Power distribution network multi-resource cooperative reactive power regulation method considering weak communication condition

Abstract

The invention relates to the technical field of reactive power regulation, and discloses a power distribution network multi-resource cooperative reactive power regulation method considering weak communication conditions. According to the method, multi-dimensional weighted missing data repair is carried out by utilizing the similarity between different transformer areas and different time sections, so that the data repair accuracy and the data processing speed are improved; a novel power distribution network multi-resource cooperative reactive power regulation model based on improved multi-agent deep reinforcement learning is constructed, the method is solved by adopting a multi-agent deep deterministic strategy gradient algorithm, and cooperative operation of four resources including a photovoltaic inverter, an energy storage inverter, a charging pile inverter and an SVG in the power distribution network is fully mobilized. The condition that the voltage exceeds the upper and lower limits is effectively governed, the network voltage is maintained at a stable level, and the voltage quality and operation economy of the power distribution network are remarkably improved. The method can be used for software development of a power distribution network control center, and quick, economical and flexible power distribution network multi-resource cooperative reactive power regulation can still be achieved under the weak communication condition.

Description

Technical Field

[0001] The present invention relates to the technical field of reactive power regulation, and particularly to a multi-resource collaborative reactive power regulation method for a distribution network considering weak communication conditions. Background Art

[0002] With the global energy transformation and upgrading, promoting green transformation and the construction of new distribution networks in the energy field provides policy guarantees for the wide access of distributed energy, and also puts forward higher requirements for improving the reactive power regulation ability of the distribution network and optimizing the operation of the power system.

[0003] Using multiple resources such as photovoltaic inverters, energy storage inverters, charging pile inverters, and static var generators (SVG) for reactive power regulation of the distribution network is a new voltage regulation method. By utilizing the reactive power regulation capabilities of resources such as photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG, the problem of voltage exceeding the upper and lower limits in the new distribution network can be effectively solved. Through a global optimization control strategy, the inverters and SVG maintain the system voltage at a stable level by absorbing / generating reactive power.

[0004] However, the existing reactive power regulation methods do not fully exploit the potential of distribution network resources. Usually, only a single distributed resource, such as photovoltaic or photovoltaic-storage system, is considered, and multiple resources such as photovoltaic, energy storage, charging pile inverters, and SVG are not used for collaborative reactive power regulation. In fact, considering the diversity and flexibility of devices such as charging pile inverters and SVG, using multiple resources for collaborative reactive power regulation will provide a more efficient and flexible solution for voltage regulation of the new distribution network.

[0005] The effectiveness of the existing reactive power regulation strategies is based on stable communication conditions. At this time, the control strategy can coordinate the operation of each substation area to achieve overall reactive power compensation of the system. However, in practical applications, there are many devices in the new distribution network, the amount of data uploaded from the intelligent fusion terminal of the substation area to the distribution network control center is large, and the geographical locations are usually relatively scattered. Communication is easily affected by factors such as bad weather, and weak communication situations are likely to occur, that is, data loss occurs during the transmission of substation area data. In contrast, the amount of data sent from the distribution network control center to the substation area is small, and the layout and maintenance of communication devices are relatively concentrated. Using communication methods such as dedicated lines, the reliability is higher. For weak communication situations, most of the existing research is based on decentralized control strategies or fixed rules. The core idea is to achieve uniform distribution or proportional adjustment of reactive power compensation through adaptive adjustment between local substation areas, so as to ensure the basic operation of the power grid. Since the substation area can only make decisions based on local information, it is difficult to judge the real-time operation status of the whole system. In the event of emergencies such as load fluctuations, the coordination between substation areas is insufficient, which may lead to uneven reactive power compensation and affect the voltage stability of the power grid.

[0006] Currently, the research on missing data repair mainly falls into three categories: interpolation method, model-based repair method, and machine learning method. The interpolation method estimates the missing values through existing adjacent data points. However, when the relationship between the missing data is complex, it is easy to introduce large errors. Both the model-based repair method and the machine learning method use the existing data to train a model for predicting missing data. Such methods require a large amount of training data, have a high computational cost, and rely on the accuracy of the model. In fact, the operating characteristics between different time sections and different substations may be similar. The missing data can be quickly repaired through similarity measurement and weighted evaluation, so as to improve the reactive power compensation coordination of the system under weak communication conditions.

