Power distribution network bearing capacity evaluation method considering electric vehicle access and terminal

Abstract

The invention discloses a power distribution network bearing capacity evaluation method considering electric vehicle access and a terminal. The method comprises the following steps: acquiring historical charging data of an electric vehicle charging pile in a target evaluation area; acquiring historical environment data in the target evaluation area based on a context sensing mechanism; inputting the historical charging data and the historical environment data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model; inputting target condition data into the target deep reinforcement learning model for prediction, and calculating a predicted charging load in the target evaluation area according to a prediction result; and inputting the predicted charging load into a pre-constructed power distribution network digital twinborn model to dynamically evaluate the bearing capacity of the power distribution network in the target evaluation area so as to determine whether the power distribution network in the target evaluation area can bear the predicted charging load under the target condition data. According to the invention, the accuracy of power distribution network bearing capacity evaluation can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of distribution network carrying capacity assessment, and in particular to a distribution network carrying capacity assessment method and terminal taking into account electric vehicle access. Background Art

[0002] Under the current background of global energy transformation and low-carbon development, the new energy vehicle industry has ushered in rapid development. At present, the number of new energy vehicles in my country is increasing year by year, and with it the rapid growth of charging facilities, which is expected to maintain a rapid growth trend in the next few years.

[0003] However, the large-scale demand for electric vehicle charging has brought unprecedented challenges to the existing distribution network. Due to the uncertainty and uneven distribution of electric vehicle charging load, higher requirements are placed on the stability and reliability of the distribution network. However, the existing distribution network does not fully consider the dynamic access of charging load, resulting in problems such as limited access or heavy overload of power facilities in high-penetration charging scenarios.

[0004] In existing research, although some scholars or research institutions have proposed methods for predicting the load of charging facilities and evaluating the carrying capacity of distribution networks, most studies use the load curve of a single charging station in a certain scenario to evaluate the carrying capacity of distribution transformers, which still does not fully consider the dynamic characteristics of the charging load. In addition, there are limitations to the carrying capacity assessment of distribution transformers only, and it cannot accurately reflect the actual carrying capacity of the distribution network under dynamically changing conditions. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a distribution network carrying capacity assessment method and terminal taking into account electric vehicle access, fully considering the impact of the dynamic characteristics of charging load on the distribution network carrying capacity, thereby improving the accuracy of the distribution network carrying capacity assessment.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] The distribution network carrying capacity assessment method considering electric vehicle access includes:

[0008] Obtain historical charging data of electric vehicle charging piles in the target assessment area;

[0009] Acquiring historical environmental data within the target assessment area based on a context-aware mechanism;

[0010] Inputting the historical charging data and the historical environmental data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model;

[0011] Inputting the target condition data into the target deep reinforcement learning model for prediction, and calculating the predicted charging load in the target evaluation area according to the prediction result;

[0012] The predicted charging load is input into a pre-built distribution network digital twin model to dynamically evaluate the distribution network carrying capacity in the target assessment area to determine whether the distribution network in the target assessment area can withstand the predicted charging load under the target condition data.

[0013] In order to solve the above technical problems, another technical solution adopted by the present invention is:

[0014] A distribution network carrying capacity assessment terminal taking into account electric vehicle access comprises a memory, a processor and a computer program stored in the memory and running on the processor. When the processor executes the computer program, each step of the above-mentioned distribution network carrying capacity assessment method taking into account electric vehicle access is implemented.

[0015] The beneficial effects of the present invention are: while obtaining the historical charging data of the electric vehicle charging pile, the historical environmental data is obtained through the context perception mechanism, and a target deep reinforcement learning model is constructed based on the historical charging data and the historical environmental data to simulate the interaction process between the electric vehicle and the power grid environment, thereby realizing the dynamic prediction of the charging behavior of the electric vehicle. When the target condition data is determined, the charging behavior of the electric vehicle can be directly predicted by the target deep reinforcement learning model, thereby calculating the predicted charging load that needs to be borne in the target evaluation area. Finally, the distribution network carrying capacity is dynamically evaluated based on the predicted charging load and the distribution network digital twin model. The present invention fully considers the impact of historical environmental data on the dynamic changes of the charging load and improves the accuracy of the charging load prediction. At the same time, the digital twin model is introduced to accurately simulate and analyze the state of the distribution network, thereby realizing an efficient dynamic evaluation of the distribution network carrying capacity. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of the steps of a method for evaluating the carrying capacity of a distribution network taking into account electric vehicle access provided by an embodiment of the present invention;

[0017] Figure 2 A schematic diagram of training a target deep reinforcement learning model provided by an embodiment of the present invention;

[0018] Figure 3 A schematic diagram of the hierarchical structure of a distribution network in a target assessment area provided by an embodiment of the present invention;

[0019] Figure 4 A schematic diagram of the structure of a distribution network carrying capacity assessment terminal taking into account electric vehicle access provided by an embodiment of the present invention;

[0020] Description of labels:

[0021] 100. A distribution network carrying capacity assessment terminal taking into account electric vehicle access; 101. A memory; 102. A processor. DETAILED DESCRIPTION

[0022] In order to explain the technical content, achieved objectives and effects of the present invention in detail, the following is an explanation in combination with the implementation modes and the accompanying drawings.

