Double-layer optimization scheduling method for source network load storage cooperative loss reduction

Abstract

The invention provides a double-layer optimization scheduling method for source network load storage collaborative loss reduction, and belongs to the technical field of source network load storage collaborative scheduling. A double-layer optimization framework comprising an upper layer planning model and a lower layer operation model is constructed, and a key node set is screened by using a node importance comprehensive scoring method based on a graph theory; the improved wolf pack algorithm is adopted to solve a double-layer coupling problem, a lower-layer operation effect is fed back to an upper layer to serve as a fitness evaluation basis, and in cooperation with population aggregation degree monitoring, a wandering wolf random jumping mechanism and a reverse learning strategy, deep collaboration and global optimal solution of a planning layer and an operation layer are achieved. The technical problem of poor network loss optimization effect caused by lack of an effective double-layer coupling solution mechanism for collaborative optimization planning and operation scheduling of the distributed photovoltaic and energy storage system in the power distribution network is solved.

Description

Technical Field

[0001] This invention belongs to the field of source-grid-load-storage coordinated scheduling technology. Specifically, it relates to a two-layer optimized scheduling method for source-grid-load-storage coordinated loss reduction. Background Technology

[0002] In the field of coordinated optimization of power generation, grid, load, and energy storage in distribution networks, traditional methods employ single-layer optimization models to handle the planning and configuration problems of distributed photovoltaic (PV) and energy storage systems. These methods typically rely on particle swarm optimization (PSO) or genetic algorithms, separating the planning and operation layers for independent optimization, or simplifying the two-layer problem into a single-objective optimization. When dealing with distribution network topology, these methods often perform indiscriminate optimization on all nodes, failing to identify critical nodes that significantly impact system performance, thus increasing computational complexity. During optimization, traditional heuristic algorithms are prone to getting trapped in local optima and fail to fully consider the coupling effect of planning layer decisions on multiple objectives of the operation layer. This results in limited reduction of network losses and insignificant improvement in voltage quality in actual operation. In existing technologies, due to the lack of an effective coordination mechanism between the planning and operation layers, and a lack of accurate identification methods for critical nodes in the distribution network, optimization algorithms exhibit poor convergence performance when solving two-layer coupled problems. In other words, existing technologies suffer from a lack of effective two-layer coupled solution mechanisms for the coordinated optimization planning and operation scheduling of distributed PV and energy storage systems in distribution networks, leading to poor network loss optimization results. Summary of the Invention

[0003] In view of this, the present invention provides a two-layer optimization scheduling method for source-grid-load-storage collaborative loss reduction, which can solve the technical problem in the prior art that the collaborative optimization planning and operation scheduling of distributed photovoltaic and energy storage systems in the distribution network lacks an effective two-layer coupling solution mechanism, resulting in poor network loss optimization effect.

[0004] This invention is implemented as follows: It provides a two-layer optimization scheduling method for source-grid-load-storage collaborative loss reduction, including acquiring basic data of the distribution network and constructing a set of key nodes; establishing a two-layer optimization model framework and applying constraints; iteratively solving the upper-layer planning model using an improved wolf pack algorithm; transmitting the optimization results of the upper-layer planning model to the lower-layer operation model and providing feedback on fitness evaluation criteria; calculating the improvement amount of the objective function and triggering an adaptive adjustment mechanism; determining convergence conditions and outputting a source-grid-load-storage collaborative scheduling scheme; the two-layer optimization model framework establishes a closed-loop feedback mechanism between the upper-layer planning layer and the lower-layer operation layer, enabling the upper-layer decision variables to be dynamically adjusted based on the lower-layer operation results; the set of key nodes identifies node locations that significantly affect system performance using a graph theory-based node importance comprehensive scoring method; and the improved wolf pack algorithm avoids getting trapped in local optima by combining population aggregation monitoring with a wandering wolf random jump mechanism and a reverse learning strategy.

[0005] This invention constructs a two-layer optimization model framework, using the planning and configuration of distributed photovoltaic and energy storage systems as the upper-level decision and system operation scheduling as the lower-level optimization. A closed-loop feedback mechanism between the planning and operation layers is established, enabling the upper-level planning decisions to be dynamically adjusted based on the lower-level operational results, achieving deep coupling and synergy between planning and operation. Through a graph theory-based node importance comprehensive scoring method, combining betweenness centrality, compact centrality, and influence ranking, key nodes with significant impact on system performance in the distribution network are accurately identified, focusing the optimization space on critical locations and significantly reducing solution complexity. An improved wolf pack algorithm is adopted, introducing population aggregation monitoring and a wandering wolf random jump mechanism, combined with a back-learning strategy, effectively avoiding getting trapped in local optima during the optimization process and improving the global optimization capability of the two-layer coupled problem. The lower-level operation model, through multi-objective weighted aggregation, integrates network loss minimization, voltage quality optimization, and load curve fluctuation minimization into the optimization framework. Its operational results are fed back to the upper layer as a fitness evaluation basis, guiding the planning layer decisions towards reducing network losses. In summary, this invention solves the technical problem mentioned in the background art: the lack of an effective two-layer coupled solution mechanism for the collaborative optimization planning and operation scheduling of distributed photovoltaic and energy storage systems in power distribution networks leads to poor network loss optimization results. Attached Figure Description

[0006] Figure 1 This is a flowchart of the method of the present invention.

[0007] Figure 2 A framework diagram of the "source-grid-load-storage" interaction mechanism for the distribution network.

[0008] Figure 3 A graph showing the change in the global optimal objective function value during the iterative process of the improved wolf pack algorithm.

[0009] Figure 4 This is a daily charge and discharge power curve of the energy storage system.