[0007] In summary, the problems of the existing technologies are as follows: 1) The existing reactive power regulation methods rarely consider the influence of weak communication conditions on the reactive power regulation decision-making of the distribution network. When communication fails and substation data is missing, under the existing methods, the substation only makes reactive power decisions based on local information, and cannot grasp the overall operating conditions of the system, resulting in insufficient coordination. The missing data repair method based on multi-dimensional weighting proposed in the present invention does not require a large amount of training data sets, can quickly and accurately repair the missing data, and greatly ensures the real-time and overall planning of the decision-making.

[0008] 2) Most of the existing reactive power regulation optimization models only perform reactive power regulation for a single distributed resource (such as photovoltaic or photovoltaic energy storage inverter), and only focus on minimizing network losses and voltage offsets. They fail to fully explore and utilize the potential of collaborative reactive power regulation of multiple resources in the new distribution network, have poor flexibility, and do not comprehensively consider the cost required for the distribution network to maintain voltage stability. Summary of the Invention

[0009] In view of the above problems, the purpose of the present invention is to provide a method for collaborative reactive power regulation of multiple resources in a distribution network considering weak communication conditions. By using the similarity between different substations and different time sections, multi-dimensional weighted missing data repair is performed, which improves the accuracy of data repair and the data processing speed; fully mobilizes the collaborative operation of four resources in the distribution network, namely photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG, effectively controls the voltage from exceeding the upper and lower limits, maintains the grid voltage at a stable level, and significantly improves the voltage quality and operating economy of the distribution network. The technical solutions are as follows: A method for collaborative reactive power regulation of multiple resources in a distribution network considering weak communication conditions, comprising the following steps: Step 1: Obtain the data of different substations and different time sections of the distribution network; Step 2: Use the hash algorithm to calculate the hash values of the data respectively at the substation intelligent fusion terminal as the sender and the distribution network control center as the receiver, and compare them to detect whether the reactive power regulation demand data is missing during the transmission process; Step 3: According to the missing situation of the data, utilize the similarity between different regions and different time sections of the distribution network, and perform multi-dimensional comprehensive weighted repair on the missing data from the dimensions of overall interconnection, regional interconnection, overall self-connection, and regional self-connection; Step 4: With the goals of minimizing the network loss of the system, the voltage offset of all nodes in the system, and the cost of purchasing active and reactive power from the upstream network by the distribution network, establish constraint conditions including distribution network constraints, rated capacity constraints, and charge and discharge constraints of the energy storage system, and construct a multi-resource collaborative reactive power regulation model for the distribution network based on improved multi-agent deep reinforcement learning; Step 5: Use the multi-agent deep deterministic policy gradient algorithm for solution; Step 6: Mobilize four types of resources, namely photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG, in the distribution network to operate collaboratively to achieve reactive power regulation and voltage optimization.

[0010] The beneficial effects of the present invention are as follows: 1) The present invention proposes a method for repairing missing data based on multi-dimensional weighting, which can perform multi-dimensional weighted repair through the similarity between different regions and different time sections. This not only avoids relying on a large amount of data sets, but also improves the accuracy and operation speed of data repair, greatly ensuring the real-time and overall planning of reactive power regulation, and avoiding the decline of the reactive power regulation effect of the distribution network caused by weak communication.

[0011] 2) The present invention proposes a new multi-resource collaborative reactive power regulation model for the distribution network based on the improved multi-agent deep reinforcement learning (Multi-Agent Deep Reinforcement Learning, MADRL) algorithm, and uses the multi-agent deep deterministic policy gradient (Multi-Agent Deep Deterministic Policy Gradient, MADDPG) algorithm to solve the proposed method. This model not only mobilizes four types of resources, namely photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG, to work collaboratively, greatly improving the flexibility and effectiveness of reactive power regulation and voltage optimization, and better maintaining the voltage stability of the distribution network, but also realizes the minimum voltage offset, minimum network loss, and lowest cost by optimizing the objective function, enhancing the operation economy of the distribution network. Description of the Drawings

[0012] Figure 1 It is a flowchart of the multi-resource collaborative reactive power regulation method for the distribution network considering weak communication conditions of the present invention. Detailed Embodiments

