A power distribution network source-grid-load multi-level collaborative planning method based on cluster division

Abstract

The application relates to a power distribution network source-grid-load multi-level collaborative planning method based on cluster division, which comprises the following steps: comprehensively considering multiple influence factors, performing cluster division on a power distribution network by using an improved adaptive K-means clustering algorithm to obtain a cluster division result; constructing a power distribution network double-layer joint planning model, wherein the power distribution network double-layer joint planning model comprises an upper layer model and a lower layer model; solving the upper layer model by using a particle swarm algorithm of an annealing strategy, and solving the lower layer model by using a deep deterministic policy gradient algorithm. The application decomposes a complex problem into a source, storage site and capacity determination problem and an emergency voltage control problem for solving, comprehensively considers the uncertainty of a source-load to obtain each site and capacity determination result in a cluster range, fully combines the characteristics of a voltage regulation device, formulates a flexible adjustment strategy to cope with problems caused by voltage fluctuation, and guarantees the safety operation of the power distribution network and the rationality of planning.

Description

Technical Field

[0001] This invention relates to the field of distribution network planning technology, and in particular to a multi-level collaborative planning method for distribution network sources-grid-load based on cluster partitioning. Background Technology

[0002] With the reform of the new energy system both domestically and internationally, China's new energy industry has developed rapidly. The integration of new energy sources, represented by a large number of distributed photovoltaic (DPV) power sources into the distribution network, has impacted the safe and stable operation of the distribution network system. This has significantly altered the power flow direction and voltage distribution of the distribution network, exacerbating the difficulty of coordinating the three elements of power source, load, and storage, and posing a severe challenge to the collaborative planning of the distribution network.

[0003] Currently, the main approach to power grid planning that includes distributed photovoltaic (DPV) systems and energy storage systems (ESS) is to expand the area and lines of the distribution network and conduct centralized optimization research on distribution network planning based on the investment and operating costs of DPV and ESS. This method is suitable for applications with low DPV penetration rates. However, when DPV penetration rates increase in the future, the scale and dimensionality of the distribution system will increase dramatically. If this method is continued for planning, it will be difficult to solve large-scale and high-dimensional distribution network scenarios.

[0004] Furthermore, based on the results of cluster partitioning, site selection and capacity planning for DPVs and ESSs within the cluster have also been conducted: a two-layer planning model based on virtual partitioning has been established. The upper layer optimizes the access location and capacity of DPVs with the goal of minimizing annual costs, while the lower layer optimizes the site selection and capacity planning of virtual partitions and ESSs with the goal of minimizing the variance of equivalent loads. Although the above studies have investigated site selection and capacity planning, effectively mitigating the impact of DPV access on the distribution network under high penetration rates, they all neglect the impact of source-load uncertainties on grid operation.

[0005] For distribution networks with high DPV penetration in the future, a zoned and hierarchical planning approach is appropriate. Therefore, it is necessary to conduct collaborative planning on a cluster basis in distribution networks containing high DPV penetration. Further in-depth research is needed on how to establish planning models that account for source-load uncertainties and achieve rationalized specification layout and collaborative operation. Summary of the Invention

[0006] To overcome the impact of source-load uncertainty on power grid operation, the present invention aims to provide a multi-level collaborative planning method for power grid source-grid-load based on cluster division. This method comprehensively considers the uncertainty of source-load to obtain the site selection and capacity determination results within the cluster, and then fully combines the characteristics of voltage regulating equipment to formulate flexible adjustment strategies to deal with the problems caused by voltage fluctuations. This ensures the safe operation of the distribution network while guaranteeing the rationality of the planning.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a multi-level collaborative planning method for power distribution network sources-grid-load based on cluster partitioning, the method comprising the following sequential steps:

[0008] (1) Taking into account various influencing factors, the improved adaptive K-means clustering algorithm is used to divide the power distribution network into clusters, and the cluster division results are obtained.

[0009] (2) Based on the cluster partitioning results obtained in step (1), construct a two-layer joint planning model for the power distribution network, which includes an upper-layer model and a lower-layer model.

[0010] (3) The upper-level model is solved by the particle swarm algorithm with annealing strategy, and the lower-level model is solved by the gradient algorithm with deep deterministic strategy.

[0011] Step (1) specifically refers to: considering multiple influencing factors, and in order to better understand the internal structure and functional characteristics of the power grid cluster while ensuring a balanced size of the power grid cluster, a comprehensive index τ is proposed, the specific formula of which is as follows:

[0012]

[0013] In the formula: ζ1, ζ2, and ζ3 represent the weighting coefficients of the indicators; λ is the load balance index. Q represents the node membership index, and Q represents the electrical module degree fusion index.

[0014] An improved adaptive K-means clustering algorithm is obtained by introducing the Calinski-Harabasz index as an indicator to evaluate the clustering effect, based on the adaptive K-means clustering algorithm. Specifically, this refers to:

[0015] Calculate the comprehensive index τ related to each node, use this comprehensive index τ as the input data of the improved adaptive K-means clustering algorithm, randomly select k samples as cluster centers, and use them as the cluster centers of the first cluster.

[0016] For each new sample point, calculate its distance to each cluster center and assign it to the cluster containing the nearest cluster center. The Euclidean distance formula is used for this calculation.

