Electric vehicle charging power aggregation distributed regulation and control method

Abstract

The invention discloses an electric vehicle charging power aggregation distributed regulation and control method, and relates to the technical field of AC charging pile charging power regulation and control, and the method specifically comprises the following steps: carrying out the cleaning and stabilization processing of power load and new energy data, employing an ARIMA model to predict the load, and employing an LSTM-Transform model to predict the new energy generation power; the upper layer aggregates the total state of the charging station, including the electricity price, the load, the total vehicle energy, the total docking power of the lower-layer single vehicle model and the battery demand; based on an MCTS enhanced DDPG algorithm, the total charging power is optimized through cost discount rewards and energy penalty; an improved ADMM algorithm distributes total power to each pile, and battery stress and heat loss are minimized; and comparing the actual battery energy with the virtual battery energy, and updating the data and re-deciding when the energy is inconsistent. According to the invention, three industry bottlenecks of poor real-time response of electric vehicle charging, slow large-scale scheduling and difficulty in consideration of economy and battery health are overcome, and power grid dynamic response acceleration, decision-making efficiency jump and comprehensive benefit breakthrough are realized.

Description

Technical Field

[0001] The present invention relates to the field of electric vehicle charging, and in particular to a method for distributed control of electric vehicle charging power aggregation. Background Art

[0002] With the rapid adoption of electric vehicles (EVs), the current mainstream "on-demand" charging strategy, due to its simplicity and crudeness (charging at full power upon vehicle connection, ignoring the dynamic balance of the grid), is increasingly straining the power grid. This strategy treats EVs as uncontrollable loads, resulting in random and spatial concentration of charging activity, which can easily lead to "peak upon peak," threatening grid security and stability and driving up costs. At the same time, the enormous gigawatt-hour aggregate energy storage potential (V2G) inherent in large-scale EV clusters is severely underestimated and wasted. These EVs could serve as efficient, distributed virtual energy storage entities, absorbing new energy during grid downturns and providing critical services such as power support and peak regulation through orderly charging or reverse discharge during peak periods, significantly improving grid efficiency and its ability to absorb new energy. Addressing this core contradiction between the "exposed drawbacks of the traditional model" and the "idle energy storage potential," this invention aims to break through the existing paradigm and design an intelligent method and system for coordinated control of EV aggregate charging power, providing key technical support for building a safe, efficient, and clean smart grid.

[0003] Current technical solutions for electric vehicle (EV) charging suffer from significant shortcomings. First, price regulation mechanisms are ineffective, electricity prices are inelastic, and user response is delayed, making them unable to match the second-by-second fluctuations in renewable energy output. For example, the price elasticity coefficient for electricity is typically below 0.1, resulting in very limited effectiveness of price adjustments in guiding user behavior. There is a lag of over 30 minutes between price release and user response, making it completely incapable of matching the second-by-second fluctuations in renewable energy output. Second, the randomness of renewable energy sources is neglected. The high volatility of local photovoltaic and wind power at charging stations leads to a disconnect between pre-set charging strategies and actual energy supply. This significant disconnect not only leads to significant deviations in electricity purchase costs but also increases the risk of transformer overload due to the lack of a source-load coordination mechanism. These issues not only impact charging efficiency and user experience, but also exacerbate grid operational pressures and limit the efficient utilization of renewable energy. Therefore, the development of more intelligent, flexible, and multi-objective coordinated charging technology solutions is urgent. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a method for distributed control of charging power aggregation of electric vehicles.