[0013] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention proposes a novel multi - resource collaborative reactive power regulation method for distribution networks considering weak communication conditions. By utilizing the similarity between different substations and different time sections, multi - dimensional weighted missing data repair is performed, which improves the accuracy of data repair and the data processing speed. A novel multi - resource collaborative reactive power regulation model for distribution networks based on an improved MADRL algorithm is constructed, and the MADDPG algorithm is used to solve the proposed method, fully mobilizing the collaborative operation of four resources in the distribution network, namely photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG, effectively controlling the situation of voltage exceeding the upper and lower limits, maintaining the grid voltage at a stable level, and significantly improving the voltage quality and operation economy of the distribution network. The present invention can be used in the software development of distribution network control centers, and can still achieve fast, economical, and flexible multi - resource collaborative reactive power regulation of distribution networks under weak communication conditions.

[0014] The algorithm flow chart of the present invention is as Figure 1 shown. The specific process is as follows: 1. Data missing assessment based on the hash algorithm: Data integrity is crucial for the reactive power regulation decision of the system. However, under weak communication conditions, data loss often occurs during data transmission, thus affecting the reactive power regulation effect. Data missing includes single - item missing and multi - item missing. For example, the missing of a certain type of data in a single substation is regarded as single - item missing, while the missing of multiple types of data in a single substation or the missing of a certain type of data in multiple substations belongs to multi - item missing. These missing data may be related to each other or independent. To accurately evaluate the data integrity, the present invention uses the hash algorithm, calculates the hash values of the data respectively at the intelligent fusion terminal (sender) of the substation and the control center (receiver) of the distribution network using the SHA - 256 algorithm, and compares them to detect whether there is missing data of reactive power regulation requirements during the transmission process. The data missing rate can be calculated by the following formula: (1); In the formula, MIS represents the data missing rate, C represents the total number of substations to be evaluated in the distribution network, B represents the total number of attributes of the reactive power regulation requirement data to be evaluated in a single substation (such as key parameters like voltage, load reactive power, load active power, etc.), N represents the number of attributes with empty values. Taking multi - item missing as an example, in a distribution network with 30 substations, each substation has 8 attributes of reactive power regulation requirement data. If a certain substation has 6 attributes with empty values, the data missing rate is 2.5%. When the data missing rate is non - zero, it is considered to be in a weak communication situation.

[0015] 2. Missing data repair method based on multi - dimensional weighting: For a certain area M with missing data, its horizontal neighborhood refers to the measurement values from different areas at the same time section, the vertical neighborhood refers to the measurement values from the same area at different time sections, and the local data window refers to the data group containing the horizontal and vertical neighborhood measurement values of M within a given time domain.

[0016] 1) Overall interconnection dimension: Under the overall interconnection dimension, the repair of missing data is achieved through the inverse distance weighting method, and the missing data is interpolated based on the neighborhood. Using the cosine similarity between variables as the weighting factor, for a certain area M with missing data, its horizontal neighborhood data is used for filling, and the formula is as follows: (2); In the formula, R em is the data repair value obtained under the overall interconnection dimension, n 1 is the number of complete data of the remaining areas in the horizontal neighborhood of M, r i is the measurement value of the complete data, λ is the weight attenuation factor based on similarity (the higher the similarity, - λ the greater the weight). c i is the cosine similarity between the vector c (x 11 , x 12 , …, x 1n ) where the area with complete data is located and the vector b (x 21 , x 22 ,..., x 2n ) where the area M with missing data is located under the same time section. The cosine similarity calculation formula is as follows: (3); In the formula, n is the number of attributes included in the area vector; and are the norms of vector c and vector b respectively.

[0017] 2) Regional interconnection dimension: Under the regional interconnection dimension, based on the user collaborative filtering recommendation method, the missing data is repaired by mining the regional mutual correlation between areas. According to the measurement values in the local data window, the similarity between the remaining areas and the area M with missing data is calculated, and the missing data is weighted and averaged with this as the weight. The formula is as follows: (4); In the formula, R pmis the data repair value obtained under the regional interconnection dimension, n 2 is the number of complete data points in the horizontal neighborhood of M in the local data window. is the similarity between the vector c where the data-complete area is located and the vector b where the area M with missing data is located in the local data window at the same time section, and the calculation formula is as follows: (5); In the formula, N 1 is the total number of time sections with complete area data in the local data window.