[0010] Figure 5 This is a comparison chart of the load power curve and the net load power curve of the distribution network system. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0012] like Figure 1 As shown, this invention provides a two-layer optimized scheduling method for source-grid-load-storage collaborative loss reduction, comprising: S01. Data Modeling, Importance Calculation, and Key Node Selection: Acquire basic data of the distribution network and construct a set of key nodes. The basic data of the distribution network includes the number of nodes, node load capacity, number of branches, branch impedance, and maximum power allowed through the branches. Abstract the distribution network into a weighted graph structure. The node weights of the weighted graph structure are the node load capacity, and the edge weights of the weighted graph structure are the branch impedance. Calculate the betweenness centrality, tight centrality, and influence ranking value of each node in the weighted graph structure. Normalize the betweenness centrality, tight centrality, and influence ranking values ​​to obtain normalized betweenness centrality, normalized tight centrality, and normalized influence ranking values. Weight the normalized betweenness centrality, normalized tight centrality, and normalized influence ranking values ​​according to the weight coefficients to obtain a comprehensive node importance score. Sort all nodes according to the comprehensive node importance score from high to low, and select the top 30% of nodes with the comprehensive node importance score as the key node set. S02. Constructing a Two-Level Optimization Model and Applying Constraints: Establish a two-level optimization model framework and apply constraints. The framework includes an upper-level planning model and a lower-level operation model. The objective function of the upper-level planning model is to minimize the total annual planning cost. The decision variables of the upper-level planning model are the location of distributed photovoltaic (PV) access, the installed capacity of distributed PV, the location of energy storage system access, and the rated capacity of energy storage system at each node in the key node set. Apply upper-level constraints to the upper-level planning model. These constraints include constraints on distributed PV power generation capacity, energy storage system installed capacity, energy storage system charge / discharge efficiency, energy storage system state, power balance of the distribution system, and branch power. The first objective function of the lower-level operation model is to minimize the distribution network loss; the second objective function is to optimize node voltage quality; and the third objective function is to minimize load curve fluctuations. The decision variables of the lower-level operation model are the operating power of distributed photovoltaic systems and the operating power of energy storage systems. The first, second, and third objective functions of the lower-level operation model are weighted and aggregated through objective function weight coefficients to obtain a lower-level single objective function. Lower-level constraints are applied to the lower-level operation model, including power flow constraints, distributed photovoltaic output constraints, energy storage system output constraints, node voltage constraints, and line power constraints. S03. Improved Wolf Pack Algorithm Initialization and Triggering Jump Mechanism: Input the decision variables and constraints of the upper-level planning model into the source-grid-load-storage collaborative scheduling optimization model, perform population initialization through the improved wolf pack algorithm, obtain the initial particle sequence, set the number of alpha wolf particles, the number of aggressive wolf particles, and the number of wandering wolf particles in the initial particle sequence, and iteratively solve the improved wolf pack algorithm. The alpha wolf particle guides the population movement direction according to the current global optimal solution, the aggressive wolf particles perform local fine search around the alpha wolf particle, and the wandering wolf particles randomly explore within the search space. Calculate the population aggregation index, determine whether the population aggregation index exceeds the aggregation threshold, and if the population aggregation index exceeds the aggregation threshold, trigger the wandering wolf random jump mechanism, randomly select some particles among the wandering wolf particles, and reinitialize the positions of the selected particles to random positions within the search space. S04. Model Solving and Feedback Result Evaluation of Fitness: The optimization results of the upper-level planning model are passed as input parameters to the lower-level operation model. The optimization results include the distributed photovoltaic installed capacity and the rated capacity of the energy storage system. Distributed photovoltaic power output data under various seasonal scenarios are obtained based on the distributed photovoltaic installed capacity. The intraday charging and discharging strategy of the energy storage system is optimized. The objective function value of the lower-level single objective function is solved. The distribution network loss value, node voltage deviation value, and load curve fluctuation variance value are calculated. The distribution network loss value, node voltage deviation value, and load curve fluctuation variance value of the lower-level operation model are fed back to the upper-level planning model as the fitness evaluation basis of the upper-level particles. The position of the alpha wolf particle, the search step size of the wolf particle, and the jump probability of the wandering wolf particle are updated according to the fitness evaluation basis. S05. Target Improvement: Reduce Step Size and Perform Backward Learning: Calculate the improvement amount of the objective function, determine if the improvement amount of the objective function is less than the improvement amount threshold for multiple consecutive iterations. If the improvement amount of the objective function is less than the improvement amount threshold for multiple consecutive iterations, reduce the search step size of the wolf particle to 0.5 times the original search step size. Perform backward learning processing on the particles with the worst fitness in the current iteration, generate symmetrical points of the positions of the particles about the center of the search space, evaluate the fitness value of the symmetrical point corresponding to the symmetrical point position, and determine if the fitness value of the symmetrical point is better than the fitness value of the original particle of the particles. If the fitness value of the symmetrical point is better than the fitness value of the original particle, replace the position of the particles with the symmetrical point position. S06. Determine if the termination condition is met: Determine if the maximum number of iterations has been reached or the convergence accuracy requirement has been met. If the maximum number of iterations has not been reached and the convergence accuracy requirement has not been met, return to step S03 to continue iterating. S07. Output the optimal collaborative scheduling scheme: When the maximum number of iterations is reached or the convergence accuracy requirement is met, output the globally optimal source-grid-load-storage collaborative scheduling scheme. The source-grid-load-storage collaborative scheduling scheme includes the optimal access location of distributed photovoltaic, the optimal installed capacity of distributed photovoltaic, the optimal access location of energy storage system, the optimal rated capacity of energy storage system, the daily charging and discharging power curve of energy storage system, and the power flow distribution results of distribution network at each time period.

[0013] The betweenness centrality value is used to measure the importance of a node as a bridge in the distribution network, representing the proportion of the number of shortest paths passing through the node to the total number of shortest paths in the distribution network; the compact centrality value is used to measure the average distance from a node to all other nodes in the distribution network, and the smaller the average distance, the higher the compact centrality value; the influence ranking value is obtained by iteratively calculating the influence propagation of a node to its neighboring nodes, reflecting the global influence of a node in the distribution network topology.

[0014] The normalization calculation steps are as follows: subtract the minimum value among the betweenness centrality values ​​from the betweenness centrality value, and then divide by the difference between the maximum and minimum values ​​among the betweenness centrality values ​​to obtain the normalized betweenness centrality value; subtract the minimum value among the tight centrality values ​​from the tight centrality value, and then divide by the difference between the maximum and minimum values ​​among the tight centrality values ​​to obtain the normalized tight centrality value; subtract the minimum value among the influence ranking values ​​from the influence ranking value, and then divide by the difference between the maximum and minimum values ​​among the influence ranking values ​​to obtain the normalized influence ranking value.

[0015] The weighting coefficients include betweenness centrality weighting coefficient, tightness centrality weighting coefficient, and influence ranking weighting coefficient. The betweenness centrality weighting coefficient is 0.4, the tightness centrality weighting coefficient is 0.3, and the influence ranking weighting coefficient is 0.3. The comprehensive score of node importance is equal to the normalized betweenness centrality value multiplied by the betweenness centrality weighting coefficient, plus the normalized tightness centrality value multiplied by the tightness centrality weighting coefficient, plus the normalized influence ranking value multiplied by the influence ranking weighting coefficient.

[0016] The total annual planned cost includes the total investment cost, system operation and maintenance cost, and system electricity purchase cost. The total annual planned cost is equal to the total investment cost multiplied by the annual equivalent factor plus the system operation and maintenance cost plus the system electricity purchase cost. The annual equivalent factor is equal to the discount rate multiplied by 1 plus the equipment lifespan raised to the power of the discount rate, and then divided by the difference between 1 plus the equipment lifespan raised to the power of the discount rate minus 1.

[0017] The distributed photovoltaic (PV) power generation capacity constraint is as follows: the sum of the distributed PV installed capacity of all nodes in the key node set is less than or equal to the product of the PV power factor and the sum of the node load power of all load nodes. The PV power factor ranges from 0.85 to 0.95, and the distributed PV installed capacity of each node should be less than the maximum distributed PV capacity allowed to be connected to the node.

[0018] The installed capacity constraint of the energy storage system is: the energy storage device at any time... The state of charge is equal to the time. The state of charge is multiplied by 1 and the charge / discharge rate is subtracted. When charging, the energy storage system charging power is multiplied by the charging efficiency and divided by the rated capacity of the energy storage system, and then multiplied by the time step. When discharging, the energy storage system discharge power is subtracted by the product of the discharge efficiency and the rated capacity of the energy storage system, and then multiplied by the time step.

[0019] The energy storage system's charge and discharge efficiency constraint is that the product of the energy storage system's charging power and its discharging power is always equal to 0.

[0020] The state constraints of the energy storage system are as follows: the state of charge should be greater than or equal to the lower limit of the state of charge and less than or equal to the upper limit of the state of charge, the lower limit of the state of charge is 0.2, the upper limit of the state of charge is 0.9, the charging power of the energy storage system should be greater than or equal to 0 and less than or equal to the maximum charging power, and the discharging power of the energy storage system should be greater than or equal to 0 and less than or equal to the maximum discharging power.

[0021] The power balance constraint of the power distribution system is: at time Photovoltaic active power and time Energy storage active power and time The sum of the power transmission contact wires equals the time. Load power and time The sum of the system's active power losses, whereby the system's active power losses are equal to the sum of the active power losses of all branches in the distribution network at time t. The summation of active power losses in the branch circuits.

[0022] The branch power constraint is: allowing power to pass through the first... The power passing through each branch should be greater than or equal to 0 and less than or equal to the maximum power allowed to pass through that branch.

[0023] The first objective function is the average of the squared differences between the net load power and the average net load power at each moment during a day's operation. The second objective function is the sum of the squared differences between the actual voltage and the rated voltage at all nodes. The third objective function is the variance between the power value and the average power value at each moment of the load curve.

[0024] The objective function weight coefficients include a first objective function weight coefficient and a second objective function weight coefficient. The sum of the first objective function weight coefficient and the second objective function weight coefficient is 1. The lower-level single objective function is equal to the first objective function multiplied by the first objective function weight coefficient plus the second objective function multiplied by the second objective function weight coefficient.

[0025] The power flow constraints are: nodes The active power is equal to the node The voltage multiplied by all the nodes Connected nodes The sum of the voltage of the node, the inter-node conductance multiplied by the cosine of the inter-node phase angle, and the inter-node susceptance multiplied by the sine of the inter-node phase angle, is the sum of the voltage of the node. The reactive power is equal to the node The voltage multiplied by all the nodes Connected nodes The sum of voltage and inter-node conductance multiplied by the sine of the inter-node phase angle minus the inter-node susceptance multiplied by the cosine of the inter-node phase angle.