[0017]

[0018] In the formula: a is the dimension of the cluster space; k and b k , are k-dimensional vectors representing the k-th attribute; D(a,b) is the distance to the cluster centers;

[0019] The cluster centers are updated using the Euclidean distance formula, which is the mean of all samples belonging to that cluster.

[0020] For each cluster, the sum of the squared distances between all sample points in the cluster and the cluster center is calculated as the sum of squared errors for that cluster. The sum of squared errors reflects the distribution of sample points within the cluster. The smaller the sum of squared errors, the closer the sample points in the cluster are to the cluster center, and the better the clustering effect.

[0021] For each cluster, the sum of the squared distances of all sample points in the cluster to the global mean is calculated as the sum of squared dispersion of the cluster. The sum of squared dispersion reflects the difference between the cluster and other clusters. The larger the sum of squared dispersion, the greater the difference between the cluster and other clusters, and the better the clustering effect.

[0022] Calculate the Calinski-Harabasz index, where a larger CH value indicates better clustering performance. The formula is as follows:

[0023]

[0024] In the formula: Tr(B k ) is the trace of the inter-cluster scatter matrix, Tr(W) k ) is the trace of the cluster scatter matrix. The total number of samples; CH is the Calinski-Harabasz index. The total number of samples;

[0025] The formulas for calculating the inter-cluster scatter matrix and the intra-cluster scatter matrix are as follows:

[0026]

[0027]

[0028] In the formula: For each sample in the i-th cluster; The mean of the samples in the i-th cluster; C is the mean of all samples; i For the i-th cluster;

[0029] The maximum number of samples is chosen as the iteration termination condition k. max ;

[0030] The maximum CH is obtained, and the optimal clustering result, i.e., the cluster partitioning result, is output.

[0031] In step (2), the objective function of the upper-level model is to minimize the investment and operating costs of distributed photovoltaic and energy storage systems, distribution network losses, and grid stability: f = min{f1,f2,f3};

[0032] The investment and operating costs of the distributed photovoltaic and energy storage system are as follows:

[0033]

[0034] In the formula, N DG The number of distributed photovoltaic units. Let N be the cost of the k-th distributed photovoltaic system. E The number of energy storage elements. The cost of the k-th energy storage element, The operating cost of the energy storage element on day r;

[0035] The network loss of the distribution network is:

[0036]

[0037] Where: N l This refers to the number of branches in the distribution network. For branch l within time period t b Line loss; T is the total time period;

[0038] The power grid stability is:

[0039]

[0040] Where: N K U represents the number of nodes in the cluster network. i U represents the actual voltage amplitude. N The rated voltage value of the nodes in the distribution network;

[0041] The constraints of the upper-level model include DPV capacity and output constraints, power balance constraints, cluster interaction power constraints, node voltage upper and lower limits constraints, and ESS charging and discharging power and state of charge constraints.

[0042] The DPV capacity and output constraints are as follows:

[0043]

[0044] In the formula: This represents the number of nodes within cluster J. p is the DPV capacity allowed to be installed on node i within cluster J. DPV,J,t For the DPV output at time t within cluster J; P DPV,J Where J is the DPV capacity in cluster J; C is the number of clusters;

[0045] The power balance constraint is:

[0046]

[0047] Where: NCP This is the number of branches on the main network; p ESS,J,t p represents the ESS charging and discharging power at time t within cluster J; ESS,J,t Positive and negative values ​​represent discharging and charging, respectively; The active power consumed by the load on node i within cluster J at time t;

[0048] The cluster interaction power constraint is:

[0049]

[0050] In the formula: p IP,l,t Let be the interaction power of branch l between clusters at time t; N represents the maximum power allowed for inter-cluster interaction branch l; Cl The total number of interaction branches between clusters;

[0051] The upper and lower limits of the node voltage are constrained as follows:

[0052]

[0053] In the formula: These are the upper and lower limits of the voltage deviation at node i, respectively;

[0054] The ESS charging and discharging power and state of charge constraints are as follows:

[0055]

[0056] In the formula: η t η is the charge / discharge efficiency of the ESS at time t. d ESS charging efficiency; η c For ESS discharge efficiency; u E,J,t The standard charging and discharging point of ESS within cluster J at time t is defined as follows: greater than 0 indicates the discharging state, less than 0 indicates the charging state, and equal to 0 indicates the floating charging state. S represents the maximum output power within cluster J; J,t Let J be the state of charge at time t within cluster J; and These represent the maximum and minimum values ​​of the state of charge; S0 is the initial state of charge quantity of the ESS; E ESS,J This represents the rated capacity within cluster J.

[0057] In step (2), the objective function of the lower-level model is to minimize the node voltage deviation:

[0058]

[0059] In the formula: U i U is the node voltage of the i-th node;N This refers to the rated voltage of the power distribution network in this area; N K This refers to the number of nodes in the cluster network.

[0060] The constraints of the lower-level model include power flow balance constraints, photovoltaic inverter constraints, energy storage output constraints, and SVC output constraints.