[0005] The technical solution adopted by the present invention is a method for distributed control of charging power aggregation of electric vehicles, which includes: Step S1: preprocess the basic power load data and the new energy power generation data. For missing data, linear interpolation is used for fitting estimation. For abnormal data, the 3-principle is used to eliminate and re-fit the estimation. For the basic power load data, the first-order difference is used to eliminate the stationarity. The ADF is used to test whether the data is stationary. If it fails, the second-order difference and third-order difference are performed until the stationarity test passes. Step S2: Model the charging station master and charging piles separately. The modeling of the charging station master needs to consider time-of-use electricity prices, basic power load, new energy generation output, the total battery energy of all electric vehicles in the station in an aggregated state, the time when the electric vehicle leaves the charging station and the corresponding total battery energy. The modeling of the charging piles needs to consider the upper-level total decision charging power, the battery energy of the connected electric vehicles, the time when the current electric vehicle leaves, and the battery energy requirement at the time of departure; Step S3: To control the total power of the charging station terminal, a deep deterministic policy gradient algorithm based on the MCTS and Actor-Critic framework is used to build a reinforcement learning environment. The environmental state includes the known time-of-use electricity price, basic power load power, new energy generation output, charging pile connection status, the current total battery energy of electric vehicles after aggregation, the maximum total charging power, the time when the electric vehicle leaves and the corresponding total battery energy. The action is the upper-level total charging power, and the cost of the charging station is introduced as a cumulative discount reward. The total battery energy that does not meet the requirements at the time of departure of the electric vehicle is introduced as a sparse negative reward. Step S4: Distributed power control of the charging piles at the charging station is performed using an improved ADMM algorithm to perform a distributed solution model, ensuring that the sum of the power decided by the lower layer is equal to the total power decided by the upper layer, with the objective function of minimizing the stress and heat loss of the battery during the charging process; Step S5: Before each scheduling prediction decision calculation, a correction fault-tolerant mechanism is inserted. After the scheduling decision is completed at the previous moment, the lower-level charging pile performs actual charging, and the upper-level simulated electric vehicle performs virtual charging. After a period of time, the battery energy of the current electric vehicle is collected and returned to the upper layer. It is compared with the battery energy of the upper-level simulated electric vehicle for consistency. If the comparison result is consistent, no processing is performed. If the comparison result is inconsistent, the actual battery energy of the lower-level electric vehicle is used to update the virtual battery energy of the upper-level simulated electric vehicle. After the correction is completed, the decision scheduling of steps S3 and S4 is performed.

[0006] Furthermore, if the test fails, the second-order difference, third-order difference, etc. are performed until the stationarity test is passed. The formula for n-order difference is:

[0007] in, represents the nth-order difference of the basic power load at time t, represents the n-1 order difference of the basic power load at time t, represents the n-1 order difference of the basic power load at time t-1. The data that pass the stationarity test are divided into three parts according to the time series order. The first 90% of the data are used for model training, and the last 10% of the data are used as model parameters ( ) test, which represent the autoregressive coefficient, moving average coefficient, and difference order respectively. Finally, the basic power load data forecast value for the next three time periods is predicted by successive extrapolation. The formula is:

[0008]

[0009]

[0010] Among them, the left side of the equation represents the predicted value, and the right side of the equation represents the calculation formula of the predicted value ARIMA model. Indicates the forecast value of basic power load in the first time period in the future. Indicates the forecast value of basic power load in the second time period in the future. It represents the forecast value of basic power load in the third time period in the future. Indicates the basic power load historical data at the 0th time point in the past, Indicates the past Basic power load historical data at a time point, Indicates the past Basic power load historical data at a time point, Indicates the past Basic power load historical data at a time point, represents the first autoregressive model parameter, Indicates the autoregressive model parameters, used to describe the relationship between the current value and the value at the kth time point in the past, represents the parameters of the first moving average model, Indicates the The parameters of the moving average model are used to describe the relationship between the current value and the error at the kth time point in the past. represents the error term at time point 1, express Error term at the time point.

[0011] Furthermore, for the renewable energy power generation data, feature processing is performed on data of different dimensions. The nonlinear relationship between light intensity and photovoltaic power generation is added, and the square term of light intensity is added as a data dimension to eliminate the stationarity of power generation data. First-order difference is performed to convert minutes into sine / cosine codes. The formula is:

[0012]

[0013] Among them, the left side of the equation Respectively represent the results of sine / cosine encoding of time, t represents the original moment of the data, the power data is normalized, and the non-power data is standardized; build an LSTM-Transform hybrid model, the number of input windows is 1440 (indicates that the input sequence contains 1440 time steps, used to capture historical data patterns), the number of output windows is 3 (indicates the prediction of renewable energy power generation in the next 3 time steps), the LSTM module input dimension is 5 (the feature dimension of each time step is 5, such as voltage or temperature, etc.), the hidden layer dimension is 64 (the hidden state size of the LSTM unit is 64, used to store time series information), the number of layers is two (two layers of LSTM are stacked to enhance the ability to extract local features), used to extract local time series dependencies (capture short-term dependencies in time series), the number of multi-head attention heads of the Transform module is 4 (the multi-head attention mechanism is divided into 4 heads to improve the modeling capabilities of different representation subspaces), the core layer depth is 12 layers (the Transformer encoder contains 12 layers to handle deep global dependencies), and the feedforward network dimension is 256 (the feedforward neural network after each attention layer). The network dimension is 256, with enhanced nonlinear transformations), and global dependencies are captured through a self-attention mechanism (modeling long-term temporal relationships and weighting key information). The data in the model first generates a hidden state sequence through an LSTM (the LSTM output sequence serves as the input feature of the Transformer). The Transformer then performs self-attention on the hidden state, enhancing the weights of key time steps. Finally, the state of the last time step is taken (the final hidden state is used as the prediction input), and the prediction result is output through a fully connected layer (the linear layer maps to the output dimension). The mean square error is used as the loss function (measures the squared difference between the predicted value and the true value), the optimizer is Adam (adaptive moment estimation optimizer, efficient gradient processing), the learning rate is 0.005 (controls the parameter update step size), and the batch size is 32 (each training batch uses 32 samples). The model is trained using data from the past month (model fitting is based on historical data) and repeated for 450 sets (450 epochs are trained to ensure full convergence of the model). The predicted model is loaded for real-time prediction to predict the renewable energy power generation power for the next three time periods.