[0026] The distributed photovoltaic (PV) output constraint is: the maximum output power of the distributed PV system is less than or equal to the installed capacity of the distributed PV system, at any given time. The actual output power of a distributed photovoltaic system should be greater than or equal to 0 and less than or equal to the value at time t. The maximum output power of the distributed photovoltaic system. The energy storage system output constraint is: both the charging power and discharging power of the energy storage system should be greater than or equal to 0 and less than or equal to the maximum charging and discharging power limit of the energy storage system. The node voltage constraint is: node... The actual voltage should be greater than or equal to the lower node voltage limit and less than or equal to the upper node voltage limit. The lower node voltage limit is equal to the rated voltage multiplied by 0.93, and the upper node voltage limit is equal to the rated voltage multiplied by 1.07. The line power constraint is: line The apparent power should be less than or equal to the maximum apparent power allowed to pass through the line.

[0027] The structure of the source-grid-load-storage coordinated scheduling optimization model includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, and an output layer. The input layer receives distribution network topology parameters, load power time-series data, distributed photovoltaic output prediction data, energy storage system parameters, and upper and lower limits of various constraints. The distribution network topology parameters include the number of nodes, branch impedance, node load capacity, distributed photovoltaic installed capacity, rated capacity of the energy storage system, and charging efficiency. The input feature vector of the input layer has a dimension of 128. The first hidden layer contains 256 neurons, and a modified linear unit activation function is used to non-activate the input feature vector. The system performs a linear transformation and extracts shallow features. The second hidden layer contains 512 neurons and uses a modified linear unit activation function. A feature fusion layer is used to perform a residual connection between the output of the first hidden layer and the input feature vector of the input layer. The third hidden layer contains 256 neurons and introduces a global feature aggregation algorithm based on self-attention pooling to calculate the attention weight at each position in the feature map and perform weighted aggregation on the feature map output by the second hidden layer. The output layer contains 64 neurons and outputs the optimal installed capacity of the distributed photovoltaic system, the optimal rated capacity of the energy storage system, the intraday charging and discharging strategy, and the predicted distribution network loss value.

[0028] The global feature aggregation algorithm based on self-attention pooling first maps the feature map output from the second hidden layer to a query vector, a key vector, and a value vector through three parallel fully connected layers. The dimensions of the query vector and the key vector are both one-eighth of the number of channels in the feature map, and the dimension of the value vector is the same as the number of channels in the feature map. The matrix product of the query vector and the transpose of the key vector is calculated to obtain an attention score matrix. Each row of the attention score matrix is ​​exponentially normalized to ensure that the sum of the elements in each row is 1, resulting in an attention weight matrix. The attention weight matrix is ​​then multiplied by the value vector to obtain a weighted feature map. This weighted feature map is mapped back to the original number of channels through a fully connected layer and the residual is added to the output of the second hidden layer to obtain the global feature aggregation result.

[0029] The self-attention pooling-based global feature aggregation algorithm adaptively learns the importance of different spatial locations in the feature map. While retaining crucial information such as node voltage deviation and branch power exceedance information for key nodes, it suppresses interference from redundant features. Compared to the traditional global average pooling method that simply averages all location features, this algorithm dynamically adjusts the weights of each location feature based on the current distribution network operating state, assigning higher weights to nodes with severe voltage exceedances and power transmission bottleneck branches. This allows the source-grid-load-storage collaborative scheduling optimization model to focus on key factors affecting the distribution network loss and node voltage quality. In the source-grid-load-storage collaborative optimization scenario, the contributions of different nodes and branches to the overall system performance vary significantly. Voltage fluctuations at key nodes have a much greater impact on the overall network voltage quality than at edge nodes, and power optimization of heavily loaded branches significantly improves loss reduction. The self-attention pooling-based global feature aggregation algorithm can automatically identify key features and strengthen their role in the decision-making process.

[0030] The attention weight matrix introduced in the feature aggregation stage of the global feature aggregation algorithm based on self-attention pooling is essentially a data-driven importance assessment method. Through training with a large amount of historical operating data, the source-grid-load-storage coordinated scheduling optimization model learns the influence patterns of each node and branch on the optimization objective under different operating conditions. In practical applications, it can quickly locate the optimization priorities under the current state, avoiding wasting computational resources on secondary features. The advantage of the global feature aggregation algorithm based on self-attention pooling lies in its adaptability to the feature space distribution. When a large number of distributed photovoltaics are connected to the distribution network, causing frequent changes in power flow direction, the fixed feature extraction mode in traditional methods is difficult to adapt to the dynamically changing operating characteristics. However, the global feature aggregation algorithm based on self-attention pooling dynamically adjusts the aggregation strategy according to the actual distribution of input features through the interactive calculation of the query vector, the key vector, and the value vector, thereby improving the robustness of the source-grid-load-storage coordinated scheduling optimization model to uncertainty and time-varying characteristics.

[0031] The residual connection enables the gradient to be directly propagated from the output layer to the first hidden layer, alleviating the gradient vanishing problem in deep network training, accelerating the convergence speed of the source-network-load-storage collaborative scheduling optimization model and improving training stability. In source-network-load-storage optimization problems involving multi-layer decision-making and complex constraints, stable gradient propagation is crucial for the source-network-load-storage collaborative scheduling optimization model to accurately learn the coupling relationship between the upper-layer planning model and the lower-layer running model.

[0032] The steps for establishing the training dataset for the source-grid-load-storage collaborative scheduling optimization model include: collecting historical distribution network operation data under different seasons and weather conditions. This historical distribution network operation data includes hourly load power of each node, actual output of distributed photovoltaic power, energy storage charging and discharging records, and distribution network power flow distribution results. The time span of the historical distribution network operation data covers at least two full years. The historical distribution network operation data is preprocessed to remove abnormal data points caused by sensor failures or communication interruptions, and missing data is filled using cubic spline interpolation. Based on load levels and distributed photovoltaic power output intensity, the historical distribution network operation data is divided into high-load, high-photovoltaic scenarios. The system is divided into three scenarios: high load with low photovoltaic (PV) capacity, low load with high PV capacity, and low load with low PV capacity. 200 sets of sample data are extracted for each scenario. For each set of sample data, the distribution network topology parameters, constraint parameters, and optimization target values ​​are calculated to construct the input feature vector and label vector. The label vector includes the optimal installed capacity of distributed PV, the optimal rated capacity of the energy storage system, the intraday charging and discharging strategy, and the distribution network loss value. The sample data is divided into a training set and a validation set at an 8:2 ratio. The training set is used for parameter learning of the source-grid-load-storage collaborative scheduling optimization model, and the validation set is used for evaluating the generalization performance of the model.

[0033] The training steps of the source-grid-load-storage collaborative scheduling optimization model include: initializing the weight parameters of each layer of the model using the Xavier initialization method to make the weight parameters follow a uniform distribution with a mean of 0 and a variance equal to the reciprocal of the input dimension; initializing the bias parameter to 0; setting the learning rate to 0.001, the batch size to 32, the maximum number of iterations to 500; using the weighted sum of mean squared error and constraint violation penalties as the loss function, whereby the constraint violation penalties include voltage over-limit penalties, power over-limit penalties, and charging / discharging mutual exclusion constraint violation penalties, with the penalty weight coefficient set to 100; and employing adaptive... The moment estimation optimizer performs gradient descent updates, setting the momentum decay coefficient to 0.9 and the second-order moment estimation decay coefficient to 0.999. The performance of the source-grid-load-storage collaborative scheduling optimization model is evaluated on the validation set every 10 training batches, and the loss value of the validation set is recorded. Training is terminated by an early stopping mechanism when the loss value of the validation set fails to decrease for 20 consecutive evaluations. After training, the weight parameters with the lowest loss value on the validation set are selected as the final model parameters. Performance testing is then performed on the test set to evaluate the optimization effect and generalization ability of the source-grid-load-storage collaborative scheduling optimization model on unseen data.