[0023] The embodiment of the present invention provides a method for evaluating the carrying capacity of a distribution network taking into account electric vehicle access, including:

[0024] Obtain historical charging data of electric vehicle charging piles in the target assessment area;

[0025] Acquiring historical environmental data within the target assessment area based on a context-aware mechanism;

[0026] Inputting the historical charging data and the historical environmental data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model;

[0027] Inputting the target condition data into the target deep reinforcement learning model for prediction, and calculating the predicted charging load in the target evaluation area according to the prediction result;

[0028] The predicted charging load is input into a pre-built distribution network digital twin model to dynamically evaluate the distribution network carrying capacity in the target assessment area to determine whether the distribution network in the target assessment area can withstand the predicted charging load under the target condition data.

[0029] From the above description, it can be seen that the beneficial effects of the present invention are: while obtaining the historical charging data of the electric vehicle charging pile, the historical environmental data is obtained through the context perception mechanism, and a target deep reinforcement learning model is constructed based on the historical charging data and the historical environmental data to simulate the interaction process between the electric vehicle and the power grid environment, thereby realizing the dynamic prediction of the charging behavior of the electric vehicle. When the target condition data is determined, the charging behavior of the electric vehicle can be directly predicted by the target deep reinforcement learning model, thereby calculating the predicted charging load that needs to be borne in the target evaluation area. Finally, the distribution network carrying capacity is dynamically evaluated based on the predicted charging load and the distribution network digital twin model. The present invention fully takes into account the impact of historical environmental data on the dynamic changes of the charging load, and improves the accuracy of the charging load prediction. At the same time, the digital twin model is introduced to accurately simulate and analyze the state of the distribution network, so as to realize the efficient dynamic evaluation of the distribution network carrying capacity.

[0030] Furthermore, the acquiring of historical environmental data in the target assessment area based on a context-aware mechanism includes:

[0031] Obtain historical climate data, historical economic data, and historical electric vehicle penetration rates within the target assessment area;

[0032] Based on the context perception mechanism, the historical climate data, the historical economic data, and the historical electric vehicle penetration rate are respectively inferred and analyzed to obtain the historical environmental data;

[0033] Among them, the historical environmental data includes weather information, population size, per capita GDP and electric vehicle penetration rate before the forecast date.

[0034] From the above description, we can see that the context-aware mechanism can perceive and identify changes in the environment in real time, so as to adjust the provided environmental data in a timely manner. Among them, weather information, population, per capita GDP and electric vehicle penetration rate before the forecast date are all likely to change in real time. Therefore, determining the above data through the context-aware mechanism can ensure the accuracy and real-time nature of the data, so that the subsequent deep reinforcement learning model can learn the influence of environmental data on electric vehicle charging behavior and realize accurate prediction of charging behavior.

[0035] Furthermore, the inputting the historical charging data and the historical environmental data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model includes:

[0036] Converting the interaction process between the historical charging data and the historical environmental data into a Markov decision process;

[0037] A deep reinforcement learning model is constructed based on the Markov decision process, and the deep reinforcement learning model is trained to obtain a target deep reinforcement learning model.

[0038] From the above description, it can be seen that the interaction between the agent's historical charging data and the agent's historical environment data is converted into a Markov decision process, so that through continuous interaction and iteration between the agent and the environment, the deep reinforcement learning model can learn the most likely charging behavior in different environments.

[0039] Furthermore, the historical charging data includes the historical quantity of each type of charging pile and the historical charging behavior data of each charging pile;

[0040] The historical charging behavior data includes historical charging time, historical usage status and charging pile type;

[0041] The Markov decision process includes a state space, a context perception space, an action space and a reward function;

[0042] The converting the interaction process between the historical charging data and the historical environmental data into a Markov decision process comprises:

[0043] Define the state space based on the historical charging time, historical usage status and charging pile type of each charging pile;

[0044] defining a context-aware space based on weather information, population, per capita GDP, and electric vehicle penetration prior to the forecast date;

[0045] Define the action space based on the charging behavior of each charging station;

[0046] A reward function is defined according to a charging cost, a charging rate, and a charging penalty corresponding to the charging behavior.

[0047] From the above description, it can be seen that the charging status of electric vehicles in the target evaluation area is determined by the historical charging time, historical usage status and charging pile type of the charging pile, so as to accurately characterize the state characteristics of the electric vehicle. At the same time, the context perception space is defined by the data that affects the charging behavior to characterize the environmental characteristics of the electric vehicle. Finally, the reward function is defined by the feedback data corresponding to each charging behavior to characterize the feedback obtained by the electric vehicle when taking the charging behavior, so as to guide the intelligent agent to learn the most likely charging behavior. In this way, the charging behavior of electric vehicles can be accurately predicted based on the charging pile data and environmental data.

[0048] Furthermore, constructing a deep reinforcement learning model based on the Markov decision process and training the deep reinforcement learning model to obtain a target deep reinforcement learning model includes:

[0049] Get the deep reinforcement learning model in the initialized state;

[0050] Inputting the state space and the context-aware space into the deep reinforcement learning model for interactive learning to determine the Q value of the action space;

[0051] The network parameters of the deep reinforcement learning model are iteratively updated based on the difference between the Q value and the temporal difference target to obtain a target deep reinforcement learning model.