[0014] Furthermore, the total battery energy of the electric vehicle at the time of leaving does not meet the requirement as a sparse negative reward, and the expression is:

[0015] Among them, t represents the current time point, and the left side of the equation Indicates the The total cost during the time period, represents the time-of-use electricity price in time period t, represents the new energy generation output, basic power load power, and total charging power of electric vehicles in the tth time period, Indicates the length of the time period, Respectively represent Penalty and penalty factor within the time period, Respectively expressed in Aggregate the total battery energy of electric vehicles and the minimum total battery energy within the time period, represents the total return corresponding to time t, represents the discount rate, , future rewards are not as good as rewards of equal value now, so, The corresponding discount rate As k increases, it gradually decreases. represents the total return in the kth future period after discount.

[0016] Furthermore, Aggregated from the battery energy boundaries of each electric vehicle during the charging process:

[0017] The left side of the equation represents the maximum and minimum values ​​of the total battery capacity, Respectively No. of charging piles connected Electric vehicles in The minimum and maximum battery energy values ​​at each moment. Since the basic power load and new energy output are irrelevant to the action decision during the state transition, the state transition is decomposed into a deterministic part and a random disturbance:

[0018] in, Indicates the state of the next time. represents a deterministic action transfer, Represents random perturbation, and introduces Monte Carlo Tree Search (MCTS) to simulate the random distribution in the above formula, initialize the Actor network, Critic network and their corresponding target network And the experience replay pool, in the state Next, generate actions through the Actor network , perform the action , get rewards and the next state ,Will Store in the experience replay pool and randomly sample small batches (batch_size) of data from the experience pool , use the sampled data and target Critic network to calculate the target value :

[0019] The left side of the equation represents the target value calculated by the target Critic network. Represents the reward obtained after the current state s and the corresponding action a, is the discount factor used to evaluate the degree to which the model considers the future. Too high may lead to instability, too low may ignore the future, Represent the network parameters of the target action (Action) network and the target evaluation (Critic) network respectively, Indicates the sampling The next state of experience, Indicates that in state s, action strategy set , target evaluation network parameters The reward value after taking the action.

[0020] Furthermore, the optimization of the target critic network aims to minimize the loss function, which is:

[0021] Among them, the left side of the equation represents the function of the network parameters of the Critic network, and the right side of the equation is the function expression. represents the network parameters of the Critic network, Indicates the sampling size, which is equal to batch_size. The Actor network optimizes the strategy by maximizing the Q value of the Critic network. Its objective function is:

[0022] The target network uses a soft update method to slowly approach the current network:

[0023] Among them, the left side of the equation represents the network parameters of the corrected Critic network and the network parameters of the target Critic network, and the right side of the equation represents the update formula of the aforementioned parameters. represents the network parameters of the Critic network, Represents the network parameters of the target Critic network, Represents the network parameters of the Actor network, The network representing the target Actor network. The parameters are to ensure that the target network is updated slowly, and the soft update coefficient The value of is generally Repeat the above steps until all sets of training are completed. Load the trained model in real-time upper-level decision-making, input the current environment state, and output the total power of the upper-level decision. .