[0061] The power flow balance constraint is:

[0062]

[0063] In the formula: The reactive power consumed by the load on node i within cluster J at time t; P represents the active power consumed by the load on node i within cluster J at time t. M (t) and Q M (t) represents the active and reactive power transmitted from the main distribution network at time t, respectively; Q DPV,J,t Q represents the distributed photovoltaic reactive power within cluster J at time t; SVC,J,t The power output is provided to the static reactive power compensation device within cluster J at time t. For branch l within time period t b Line loss; p DPV,J,t For the DPV output at time t within cluster J; p ESS,J,t The charging and discharging power of the ESS within cluster J at time t; Let be the reactive power loss of branch lb during time period t;

[0064] The constraints of the photovoltaic inverter are:

[0065]

[0066] Q i,PV 2 (t)≤S inv 2 -P i,PV 2 (t)

[0067] In the formula: These are the upper and lower limits of the active power output of the photovoltaic inverter, S inv 2 Indicates the capacity of the inverter; P i,PV (t) represents the active power output of node i at time t; Q i,PV (t) represents the reactive power output of node i at time t;

[0068] The energy storage output constraint:

[0069]

[0070]

[0071] In the formula, ΔP i ESS (t) represents the change in active power output of the energy storage system at node i at time t; and These represent the minimum and maximum active power output of the energy storage at node i at time t, respectively. and P represents the minimum and maximum changes in active power output at node i at time t, respectively; i,ESS (t) is the active power output of the energy storage system at node i at time t;

[0072] The SVC output constraint is:

[0073]

[0074]

[0075] Where: ΔQ i SVC (t) represents the reactive power change of the SVC output at node i; and These represent the upper and lower limits of reactive power output by the SVC at node i, respectively. and These represent the upper and lower limits of reactive power ramping at node i of the SVC output at time t; Q i,SVC (t) is the reactive power output of the SVC at node i at time t.

[0076] Step (3) specifically refers to solving the upper-level model using the particle swarm optimization algorithm with annealing strategy, which includes the following steps:

[0077] (3a) Set the initial population size, maximum weight coefficient, minimum weight coefficient, learning factor, initial temperature and cooling rate;

[0078] (3b) Calculate the swarm fitness using the power flow algorithm, set the current position pBest and the optimal position gBest of the particles, and determine whether the convergence criteria of the particle swarm algorithm with annealing strategy are met based on the constraints and objective function of the upper-level model. If they are met, output gBest directly; otherwise, proceed to step (3c).

[0079] (3c) Update particle velocity and position, update current particle position and optimal position, and calculate fitness difference;

[0080] (3d) Determine if the cooling rate is satisfied. If it is, output the best position gBest. Otherwise, return to step (3c) and continue to update the particle velocity and position. Repeat the calculation until the optimal solution is satisfied and output gBest.

[0081] The use of a deep deterministic policy gradient algorithm to solve the lower-level model specifically refers to: first, setting the number of episodes for the deep deterministic policy gradient algorithm, training each episode a certain number of times, and after a certain number of episodes of training, when the agent's reward value stabilizes, indicating that the deep deterministic policy gradient algorithm has converged; then, adjusting the reactive power and energy storage of the photovoltaic system, reducing the active power, and using the operation of the photovoltaic inverter to quickly control the bus voltage within the safety threshold.

[0082] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, based on the division of the distribution network cluster, the present invention proposes a two-layer joint planning model for the distribution network, including an upper-layer model and a lower-layer model; then, the two-layer joint planning model is solved by the particle swarm optimization algorithm with annealing strategy and the gradient algorithm with deep deterministic strategy; Second, the present invention decomposes the complex problem into source and storage addressing and capacity determination problems and emergency voltage control problems, comprehensively considers the uncertainty of source-load to obtain the addressing and capacity determination results within the cluster; and further, by fully combining the characteristics of voltage regulating equipment, a flexible adjustment strategy is formulated to deal with the problems caused by voltage fluctuations, ensuring the safe operation of the distribution network while ensuring the rationality of the planning. Attached Figure Description

[0083] Figure 1 This is a flowchart of the method of the present invention;

[0084] Figure 2 The flowchart of the particle swarm optimization algorithm for annealing strategy;

[0085] Figure 3 This is a framework diagram of the deep deterministic policy gradient algorithm;

[0086] Figure 4 A geographical topology map;

[0087] Figure 5 This is a schematic diagram of the cluster partitioning result based on a cluster modularity function that takes flexibility into account.

[0088] Figure 6 This is a schematic diagram of the cluster partitioning result of the improved adaptive K-means algorithm in this invention;

[0089] Figure 7 Schematic diagram for DPV site selection and capacity determination;

[0090] Figure 8 Schematic diagram for ESS site selection and capacity determination. Detailed Implementation

[0091] like Figure 1As shown, a multi-level collaborative planning method for power distribution networks based on cluster partitioning is proposed. This method includes the following sequential steps:

[0092] (1) Taking into account various influencing factors, the improved adaptive K-means clustering algorithm is used to divide the power distribution network into clusters, and the cluster division results are obtained.

[0093] (2) Based on the cluster partitioning results obtained in step (1), construct a two-layer joint planning model for the power distribution network, which includes an upper-layer model and a lower-layer model.

[0094] (3) The upper-level model is solved by the particle swarm algorithm with annealing strategy, and the lower-level model is solved by the gradient algorithm with deep deterministic strategy.