[0052] From the above description, we can see that the Q value represents the total reward expected from taking a specific action in a given state. The network parameters are iteratively updated by the difference between the Q value and the temporal difference target, so that the target deep reinforcement learning model can adjust the agent's estimate of the value of its action to reduce the prediction error, so that the agent can make better action choices when facing similar states in the future, thereby improving long-term rewards.

[0053] Further, the deep reinforcement learning model includes a Q-value network and a target network; the state space includes a plurality of state vectors; the context-aware space includes a plurality of context-aware vectors;

[0054] The step of inputting the state space and the context-aware space into the deep reinforcement learning model for interactive learning to determine the Q value of the action space includes:

[0055] Inputting a next state vector and a next context perception vector into the target network to predict a next action vector corresponding to the next state vector and the next context perception vector, and calculating a maximum Q value of the next action vector;

[0056] The maximum Q value and the current state vector and the current context perception vector corresponding to the next action vector are input into the Q value network, the current action vector corresponding to the current state vector and the current context perception vector is selected based on the ε-greedy algorithm, and the Q value of the current action vector is calculated according to the maximum Q value.

[0057] From the above description, we can know that the target network is used to predict the next action and the maximum Q value corresponding to the next state vector, and the Q value network is used to determine the current action and the current Q value according to the maximum Q value predicted by the target network and the current state vector. This realizes the interactive learning of the agent and the environment information.

[0058] Furthermore, the step of inputting the target condition data into the target deep reinforcement learning model for prediction, and calculating the predicted charging load in the target evaluation area according to the prediction result includes:

[0059] Input the target condition data into the target deep reinforcement learning model for prediction to obtain the optimal charging behavior of each charging pile;

[0060] Calculating the charging power of each charging pile according to the optimal charging behavior and the context perception space corresponding to the optimal charging behavior;

[0061] The predicted charging load is determined according to the sum of the charging powers of all charging piles in the target assessment area.

[0062] From the above description, it can be seen that when the target condition data is determined, the charging behavior of the charging pile can be predicted based on the trained target deep reinforcement learning model to characterize the charging behavior of the electric vehicle, and then the charging power can be determined based on the charging behavior, so as to calculate the charging load in the target evaluation area. Since electric vehicles may move with the transfer of user usage, and the location of the charging pile is fixed, the distribution network carrying capacity of the target evaluation area is evaluated through the charging behavior of the charging pile, which can ensure the pertinence of the prediction data.

[0063] Furthermore, before acquiring the historical charging data of the electric vehicle charging piles in the target evaluation area, the method further includes:

[0064] Acquire distribution network data of the target assessment area, wherein the distribution network data includes distribution network topology, distribution network equipment parameters and historical operation data;

[0065] Constructing a distribution network physical model according to the distribution network topology;

[0066] Constructing a distribution network system function model according to the distribution network equipment parameters;

[0067] constructing a distribution network neural network model based on the historical operation data;

[0068] A distribution network digital twin model is constructed based on the distribution network physical model, the distribution network system function model and the distribution network neural network model.

[0069] From the above description, it can be seen that through the distribution network topology, equipment parameters and historical operation data, a virtual digital twin model that is synchronized with the actual distribution network in real time is created to accurately simulate and analyze the status of the distribution network under different operating conditions and charging scenarios, thereby improving the accuracy of the evaluation.

[0070] Furthermore, the step of inputting the predicted charging load into a pre-built distribution network digital twin model to dynamically evaluate the distribution network carrying capacity in the target evaluation area includes:

[0071] Input the predicted charging load into the distribution network digital twin model, and calculate the first carrying capacity of the distribution network transformer, the second carrying capacity of each preset line, and the third carrying capacity of each preset substation in the target assessment area;

[0072] The distribution network carrying capacity is dynamically evaluated according to the first carrying capacity, the second carrying capacity and the third carrying capacity.

[0073] From the above description, it can be seen that by evaluating and calculating the distribution network transformers, preset lines and substations respectively, the comprehensiveness of the carrying capacity assessment can be improved, thereby accurately reflecting the real carrying capacity of the distribution network under dynamically changing conditions.

[0074] Another embodiment of the present invention provides a distribution network carrying capacity assessment terminal taking into account electric vehicle access, comprising a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor implements each step of the above-mentioned distribution network carrying capacity assessment method taking into account electric vehicle access when executing the computer program.

[0075] From the above description, it can be seen that the beneficial effects of the present invention are: while obtaining the historical charging data of the electric vehicle charging pile, the historical environmental data is obtained through the context perception mechanism, and a target deep reinforcement learning model is constructed based on the historical charging data and the historical environmental data to simulate the interaction process between the electric vehicle and the power grid environment, thereby realizing the dynamic prediction of the charging behavior of the electric vehicle. When the target condition data is determined, the charging behavior of the electric vehicle can be directly predicted by the target deep reinforcement learning model, thereby calculating the predicted charging load that needs to be borne in the target evaluation area. Finally, the distribution network carrying capacity is dynamically evaluated based on the predicted charging load and the distribution network digital twin model. The present invention fully takes into account the impact of historical environmental data on the dynamic changes of the charging load, and improves the accuracy of the charging load prediction. At the same time, the digital twin model is introduced to accurately simulate and analyze the state of the distribution network, so as to realize the efficient dynamic evaluation of the distribution network carrying capacity.