[0024] Furthermore, the objective function is to minimize the stress and heat loss of the battery during charging:

[0025] in, represents the decision step length, Indicates the current moment, express No. of charging piles connected Electric vehicles in Charging power during the time period, Indicates the charging efficiency of the charging pile, express The maximum energy of electric vehicle battery represents the penalty parameter, Too small will cause the original residual to decrease too slowly. Too large will lead to dual residual oscillation, and it will be difficult to approach the convergence solution. Respectively No. charging pile corresponding The global consistent variables and Lagrange multipliers of the time period, in the first iteration of the lower layer, Initialized as:

[0026] The left side of the equation represents the parameters to be initialized in the thousand-one formula. Indicates the plan for all charging piles to ensure that the battery energy of electric vehicles meets the needs of car owners when they leave the charging station. The charging power of the charging pile is restricted:

[0027] in, Respectively Charging pile The minimum / maximum charging power at the moment can quickly draw the energy change trajectory of the electric vehicle battery according to the current electric vehicle battery energy and the minimum / maximum charging power , while ensuring that the battery energy of electric vehicles is within a safe range:

[0028] in, express Electric car Battery energy at all times, They represent the maximum and minimum values ​​of the battery energy safety range respectively. After the model is built, iterative convergence calculation is performed:

[0029] The left side of the equation represents the charging power value after m iterations, and the right side of the equation represents the iteration formula. Transmitted back to the upper layer To update:

[0030] The left side of the above equation represents the parameters after m iterations, and the right side represents the update formula. represents the number of iterations, Indicates the The raw error of the iteration, Indicates the The dual error of the iteration, Represent the up / down scheduling coefficients respectively, and take , Indicates the tolerance coefficient, which ranges from 1 to 20. The optimization calculation is carried out again in the lower layer, and the above steps are repeated until the iteration converges. The convergence judgment condition is:

[0031] The left side of the inequality represents the deviation between the sum of the lower-layer charging power and the upper-layer charging power, and the right side of the inequality represents the maximum allowable deviation. Represents the total power result of the upper layer decision, Indicates the After iterations Charging pile Charging power of electric vehicles, Indicates the allowable error, the value is .

[0032] Beneficial effects: The present invention proposes a method for distributed control of electric vehicle charging power aggregation, which has the following advantages: 1. High prediction accuracy. The most suitable model is selected based on the characteristics of different data. Through layered prediction using multiple models, the accuracy of basic power load forecasts exceeds 90%, and the accuracy of new energy power generation output forecasts exceeds 85%.

[0033] 2. Optimal decision-making. The upper layer uses the DDPG deep deterministic policy gradient algorithm based on Monte Carlo tree search and actor-critic algorithms. This fully accounts for random perturbations in state transitions, ensuring that the trained model is realistic and that the total power meets the energy requirements of electric vehicle batteries. The lower layer uses a model optimization solver to ensure that each electric vehicle's charging needs are met while also reducing thermal stress and heat loss during charging, thereby extending battery life. The resulting decision is the optimal one, taking multiple objectives into account.

[0034] 3. Short computation time, meeting real-time requirements. By strengthening the model's day-ahead / month-ahead training and distributed solution, the computation time can be significantly reduced in large-scale electric vehicle charging stations compared to traditional overall model optimization and solution. For large-scale charging stations with 50 charging piles, the average single decision time is reduced from 10 seconds to 0.6 seconds.

[0035] 4. Strong robustness, using a correction fault-tolerant mechanism. It detects upper and lower layer decision mismatches caused by data transmission delays or packet loss and corrects the upper layer virtual data.

[0036] 5. Wide applicability and strong generalization capabilities. Prediction and upper-level decision-making utilize model learning. To accommodate different factors, such as electricity pricing, charging station sizes, and renewable energy generation equipment, simply modifying some parameters or retraining the model based on historical training data eliminates the need for major, complex modifications. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the method steps of the present invention; Figure 2 Establish a prediction flow chart for the ARIMA model of the present invention; Figure 3 This is the LSTM-Transformer model training and prediction flow chart of the present invention; Figure 4 This is the improved ADMM distributed algorithm diagram of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that, unless there is a conflict, the embodiments in this application and the features described in the embodiments can be combined with each other. The application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, a method for distributed control of charging power aggregation of electric vehicles includes: Step S1: pre-process the basic power load data and renewable energy power generation data (including non-power data such as light intensity and wind speed, all with a unified sampling frequency of once per minute), use linear interpolation to fit the missing data, and use 3D to fit the abnormal data. The principle is to eliminate and refit the estimate. Figure 2 As shown, an ARIMA model is built. For basic power load data, the first-order difference is used to eliminate stationarity, and the ADF test is used to check whether the data is stationary. If it fails, the second-order difference, third-order difference, etc. are performed until the stationarity test is passed. The formula for the n-order difference is:

[0040] in, Represents the nth-order difference of the basic power load at time t. The data that pass the stationarity test are divided into three parts according to the time series order. The first 90% of the data are used for model training, and the last 10% of the data are used as model parameters ( ) test, which represent the autoregressive coefficient, moving average coefficient, and difference order respectively. Finally, the basic power load data forecast value for the next three time periods is predicted by successive extrapolation. The formula is:

[0041]

[0042]

[0043] in, It represents the forecast value of basic power load in the next n time periods. Indicates the past Basic power load historical data at a time point, represents the white noise error term.