[0034] The initial particle sequence is generated through chaotic mapping. Two chaotic variables with different initial values ​​are set. The values ​​of the chaotic variables range from -1 to +1 and are not equal to 0. A chaotic sequence is generated using a sine function. The chaotic variables are in the first... The value of the second iteration is equal to 1.18 multiplied by 2 divided by the first iteration. The sine value of the chaotic variable in the next iteration; the chaotic variable is transformed into the range of particle position and particle velocity through a linear mapping, wherein the particle position is in the [nth iteration]. The coordinates of the dimensional variable are equal to the chaotic variable plus 1 multiplied by the particle position at the dimensional variable. The maximum value of the dimension minus the particle position at the th dimension The difference between the minimum values ​​of the dimensions divided by 2, plus the particle position at the th dimension, is used to determine the minimum value of the dimensional values. The minimum value of the dimension, the particle velocity in the th dimension The component of the dimension is equal to the chaotic variable plus 1 multiplied by the particle velocity in the th dimension. The maximum value of the dimension minus the particle velocity in the th dimension The difference between the minimum values ​​of the two dimensions divided by 2, plus the particle velocity at the 1st dimension, is used to calculate the particle velocity at the 1st dimension. The minimum value of the dimension.

[0035] The number of alpha wolf particles accounts for 10% of the total number of particles in the initial particle sequence, the number of ferocious wolf particles accounts for 30% of the total number of particles in the initial particle sequence, and the number of wandering wolf particles accounts for 60% of the total number of particles in the initial particle sequence.

[0036] The population aggregation index is calculated as follows: The average Euclidean distance from the position of each particle to the population center is calculated, where the population center is the arithmetic mean of the coordinates of all particle positions. This average Euclidean distance is then normalized by dividing the average distance by the diagonal length of the search space to obtain the normalized population aggregation index. The normalized population aggregation index ranges from 0 to 1, with a smaller value indicating a more aggregated population. The aggregation threshold is 0.15. When the normalized population aggregation index is less than this threshold, the wandering wolf random jump mechanism is triggered, randomly selecting 20% ​​of the wandering wolf particles as the selected subset. The objective function improvement is calculated as follows: The difference between the current iteration's global optimal objective function value and the previous iteration's global optimal objective function value is recorded. The absolute value of this difference is then divided by the previous iteration's global optimal objective function value to obtain the normalized objective function improvement rate. The improvement threshold is 0.001. When the improvement rate of the normalized objective function is less than the improvement threshold for five consecutive iterations, the improvement of the objective function is deemed to have stalled. The original search step size is the search step size used by the wolf particle in the previous iteration.

[0037] The reverse learning process includes the following steps: after each iteration, statistically analyzing the fitness evaluation criteria of all particles, sorting them from worst to best according to the fitness evaluation criteria, and selecting the 10% of particles with the worst fitness evaluation criteria as the worst-fitting particles; for each worst-fitting particle, generating a symmetric point position of the worst-fitting particle about the center of the search space, wherein the symmetric point position is at the 1st iteration. The coordinates of the dimensional particle are equal to the particle's position at the dimensional coordinates. The maximum value of the dimension plus the particle position at the th dimension The minimum value of the dimension minus the worst-fit particles in the th dimension Dimensional coordinates.

[0038] The convergence accuracy requirement is as follows: when the absolute value of the change in the global optimal objective function value is less than 0.0001 for 30 consecutive iterations, the improved wolf pack algorithm is considered to have converged to a stable solution, thus satisfying the convergence accuracy requirement. The maximum number of iterations is set to 1000. When the number of iterations reaches the maximum number of iterations, the improved wolf pack algorithm is terminated regardless of whether the convergence accuracy requirement is met, and the current global optimal solution is output.

[0039] The output of the source-grid-load-storage coordinated dispatch scheme includes: whether each node in the key node set is connected to distributed photovoltaic (PV) power, the optimal installed capacity of the distributed PV power corresponding to the node connected to PV power, whether each node in the key node set is connected to an energy storage system, and the optimal rated capacity of the energy storage system corresponding to the node connected to the energy storage system; the charging power and discharging power of the energy storage system for each hour within 24 hours, constituting the daily charging and discharging power curve of the energy storage system; the active power transmission value and reactive power transmission value of each branch of the distribution network within 24 hours, the voltage amplitude and phase angle of each node, used to evaluate the rationality and safety of the power flow distribution results of the distribution network in each time period; and the calculated annual total planning cost, the total annual network loss value of the distribution network, the number of times the node voltage exceeds the limit, and the variance value of the load curve fluctuation, which serve as comprehensive evaluation indicators of the economic and technical aspects of the source-grid-load-storage coordinated dispatch scheme.

[0040] The present invention also provides a two-layer optimized scheduling system for source-grid-load-storage collaborative loss reduction implemented by a computer. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-mentioned two-layer optimized scheduling method for source-grid-load-storage collaborative loss reduction.

[0041] It should be noted that this invention also addresses the following technical problem: Traditional optimization methods fail to effectively identify nodes in the distribution network that significantly impact system performance when handling source-grid-load-storage coordinated scheduling, leading to dispersed optimization resources, low computational efficiency, and difficulty in focusing on key loss-reduction aspects. This invention establishes a graph-theory-based comprehensive node importance scoring system, abstracting the distribution network into a weighted graph structure. It calculates the betweenness centrality, tight centrality, and influence ranking value of each node. After normalizing these three indicators, they are weighted and summed according to their respective weight coefficients to obtain a comprehensive node importance score. The top 30% of nodes in the score are selected as the key node set. This method quantifies the bridging role of nodes in the power transmission path, their central position in the network, and their influence propagation ability on adjacent nodes at the topological level. It limits the optimization configuration space of distributed photovoltaic and energy storage systems to the positions that contribute the most to loss reduction, ensuring both optimization effectiveness and significantly reducing computational complexity. This allows the two-layer optimization model to converge to a high-quality solution within a reasonable timeframe.

[0042] It should be noted that this invention also solves the following technical problems: traditional heuristic optimization algorithms are prone to getting trapped in local optima when solving the dual-layer coupling optimization problem of power distribution network sources, grids, loads, and storage, and suffer from slow convergence speed and loss of population diversity in the later stages of optimization. This invention introduces a population aggregation monitoring mechanism by improving the wolf pack algorithm. It calculates and normalizes the average Euclidean distance from all particle positions to the population center. When the normalized population aggregation index is less than a set threshold, a random jumping mechanism of wandering wolves is triggered, randomly selecting 20% ​​of the wandering wolf particles and reinitializing their positions to random positions within the search space, thus breaking the excessive aggregation state of the population and maintaining the diversity of solutions. Meanwhile, this invention sets up a monitoring mechanism for the improvement of the objective function. When the improvement rate of the normalized objective function is less than the improvement threshold for five consecutive iterations, the search step size of the wolf particles is reduced to 0.5 times the original step size. The 10% of particles with the worst fitness are subjected to reverse learning to generate their symmetrical point positions about the center of the search space. If the fitness of the symmetrical point is better than that of the original particle, it is replaced. These mechanisms work together to enable the algorithm to actively jump out of local optima and re-explore better solution regions, ensuring the global convergence performance of the two-layer optimization model.