[0076] The embodiment of the present invention provides a distribution network carrying capacity assessment method and terminal taking into account electric vehicle access, which can be applied to the dynamic assessment of the distribution network carrying capacity in a specified area, and fully considers the impact of the dynamic characteristics of the charging load on the distribution network carrying capacity, thereby improving the accuracy of the distribution network carrying capacity assessment. The following is an explanation through a specific embodiment:

[0077] Please refer to Figure 1 to Figure 2 , Embodiment 1 of the present invention is:

[0078] The distribution network carrying capacity assessment method considering electric vehicle access includes:

[0079] S1. Obtain the historical charging data of electric vehicle charging piles in the target assessment area.

[0080] The historical charging data includes the historical quantity of each type of charging pile and the historical charging behavior data of each charging pile; the historical charging behavior data includes the historical charging time, historical usage status and charging pile type.

[0081] In some embodiments, the target evaluation area is selected according to the administrative area or the power supply area boundary. After determining the target evaluation area, the number of substations contained in the target evaluation area, the number of lines corresponding to each substation, and the number of distribution transformers corresponding to each line are determined respectively. For example, the number of 110KV substations contained in the target evaluation area is currently determined to be N_sub, where the number of 10KV lines corresponding to the i-th substation is N_line(i), and the number of distribution transformers corresponding to the j-th 10KV line is N_xfmr(j).

[0082] In some embodiments, the types of charging piles include private charging piles for residents, public charging piles in cities, internal charging piles for units, bus passenger charging piles, logistics vehicle charging piles, and other special vehicle charging piles. The historical charging time includes the charging time and the charging date; the charging date includes the day of the week when charging and whether it is a holiday. The historical usage status includes in use, idle, and under maintenance.

[0083] S2. Acquire historical environmental data within the target assessment area based on a context-aware mechanism.

[0084] The step S2 comprises:

[0085] S21. Obtain historical climate data, historical economic data, and historical electric vehicle penetration rate in the target assessment area.

[0086] S22. Based on the context perception mechanism, the historical climate data, the historical economic data and the historical electric vehicle penetration rate are respectively inferred and analyzed to obtain the historical environmental data.

[0087] Among them, the historical environmental data includes weather information, population size, per capita GDP and electric vehicle penetration rate before the forecast date.

[0088] In some embodiments, weather information mainly includes temperature and precipitation.

[0089] In some embodiments, after the historical charging data and the historical environmental data are obtained, the historical charging data and the historical environmental data are cleaned to ensure the integrity of the data, while processing missing values ​​or outliers, and removing invalid or erroneous charging records. In addition, feature extraction is performed on the historical charging data and the historical environmental data. For example, the historical charging data extracts time features such as hours, weeks, and dates from the timestamps, and extracts charging habit features from the user data, such as charging frequency and preferred charging time. Data normalization is performed on the historical charging data and the historical environmental data, and all continuous features are processed so that their value range is between 0 and 1 so that the model can process more efficiently.

[0090] In some embodiments, the historical charging data and the historical environmental data are divided into a training set, a validation set, and a test set to facilitate subsequent training and evaluation of the model.

[0091] S3. Input the historical charging data and the historical environmental data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model.

[0092] The step S3 comprises:

[0093] S31, converting the interaction process between the historical charging data and the historical environmental data into a Markov decision process.

[0094] The Markov decision process includes a state space, a context-aware space, an action space, and a reward function, and step S31 includes:

[0095] S311. Define a state space according to the historical charging time, historical usage status, and charging pile type of each charging pile.

[0096] In some embodiments, each state vector S in the state space t Specifically:

[0097] S t =[hour,day_of_week,is_holiday,charger_status,type];

[0098] Among them, hour represents the number of hours of charging pile data collection time, ranging from 0 to 23; day_of_week represents the day of the week, ranging from 1 (Monday) to 7 (Sunday); is_holiday represents whether it is a holiday, with 0 representing non-holidays and 1 representing holidays; charger_status represents the status of the charging pile, 0 represents idle, 1 represents in use, and 2 represents under maintenance; type represents the type of charging pile, such as private charging piles for residents, public charging piles in cities, internal charging piles for units, bus passenger charging piles, logistics vehicle charging piles, other special vehicle charging piles, etc., and its range starts from 0; t belongs to the integer set and represents a time interval of 15 minutes.

[0099] Specifically, a private charging pile of a resident is connected to an electric vehicle for charging from 12:00 to 20:00 on August 20, 2023 (Sunday). If the data collection time is 16:00, the state vector S of the charging pile is t =[16, 7, 0, 1, 0].

[0100] S312: Define a context-aware space according to weather information, population, per capita GDP, and electric vehicle penetration rate before the forecast date.

[0101] In some embodiments, each context-aware vector CA in the context-aware space t Specifically:

[0102] CA t =[Temperatur,Speed,Precipitation,Population,GDPc,PNEV];

[0103] Among them, Temperatur represents the historical temperature data in the target assessment area; Precipitation represents the historical precipitation data in the target assessment area; Population represents the population in the target assessment area; GDPc represents the per capita gross domestic product in the target assessment area; PNEV represents the electric vehicle penetration rate in the target assessment area; t belongs to the integer set and represents a 15-minute time interval.

[0104] S313: Define an action space according to the charging behavior of each charging pile.

[0105] In some embodiments, each action a in the action space t Belongs to {start, stop, adjust(P)}, where start means starting charging, stop means stopping charging, and adjust(P) means adjusting the charging power to P charging.