[0044] For renewable energy power generation data, feature processing is performed on data of different dimensions. Due to the nonlinear relationship between light intensity and photovoltaic power generation, a new data dimension, the square term of light intensity, is added. To eliminate the stationarity of power generation data, first-order difference is performed on it. To retain the periodicity of time, minutes are converted into sine / cosine codes. The formula is:

[0045]

[0046] in, Represents sine / cosine code respectively, and t represents the original time of data. The electric quantity data is normalized, and the non-electric quantity data is standardized (after solving the level value and standard deviation). Figure 3As shown in the figure, a hybrid LSTM-Transform model is constructed with 1440 input windows and 3 output windows. The LSTM module has an input dimension of 5, a hidden layer dimension of 64, and two layers to extract local temporal dependencies. The Transform module uses 4 multi-head attention heads, a core layer depth of 12 layers, and a feedforward network dimension of 256. Global dependencies are captured through a self-attention mechanism. In this model, data is first generated through an LSTM to generate a sequence of hidden states. The Transformer then performs self-attention on these hidden states, enhancing the weights of key time steps. Finally, the state of the last time step is taken and the prediction result is output through a fully connected layer. The mean squared error loss function is used, the Adam optimizer is used, the learning rate is 0.005, and the batch size is 32 (with the sample order shuffled). The model is trained on data from the past month and repeated 450 times. During real-time prediction, the predicted model is loaded to predict renewable energy power generation for three future time periods.

[0047] Step S2: Model the charging station main station and charging piles separately. The modeling of the charging station main station needs to consider the time-of-use electricity price, basic power load, new energy power generation output, the total battery energy of all electric vehicles in the station in the aggregated state, the time when the electric vehicle leaves the charging station and the corresponding total battery energy. The modeling of the charging piles needs to consider the upper-level total decision charging power, the battery energy of the connected electric vehicles, the current electric vehicle departure time and the battery energy requirement when leaving.

[0048] Step S3: For the total power control of the charging station terminal (i.e., the upper layer), a deep deterministic policy gradient algorithm based on Monte Carlo Tree Search (MCTS) and Actor-Critic framework is used. First, a reinforcement learning environment is built, and the environment state Including known time-of-use electricity prices, basic power load power, new energy generation output, electric pile connection status, current total battery energy after electric vehicle aggregation, maximum total charging power, electric vehicle departure time and corresponding total battery energy, action The total charging power of the upper layer is introduced, and the cost of the charging station is introduced as the cumulative discount reward. The total battery energy of the electric vehicle at the time of departure does not meet the requirements as a sparse negative reward:

[0049] in, Indicates the The total cost during the time period, represents the time-of-use electricity price in time period t, represents the new energy generation output, basic power load power, and total charging power of electric vehicles in the tth time period, Indicates the length of the time period, Respectively represent Penalty and penalty factor within the time period, Respectively expressed in Aggregate the total battery energy of electric vehicles and the minimum total battery energy within the time period, Aggregated from the battery energy boundaries of each electric vehicle during the charging process:

[0050] in, Respectively No. of charging piles connected Electric vehicles in The minimum and maximum battery energy values ​​at the moment. Since the basic power load and new energy output are irrelevant to the action decision during the state transition process, the state transition is decomposed into a deterministic part and a random disturbance:

[0051] in, represents a deterministic action transfer, Denotes random perturbations, and Monte Carlo Tree Search (MCTS) is introduced to simulate the random branches in the above formula. Next, initialize the Actor network (input layer is state dimension, output layer is action dimension), the Critic network (input layer is state + action, output layer is the action value Q of the action) and their corresponding target networks And the experience replay pool. Next, generate actions through the Actor network , perform the action , get rewards and the next state ,Will Store in the experience replay pool and randomly sample small batches (batch_size) of data from the experience pool , use the sampled data and target Critic network to calculate the target value :