[0018] 3) Overall self-connection dimension: Under the overall self-connection dimension, through the idea of simple exponential smoothing, using the autocorrelation of a certain area M with missing data, and based on the historical data before the occurrence of the missing data, the missing data is repaired with the values of other time sections. The formula is as follows: (6); In the formula, R es is the data repair value obtained under the overall self-connection dimension, n 3 is the number of complete data in the longitudinal neighborhood of area M, r u is at the time section u the measured value of the complete data, γ` is the weight decay factor based on time series (the shorter the time interval, the smaller this value, and the value range is 0-1), f u is the difference between the complete data of this area and the time section where the missing value is located.

[0019] 4) Regional self-connection dimension: The data repair under the regional self-connection dimension is similar to the regional interconnection dimension. Using the measured values in the local data window, calculate the similarity between different time sections, and use this as the weight to interpolate the missing data. The formula is as follows: (7); In the formula, R ps is the data repair value obtained under the regional self-connection dimension, n 4 is the number of complete data points in the longitudinal neighborhood of M in the local data window. δ u is the similarity between different time sections of the area with missing data in the local data window, and the calculation formula is as follows: (8); In the formula, N 2 is the number of parameters that are complete under time sections u and q in the local data window; r uv is the v th parameter's measured value at time section u , r qv is the v th parameter's measured value at time section q .

[0020] 5) Multi-dimensional data repair weighted calculation: Integrate the calculation results of the above four dimensions and calculate the data repair value according to the following formula: (9); In the formula, , , and are the weight values assigned to the four dimensions, y is the residual. By minimizing the squared error between the predicted value and the actual value, each substation area variable is trained to determine the optimal weight value.

[0021] 3. New type of distribution network multi-resource collaborative reactive power regulation model based on improved MADRL algorithm: In the new type of distribution network multi-resource collaborative reactive power regulation, the target decision-making layer and the optimization control layer will act as independent computational intelligent agents and jointly participate in system optimization.

[0022] 1) Objective function: (10); In the formula, N d is the number of decision instructions, , and are the regulation coefficients of each optimization objective, is the network loss of the system, is the voltage deviation of all nodes in the system, is the cost of purchasing active and reactive power by the distribution network from the upstream network.

[0023] 2) Constraint conditions: ① Distribution network constraints: (11); (12); (13); In the formula, N bus refers to the number of system buses, and are respectively t the active power and reactive power obtained from the upstream network at time and are respectively t the active power and reactive power generated by the photovoltaic system on bus j at time and are respectively t the active power and reactive power generated by the energy storage system on bus j at time and are respectively t the active power and reactive power generated by the charging pile on bus j at time is t the reactive power generated by the SVG on bus j at time and are respectively t the active load and reactive load of bus j at time and are respectively t the active power loss and reactive power loss of the network at time U j is the voltage on bus j When multiple resources are used for collaborative power generation, it is necessary to ensure the balance of supply and demand for each bus, which is expressed as follows:

[0024] (14); (15); (15); In the formula, H is a set of buses connected to bus j , G jp is the conductance from bus j to bus p , B jp is the susceptance from bus j to bus p , θ t,jp is the voltage phase difference between bus j and bus p , and are respectively t the voltages of bus j and busp voltage

[0025] ② Rated capacity constraint: The reactive power generated by the SVG needs to be less than its rated capacity. Similarly, the active and reactive powers generated by the PV, energy storage, and charging pile also need to be less than their rated capacities, as shown below: (16); (17); (18); (19); In the formula, is the rated capacity of the SVG at bus j , is the rated capacity of the PV at bus j , is the rated capacity of the energy storage at bus j , is the rated capacity of the charging pile at bus j .

[0026] ③ Energy storage system charge and discharge constraint: When the energy storage system is charging and discharging, the state of charge of the energy storage battery on bus j needs to meet the following conditions: (20); (21); In the formula, is t the state of charge of the energy storage system on bus j at time . Considering the service life of the energy storage battery, the minimum state of charge is set to 0.3, and the maximum state of charge

[0027] is set to 0.9. To effectively solve the problem of coordinated reactive power regulation of multiple resources in the distribution network, the optimization problem can be constructed as a Markov decision process and solved using an improved MADRL algorithm.