[0106] S314: Define a reward function according to the charging cost, charging rate, and charging penalty corresponding to the charging behavior.

[0107] In some embodiments, the reward function includes a reward term and a penalty term, then the reward function R(s t ,a t ) is specifically: R(s t ,a t )=w1×cost(a t )-w2×efficiency(a t )-w3×penalty(s t ), where w1, w2, and w3 represent the weights of cost, rate, and penalty, respectively, to balance the importance of different factors; cost(a t ) indicates that the action a t The relevant costs, such as the electricity price multiplied by the charging power and charging time; efficiency (at) represents the action a t The rate of charge, such as the charging rate or the inverse of the time required to complete charging; penalty(st) represents the penalty according to the state vector S t Calculated penalties, such as charging during peak hours or periods of high electricity prices may incur additional penalties.

[0108] S32. Construct a deep reinforcement learning model based on the Markov decision process, and train the deep reinforcement learning model to obtain a target deep reinforcement learning model.

[0109] The step S32 comprises:

[0110] S321. Obtain a deep reinforcement learning model in an initialized state.

[0111] In some embodiments, the deep reinforcement learning model is specifically expressed as π*=argmax π E[R]; where π* represents the optimal strategy, which is the best way for the agent to choose an action in a given state and can bring the maximum expected return; argmax π It means finding the strategy π that maximizes the following expression among all strategies; E[R] represents the expected value of the reward obtained by the agent following the strategy π, where E is the expected value and R is the reward value.

[0112] S322: Input the state space and the context-aware space into the deep reinforcement learning model for interactive learning to determine the Q value of the action space.

[0113] In some embodiments, the state vector S of the state space and the context-aware vector CA of the context-aware space are input into a deep reinforcement learning model to obtain the Q-value Q-Value(S, a) corresponding to each action a.

[0114] Among them, the deep reinforcement learning model includes a Q-value network and a target network; the state space includes multiple state vectors; and the context-aware space includes multiple context-aware vectors.

[0115] In some embodiments, the deep reinforcement learning model is initialized, that is, the weights and bias values ​​of the convolutional layers and the fully connected layers in the Q-value network are initialized. In addition, the relevant parameters of the target network are generally the same as those of the Q-value network.

[0116] The step S322 includes:

[0117] S3221. Input a next state vector and a next context-aware vector into the target network to predict a next action vector corresponding to the next state vector and the next context-aware vector, and calculate a maximum Q value of the next action vector.

[0118] In some embodiments,

[0119] S3222. Input the maximum Q value and the current state vector and current context perception vector corresponding to the next action vector into the Q value network, select the current action vector corresponding to the current state vector and the current context perception vector based on the ε-greedy algorithm, and calculate the Q value of the current action vector according to the maximum Q value.

[0120] In some embodiments, Figure 2 As shown in the figure, the state vector and the context perception vector are input into the deep reinforcement learning model to output the experience tuple of the corresponding action. The specific experience tuple is (S t , at , C.A. t ), which represents the state vector S t With context-aware vector CA t The corresponding action is a t The experience tuples are stored in the buffer Replay Memory. After a certain number of experience tuples are stored, a batch of experience tuples are randomly sampled from the buffer Replay Memory for training the deep reinforcement learning model. For the batch of experience tuples, the current action vector corresponding to the current state vector and the current context perception vector is selected based on the ε-greedy algorithm in step S3222, and when executing action a t After that, calculate the current instant reward r obtained by the agent from the environment t+1 , then the Q value of the current action vector calculated according to the maximum Q value in step S3222 is specifically:

[0121] Q(S t ,a t ,CA t )←Q(S t ,a t ,CA t )+α[r t+1 +γmax a' Q(S t+1 ,a',CA t+1 )-Q(S t ,a t ,CA t )];

[0122] Among them, max a' Q(S t+1 ,a',CA t+1 ) is the maximum Q value that can be obtained by taking action a' at the next moment, which is predicted based on the target network; α is the learning rate of the objective function, 0≤α≤1, which indicates the degree to which the prior data is retained when updating the Q value; γ is the discount factor, 0≤γ≤1, which indicates the importance of future rewards relative to immediate rewards.

[0123] Then return to the step of selecting the current action vector corresponding to the current state vector and the current context perception vector based on the ε-greedy algorithm in step S3222, iteratively update the Q value until the difference between two consecutive updated Q values ​​is less than a preset threshold θ, the interval range of the preset threshold θ is specifically [0.001, 0.01], Q(S t ,a t ,CA t )-Q'(S t ,a t ,CA t )<θ, where Q'(St ,a t ,CA t ) is the updated Q value.

[0124] In some embodiments, the timing difference target TD Target is specifically:

[0125] TD Target = r t+1 +γMaxQ t+1 ; Among them, MaxQ t+1 =max a' Q(S t+1 ,a',CA t+1 ).

[0126] Therefore, the formula for calculating the Q value of the current motion vector based on the maximum Q value can be expressed as Q(S t ,a t ,CA t )←Q(S t ,a t ,CA t )+α[TD Target-Q(S t ,a t ,CA t )], representing the estimated Q value in the Q value network that iteratively updates the Q value based on the temporal difference target.

[0127] S323. Iteratively update the network parameters of the deep reinforcement learning model based on the difference between the Q value and the temporal difference target to obtain a target deep reinforcement learning model.