[0095] Step (1) specifically refers to: considering multiple influencing factors, and in order to better understand the internal structure and functional characteristics of the power grid cluster while ensuring a balanced size of the power grid cluster, a comprehensive index τ is proposed, the specific formula of which is as follows:

[0096]

[0097] In the formula: ζ1, ζ2, and ζ3 represent the weighting coefficients of the indicators; λ is the load balance index. Q represents the node membership index, and Q represents the electrical module degree fusion index.

[0098] An improved adaptive K-means clustering algorithm is obtained by introducing the Calinski-Harabasz index as an indicator to evaluate the clustering effect, based on the adaptive K-means clustering algorithm. Specifically, this refers to:

[0099] Calculate the comprehensive index τ related to each node, use this comprehensive index τ as the input data of the improved adaptive K-means clustering algorithm, randomly select k samples as cluster centers, and use them as the cluster centers of the first cluster.

[0100] For each new sample point, calculate its distance to each cluster center and assign it to the cluster containing the nearest cluster center. The Euclidean distance formula is used for this calculation.

[0101]

[0102] In the formula: a is the dimension of the cluster space; k and b k , are k-dimensional vectors representing the k-th attribute; D(a,b) is the distance to the cluster centers;

[0103] The cluster centers are updated using the Euclidean distance formula, which is the mean of all samples belonging to that cluster.

[0104] For each cluster, the sum of the squared distances between all sample points in the cluster and the cluster center is calculated as the sum of squared errors for that cluster. The sum of squared errors reflects the distribution of sample points within the cluster. The smaller the sum of squared errors, the closer the sample points in the cluster are to the cluster center, and the better the clustering effect.

[0105] For each cluster, the sum of the squared distances of all sample points in the cluster to the global mean is calculated as the sum of squared dispersion of the cluster. The sum of squared dispersion reflects the difference between the cluster and other clusters. The larger the sum of squared dispersion, the greater the difference between the cluster and other clusters, and the better the clustering effect.

[0106] Calculate the Calinski-Harabasz index, where a larger CH value indicates better clustering performance. The formula is as follows:

[0107]

[0108] In the formula: Tr(B k ) is the trace of the inter-cluster scatter matrix, Tr(W) k ) is the trace of the cluster scatter matrix. The total number of samples; CH is the Calinski-Harabasz index. The total number of samples;

[0109] The formulas for calculating the inter-cluster scatter matrix and the intra-cluster scatter matrix are as follows:

[0110]

[0111]

[0112] In the formula: For each sample in the i-th cluster; The mean of the samples in the i-th cluster; C is the mean of all samples; i For the i-th cluster;

[0113] The maximum number of samples is chosen as the iteration termination condition k. max ;

[0114] The maximum CH is obtained, and the optimal clustering result, i.e., the cluster partitioning result, is output, such as... Figure 6 As shown; by Figure 5 As shown, in the cluster partitioning based on the cluster modularity function considering flexibility, isolated nodes formed separate clusters (cluster 6), resulting in uneven cluster load. This is compared to... Figure 5 , Figure 6It can be seen that the present invention can make the number of cluster nodes more balanced. This balanced distribution is beneficial to subsequent cluster control and operation, because the number of nodes in each cluster is relatively balanced, making cluster control and management easier.

[0115] In step (2), the objective function of the upper-level model is to minimize the investment and operating costs of distributed photovoltaic and energy storage systems, distribution network losses, and grid stability: f = min{f1,f2,f3};

[0116] The investment and operating costs of the distributed photovoltaic and energy storage system are as follows:

[0117]

[0118] In the formula, N DG The number of distributed photovoltaic units. Let N be the cost of the k-th distributed photovoltaic system. E The number of energy storage elements. The cost of the k-th energy storage element, The operating cost of the energy storage element on day r;

[0119] The network loss of the distribution network is:

[0120]

[0121] Where: N l This refers to the number of branches in the distribution network. For branch l within time period t b Line loss; T is the total time period;

[0122] The power grid stability is:

[0123]

[0124] Where: N K U represents the number of nodes in the cluster network. i U represents the actual voltage amplitude. N The rated voltage value of the nodes in the distribution network;

[0125] The constraints of the upper-level model include DPV capacity and output constraints, power balance constraints, cluster interaction power constraints, node voltage upper and lower limits constraints, and ESS charging and discharging power and state of charge constraints.

[0126] The DPV capacity and output constraints are as follows:

[0127]

[0128] In the formula: This represents the number of nodes within cluster J. p is the DPV capacity allowed to be installed on node i within cluster J. DPV,J,t For the DPV output at time t within cluster J; P DPV,J Where J is the DPV capacity in cluster J; C is the number of clusters;

[0129] The power balance constraint is:

[0130]

[0131] Where: N CP This is the number of branches on the main network; p ESS,J,t p represents the ESS charging and discharging power at time t within cluster J; ESS,J,t Positive and negative values ​​represent discharging and charging, respectively; The active power consumed by the load on node i within cluster J at time t;

[0132] The cluster interaction power constraint is:

[0133]

[0134] In the formula: p IP,l,t Let be the interaction power of branch l between clusters at time t; N represents the maximum power allowed for inter-cluster interaction branch l; Cl The total number of interaction branches between clusters;