[0028] 1) State space: The state space S is used to represent all possible states of the system at each moment. For any state s k ∈ S , it can be represented in the following way: (22); In the formula, is kThe voltage value of the bus in the distribution network at a certain moment j ; , , are respectively k the active power of the PV inverter, energy storage inverter, and charging pile inverter on the bus at a certain moment j ; , , , are respectively k the reactive power of the PV inverter, energy storage inverter, charging pile inverter, and SVG on the bus at a certain moment j ; is k the state of charge of the energy storage battery on the bus at a certain moment j ; is the current network loss of the distribution network; and are respectively the active power demand information and reactive power demand information of the distribution network, and are respectively k the active demand and reactive demand for the upstream network at a certain moment.

[0029] 2) Action space: The action space A is composed of all possible actions that the agent can take. For any action a j,k ∈ A , it can be represented in the following way: (23); In the formula, , and are k the active power that the PV inverter, energy storage inverter, and charging pile inverter on the bus at a certain moment j need to emit; , , and are k the reactive power that the PV inverter, energy storage inverter, charging pile inverter, and SVG on the bus at a certain moment j need to emit; and are k the active power and reactive power purchased from the upstream network at a certain moment.

[0030] 3) State transition matrix: The state transition matrix T describes the probability of the system state changing over time after executing a certain action and can be expressed as: (24); Wherein, p ([[]] s k+1 | s k , a j,k ) refers to the original state s k after taking an action a j,k and transferring to the state s k+1 with probability.

[0031] 4) Reward function and discount factor: In order to minimize the voltage deviation, network loss, and active and reactive power purchase costs, the objective function shown in formula (10) can be transformed into a single-step reward function R i , that is: (25); Wherein, ε i is a penalty factor used to measure whether the relevant indicators exceed the constraint value. When the system indicators exceed the constraint value, the reward function will increase the penalty, and vice versa, to ensure that the agent provides sufficient safety margin for the operation of the new distribution network while meeting the system constraints.

[0032] In reinforcement learning, the discount factor determines the weight of the long-term reward and reflects the importance of future rewards. When the discount factor is close to 1, it emphasizes future rewards, and when it is close to 0, it emphasizes the current reward. Therefore, in the reactive power optimization scenario, the discount factor γ can be calculated using the following formula: (26); Wherein, is the attenuation coefficient of reactive power. The smaller the value, the slower the attenuation speed of reactive power, which means that the long-term reward occupies a greater weight in the decision-making.

[0033] 5) Agent policy: The agent policy defines the actions taken by the agent in a specified state. In the multi-resource collaborative reactive power regulation problem proposed in the present invention, the agent policy can be expressed in the following manner: (27); Wherein, F ([[]] s k , a j,k ) represents in the states k Take actions below a j,k Expected return after softmax The function can be transformed into a probability distribution.

[0034] 4. Solving the MADDPG algorithm: In solving the reactive power optimization problem, the MADDPG algorithm constructs two neural networks: an actor network and a critic network. The actor network takes the current state as input and outputs the corresponding action. The critic network takes the current state and action as input, calculates the temporal difference error, and outputs the value corresponding to that state-action. A neural network with three fully connected layers is constructed as the actor network. The dimension of the input layer corresponds to the state matrix, and the dimension of the output layer corresponds to the number of reactive power regulation resources in the distribution network. To introduce non-linear relationships, the ReLU function is used as the activation function. The critic network is also composed of three fully connected networks, and the value in different states can be evaluated using the following formula: (28); In the formula, θ i are the parameters of the value network, s is the overall state, a is the overall action; s i is the i observation of the F i ( s , a 1 , a 2 ,..., a n ) is the evaluation function, taking the action a i and the state s as input values and the action-value function as the output value, is the action policy, used to select actions during the process of learning the target policy, is the partial derivative vector of θ i with respect to is the parameter θ i the performance metric corresponding to the policy, E denotes taking the expected value.

[0035] Selecting actions according to probability will result in the problem of low convergence efficiency. Therefore, the MADDPG algorithm replaces the stochastic policy with a deterministic policy, and the gradient of the agent can be expressed by the following formula: (29); Wherein, is the strategy of the i th agent, D ([ s , a , r , s` ) is the experience replay buffer for recording the experiences of all agents, is the state s when taking the action a visible feature vector, s` is the state after taking the action.

[0036] The loss function of the critic network can be expressed in the following way: (30); Wherein, is the delay parameter, s k+1 and a k+1 are k the state and action corresponding to the +1 moment.