[0128] In some embodiments, the error between the Q value output by the current network and the temporal difference target is calculated, and the mean square error is used as the loss function:

[0129]

[0130] Where N represents the number of experience tuples in the batch sampled from the buffer Replay Memory, that is, the number of samples used for training in one iteration; y i represents the time difference target of the i-th sample; Q(S t ,a t ,CA t ; Θ) represents the i-th sample in the state vector S t , context-aware vector CA t And the action vector a t In this case, the Q-value predicted by a deep reinforcement learning model with parameters Θ, where Θ includes the weights and biases in the network.

[0131] Based on the above loss function, the gradient of the loss relative to the Q value network parameter Θ is calculated by the reverse propagation algorithm Then use the gradient descent method to update the Q value network parameter Θ according to the calculated gradient. The updated Q value network parameter Among them, η represents the learning rate, which controls the amplitude of parameter update, and its value range is [0.001, 0.01].

[0132] To ensure the target network parameter Θ - To ensure the stability of the network, the target network parameters Θ are updated regularly through soft updates. - ←τ·Θ - +(1-τ)·Θ, where τ is a decimal close to 1, which controls the update speed of the target network parameters.

[0133] After an iterative update, a new batch of experience tuples is sampled from the buffer, and the network parameters are re-evaluated and updated until the model converges or reaches the preset training cycle to obtain the target deep learning model.

[0134] S4. Input the target condition data into the target deep reinforcement learning model for prediction, and calculate the predicted charging load in the target evaluation area based on the prediction results.

[0135] The step S4 comprises:

[0136] S41. Input the target condition data into the target deep reinforcement learning model for prediction to obtain the optimal charging behavior of each charging pile.

[0137] In some embodiments, the target condition data includes the current number of charging piles of each type in the target evaluation area, the current status of the charging piles, the current time information (hours, weeks, and dates), the current environmental factors (temperature and precipitation), the current socioeconomic indicators (population size and per capita GDP), and the current electric vehicle penetration rate.

[0138] S42: Calculate the charging power of each charging pile according to the optimal charging behavior and the context-aware space corresponding to the optimal charging behavior.

[0139] In some embodiments, for each charging pile, the charging power P i (S t ,a t ,CA t ; θ) specifically:

[0140] P i (S t ,a t ,CA t ; Θ) = f(Q(S t ,at ,CA t ; Θ));

[0141] Among them, function f represents a mapping process from Q value to charging power, which can convert the Q value into specific charging power.

[0142] S43: Determine the predicted charging load according to the sum of the charging powers of all charging piles in the target evaluation area.

[0143] In some embodiments, the charging power of n charging piles in the target evaluation area is accumulated to obtain the total predicted charging load P at time t. pred,t Specifically:

[0144]

[0145] S5. Input the predicted charging load into a pre-built distribution network digital twin model to dynamically evaluate the distribution network carrying capacity within the target assessment area to determine whether the distribution network within the target assessment area can withstand the predicted charging load under the target condition data.

[0146] Please refer to Figure 3 , Embodiment 2 of the present invention is:

[0147] The method for evaluating the carrying capacity of a distribution network taking into account electric vehicles access is different from the first embodiment in that, before step S1, it also includes:

[0148] S101. Acquire distribution network data of the target assessment area, wherein the distribution network data includes distribution network topology, distribution network equipment parameters, and historical operation data.

[0149] In some embodiments, the acquired data is cleaned, standardized, and feature extracted to ensure data quality in preparation for model training and simulation.

[0150] S102: construct a distribution network physical model according to the distribution network topology structure.

[0151] In some embodiments, circuit models and mathematical expressions at various levels are constructed based on the physical characteristics of the distribution network, namely the circuit topology, including parameters such as the load status and load capacity of transformers, lines and substations.

[0152] S103: constructing a distribution network system function model according to the distribution network equipment parameters.

[0153] In some embodiments, the system function model can calculate the response of the distribution network based on input parameters, such as voltage, current, power, etc.

[0154] S104: constructing a distribution network neural network model according to the historical operation data.

[0155] In some embodiments, the historical operating data of each level is used to train a neural network model to learn the load characteristics and response patterns of different levels.

[0156] S105. Construct a distribution network digital twin model according to the distribution network physical model, the distribution network system function model and the distribution network neural network model.

[0157] In some embodiments, the physical model, system function model and neural network model are integrated into a unified digital twin framework to ensure the coordination and consistency of models at all levels and realize the simulation of the entire distribution network.

[0158] In some embodiments, the accuracy of the digital twin model can be verified by comparing the simulation data output by the digital twin model with the actual operation data, and the parameters can be calibrated to ensure that the model can truly reflect the state of the distribution network. At the same time, a carrying capacity assessment module is developed according to the different hierarchical structures of the distribution network, thereby providing real-time prediction analysis for different hierarchical structures.

[0159] The step S5 comprises:

[0160] S51. Input the predicted charging load into the distribution network digital twin model, and calculate the first carrying capacity of the distribution network transformer in the target assessment area, the second carrying capacity of each preset line, and the third carrying capacity of each preset substation.