[0043] Specifically, the principle of this invention is as follows: The reason this invention can solve the technical problem lies in the fact that the two-layer optimization model framework establishes an information transmission path between the planning layer and the operation layer. The distributed photovoltaic installed capacity and the rated capacity of the energy storage system output by the upper-layer planning model serve as input parameters for the lower-layer operation model. The distribution network loss value, node voltage deviation value, and load curve fluctuation variance value calculated by the lower-layer operation model are fed back to the upper layer as the basis for fitness evaluation, forming a closed-loop optimization mechanism of decision-making-evaluation-feedback. The construction of the key node set is based on graph theory analysis of the distribution network topology. Betweenness centrality measures the importance of a node as a bridge for power transmission, tight centrality reflects the central position of a node in the network, and influence ranking reflects the ability of a node to radiate influence on neighboring nodes. The weighted fusion of these three factors can comprehensively evaluate the contribution of nodes to the overall system performance, concentrating optimization resources on the positions with the greatest potential for loss reduction. The improved wolf pack algorithm guides the global search direction through the alpha wolf particle, while the aggressive wolf particles perform local fine-grained searches and the wandering wolf particles conduct random explorations. The collaborative work of these three types of particles ensures sufficient coverage of the search space. When excessive population aggregation is detected, the wandering wolf is triggered to randomly jump, redistributing some particles to different regions of the search space and breaking the local convergence state. The reverse learning strategy generates symmetrical points about the center of the search space for particles with poor fitness, expanding the diversity of solutions. These mechanisms together ensure the global optimal solution capability of the two-layer coupled optimization problem, enabling the planning and configuration scheme to truly achieve the goal of minimizing network loss during operation.

[0044] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0045] The specific implementation of step S01 involves first obtaining basic data of the distribution network, including the number of nodes, node load capacity, number of branches, branch impedance, and maximum power allowed through each branch. The distribution network is then abstracted into a weighted graph structure, where the node weights represent node load capacity and the edge weights represent branch impedance. Betweenness centrality value. The formula used to measure the importance of a node as a bridge in a power distribution network is expressed as follows: ; In the formula, For nodes The betweenness centrality value is dimensionless; For nodes To the node The total number of shortest paths, dimensionless; For nodes To the node The shortest path passes through the nodes The number of paths, dimensionless; and In the distribution network, except for nodes Any two nodes outside the range. Tight centrality value. The formula used to measure the average distance from a node to all other nodes in a distribution network is expressed as follows: ; In the formula, For nodes The compact centrality value is dimensionless; The total number of nodes in the distribution network is dimensionless. For nodes To the node The shortest path length, dimensionless, representing the number of edges traversed by the path. Influence ranking value. This is obtained by iteratively calculating the propagation of a node's influence on its neighboring nodes, reflecting the node's global influence in the distribution network topology. The calculation formula is expressed as follows: ; In the formula, For nodes The influence ranking value is dimensionless; The largest eigenvalue of the adjacency matrix is ​​dimensionless. For nodes The set of adjacent nodes; Adjacent nodes The influence ranking value, dimensionless. Normalized betweenness centrality value. The calculation formula is expressed as follows: ; In the formula, The normalized betweenness centrality value is dimensionless. It is the minimum value of the betweenness centrality of all nodes, and is dimensionless. The maximum value of betweenness centrality among all nodes, dimensionless. Normalized compact centrality value. The calculation formula is expressed as follows: ; In the formula, The normalized compact centrality value is dimensionless; It is the minimum value of the tight centrality of all nodes, and is dimensionless. The maximum value among all node compact centrality values, dimensionless. Normalized influence ranking value. The calculation formula is expressed as follows: ; In the formula, The normalized influence ranking value is dimensionless. The minimum value among all node influence ranking values, dimensionless; The maximum value among all node influence rankings, dimensionless. Overall node importance score. The calculation formula is expressed as follows: ; In the formula, For nodes The importance score is dimensionless; 0.4 is the betweenness centrality weight coefficient, dimensionless; 0.3 is the tight centrality weight coefficient, dimensionless; 0.3 is the influence ranking weight coefficient, dimensionless. All nodes are sorted from high to low according to their overall importance score, and the top 30% of nodes by overall importance score are selected as the key node set.

[0046] The specific implementation of step S02 involves establishing a two-layer optimization model framework, including an upper-layer planning model and a lower-layer execution model. The mathematical expression of the two-layer optimization model is: ; ; ; ; In the formula, The objective function is the function of the upper layer. The lower-level objective function; These are the constraints for the upper-level objective function; These are the constraints for the lower-level objective function; Decision variables for higher-level objectives; These are the decision variables for the lower-level objectives. The objective function of the upper-level planning model is the total annual planning cost. Minimization, its calculation formula is as follows: ; In the formula, The total annual planning cost is expressed in yuan. Total investment cost, in yuan; System operation and maintenance costs, in yuan; The system's electricity purchase cost is expressed in yuan. This is the annual equivalent coefficient, dimensionless. Annual equivalent coefficient The calculation formula is expressed as follows: ; In the formula, The discount rate is dimensionless and has an empirical value of 0.05. The lifespan of the equipment is expressed in years, typically 20 years. The formula for the capacity constraint of distributed photovoltaic power generation is as follows: ; In the formula, For nodes in the key node set Distributed photovoltaic grid connection capacity, in kW; The photovoltaic power factor is dimensionless and ranges from 0.85 to 0.95. For nodes The load power, in kW; The number of all load nodes, dimensionless; This represents the set of key nodes. The formula for the distributed photovoltaic installed capacity constraint of each node is as follows: ; In the formula, For nodes The maximum permitted distributed photovoltaic capacity, in kW. Energy storage devices at any time... State of charge The calculation formula is expressed as follows, during charging: ; During discharge: ; In the formula, For a moment The state of charge is dimensionless; For a moment The state of charge is dimensionless; This is the charge / discharge rate, dimensionless, with an empirical value of 0.01; For a moment The charging power of the energy storage system, in kW; The charging efficiency is dimensionless and is typically taken as 0.95. The rated capacity of the energy storage system is expressed in kWh. The time step is in hours (h), with a default value of 1 hour. For a moment The discharge power of the energy storage system, in kW; The discharge efficiency is dimensionless and typically taken as 0.95. The formula for the charge / discharge efficiency constraint of the energy storage system is as follows: ; In the formula, and Cannot be greater than 0 at the same time, unit is The formula for the state constraints of the energy storage system is as follows: ; ; ; In the formula, Maximum charging power, in kW; The maximum discharge power is expressed in kW. The formula for the power balance constraint of the power distribution system is as follows: ; In the formula, For a moment The photovoltaic active power is expressed in kW. For a moment The energy storage active power, measured in kW; For a moment The power transmission contact wire, in kW; For a moment The load power, in kW; For a moment The system active power loss is expressed in kW. The calculation formula is expressed as follows: ; In the formula, For a moment branch road Active power loss, in kW; Let be the total number of branches, dimensionless. The formula for branch power constraints is as follows: ; In the formula, For the first The power carried by each branch is expressed in kW. branch road The maximum allowed power, in kW. The first objective function of the lower-level operational model. It is expressed as the average of the squares of the differences between the net load power and the average net load power at each moment during a day's operation. The calculation formula is as follows: ; In the formula, The objective function is to minimize the distribution network loss, with units of . ; The running time is one day, in hours, and the value is 24 hours. For a moment The net load power, in kW. The second objective function of the lower-level operation model. It is expressed as the sum of the squares of the differences between the actual voltage and the rated voltage at all nodes, and its calculation formula is as follows: ; In the formula, The objective function for optimizing node voltage quality is given in units of ; The total number of nodes in the distribution network is dimensionless. For nodes The actual voltage, in kV; The rated voltage of the distribution network is expressed in kV. This is the third objective function of the lower-level operation model. This is expressed as the variance between the power value and the average power value at each moment of the load curve, and its calculation formula is as follows: ; In the formula, The objective function is to minimize the load curve fluctuation, with units of ; For a moment The load power, in kW. Lower-level single-objective function. The calculation formula is expressed as follows: ; In the formula, This is a single-objective function at the lower level, and it is dimensionless. These are the weight coefficients of the first objective function, and are dimensionless. These are the weight coefficients of the second objective function, and are dimensionless. ; This is the power normalization factor, in kW, with a value of 1000kW; This is the voltage normalization coefficient, in kV, with a value of 10kV. (Nodes in power flow constraints) active power The calculation formula is expressed as follows: ; In the formula, For nodes The active power, expressed in kW; For nodes The voltage, in kV; For nodes Connected nodes The voltage, in kV; For nodes With nodes The electrical conductance between them is expressed in s (s). For nodes With nodes The susceptance between these values ​​is expressed in s (s). For nodes With nodes The phase angle between them, in rad; For nodes The set of all connected nodes. Nodes reactive power The calculation formula is expressed as follows: ; In the formula, For nodes The reactive power is expressed in kVar. The formula for the output constraint of distributed photovoltaic power is as follows: ; ; In the formula, This represents the maximum output power of a distributed photovoltaic system, expressed in kW. This refers to the installed capacity of distributed photovoltaic (PV) systems, expressed in kW. For a moment The actual output power of the distributed photovoltaic system, in kW; For a moment The maximum output power of the distributed photovoltaic system is expressed in kW. The formula for the output constraint of the energy storage system is as follows: ; ; In the formula, The maximum power limit for charging and discharging the energy storage system, in kW. The formula for the node voltage constraint is as follows: ; In the formula, This is the lower limit of the node voltage, in kV, and is calculated as follows: ; This represents the upper limit of the node voltage, in kV, and is calculated as follows: The formula for line power constraints is as follows: ; In the formula, For the line Apparent power, in kVA; This represents the maximum apparent power allowed to pass through the line, expressed in kVA.