[0037] In summary, the present invention proposes a multi - resource collaborative reactive power regulation method for distribution networks considering weak communication conditions. By using the similarity between different regions and different time sections of the distribution network, multi - dimensional comprehensive weighted repair of missing data is carried out from the dimensions of overall interconnection, regional interconnection, overall self - connection, and regional self - connection. It does not require a large amount of data sets as support, can effectively improve the accuracy of data repair and the operation speed; with the minimum voltage deviation, minimum network loss, and minimum active and reactive power purchase cost as the objective function, a new multi - resource collaborative reactive power regulation model for distribution networks based on the improved MADRL algorithm is constructed, and the MADDPG algorithm is used to solve the proposed method, fully mobilizing four resources in the distribution network, namely photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG, realizing reactive power regulation and voltage optimization, improving the flexibility and effectiveness of decision - making, and enhancing the operation economy of the distribution network.

Claims

1. A distribution network multi-resource coordinated reactive power regulation method considering weak communication conditions, characterized in that: The following steps are involved: Step 1: Obtain data from different areas and time sections of the distribution network; Step 2: Use the hash algorithm to calculate the hash value of the data at the intelligent fusion terminal of the substation as the sending end and the distribution network control center as the receiving end, and compare them to detect whether the reactive power regulation demand data is missing during the transmission process; Step 3: According to the missing data, the similarities between different distribution areas and different time sections of the distribution network are used to perform multi-dimensional comprehensive weighted repair of the missing data from the overall interconnection dimension, regional interconnection dimension, overall self-connection dimension and regional self-connection dimension; Step 4: With the goal of minimizing the network loss of the system, the voltage offset of all nodes in the system, and the cost of the distribution network purchasing active and reactive power from the upstream network, establish constraints including distribution network constraints, rated capacity constraints, and energy storage system charging and discharging constraints, and build a distribution network multi-resource collaborative reactive power regulation model based on improved multi-agent deep reinforcement learning; Step 5: Use the multi-agent deep deterministic policy gradient algorithm to solve; Step 6: Mobilize the four resources of photovoltaic inverters, energy storage inverters, charging pile inverters and SVG in the distribution network to work together to achieve reactive power regulation and voltage optimization.

2. A distribution network multi-resource coordinated reactive power regulation method considering weak communication conditions according to claim 1, characterized in that: In step 2, the specific process of detecting whether the reactive power regulation demand data is missing during the transmission process is: The SHA-256 algorithm is used to calculate the hash value of the data in the intelligent fusion terminal of the substation area and the control center of the distribution network, and the data missing rate is calculated: （1）； In the formula, MIS represents the data missing rate, C Indicates the total number of distribution network areas to be evaluated. B Indicates the total number of attributes of reactive power regulation demand data to be evaluated in a single substation. N Indicates the number of attributes whose values ​​are empty; When the data missing rate is non-zero, it is considered to be a weak communication situation and the reactive power regulation demand data is missing.