[0135] The upper and lower limits of the node voltage are constrained as follows:

[0136]

[0137] In the formula: These are the upper and lower limits of the voltage deviation at node i, respectively;

[0138] The ESS charging and discharging power and state of charge constraints are as follows:

[0139]

[0140] In the formula: η t η is the charge / discharge efficiency of the ESS at time t. d ESS charging efficiency; η c For ESS discharge efficiency; u E,J,t The standard charging and discharging point of ESS within cluster J at time t is defined as follows: greater than 0 indicates the discharging state, less than 0 indicates the charging state, and equal to 0 indicates the floating charging state. S represents the maximum output power within cluster J; J,t Let J be the state of charge at time t within cluster J; and These represent the maximum and minimum values ​​of the state of charge; S0 is the initial state of charge quantity of the ESS; EESS,J This represents the rated capacity within cluster J.

[0141] In step (2), the objective function of the lower-level model is to minimize the node voltage deviation:

[0142]

[0143] In the formula: U i U is the node voltage of the i-th node; N This refers to the rated voltage of the power distribution network in this area; N K This refers to the number of nodes in the cluster network.

[0144] The constraints of the lower-level model include power flow balance constraints, photovoltaic inverter constraints, energy storage output constraints, and SVC output constraints.

[0145] The power flow balance constraint is:

[0146]

[0147] In the formula: The reactive power consumed by the load on node i within cluster J at time t; P represents the active power consumed by the load on node i within cluster J at time t. M (t) and Q M (t) represents the active and reactive power transmitted from the main distribution network at time t, respectively; Q DPV,J,t Q represents the distributed photovoltaic reactive power within cluster J at time t; SVC,J,t The power output is provided to the static reactive power compensation device within cluster J at time t. For branch l within time period t b Line loss; p DPV,J,t For the DPV output at time t within cluster J; p ESS,J,t The charging and discharging power of the ESS within cluster J at time t; Let be the reactive power loss of branch lb during time period t;

[0148] The constraints of the photovoltaic inverter are:

[0149]

[0150] Q i,PV 2 (t)≤S inv 2 -P i,PV 2 (t)

[0151] In the formula: These are the upper and lower limits of the active power output of the photovoltaic inverter, S inv 2Indicates the capacity of the inverter; P i,PV (t) represents the active power output of node i at time t; Q i,PV (t) represents the reactive power output of node i at time t;

[0152] The energy storage output constraint:

[0153]

[0154]

[0155] In the formula, ΔP i ESS (t) represents the change in active power output of the energy storage system at node i at time t; and These represent the minimum and maximum active power output of the energy storage at node i at time t, respectively. and P represents the minimum and maximum changes in active power output at node i at time t, respectively; i,ESS (t) is the active power output of the energy storage system at node i at time t;

[0156] The SVC output constraint is:

[0157]

[0158]

[0159] In the formula: The table shows the reactive power change of the SVC output at node i. and These represent the upper and lower limits of reactive power output by the SVC at node i, respectively. and These represent the upper and lower limits of reactive power ramping at node i of the SVC output at time t; Q i,SVC (t) is the reactive power output of the SVC at node i at time t.

[0160] like Figure 2 As shown, step (3) specifically refers to solving the upper-level model using the particle swarm optimization algorithm with annealing strategy, which includes the following steps:

[0161] (3a) Set the initial population size, maximum weight coefficient, minimum weight coefficient, learning factor, initial temperature and cooling rate;

[0162] (3b) Calculate the swarm fitness using the power flow algorithm, set the current position pBest and the optimal position gBest of the particles, and determine whether the convergence criteria of the particle swarm algorithm with annealing strategy are met based on the constraints and objective function of the upper-level model. If they are met, output gBest directly; otherwise, proceed to step (3c).

[0163] (3c) Update particle velocity and position, update current particle position and optimal position, and calculate fitness difference;

[0164] (3d) Determine if the cooling rate is satisfied. If it is, output the best position gBest. Otherwise, return to step (3c) and continue to update the particle velocity and position. Repeat the calculation until the optimal solution is satisfied and output gBest.

[0165] like Figure 3 As shown, the use of the deep deterministic policy gradient algorithm to solve the lower-level model specifically refers to: first, setting the number of episodes for the deep deterministic policy gradient algorithm, training each episode a certain number of times, and after training a certain number of episodes, when the agent's reward value stabilizes, it indicates that the deep deterministic policy gradient algorithm has converged; then, adjusting the reactive power and energy storage of the photovoltaic system, reducing the active power, and using the operation of the photovoltaic inverter to quickly control the bus voltage within the safety threshold.

[0166] By combining the Particle Swarm Optimization (PSO) algorithm with Simulated Annealing (SA), an annealing-strategy PSO is obtained. This leverages the characteristic of SA to temporarily accept some inferior solutions under certain probabilistic control, improving the local optimization capability of the PSO algorithm. The specific process is as follows: Figure 2 As shown.

[0167] The improved algorithm, the annealing-based particle swarm optimization, observes that during particle movement, if the next position is better than the current position, the particle moves to the next position; conversely, if the next position is worse than the current position, the particle does not move directly but moves with a certain probability, controlled by the initial and current temperatures. This prevents particles from blindly rushing to the next position, instead "testing" it with a certain probability before acting. Therefore, when the temperature decreases sufficiently, particles will not easily jump out of the "hopeful" search region, thus enhancing the particle's local search capability.