[0047] The specific implementation of step S03 is as follows: first, an initial particle sequence is generated through chaotic mapping, and two chaotic variables with different initial values ​​are set. and The chaotic variable takes values ​​from -1 to +1 and is not equal to 0. The iterative formula for generating a chaotic sequence using a sine function is as follows: ; ; In the formula, For the first group of chaotic variables in the th The value of the next iteration is dimensionless. For the second group of chaotic variables in the th The value of the next iteration is dimensionless. The number of iterations is dimensionless. and ; and ; This is a dimensionless index for the variable dimensions. The chaotic variables are transformed to the range of particle position and velocity using a linear mapping, and the calculation formula is as follows: ; ; In the formula, For particles In the Dimensional position coordinates; For the particle position at the th The maximum value of the dimension; For the particle position at the th The minimum value of the dimension; For particles The speed in the first Dimensional components; For the particle velocity in the first... The maximum value of the dimension; For the particle velocity in the first... The minimum value of the dimension; The particle index is dimensionless. In the initial particle sequence, the alpha wolf particles account for 10% of the total, the fierce wolf particles account for 30%, and the wandering wolf particles account for 60%. The population aggregation index is calculated by first determining the population center location. The calculation formula is expressed as follows: ; In the formula, For the population center in the first Dimensional position coordinates; The total number of particles is dimensionless. For particles In the The position coordinates of each particle. Calculate the average Euclidean distance from all particle positions to the population center. The calculation formula is expressed as follows: ; In the formula, This represents the average Euclidean distance. For particles The Euclidean distance from the location of the population center is calculated as follows: ; For particles The position vector; The vector representing the population center location; The search space dimension is dimensionless. The length of the diagonal of the search space is... The calculation formula is expressed as follows: ; In the formula, The length of the diagonal of the search space. Normalized population aggregation index. The calculation formula is expressed as follows: ; In the formula, The normalized population aggregation index is dimensionless and ranges from 0 to 1. When the normalized population aggregation index is less than the aggregation threshold of 0.15, the wandering wolf random jump mechanism is triggered, randomly selecting 20% ​​of the wandering wolf particles and re-initializing their positions to random positions within the search space.

[0048] The specific implementation of step S04 is the same as described above, and will not be repeated in detail here.

[0049] The specific implementation of step S05 involves calculating the improvement amount of the objective function and normalizing the improvement rate of the objective function. The calculation formula is expressed as follows: ; In the formula, The improvement rate of the normalized objective function is dimensionless. For the current iteration The global optimal objective function value of the order of magnitude; For the previous iteration The global optimal objective function value of the order of magnitude; This is an iterative round index, dimensionless. When the improvement rate of the normalized objective function is less than the improvement threshold of 0.001 for five consecutive iterations, the objective function is considered to have stalled, and the search step size of the wolf particle is adjusted. The calculation formula for reducing the search step size to 0.5 times the original step size is as follows: ; In the formula, The updated Wolf Particle search step size; This is the original search step size, i.e., the search step size used by the wolf particle in the previous iteration. The location of the symmetric point. The calculation formula is expressed as follows: ; In the formula, For particles In the The coordinates of the symmetrical point in the dimension; For the particles with the worst fitness, in the first... Dimensional coordinates. When the fitness value of a symmetric point is better than that of the original particle, the positions of some particles are replaced with the positions of the symmetric point.

[0050] The specific implementation of step S06 is to determine whether the maximum number of iterations of 1000 has been reached or whether the convergence accuracy requirement has been met. The convergence accuracy requirement is that the absolute value of the change in the global optimal objective function value over 30 consecutive iterations is less than 0.0001. If the maximum number of iterations has not been reached and the convergence accuracy requirement has not been met, then return to step S03 to continue iterating. The specific implementation of step S07 is the same as described above, and will not be described in detail here.

[0051] It should be noted that, during the training process of the source-grid-load-storage collaborative scheduling optimization model, the weight parameters are initialized using the Xavier method, with initial weights... It follows a uniform distribution, and its calculation formula is as follows: ; In the formula, These are the initial weight parameters, which are dimensionless. Indicates a uniform distribution; The input dimension is dimensionless. The output dimension is dimensionless. Loss function. The weighted sum of mean square error and constraint violation penalty term is expressed by the following formula: ; In the formula, The value of the loss function; The sample size is dimensionless. For the first The true label value of each sample; For the first Predicted values ​​for each sample; The penalty weighting coefficient is dimensionless and has a value of 100. The total number of constraints is dimensionless. For the first The amount of violation of each constraint. In a global feature aggregation algorithm based on self-attention pooling, the query vector... Key vector Sum value vector The calculation formula is expressed as follows: ; ; ; In the formula, For query vector; The key vector; It is a value vector; , and These are the weight matrices for the query, key, and value, respectively, and are dimensionless. This is the feature map output by the second hidden layer. Attention score matrix. The calculation formula is expressed as follows: ; In the formula, This is the attention score matrix; This is the transpose of the key vector. Attention weight matrix. Middle elements The calculation formula is expressed as follows: ; In the formula, The first element in the attention weight matrix Line number The elements of the column are dimensionless; The first in the attention score matrix Line number Column elements; This represents the total number of spatial locations in the feature map, dimensionless. The weighted feature map... The calculation formula is expressed as follows: ; In the formula, This is the weighted feature map. Global feature aggregation result. The calculation formula is expressed as follows: ; In the formula, This is the result of global feature aggregation; This is the dimensionless weight matrix for the output mapping. (Inertia weights) The adaptive adjustment formula is expressed as follows: ; In the formula, For the first The inertial weight of each particle, dimensionless; This is the maximum value of the inertial weight, which is dimensionless and usually takes the value of 0.9. This is the minimum inertial weight, dimensionless, and typically takes the value of 0.4. The current iteration number is dimensionless. The total number of iterations is dimensionless. Indicator factor, dimensionless. The calculation formula is expressed as follows: ; In the formula, For the first The index factor of an individual particle, dimensionless; The fitness value is the globally optimal value in the population. For the first The individual optimal fitness value of each particle; This represents the average fitness value of all particles in the population. This is the arctangent function, with dimensionless input and output in rad. It is a function of the natural logarithm and is dimensionless. Pi is a dimensionless mathematical constant, with a value of approximately 3.14159. The particle velocity update formula is expressed as follows: ; In the formula, For particles In the Vidi The speed of each iteration; For particles In the Vidi The speed of each iteration; and The learning factor is dimensionless and typically takes a value of 2. and These are uniformly distributed random numbers, dimensionless, and range from 0 to 1. For particles In the The optimal position of an individual in dimension; For the population in the first The global optimal position of the dimension; The iteration number index is dimensionless. The particle position update formula is expressed as follows: ; In the formula, For particles In the Vidi The position of the next iteration; These are the position update coefficients, which are dimensionless. The rule for determining the value is: when hour, ;when hour, In the formula, Let Levy's flight distribution function be dimensionless. The calculation formula is expressed as follows: ; In the formula, This is the step scaling factor, dimensionless, with an empirical value of 0.01; The random numbers are normally distributed, with a mean of 0 and a standard deviation of . Dimensionless; These are random numbers that follow a normal distribution, with a mean of 0, a standard deviation of 1, and are dimensionless. This is a stability parameter, dimensionless, and typically takes a value of 1.5. The calculation formula is In the formula, Let gamma be the dimensionless gamma function. The formula for the random reset of out-of-bounds particles is expressed as follows: ; ; In the formula, and A dimensionless random number between 0 and 1.