[0161] In some embodiments, combined with the historical operation data of each distribution transformer in the target evaluation area, the maximum load day is selected as the typical day, and the 96-point load data at every 15-minute time interval is used as the evaluation status. The growth of conventional power load and the predicted additional charging load are synchronized to the digital twin load capacity evaluation module of the distribution network transformer to evaluate the first load capacity of the distribution network transformer. Among them,

[0162] The mathematical model of the planned annual total load of the digital twin carrying capacity assessment module of the distribution network transformer is as follows:

[0163] L future,t =L list,t ×(1+γ)+P pred,t Where, L list is the conventional power load data of the distribution network transformer at 96 o’clock on the maximum load day, and γ is the average annual growth rate of conventional power load calculated according to different scenario characteristics and historical annual load levels.

[0164] The load rate mathematical model of the digital twin load capacity assessment module of the distribution network transformer is as follows:

[0165]

[0166] Among them, P rated Indicates the rated power of the distribution network transformer. If the load rate of the distribution network transformer is greater than 80%, capacity increase or related measures should be taken. max(L future,t ) represents the maximum total load of distribution network transformers in the planned year.

[0167] In some embodiments, the predicted loads of all distribution network transformers under each 10kV line and the simultaneity factor between the distribution network transformers are input into the digital twin carrying capacity evaluation module of the 10kV line to evaluate the second carrying capacity of the preset line.

[0168] The mathematical model of the planned annual total load of the digital twin carrying capacity assessment module of the 10kV line is as follows:

[0169] Where, L future(j) It represents the maximum power load of each distribution network transformer under the 10kV line, and η represents the simultaneity factor between the distribution network transformers calculated based on historical data.

[0170] Based on the rated load, wiring mode and planned annual total power load of the 10kV line, the carrying capacity of the 10kV line in the planned year is calculated and its ability to access the charging load is evaluated.

[0171] The load rate mathematical model of the digital twin carrying capacity assessment module of the 10kV line is as follows:

[0172] Where P 10KV Indicates the rated transmission capacity of the 10kV line. If the load rate of the 10kV line exceeds 50% (or the grid structure of the actual line meets the "N-1" requirement), it means that the line may not be able to operate safely during the planning period and line modification or related measures are needed.

[0173] In some embodiments, the predicted loads of all 10kV lines under each 110kV substation and the simultaneity coefficients between the 10kV lines are input into the digital twin capacity assessment module of the 110kV substation to assess the third capacity of the preset substation line.

[0174] The mathematical model of the planned annual total load of the digital twin carrying capacity assessment module of the 110kV substation is as follows:

[0175] Where, L 10kV (i) represents the power load of each 10kV line under the 110kV substation; It indicates the simultaneity factor between 10kV lines calculated based on historical data.

[0176] Calculate the carrying capacity of the 110kV substation in the planning year, benchmark the reasonable capacity-load ratio range in the planning year, and evaluate the carrying capacity of the substation.

[0177] The load rate mathematical model of the digital twin carrying capacity assessment module of the 110kV substation is as follows:

[0178] Where P 110KV Indicates the rated load of the 110kV substation. If the load rate of the 110kV substation exceeds 50% (or the actual number of transformers in the station does not meet the "N-1" requirement), it means that the substation may not be able to operate safely during the planning period, and substation reconstruction or related measures are needed.

[0179] S52: Dynamically evaluate the distribution network carrying capacity according to the first carrying capacity, the second carrying capacity, and the third carrying capacity.

[0180] A digital twin model of the distribution network is constructed based on the network topology of each level of distribution network in the target assessment area, and a high-precision digital model is used to achieve real-time monitoring and simulation of the physical power grid. The rated parameters, load data, voltage data, current data, equipment status and other information of each level of distribution network are input into the digital twin model. At the same time, the charging load predicted by the context-aware deep reinforcement learning model is connected to the load end of the digital twin model to simulate the actual operating status of the distribution network and dynamically evaluate the carrying capacity of distribution transformers, 10kV lines and substations. Based on the application of digital twin technology, the carrying capacity assessment of the distribution network is more accurate and real-time, which helps to improve the safety and economy of power grid operation.

[0181] Please refer to Figure 4 , Embodiment 3 of the present invention is:

[0182] The distribution network carrying capacity assessment terminal 100 taking into account electric vehicle access includes a memory 101, a processor 102, and a computer program stored in the memory 101 and running on the processor 102. When the processor 102 executes the computer program, each step of the distribution network carrying capacity assessment method taking into account electric vehicle access of the above-mentioned embodiment 1 is implemented.

[0183] In summary, the distribution network carrying capacity assessment method and terminal taking into account electric vehicle access provided by the present invention, while obtaining the historical charging data of the electric vehicle charging pile, obtains the environmental data before the prediction date through the context perception mechanism, and constructs a target deep reinforcement learning model based on the historical charging data and the historical environmental data to simulate the interaction process between the electric vehicle and the power grid environment, thereby realizing the dynamic prediction of the charging behavior of the electric vehicle. At the same time, since the present invention predicts the charging behavior of the electric vehicle based on the charging pile data of the target assessment area, it solves the problem that the distribution network carrying capacity can only be evaluated in a single charging station or charging scenario. When the target condition data is determined, the charging behavior of the electric vehicle can be directly predicted by the target deep reinforcement learning model, thereby calculating the predicted charging load that needs to be borne in the target assessment area, fully considering the impact of environmental factors on the charging behavior, and improving the accuracy of the charging load prediction. Finally, the distribution network carrying capacity is dynamically evaluated based on the predicted charging load and the distribution network digital twin model to accurately simulate and analyze the state of the distribution network, thereby realizing an accurate evaluation of the distribution network carrying capacity.