[0168] Deep Deterministic Policy Gradient (DDPG) is a deep reinforcement learning algorithm used to solve problems in continuous action spaces and participates in Markov decision processes. The DDPG algorithm framework is as follows: Figure 3 As shown.

[0169] Example 1

[0170] To verify the effectiveness of the above theory, this invention selects a real 35kV / 10kV distribution network for simulation verification. The geographical topology map of the distribution network to be planned for the nodes and lines is as follows: Figure 4 As shown in Table 1, there are 33 nodes in this network structure, including 24 power load nodes. The node power parameters are shown in Table 2. The total length of the planned network is 91.07 km.

[0171] Table 1 Line Parameters

[0172]

[0173]

[0174] Table 2 Node Power Parameters

[0175]

[0176] In this invention, when considering the operational cost, distribution network losses, and voltage stability issues of the particle swarm optimization algorithm employing an annealing strategy, the initial parameters are set as follows: initial population size is 20; maximum number of iterations is 80; inertia weight coefficient is 0.9; minimum inertia weight coefficient is 0.4; learning factor is 2 for all; initial temperature T = 1000℃; cooling rate is 0.9. To highlight the advantages of the proposed hierarchical planning strategy, this invention constructs three planning schemes and compares and analyzes the DPV and ESS location and capacity determination problems of the distribution network under different conditions using three different schemes.

[0177] Option 1: Without dividing the distribution network into clusters, a traditional heuristic algorithm is used for site selection and capacity determination based on the entire distribution network. The capacity of DPV and ESS connected to each node is directly planned to determine the candidate installation nodes and capacities of DPV and ESS.

[0178] Option 2: Based on the cluster partitioning results considering the flexibility of the cluster modularity function, establish the capacity and addressing models of DPV and ESS within the cluster, and use a genetic algorithm to solve them, thereby determining the candidate installation nodes and capacity of DPV and ESS.

[0179] Solution 3: Based on the cluster partitioning results proposed in this invention, a multi-objective addressing and capacity-determining model is established at the upper layer, and the particle swarm optimization algorithm with annealing strategy is used to solve the problem, thereby optimizing the access positions of each DPV and ESS within the cluster.

[0180] Under different schemes, the DPV and ESS site selection planning and capacity of each cluster access are as follows: Figure 7 , Figure 8As shown in Table 3, the total branch losses, total costs, and average voltage deviation of the distribution network for each scheme are as follows: Table 3 Optimization results of different schemes for DPV and ESS

[0181]

[0182] Comparing schemes 1 and 3, the cluster site selection planning results are analyzed. Figure 7 , Figure 8 As shown in Table 3, the two schemes differ significantly:

[0183] 1) In Scheme 3, the total capacity of DPV connected to the distribution network and the total capacity of ESS energy storage system are 1.09 and 0.65 times that of Scheme 1, respectively;

[0184] 2) Option 1 has a higher ESS capacity and a more concentrated distribution;

[0185] 3) In Scheme 3, the branch road loss was reduced by 52.62% compared with the unplanned version, and in Scheme 1, the branch road loss was reduced by 17.21% compared with the unplanned version.

[0186] 4) In Scheme 3, the average voltage deviation of the power distribution system improved by 78.58% compared with the unplanned system; in Scheme 1, the average voltage deviation of the power distribution system improved by 53.57% compared with the unplanned system.

[0187] 5) Option 1 has a lower investment cost, which is RMB 15,100 less than Option 3, but Option 3 has a higher DPV penetration rate.

[0188] Comparing schemes 2 and 3, by Figure 7 , Figure 8 As shown in Table 3, the following differences exist:

[0189] 1) The number of ESS connected to the distribution network in Scheme 2 is 0.33 times that in Scheme 3;

[0190] 2) In Scheme 2, the DPV capacity is concentrated in cluster 1, while the installed capacity of the other clusters is relatively low;

[0191] 3) Compared with Scheme 3, Scheme 2 has a 67.30% increase in branch losses;

[0192] 4) Compared with Scheme 2, Scheme 3 improved the average voltage deviation of the power distribution system by 183.33%;

[0193] 5) Scheme 2 has a lower investment cost compared to Scheme 3 because Scheme 2 involves fewer ESS (Emergency Safe Module) installations. However, it cannot cope with the grid interference caused by the uncertainty of DPV output under load and high penetration rates. Overall, Scheme 3, based on cluster division for DPV and ESS planning, increases photovoltaic penetration and improves the distribution of DPV and ESS across clusters, meeting the requirements of future new energy development.

[0194] When a power disturbance is introduced into the IEEE 33-node distribution network at a certain moment, some nodes exceed the safety limit. Therefore, further processing of the area where the nodes are located is required using the DDPG algorithm.

[0195] The DDPG agent was trained for 300 episodes, with each episode consisting of 10 training samples. After 60 episodes, the reward value of the DDPG agent stabilized, indicating that the algorithm had converged. The reactive power and energy storage of the photovoltaic system need to be adjusted first. Since the node voltage in the region is not under control, active power needs to be further reduced. At this point, by operating the photovoltaic inverter, the bus node voltage can be quickly controlled within a safe threshold.