3. A distribution network multi-resource coordinated reactive power regulation method considering weak communication conditions according to claim 1, characterized in that: In step 3, the specific steps of the multi-dimensional comprehensive weighted repair include: Step 3.1: Calculate the data patch value under the overall interconnection dimension: In the overall interconnected dimension, missing data is repaired by the inverse distance weighted method, and the missing data is interpolated based on the neighborhood. The cosine similarity between variables is used as a weighting factor. For a station M with missing data, its horizontal neighborhood data is used to fill it. The formula is as follows: （2）； In the formula, R em It is the data repair value obtained under the overall interconnection dimension. n 1 is the number of complete data of other stations in the horizontal neighborhood of station M, r i is the measurement value of the complete data, λ is a weight decay factor based on similarity; c i is the vector c (x 11 , x 12 , ..., x 1n ) and the vector b(x 21 , x 22 , ..., x 2n ), the cosine similarity between the same time section; the cosine similarity calculation formula is as follows: （3）； In the formula, n is the number of attributes contained in the area vector; and are the magnitudes of vector c and vector b respectively; and are the elements in vector c and vector b respectively; Step 3.2: Calculate the data patch value under the regional interconnection dimension: In the dimension of regional interconnection, based on the user collaborative filtering recommendation method, the missing data is repaired by mining the regional correlation between each substation. According to the measurement value in the local data window, the similarity between the remaining substations and the substation M with missing data is calculated, and the missing data is weighted averaged with this as the weight. The formula is as follows: （4）； In the formula, R pm It is the data repair value obtained under the regional interconnection dimension. n 2 is the number of complete data points in the lateral neighborhood of station M in the local data window; is the vector c (x 11 , x 12 , ...,x 1n ) and the vector b(x 21 , x 22 , ..., x 2n ), the similarity between the same time section is calculated as follows: （5）； In the formula, N 1 is the total number of complete time sections of the station area data in the local data window; Step 3.3: Calculate the data patch value under the overall self-connection dimension: Under the overall self-connected dimension, the idea of ​​simple exponential smoothing is used to utilize the autocorrelation of a certain area M with missing data. Based on the historical data before the missing data occurs, the missing data is repaired with the values ​​of other time sections. The formula is as follows: （6）； In the formula, R es It is the data repair value obtained under the overall self-connection dimension. n 3 is the number of complete data in the vertical neighborhood of station M, r u It is in the time section u The measured value of the complete data is γ` is a time-based weight decay factor, f u It is the difference between the time section where the complete data of station area M is located and the time section where the missing value is located; Step 3.4: Calculate the data patch value under the regional self-connection dimension: In the regional self-connection dimension, the similarity between different time sections is calculated using the measured values ​​in the local data window, and the missing data is interpolated based on this as the weight. The formula is as follows: （7）； In the formula, R ps This is the result obtained under the regional self-connection dimension. n 4 is the number of complete data points in the longitudinal neighborhood of station M in the local data window; δ u It is the similarity between different time sections of the data missing area in the local data window. The calculation formula is as follows: （8）； In the formula, N 2 is the local data window, in the time section u and time section q The number of complete parameters under average; It is v Parameters in time section u The measured value of r qv It is v Parameters in time section q The measured value of Step 3.5: Multi-dimensional data patching weighted calculation: Integrate the calculation results of the above four dimensions and calculate the final data repair value according to the following formula: : （9）； In the formula, , , and are the weights assigned to the four dimensions, y is the residual; Each station variable is trained by minimizing the square error between the predicted value and the actual value to determine the optimal weight.

4. A distribution network multi-resource coordinated reactive power regulation method considering weak communication conditions according to claim 1, characterized in that: Step 4: The specific process is as follows: Step 4.1: Determine the objective function F : （10）； In the formula, N d is the number of decision instructions, is the network loss of the system, is the voltage offset of all nodes in the system, It is the cost of the distribution network purchasing active and reactive power from the upstream network; , and is the adjustment coefficient of each optimization objective; Step 4.2: Determine the constraints: 1) Grid constraints: （11）； （12）； （13）； In the formula, N bus Refers to the number of system buses, and They are t The active power and reactive power obtained from the upstream network at all times, and They are t Moment bus j Active power and reactive power generated by photovoltaics; and They are t Moment bus j The active power and reactive power generated by the upper energy storage system; and They are t Moment bus j The active power and reactive power generated by the charging pile, yes t Moment bus j Reactive power generated by SVG; and They are t Moment bus j Active load and reactive load; and They are t The active power loss and reactive power loss of the network at all times, U j It is a busbar j The voltage on In order to ensure the supply and demand balance of each busbar when multiple resources are coordinated for power generation, the constraints are expressed as follows: （14）； （15）； In the formula, H Is connected to the bus j A set of busbars, G jp It is a busbar j To busbar p The conductivity, B jp It is a busbar j To busbar p of electrical susceptance; θ t,jp is busbar j and busbar p The voltage phase difference between and They are t Moment bus j and busbar p Voltage; 2) Rated capacity constraints: The reactive power generated by SVG is less than its rated capacity. The active power and reactive power generated by photovoltaic, energy storage and charging piles are also less than their rated capacity, as shown below: （16）； （17）； （18）； （19）； In the formula, It is a busbar j Rated capacity of SVG, It is a busbar j The PV rated capacity, It is a busbar j The rated capacity of energy storage is It is a busbar j The rated capacity of the charging pile; 3) Energy storage system charging and discharging constraints: When the energy storage system is charging and discharging, the bus j State of charge of the upper energy storage battery The following conditions are met: （20）； （21）； In the formula, yes t Time busbar j The state of charge of the energy storage system, taking into account the service life of the energy storage battery; is the minimum state of charge, is the maximum state of charge; Step 4.3: The optimization problem of solving the coordinated reactive power regulation of multiple resources in the distribution network is constructed as a Markov decision process and solved using an improved multi-agent deep reinforcement learning algorithm.