[0052] To better understand and implement this invention, a specific application scenario of this invention is provided below as an example 2: To verify the effect of this invention, technicians built a numerical simulation analysis environment. The framework diagram of the "source-grid-load-storage" interaction mechanism of the distribution network throughout the process is shown below. Figure 2 As shown, technicians selected a typical urban power distribution network system as the research object. This system comprises 33 nodes and 32 branches, with a rated voltage of 12.66 kV and a base capacity of 100 MVA. Distributed photovoltaic power generation systems are connected to nodes 2, 6, 12, 18, and 25, and energy storage systems are configured at nodes 8, 15, and 28. The simulation time span is one complete operating cycle of 24 hours, with a time step of 1 hour.

[0053] Technicians first obtained basic data for the distribution network system, including the node load capacity, branch impedance, and maximum allowable power of each branch. The node load capacity reaches its maximum during peak hours, with node 5 having a peak load of 1850kW, node 10 2140kW, node 20 1920kW, and node 30 2380kW. The branch impedances used standard distribution network parameters, with an average impedance of 0.28Ω for branches 1 to 10. / km, the average impedance of branch lines 11 to 20 is 0.35 / km, the average impedance of branch lines 21 to 32 is 0.42 / km. The maximum allowable power for the branch lines is determined based on the conductor cross-section and operating temperature. The maximum allowable power for the main line is 4500kW, and the maximum allowable power for the branch lines is 2800kW.

[0054] Technicians abstracted the distribution network system into a weighted graph structure and calculated the betweenness centrality, tight centrality, and influence ranking values ​​of each node. The betweenness centrality calculation results showed that nodes 8, 15, 22, and 28 are key hub nodes in the distribution network, with betweenness centrality values ​​of 0.186, 0.174, 0.162, and 0.179, respectively. The tight centrality calculation results showed that nodes 12, 18, and 25 are located at the center of the distribution network, with tight centrality values ​​of 0.412, 0.438, and 0.425, respectively. The influence ranking values ​​were obtained through iterative calculation, with nodes 5, 10, and 20 having influence ranking values ​​of 0.224, 0.218, and 0.206, respectively. Technicians normalized the betweenness centrality value, the tight centrality value, and the influence ranking value, and then performed a weighted summation with weight coefficients of 0.4, 0.3, and 0.3 to obtain a comprehensive node importance score. The top 10 nodes in terms of the comprehensive node importance score were selected as the key node set, which includes nodes 5, 8, 10, 12, 15, 18, 20, 22, 25, and 28.

[0055] Technical personnel established a two-layer optimization model framework. The objective function of the upper-layer planning model is to minimize the total annual planning cost, with a discount rate of 0.08 and an equipment lifespan of 20 years, resulting in an annual equivalent coefficient of 0.1019. The total investment cost includes the unit capacity investment cost of the distributed photovoltaic system (4200 yuan / kW) and the unit capacity investment cost of the energy storage system (5800 yuan / kW). The system operation and maintenance cost is calculated at 2.5% of the installed capacity. The system electricity purchase cost adopts a time-of-use pricing mechanism: peak hour price is 0.95 yuan / kWh, normal hour price is 0.62 yuan / kWh, and off-peak hour price is 0.38 yuan / kWh. The lower-layer operation model has three objective functions: minimizing distribution network losses, optimizing node voltage quality, and minimizing load curve fluctuations. The weight coefficients for the first and second objective functions are set to 0.5 each. The weighted aggregation of the first and second objective functions yields the lower-layer single objective function.

[0056] Technicians imposed upper-level constraints on the upper-level planning model, setting the photovoltaic power factor to 0.9, and ensuring that the sum of the distributed photovoltaic installed capacity of all nodes in the key node set does not exceed 40% of the total load capacity of the distribution network. The energy storage system has a minimum state of charge (SOC) of 0.2, an upper SOC of 0.9, a charging efficiency of 0.95, a discharging efficiency of 0.93, and a charge / discharge ratio of 0.005. Technicians also imposed lower-level constraints on the lower-level operation model, setting the node voltage lower limit to 11.78kV, the node voltage upper limit to 13.54kV, and the maximum output power of the distributed photovoltaic system determined based on irradiance and installed capacity, with peak irradiance at 850W / m². The corresponding photovoltaic system capacity factor is 0.85.

[0057] Technicians input the decision variables and constraints of the upper-level planning model into the source-grid-load-storage collaborative scheduling optimization model. The initial particle sequence was set to a total of 60 particles, with 6 alpha particles, 18 aggressive particles, and 36 wandering particles. The initial values ​​of the chaotic variables were set to 0.4 and 0.6, respectively. The initial particle sequence was generated through chaotic mapping. The particle positions ranged from 0 to 5000 kW in the distributed photovoltaic installed capacity dimension and from 0 to 3000 kWh in the energy storage system rated capacity dimension. Technicians set the aggregation threshold to 0.15, the improvement threshold to 0.001, the maximum number of iterations to 1000, and the convergence accuracy requirement to 0.0001.

[0058] Technicians iteratively solved the problem by improving the wolf pack algorithm. The alpha wolf particle guides the population's movement direction based on the current global optimum, the aggressive wolf particles perform a fine-grained local search around the alpha wolf particle, and the wandering wolf particles randomly explore the search space. In the 58th iteration, the normalized population aggregation index dropped to 0.142, triggering the wandering wolf random jump mechanism, randomly selecting 7 wandering wolf particles to reinitialize their positions. In the 125th iteration, the improvement rate of the normalized objective function was less than the improvement threshold for 5 consecutive iterations, so the search step size of the aggressive wolf particles was reduced to 0.5 times the original search step size to enhance local search capabilities. Technicians then performed back-learning on the 6 particles with the worst fitness in the current iteration to generate the symmetric point positions and evaluate the fitness values ​​of the symmetric points. The fitness values ​​of 4 of the symmetric points were better than the original particle fitness values, and the original particle positions were replaced by the symmetric point positions.

[0059] like Figure 3As shown, the improved wolf pack algorithm exhibits a rapid decline followed by a gradual convergence in its global optimal objective function value during the iteration process. In the first 200 iterations, the objective function value decreases from an initial 8.42 × 10⁻⁶. The yuan dropped rapidly to 4.15× After initialization, a slow optimization phase begins. The convergence accuracy requirement is met on the 687th iteration, outputting the globally optimal source-grid-load-storage coordinated scheduling scheme. This scheme determines nodes 8, 15, and 28 as the optimal access points for distributed photovoltaic (PV) systems, with optimal installed capacities of 1850kW, 1620kW, and 1950kW, respectively. Nodes 10 and 22 are designated as the optimal access points for energy storage systems, with optimal rated capacities of 1280kWh and 1450kWh, respectively.

[0060] Technicians transmit the optimal installed capacity of the distributed photovoltaic system and the optimal rated capacity of the energy storage system as input parameters to the lower-level operation model, and obtain distributed photovoltaic power output data for each season based on historical meteorological data. On a typical spring day, photovoltaic power output peaks between 10:00 and 15:00, with a maximum output of 4185kW. On a typical summer day, the peak output period extends to 9:00 to 16:00, with a maximum output of 4852kW. On a typical autumn day, the peak output period occurs between 11:00 and 14:00, with a maximum output of 3976kW. On a typical winter day, the peak output period shortens to 11:00 to 13:00, with a maximum output of 3254kW. Figure 4 As shown in the figure, the daily charge and discharge power curve of the energy storage system shows that the energy storage system is in the charging state from 0:00 to 7:00, and the charging power of the energy storage system reaches a peak of 1420kW at 3:00. It is in the discharging state from 8:00 to 11:00 and from 17:00 to 22:00, and the discharging power of the energy storage system reaches a peak of 1650kW at 19:00. It is in the standby state from 12:00 to 16:00 and 23:00.