[0196] To verify the superiority of the proposed method, it was compared with the traditional Sparrow Optimization Algorithm (SSA) and Particle Swarm Optimization (POS) algorithms. The voltage control results under different strategies are shown in Table 4. SSA achieved a response time of 5.52 s and an average voltage deviation of 0.0029 ms. p.u. The POS response time for control adjustment is 4.84 seconds, and the average voltage deviation is 0.0031. p.u. When using reinforcement learning control, the response time is 0.11s and the average voltage deviation is 0.0023. p.u. As can be seen, although the average voltage deviation of the voltage control strategy of the present invention is not much different, it greatly shortens the response time of the algorithm, which fully demonstrates the efficiency and accuracy of the reinforcement learning algorithm.

[0197] Table 4. Voltage control results under different strategies

[0198]

[0199] In summary, this invention decomposes the complex problem into source and storage addressing and capacity determination problems and emergency voltage control problems. It comprehensively considers the uncertainties of source-load to obtain the addressing and capacity determination results within the cluster. Furthermore, it fully integrates the characteristics of voltage regulating equipment to formulate flexible adjustment strategies to deal with the problems caused by voltage fluctuations, ensuring the safe operation of the distribution network while ensuring the rationality of the planning.

Claims

1. A multi-level collaborative planning method for power distribution network sources-grid-load based on cluster partitioning, characterized in that: The method includes the following steps in sequence: (1) Taking into account multiple influencing factors, the improved adaptive K-means clustering algorithm is used to divide the distribution network into clusters, and the cluster division results are obtained. Considering multiple influencing factors, in order to better understand the internal structure and functional characteristics of the power grid clusters while ensuring the balance of power grid cluster size, a comprehensive index is proposed. The specific formula is as follows: ； In the formula: , and These represent the weighting coefficients of the indicators; As an indicator of load balance, This is a node membership index. For electrical modularity integration indicators; An improved adaptive K-means clustering algorithm is obtained by introducing the Calinski-Harabasz index as an indicator to evaluate the clustering effect based on the adaptive K-means clustering algorithm. (2) Based on the cluster partitioning results obtained in step (1), construct a two-layer joint planning model for the power distribution network, which includes an upper-layer model and a lower-layer model. The objective function of the upper-level model is to minimize the investment and operating costs of distributed photovoltaic and energy storage systems, distribution network losses, and grid stability. ; The investment and operating costs of the distributed photovoltaic and energy storage system are as follows: ； In the formula, The number of distributed photovoltaic units. Let the cost of the k-th distributed photovoltaic power generation be... The number of energy storage elements. The cost of the k-th energy storage element, The operating cost of the energy storage element on day r; The network loss of the distribution network is: ； In the formula: This refers to the number of branches in the distribution network. for Branch roads during the period Line loss; This is the total time period; The power grid stability is: ； In the formula: This refers to the number of nodes in the cluster network. This represents the actual voltage amplitude. The rated voltage value of the nodes in the distribution network; The constraints of the upper-level model include DPV capacity and output constraints, power balance constraints, cluster interaction power constraints, node voltage upper and lower limits constraints, and ESS charging and discharging power and state of charge constraints. The objective function of the lower-level model is to minimize the node voltage deviation. ； In the formula: For the first The node voltage of each node; This refers to the rated voltage value of the power distribution network in this area; This refers to the number of nodes in the cluster network. The constraints of the lower-level model include power flow balance constraints, photovoltaic inverter constraints, energy storage output constraints, and SVC output constraints. (3) The upper-level model is solved by the particle swarm algorithm with annealing strategy, and the lower-level model is solved by the gradient algorithm with deep deterministic strategy.

2. The multi-level collaborative planning method for power distribution network sources-grid-load based on cluster partitioning as described in claim 1, characterized in that: Step (1) specifically refers to: the improved adaptive K-means clustering algorithm specifically refers to: Calculate the comprehensive indicators related to each node. This comprehensive indicator As input data for the improved adaptive K-means clustering algorithm, randomly selected One sample is used as the cluster center, and it is used as the cluster center of the first cluster; For each new sample point, calculate its distance to each cluster center and assign it to the cluster containing the nearest cluster center. The Euclidean distance formula is used for this calculation. ； In the formula: Let be the dimension of the cluster space; and They are respectively A dimensional vector, used to represent the first dimension One attribute; The distance to the cluster center; The cluster centers are updated using the Euclidean distance formula, which is the mean of all samples belonging to that cluster. For each cluster, the sum of the squared distances between all sample points in the cluster and the cluster center is calculated as the sum of squared errors for that cluster. The sum of squared errors reflects the distribution of sample points within the cluster. The smaller the sum of squared errors, the closer the sample points in the cluster are to the cluster center, and the better the clustering effect. For each cluster, the sum of the squared distances of all sample points in the cluster to the global mean is calculated as the sum of squared dispersion of the cluster. The sum of squared dispersion reflects the difference between the cluster and other clusters. The larger the sum of squared dispersion, the greater the difference between the cluster and other clusters, and the better the clustering effect. Calculate the Calinski-Harabasz index, and The larger the value, the better the clustering effect. The calculation formula is as follows: ； In the formula: The trace of the inter-cluster scatter matrix. The trace of the cluster scatter matrix. The total number of samples; The Calinski-Harabasz index. The total number of samples; The formulas for calculating the inter-cluster scatter matrix and the intra-cluster scatter matrix are as follows: ； In the formula: For the first Samples from each cluster; For the first The sample mean in each cluster; The mean of all samples; For the first A cluster; The maximum number of samples was chosen as the iteration termination condition. ; The maximum CH is obtained, and the optimal clustering result, i.e., the cluster partitioning result, is output.