5. A distribution network multi-resource coordinated reactive power regulation method considering weak communication conditions according to claim 4, characterized in that: The solution process of step 4.3 is as follows: Step 4.3.1: Determine the state space: Using state space S Represents all possible states of the system at each moment. For any state s k ∈ S , expressed as follows: （22）； In the formula, yes k The voltage value of bus j in the distribution network at the moment, , , They are k Moment bus j Active power of photovoltaic inverter, energy storage inverter and charging pile inverter; , , , They are k Moment bus j Reactive power of photovoltaic inverters, energy storage inverters, charging pile inverters, and SVG; yes k Moment bus j The state of charge of the upper energy storage battery; is the current network loss of the distribution network; and They are the active power demand information and reactive power demand information of the distribution network, and They are k Always monitor the active and reactive power demands of the upstream network; Step 4.3.2: Determine the action space: Use action space A Represents all possible actions taken by the agent. For any action a j,k ∈ A , expressed as follows: （23）； In the formula, , and yes k Moment bus j The active power that needs to be generated by the photovoltaic inverter, energy storage inverter and charging pile inverter; , , and yes k Moment bus j The reactive power that needs to be generated by the photovoltaic inverter, energy storage inverter, charging pile inverter and SVG; and They are k Active power and reactive power purchased from the upstream network at all times; Step 4.3.3: Determine the state transition matrix: Using the state transfer matrix T Describes the probability of the system state changing over time after performing an action, expressed as: （24）； In the formula, p ( s k+1 |s k ,a j,k ) refers to the original state s k Taking Action a j,k After transfer to state s k+1 probability; Step 4.3.4: Determine the reward function and discount factor: In order to minimize the voltage offset, network loss, and active and reactive power purchase costs, the objective function shown in formula (10) is transformed into a single-step reward function: R i ,Right now: （25）； In the formula, ε i It is a penalty factor used to measure whether the relevant indicator exceeds the constraint value; In the reactive power optimization scenario, the discount factor γ Calculated using the following formula: （26）； In the formula, is the attenuation coefficient of reactive power; Step 4.3.5: Determine the agent strategy: The agent strategy defines the actions taken by the agent in a given state and is expressed as follows: （27）； In the formula, F ( s k , a j,k ) indicates that in the state s k Take action a j,k The expected return after softmax The function can be converted into a probability distribution.

6. A distribution network multi-resource coordinated reactive power regulation method considering weak communication conditions according to claim 5, characterized in that: The specific process of step 5 is as follows: In solving the reactive optimization problem, the multi-agent deep deterministic policy gradient algorithm constructs two neural networks, the actor and the critic; The actor neural network accepts the current state as input and outputs the corresponding action. The critic neural network accepts the current state and action as input, and outputs the value corresponding to the state-action by calculating the time difference error. A neural network with three fully connected layers is constructed as the actor neural network. The input layer dimension corresponds to the state matrix, and the output layer dimension corresponds to the number of reactive regulation resources in the distribution network. In order to introduce nonlinear relationships, the ReLU function is used as the activation function. The critic neural network also includes a three-layer fully connected network, and the following formula is used to evaluate the value under different states: （28）； In the formula, θ i are the parameters of the value network, s is the overall state, a It is an overall action; s i It is i The observation value of an agent, F i ( s , a 1, a 2,..., a n ) is the evaluation function, taking action a i and overall status s As input value, the action-value function is the output value; is the action strategy, which is used to select actions in the process of learning the target strategy; yes right θ i The partial differential vector of For parameters θ i Performance measures of corresponding strategies; E It means taking the expected value; Replacing the random strategy with a deterministic strategy, the agent's gradient is expressed as follows: （29）； In the formula, It is i The strategy of each agent; D ( s , a , r , s `) is the experience returned to the buffer area, which is used to record the experience of all agents. r For reward; Yes Status s Take Action a The visible eigenvector of ; s` is the state after taking action; The loss function of the critic neural network is expressed as follows: （30）； In the formula, It is a delay parameter. s k+1 and a k+1 yes k +1 The state and action corresponding to the moment.

Images