adaptive learning mechanism and a constraint satisfaction strategy include: For cost-based metrics, the solution in the Pareto front should be standardized as follows: ； in, It is the first The solution is the first The standardized value of the objective function. It is the first The solution is the first The original values ​​of the objective function. and They represent the first Minimum and maximum values ​​of the objective function; The standard deviation is calculated as follows: ； in, It is the first The standard deviation of each objective It is the size of the Pareto front. It is the first The average of the targets; Calculation target and The correlation coefficient is: ； in, The goal and The correlation coefficient; No. The solution of the first The information content of each target is calculated as follows: ； in, It's about information content. The number of targets; Weight of each objective The calculation is as follows: ； in, For the target sequence number, Indicates the target sequence number is The information content of the target; Then, the weighted standardized decision matrix is ​​constructed as follows: ； in, It is the first The solution is the first The weighted standardized value of each indicator; Then the positive ideal solution and the negative ideal solution and yes: ； in, This represents the ideal solution for the last objective. , ; The distance between the solution and the ideal solution is calculated as follows: ； in, It is the distance between the solution and the ideal solution. It is the distance between the solution and the negative ideal solution; Relative closeness of each solution The calculation is as follows: ； The larger, the solution The better.

Images