[0052] in, is the discount factor used to evaluate the degree to which the model considers the future. Too much may lead to instability, too little may ignore the future. Represent the network parameters of the target action (Action) network and the target evaluation (Critic) network respectively, Indicates the sampling The optimization of the target critic network aims to minimize the loss function, which is:

[0053] in, represents the network parameters of the Critic network, Indicates the sampling size, which is equal to batch_size. The Actor network optimizes the strategy by maximizing the Q value of the Critic network. Its objective function is:

[0054] The target network uses a soft update method to slowly approach the current network:

[0055] Among them, in order to ensure the slow update of the target network, the soft update coefficient The value of is generally Repeat the above steps until all sets of training are completed. Load the trained model in real-time upper-level decision-making, input the current environment state, and output the total power of the upper-level decision .

[0056] Step S4: Distributed power control of charging piles (i.e., lower layer) at charging stations, such as Figure 4 As shown in the figure, the improved ADMM algorithm is used for distributed solution model to ensure that the total power of the lower layer decision is equal to the total power of the upper layer decision, with the objective function of minimizing the stress and heat loss of the battery during the charging process:

[0057] in, represents the decision step length, Indicates the current moment, express No. of charging piles connected Electric vehicles in Charging power during the time period, Indicates the charging efficiency of the charging pile, express The maximum energy of electric vehicle battery represents the penalty parameter, Too small will cause the original residual to decrease too slowly. Too large will lead to dual residual oscillation, and it will be difficult to approach the convergence solution. Respectively No. charging pile corresponding The global consistent variables and Lagrange multipliers of the time period, in the first iteration of the lower layer, Initialized as:

[0058] in, Indicates the plan of all charging piles. To ensure that the battery energy of electric vehicles meets the needs of owners when they leave the charging station, The charging power of the charging pile is restricted:

[0059] in, Respectively Charging pile The minimum / maximum charging power at the moment. According to the current electric vehicle battery energy and the minimum / maximum charging power, the energy change trajectory of the electric vehicle battery can be quickly drawn. At the same time, it is necessary to ensure that the battery energy of electric vehicles is within a safe range:

[0060] in, express Electric car Battery energy at all times, Respectively represent the maximum and minimum values ​​of the battery energy safety range. After the model is built, iterative convergence calculation is performed:

[0061] Will Transmitted back to the upper layer To update:

[0062] in, represents the number of iterations, Indicates the The raw error of the iteration, Indicates the The dual error of the iteration, Represent the up / down scheduling coefficients respectively, and take , represents the tolerance coefficient, taking . The updated The result is passed back to the lower layer for optimization calculation again. Repeat the above steps until the iteration converges. The convergence judgment condition is:

[0063] in, Indicates the allowable error, generally taken as .

[0064] Step S5. To prevent data inconsistencies between the upper and lower layers due to data transmission delays or packet loss, a correction and fault-tolerance mechanism is implemented before each scheduling prediction and decision calculation. After the previous scheduling decision is made, the lower-layer charging pile performs actual charging, while the upper layer performs virtual charging of the simulated electric vehicle. After a period of time, the battery energy of the current electric vehicle is collected and returned to the upper layer for consistency comparison with the battery energy of the simulated electric vehicle in the upper layer. If the comparison results are consistent, no further processing is performed. If the comparison results are inconsistent, the actual battery energy of the lower-layer electric vehicle is used to update the virtual battery energy of the simulated electric vehicle in the upper layer. After completing these corrections, the decision-making and scheduling process proceeds to steps S3 and S4.