3. The multi-level collaborative planning method for power distribution network sources-grid-load based on cluster partitioning as described in claim 1, characterized in that: In step (2), the DPV capacity and output constraints are as follows: ； In the formula: For clusters The number of nodes within; For clusters Nodes within Allowable DPV capacity to be installed; For clusters Inside DPV output at any given moment; The DPV capacity in cluster J; Number of clusters; The power balance constraint is: ； In the formula: This is the number of branches on the main network; For clusters Inside ESS charging and discharging power at any given time; Positive and negative values ​​represent discharging and charging, respectively; For clusters internal nodes On the load The active power consumed at any given moment; The cluster interaction power constraint is: ； In the formula: For inter-cluster branches exist Interaction power at any given moment; For inter-cluster interaction branches Maximum power exchange allowed to pass; The total number of interaction branches between clusters; The upper and lower limits of the node voltage are constrained as follows: ； In the formula: , They are nodes The upper and lower limits of voltage deviation; The ESS charging and discharging power and state of charge constraints are as follows: ； In the formula: for The charge / discharge efficiency of the time-limited ESS; Improve ESS charging efficiency; For ESS discharge efficiency; For cluster Inside The ESS charging and discharging standard point at any given time indicates the state of discharge if the value is greater than 0, the state of charge if the value is less than 0, and the state of float charge if the value is equal to 0. For cluster Maximum output power within; For cluster Inside The state of charge at any given moment; and These are the maximum and minimum values ​​of the state of charge, respectively; For the initial state of charge of the ESS; For cluster Rated capacity within.

4. The multi-level collaborative planning method for power distribution network sources-grid-load based on cluster partitioning as described in claim 1, characterized in that: In step (2), the power flow balance constraint is: ； In the formula: For cluster internal nodes On the load The constant consumption of reactive power; For cluster internal nodes On the load The active power consumed at any given moment; and They are respectively Active and reactive power transmitted from the main distribution network at all times; For cluster Inside Distributed photovoltaic reactive power at any given time; For cluster Inside Output of the static reactive power compensation device at any given moment; for Branch roads during the period Line loss; For cluster Inside DPV output at any given moment; For cluster Inside ESS charging and discharging power at any given time; Let be the reactive power loss of branch lb during time period t; The constraints of the photovoltaic inverter are: ； ； In the formula: , These are the upper and lower limits of the active power output of the photovoltaic inverter. Indicates the capacity of the inverter; This represents the active power output of node i at time t; This represents the reactive power output of node i at time t; The energy storage output constraint: ； ； In the formula, Represented as Time Node The change in active power output of the energy storage system; and They are respectively represented as Time Node The minimum and maximum active power output of the upper energy storage; and They are respectively represented as Time Node The minimum and maximum changes in active power output of the upper energy storage system; It is the active power output of the energy storage system at node i at time t; The SVC output constraint is: ； ； In the formula: The table shows nodes. The SVC outputs the reactive power change. and They are nodes The upper and lower limits of reactive power output by the upper SVC; and They are respectively Time Node The upper and lower limits of reactive power ramping output of the upper SVC; yes Time Node The SVC outputs reactive power.

5. The multi-level collaborative planning method for power distribution network sources-grid-load based on cluster partitioning as described in claim 1, characterized in that: Step (3) specifically refers to solving the upper-level model using the particle swarm optimization algorithm with annealing strategy, which includes the following steps: (3a) Set the initial population size, maximum weight coefficient, minimum weight coefficient, learning factor, initial temperature and cooling rate; (3b) Calculate the swarm fitness using the power flow algorithm, set the current position pBest and the optimal position gBest of the particles, and determine whether the convergence criteria of the particle swarm algorithm with annealing strategy are met based on the constraints and objective function of the upper-level model. If they are met, output gBest directly; otherwise, proceed to step (3c). (3c) Update particle velocity and position, update current particle position and optimal position, and calculate fitness difference; (3d) Determine whether the cooling rate is satisfied. If it is, output the best position gBest. Otherwise, return to step (3c) and continue to update the particle velocity and position. Repeat the calculation until the optimal solution is satisfied and output gBest. The use of a deep deterministic policy gradient algorithm to solve the lower-level model specifically refers to: first, setting the number of episodes for the deep deterministic policy gradient algorithm, training each episode a certain number of times, and after a certain number of episodes of training, when the agent's reward value stabilizes, indicating that the deep deterministic policy gradient algorithm has converged; then, adjusting the reactive power and energy storage of the photovoltaic system, reducing the active power, and using the operation of the photovoltaic inverter to quickly control the bus voltage within the safety threshold.

Images