[0061] Technicians solved for the objective function value of the lower-level single-objective function, calculating the distribution network loss to be 178.6 kW, a significant reduction compared to the initial loss of 285.3 kW before optimization. The node voltage deviation was 0.0324, and the load curve fluctuation variance was 0.0186. Technicians analyzed the power flow distribution results of the distribution network at different time periods. The active power transmission values ​​of branches 1 to 8 were 2850 kW to 3680 kW during peak periods; the active power transmission values ​​of branches 9 to 16 were 1920 kW to 2540 kW during peak periods; and the active power transmission values ​​of branches 17 to 32 were 1150 kW to 1780 kW during peak periods. The voltage amplitude of each node remains between the lower limit and the upper limit of the node voltage throughout the 24 hours of the day. The voltage amplitude of node 5 during the peak period is 12.28kV, the voltage amplitude of node 15 during the peak period is 12.41kV, and the voltage amplitude of node 25 during the peak period is 12.35kV.

[0062] like Figure 5 As shown, a comparison of the load power curve and net load power curve of the distribution network system over a 24-hour period reveals that, through the optimization of the source-grid-load-storage coordinated dispatch scheme, the fluctuation amplitude of the net load power curve is significantly reduced, with the peak-to-valley difference decreasing from 4280kW before optimization to 2650kW after optimization. The load power curve exhibits significant load peaks between 8:00-9:00 and 18:00-20:00, with peak loads of 6850kW and 7420kW respectively. Through the output of the distributed photovoltaic system and the discharge power adjustment of the energy storage system, the peak values ​​of the net load power curve are reduced to 5320kW and 5780kW respectively. The load power curve is at a low point between 0:00 and 6:00, with a valley load of 3140kW. Through the absorption of charging power by the energy storage system, the valley value of the net load power curve is increased to 3920kW, achieving the effect of peak shaving and valley filling.

[0063] Technicians calculated the total annual planned cost, which is 2.29 × the investment cost of the distributed photovoltaic system. Yuan plus the investment cost of the energy storage system is 1.58 × Yuan, totaling 3.87× Multiplying the amount by the annual equivalent factor of 0.1019 yields an annualized investment cost of 3.94 × The system operation and maintenance cost is 2.5% of the total investment cost, calculated to be 9.68 × The system's electricity purchase cost is calculated based on annual electricity consumption over 8760 hours and time-of-use pricing. The optimized annual electricity purchase cost is 2.85 × The total annual planning cost is 3.34 × yuan. Yuan. Technical personnel compared and analyzed the technical and economic indicators before and after optimization. Before optimization, the total annual network loss of the distribution network was 1.56 × kWh, reduced to 9.84× after optimization. The number of node voltage overruns in kWh decreased from 187 to 0, and the variance of the load curve fluctuation decreased from 0.0425 to 0.0186.

[0064] Furthermore, this invention introduces a source-grid-load-storage collaborative scheduling optimization model, employing a global feature aggregation algorithm based on self-attention pooling. Through the interactive calculation of query vectors, key vectors, and value vectors, it adaptively learns the importance of different spatial locations in the feature map, automatically assigning higher weights to nodes with severe voltage exceedances and power transmission bottleneck branches. This enables the model to dynamically adjust the weights of features at each location according to the current operating status of the distribution network. Even when the distribution network is connected to a large number of distributed photovoltaic systems, causing frequent changes in power flow direction, it can still maintain stable optimization performance and strong robustness.

[0065] It should be noted that the variables involved in this invention are explained in detail in Tables 1, 2, 3, and 4.

[0066] Table 1. Variable Explanation Table (Part 1)

[0067] Table 2. Variable Explanation Table (Part Two)

[0068] Table 3. Variable Explanation Table (Part 3)

[0069] Table 4. Variable Explanation Table (Part Four)

Claims

1. A two-layer optimized scheduling method for source-grid-load-storage collaborative loss reduction, characterized in that, This includes acquiring basic data of the distribution network and constructing a set of key nodes, establishing a two-layer optimization model framework and applying constraints, iteratively solving the upper-layer planning model using an improved wolf pack algorithm, transmitting the optimization results of the upper-layer planning model to the lower-layer operation model and providing feedback on fitness evaluation criteria, calculating the improvement amount of the objective function and triggering an adaptive adjustment mechanism, determining convergence conditions, and outputting a source-grid-load-storage coordinated scheduling scheme. The two-layer optimization model framework establishes a closed-loop feedback mechanism between the upper-layer planning layer and the lower-layer operation layer, enabling the upper-layer decision variables to be dynamically adjusted based on the lower-layer operation results. The set of key nodes identifies the locations of nodes that significantly affect system performance using a graph theory-based node importance comprehensive scoring method. The improved wolf pack algorithm avoids getting trapped in local optima by using population aggregation monitoring and a wandering wolf random jumping mechanism in conjunction with a back-learning strategy.

2. The method according to claim 1, characterized in that, The steps for constructing the set of key nodes specifically involve abstracting the distribution network into a weighted graph structure, calculating the betweenness centrality value, tight centrality value, and influence ranking value of each node in the weighted graph structure, normalizing the betweenness centrality value, tight centrality value, and influence ranking value respectively, and then weighting and summing them according to the weight coefficients to obtain a comprehensive score of node importance.

3. The method according to claim 2, characterized in that, The normalization process specifically involves subtracting the minimum value from the betweenness centrality value and then dividing by the difference between the maximum and minimum values ​​to obtain the normalized betweenness centrality value, and subtracting the minimum value from the compact centrality value and then dividing by the difference between the maximum and minimum values ​​to obtain the normalized compact centrality value.

4. The method according to claim 3, characterized in that, The comprehensive score for node importance is specifically calculated as the normalized betweenness centrality value multiplied by the betweenness centrality weight coefficient, plus the normalized tight centrality value multiplied by the tight centrality weight coefficient, plus the normalized influence ranking value multiplied by the influence ranking weight coefficient.

5. The method according to claim 4, characterized in that, The two-layer optimization model framework includes an upper-layer planning model and a lower-layer operation model. The objective function of the upper-layer planning model is to minimize the total annual planning cost. The decision variables of the upper-layer planning model are the location of distributed photovoltaic access, the installed capacity of distributed photovoltaic, the location of energy storage system access, and the rated capacity of energy storage system at each node in the set of key nodes.

6. The method according to claim 5, characterized in that, The first objective function of the lower-level operation model is to minimize the distribution network loss, the second objective function is to optimize the node voltage quality, and the third objective function is to minimize the load curve fluctuation. The first, second, and third objective functions are weighted and aggregated through objective function weight coefficients to obtain the lower-level single objective function.

7. The method according to claim 6, characterized in that, The improved wolf pack algorithm's population initialization specifically involves generating an initial particle sequence through chaotic mapping, setting the number of alpha wolf particles, aggressive wolf particles, and wandering wolf particles in the initial particle sequence, with the alpha wolf particles guiding the population's movement direction based on the current global optimal solution, aggressive wolf particles performing local fine-grained searches around the alpha wolf particles, and wandering wolf particles randomly exploring within the search space.

8. The method according to claim 7, characterized in that, The chaotic mapping generates an initial particle sequence by setting two chaotic variables with different initial values ​​and using a sine function to generate a chaotic sequence. The chaotic variable is at the [missing information - likely a specific value]. The value of the second iteration is equal to 1.18 multiplied by 2 divided by the first iteration. The sine value of the chaotic variable in the next iteration is used to transform the chaotic variable into the range of particle position and particle velocity values ​​through a linear mapping.

9. The method according to claim 8, characterized in that, The population aggregation monitoring specifically involves calculating the average Euclidean distance from the position of all particles to the population center, dividing the average Euclidean distance by the diagonal length of the search space to normalize it, and obtaining a normalized population aggregation index. It is then determined whether the normalized population aggregation index exceeds the aggregation threshold.

10. The method according to claim 9, characterized in that, The random jumping mechanism of the wandering wolf is triggered when the normalized population aggregation index exceeds the aggregation threshold. It randomly selects some particles from the wandering wolf particles and reinitializes the positions of these particles to random positions within the search space.

Images