[0065] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "disposed," "installed," "connected," "connected," and "fixed" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to mechanical connections or electrical connections; they may refer to direct connections or indirect connections through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0066] While embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for distributed control of electric vehicle charging power aggregation, characterized in that: The method includes: Step S1: preprocess the basic power load data and the new energy power generation data. For missing data, linear interpolation is used for fitting estimation. For abnormal data, the 3-principle is used to eliminate and re-fit the estimation. For the basic power load data, the first-order difference is used to eliminate the stationarity. The ADF is used to test whether the data is stationary. If it fails, the second-order difference and third-order difference are performed until the stationarity test passes. Step S2: Model the charging station master and charging piles separately. The modeling of the charging station master needs to consider time-of-use electricity prices, basic power load, new energy generation output, the total battery energy of all electric vehicles in the station in an aggregated state, the time when the electric vehicle leaves the charging station and the corresponding total battery energy. The modeling of the charging piles needs to consider the upper-level total decision charging power, the battery energy of the connected electric vehicles, the time when the current electric vehicle leaves, and the battery energy requirement at the time of departure; Step S3: To control the total power of the charging station terminal, a deep deterministic policy gradient algorithm based on the MCTS and Actor-Critic framework is used to build a reinforcement learning environment. The environmental state includes the known time-of-use electricity price, basic power load power, new energy generation output, charging pile connection status, the current total battery energy of electric vehicles after aggregation, the maximum total charging power, the time when the electric vehicle leaves and the corresponding total battery energy. The action is the upper-level total charging power, and the cost of the charging station is introduced as a cumulative discount reward. The total battery energy that does not meet the requirements at the time of departure of the electric vehicle is introduced as a sparse negative reward. Step S4: Distributed power control of the charging piles at the charging station is performed using an improved ADMM algorithm to perform a distributed solution model, ensuring that the sum of the power decided by the lower layer is equal to the total power decided by the upper layer, with the objective function of minimizing the stress and heat loss of the battery during the charging process; Step S5: Before each scheduling prediction decision calculation, a correction fault-tolerant mechanism is inserted. After the scheduling decision is completed at the previous moment, the lower-level charging pile performs actual charging, and the upper-level simulated electric vehicle performs virtual charging. After a period of time, the battery energy of the current electric vehicle is collected and returned to the upper layer. It is compared with the battery energy of the upper-level simulated electric vehicle for consistency. If the comparison result is consistent, no processing is performed. If the comparison result is inconsistent, the actual battery energy of the lower-level electric vehicle is used to update the virtual battery energy of the upper-level simulated electric vehicle. After the correction is completed, the decision scheduling of steps S3 and S4 is performed.

2. The method for distributed control of charging power of electric vehicles according to claim 1, characterized in that: If the test fails, the second-order difference, third-order difference, etc. are performed until the stationarity test is passed. The formula for n-order difference is: in, represents the nth-order difference of the basic power load at time t, represents the n-1 order difference of the basic power load at time t, represents the n-1 order difference of the basic power load at time t-1. The data that pass the stationarity test are divided into three parts according to the time series order. The first 90% of the data are used for training, and the last 10% of the data are used as parameters ( ) test, which represent the autoregressive coefficient, moving average coefficient, and difference order respectively. Finally, the basic power load data forecast value for the next three time periods is predicted by successive extrapolation. The formula is: in, Indicates the forecast value of basic power load in the first time period in the future. Indicates the forecast value of basic power load in the second time period in the future. It represents the forecast value of basic power load in the third time period in the future. Indicates the basic power load historical data at the 0th time point in the past, Indicates the past Basic power load historical data at a time point, Indicates the past Basic power load historical data at a time point, Indicates the past Basic power load historical data at a time point, represents the first autoregressive model parameter, Indicates the autoregressive model parameters, used to describe the relationship between the current value and the value at the kth time point in the past, represents the parameters of the first moving average model, Indicates the The parameters of the moving average model are used to describe the relationship between the current value and the error at the kth time point in the past. represents the error term at time point 1, express Error term at the time point.

3. The method for distributed control of charging power of electric vehicles according to claim 1, characterized in that: The new energy power generation data is processed for different dimensions of data. The nonlinear relationship between light intensity and photovoltaic power generation is added. The square term of light intensity is added as a data dimension to eliminate the stationarity of power generation data. The first-order difference is performed to convert minutes into sine / cosine codes. The formula is: in, Represent the results of sine / cosine encoding of time, t represents the original time of the data, the power data is normalized, and the non-power data is standardized; build an LSTM-Transform hybrid model with 1440 input windows and 3 output windows. The LSTM module input dimension is 5, the hidden layer dimension is 64, and there are two layers to extract local temporal dependencies. The Transform module has 4 multi-head attention heads, a core layer depth of 12 layers, and a feedforward network dimension of 256. Global dependencies are captured through the self-attention mechanism. The data first generates a hidden state sequence through LSTM, and then the Transformer performs self-attention calculation on the hidden state to enhance the weight of the key time step. Finally, the state of the last time step is taken and the prediction result is output through the fully connected layer; the mean square error is used as the loss function, the optimizer is Adam, the learning rate is 0.005, and the batch size is 32. The data from the past month is used for training, and 450 sets are repeated. The predicted model is loaded during real-time prediction to predict the new energy power generation power in the next three time periods.