[0184] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's specification and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for evaluating the carrying capacity of a distribution network taking into account electric vehicles, characterized in that: include: Obtain historical charging data of electric vehicle charging piles in the target assessment area; Acquiring historical environmental data within the target assessment area based on a context-aware mechanism; Inputting the historical charging data and the historical environmental data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model; Inputting the target condition data into the target deep reinforcement learning model for prediction, and calculating the predicted charging load in the target evaluation area according to the prediction result; The predicted charging load is input into a pre-built distribution network digital twin model to dynamically evaluate the distribution network carrying capacity in the target assessment area to determine whether the distribution network in the target assessment area can withstand the predicted charging load under the target condition data.

2. The method according to claim 1, characterized in that The acquiring of historical environmental data within the target assessment area based on a context-aware mechanism includes: Obtain historical climate data, historical economic data, and historical electric vehicle penetration rates within the target assessment area; Based on the context perception mechanism, the historical climate data, the historical economic data, and the historical electric vehicle penetration rate are respectively inferred and analyzed to obtain the historical environmental data; Among them, the historical environmental data includes weather information, population size, per capita GDP and electric vehicle penetration rate before the forecast date.

3. The method according to claim 2, characterized in that The inputting the historical charging data and the historical environment data into a deep reinforcement learning model for training to obtain a target deep reinforcement learning model comprises: Converting the interaction process between the historical charging data and the historical environmental data into a Markov decision process; A deep reinforcement learning model is constructed based on the Markov decision process, and the deep reinforcement learning model is trained to obtain a target deep reinforcement learning model.

4. The method according to claim 3, characterized in that The historical charging data includes the historical quantity of each type of charging pile and the historical charging behavior data of each charging pile; The historical charging behavior data includes historical charging time, historical usage status and charging pile type; The Markov decision process includes a state space, a context perception space, an action space and a reward function; The converting the interaction process between the historical charging data and the historical environmental data into a Markov decision process comprises: Define the state space based on the historical charging time, historical usage status and charging pile type of each charging pile; defining a context-aware space based on weather information, population, per capita GDP, and electric vehicle penetration prior to the forecast date; Define the action space based on the charging behavior of each charging station; A reward function is defined according to a charging cost, a charging rate, and a charging penalty corresponding to the charging behavior.

5. The method according to claim 4, characterized in that The step of constructing a deep reinforcement learning model based on the Markov decision process and training the deep reinforcement learning model to obtain a target deep reinforcement learning model includes: Get the deep reinforcement learning model in the initialized state; Inputting the state space and the context-aware space into the deep reinforcement learning model for interactive learning to determine the Q value of the action space; The network parameters of the deep reinforcement learning model are iteratively updated based on the difference between the Q value and the temporal difference target to obtain a target deep reinforcement learning model.

6. The method according to claim 4, characterized in that The deep reinforcement learning model includes a Q-value network and a target network; the state space includes a plurality of state vectors; the context-aware space includes a plurality of context-aware vectors; The step of inputting the state space and the context-aware space into the deep reinforcement learning model for interactive learning to determine the Q value of the action space includes: Inputting a next state vector and a next context perception vector into the target network to predict a next action vector corresponding to the next state vector and the next context perception vector, and calculating a maximum Q value of the next action vector; The maximum Q value and the current state vector and the current context perception vector corresponding to the next action vector are input into the Q value network, the current action vector corresponding to the current state vector and the current context perception vector is selected based on the ε-greedy algorithm, and the Q value of the current action vector is calculated according to the maximum Q value.

7. The method according to claim 4, characterized in that The step of inputting the target condition data into the target deep reinforcement learning model for prediction, and calculating the predicted charging load in the target evaluation area according to the prediction result includes: Input the target condition data into the target deep reinforcement learning model for prediction to obtain the optimal charging behavior of each charging pile; Calculating the charging power of each charging pile according to the optimal charging behavior and the context perception space corresponding to the optimal charging behavior; The predicted charging load is determined according to the sum of the charging powers of all charging piles in the target assessment area.

8. The method according to claim 1, characterized in that Before obtaining the historical charging data of the electric vehicle charging piles in the target assessment area, the method further includes: Acquire distribution network data of the target assessment area, wherein the distribution network data includes distribution network topology, distribution network equipment parameters and historical operation data; Constructing a distribution network physical model according to the distribution network topology; Constructing a distribution network system function model according to the distribution network equipment parameters; constructing a distribution network neural network model based on the historical operation data; A distribution network digital twin model is constructed based on the distribution network physical model, the distribution network system function model and the distribution network neural network model.

9. The method according to claim 8, characterized in that The step of inputting the predicted charging load into a pre-built distribution network digital twin model to dynamically evaluate the distribution network carrying capacity within the target evaluation area includes: Input the predicted charging load into the distribution network digital twin model, and calculate the first carrying capacity of the distribution network transformer, the second carrying capacity of each preset line, and the third carrying capacity of each preset substation in the target assessment area; The distribution network carrying capacity is dynamically evaluated according to the first carrying capacity, the second carrying capacity and the third carrying capacity.

10. A distribution network carrying capacity assessment terminal taking into account electric vehicles, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, each step of the method for evaluating the carrying capacity of a distribution network taking into account electric vehicle access as described in any one of claims 1 to 9 is implemented.

Images