4. The method for distributed control of charging power of electric vehicles according to claim 1, characterized in that: The total battery energy of the electric vehicle does not meet the requirement when leaving as a sparse negative reward, and the expression is: Where t represents the current time point, Indicates the The total cost during the time period, represents the time-of-use electricity price in time period t, represents the new energy generation output, basic power load power, and total charging power of electric vehicles in the tth time period, Indicates the length of the time period, Respectively represent Penalty and penalty factor within the time period, Respectively expressed in Aggregate the total battery energy of electric vehicles and the minimum total battery energy within the time period, represents the total return corresponding to time t, represents the discount rate, , The corresponding discount rate As k increases, it gradually decreases. represents the total return in the kth future period after discount.

5. The method for distributed control of charging power of electric vehicles as claimed in claim 4, characterized in that: Aggregated from the battery energy boundaries of each electric vehicle during the charging process: in, Respectively No. of charging piles connected Electric vehicles in The minimum and maximum battery energy values ​​at each moment. During the state transition process, the basic power load and the output of new energy sources are irrelevant to the action decision. The state transition is decomposed into a deterministic part and a random disturbance: in, Indicates the state of the next time. represents a deterministic action transfer, Represents random perturbation, and introduces Monte Carlo tree search to simulate the random distribution in the above formula, initialize the Actor network, Critic network and their corresponding target network And the experience replay pool, in the state Next, generate actions through the Actor network , perform the action , get rewards and the next state ,Will Store in the experience replay pool and randomly sample small batches of data from the experience pool , use the sampled data and target Critic network to calculate the target value : in, Represents the reward obtained after the current state s and the corresponding action a, is the discount factor used to assess the degree of consideration for the future, Too high may lead to instability, too low may ignore the future, represent the network parameters of the target action network and the target evaluation network respectively, Indicates the sampling The next state of experience, Indicates that in state s, action strategy set , target evaluation network parameters The reward value after taking the action.

6. The method for distributed control of electric vehicle charging power aggregation according to claim 5, characterized in that: The optimization of the target critic network aims to minimize the loss function, which is: in, represents the network parameters of the Critic network, Indicates the sampling size, which is equal to batch_size. The Actor network optimizes the strategy by maximizing the Q value of the Critic network. Its objective function is: The target network uses a soft update method to slowly approach the current network: in, Represents the network parameters of the target Critic network, Represents the network parameters of the Actor network, The network representing the target Actor network. The parameters are to ensure that the target network is updated slowly, and the soft update coefficient The value of Repeat until all sets of training are completed, load the trained model in real-time upper-level decision-making, input the current environment state, and output the total power of the upper-level decision .

7. The method for distributed control of charging power of electric vehicles according to claim 1, characterized in that: The objective function is to minimize the stress and heat loss of the battery during charging, and the expression is: in, represents the decision step length, Indicates the current moment, express No. of charging piles connected Electric vehicles in Charging power during the time period, Indicates the charging efficiency of the charging pile, express The maximum energy of electric vehicle battery represents the penalty parameter, Too small will cause the original residual to decrease too slowly. Too large will lead to dual residual oscillation, and it will be difficult to approach the convergence solution. Respectively No. charging pile corresponding The global consistent variables and Lagrange multipliers of the time period, in the first iteration of the lower layer, Initialized as: in, Indicates the plan for all charging piles to ensure that the battery energy of electric vehicles meets the needs of car owners when they leave the charging station. The charging power of the charging pile is restricted: in, Respectively Charging pile The minimum / maximum charging power at the moment can quickly draw the energy change trajectory of the electric vehicle battery according to the current electric vehicle battery energy and the minimum / maximum charging power , while ensuring that the battery energy of electric vehicles is within a safe range: in, express Electric car Battery energy at all times, Respectively represent the maximum and minimum values ​​of the battery energy safety range, and perform iterative convergence calculation: The left side of the equation represents the charging power value after m iterations, and the right side of the equation represents the iteration formula. Transmitted back to the upper layer To update: in, represents the number of iterations, Indicates the The raw error of the iteration, Indicates the The dual error of the iteration, Represents the up / down scheduling coefficient, and the value is , Indicates the tolerance coefficient, which ranges from 1 to 20. The optimization calculation is repeated until the iteration converges. The convergence criteria are: in, Represents the total power result of the upper layer decision, Indicates the After iterations Charging pile Charging power of electric vehicles, Indicates the allowable error, the value is .

Images