An electric vehicle non-inductive charging control method and system based on lightweight artificial intelligence technology

Abstract

This invention discloses a method and system for contactless charging control of electric vehicles based on lightweight artificial intelligence technology, belonging to the field of electric vehicle charging control technology. Addressing the problems of poor adaptability, weak scenario generalization ability, insufficient decision-making accuracy, and difficulty in balancing multiple objectives in existing contactless charging control models, this method designs a lightweight scheduling control model to achieve point-to-point contactless charging control of electric vehicles through pre-training of data, conversion of decoupled high-dimensional features into specific scheduling strategy parameters, construction of multi-dimensional scenario vectors, and generation of typical scenarios. The corresponding system includes four main modules: data processing, sample decoupling, scenario generation, and contactless control, each carrying the aforementioned core functions. This invention reduces model resource consumption and improves response efficiency through lightweight technology, enhances strategy adaptability through multi-dimensional scenario modeling, and improves decision-making accuracy through sample optimization and multi-objective constraint design, effectively balancing charging efficiency, grid stability, and equipment utilization, ensuring the safe and efficient operation of the contactless charging process.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging control technology, and specifically to a method and system for contactless charging control of electric vehicles based on lightweight artificial intelligence technology. Background Technology

[0002] With the rapid development of the electric vehicle industry and the new energy power grid, seamless charging technology has become a hot topic in the industry due to its advantages of requiring no human intervention and automating the process. Its core is to achieve efficient scheduling while ensuring charging efficiency, grid stability, and equipment utilization. However, existing technologies have obvious shortcomings: redundant control model parameters make it difficult to adapt to lightweight hardware scenarios; insufficient integration of multi-dimensional scenario information results in weak generalization ability; inefficient sample processing and severe feature coupling affect the accuracy of decision-making; and it is difficult to balance multiple objective requirements, which can easily lead to problems such as low charging efficiency and grid fluctuations.

[0003] Therefore, there is an urgent need for a high-efficiency, contactless charging control solution for electric vehicles based on lightweight artificial intelligence technology. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention proposes a method and system for contactless charging control of electric vehicles based on lightweight artificial intelligence technology. This solves the technical problems in existing contactless charging control of electric vehicles, such as the difficulty in adapting models to lightweight hardware, weak scenario generalization ability, insufficient decision-making accuracy, and difficulty in balancing multiple objective requirements.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology is provided, comprising the following steps:

[0006] Step 1: Collect raw data from the power-transportation coupled network, compress sample storage space using INT8 quantization technology and evaluate quantization error to complete the pre-training preparation of the data;

[0007] Step 2: Perform mutual information decoupling and scheduling policy mapping operations on the pre-trained stored samples. Through a lightweight gating mapping function, the decoupled high-dimensional features are converted into specific scheduling policy parameters.

[0008] Step 3: Construct multi-dimensional scene vectors, model and generate typical scenes, and cluster historical scenes to generate multi-scene data;

[0009] Step 4: Design a lightweight electric vehicle seamless charging scheduling and control model, and implement seamless charging control for the electric vehicle.

[0010] Furthermore, in step one, the specific steps for collecting and pre-training the raw data in the power-transportation coupled network are as follows:

[0011] Step 1: Collect data samples from the power-transportation coupled network;

[0012] Through monitoring equipment and dispatching systems, real-time data including power grid operation parameters, electric vehicle behavior parameters, market transaction parameters, and time characteristic parameters are acquired. The collected power grid operation parameters... Electric vehicle behavior parameters Market trading parameters Time characteristic parameters Scene vectors Format the samples and store them uniformly as a set of original samples to be processed. :

[0013] (1)

[0014] Step 2: Quantify the power grid operating status (reflecting power grid stability and load conditions) according to the following formula, and construct the power grid operating parameters. :

[0015] (2)

[0016] in, , , and These are node voltage deviation, line load rate, transformer capacity utilization rate, and reactive power deficit, respectively.

[0017] Step 3: Construct electric vehicle behavior parameters based on the following formula, which describes the charging behavior characteristics of electric vehicles (affecting charging demand and scheduling priority). :

[0018] (3)

[0019] in, , , and These are the current charging power, state of charge, estimated charging time, and vehicle-to-charging station connection topology matrix, respectively.

[0020] Step 4: Construct market trading parameters based on the following formula To obtain the market-level load fluctuation forecast status:

[0021] (4)

[0022] in, and This represents the predicted charging load fluctuation values ​​for the next 15 minutes and 1 hour;

[0023] Step 5: Based on the following formula, capture the impact of time factors on charging scenarios (such as peak hours, seasonal differences, etc.) and construct time feature parameters. :

[0024] (5)

[0025] in, , , These are time, date type, and season, respectively.

[0026] Step 6: Compress the storage space of the sample data in the original sample set according to the following formula to obtain the quantized INT8 integer q:

[0027] (6)

[0028] in, This represents the clipping function. This represents the floor function. These are the original floating-point weights; Let be the weighted mean and standard deviation, respectively. This is the scaling factor;

[0029] Step 7: Evaluate the errors generated during the quantization process using the following formula, and calculate the quantization error of the original data sample. :

[0030] (7)

[0031] Where N is the total number of parameters; This is the quantization step size; This represents the i-th original floating-point weight parameter in the model; It is the i-th INT8 integer after quantization;

[0032] Step 8: Calculate the similarity between each sample and the existing centers using the following formula. Up to recently, there are already centers distance :

[0033] (8)

[0034] Where M is the feature dimension; These are the k-th feature values ​​of the sample and the center, respectively;

[0035] Step 9: Determine the probability distribution of the initial center according to the following formula, and calculate the initial center probability. :

[0036] (9)

[0037] Where Y is the sample set;

[0038] Step 10: Based on the following formula, through the time factor Dynamically eliminate outdated samples:

[0039] (10)

[0040] in, As the initial time factor, The time decay coefficient, Generate intervals for the scene. For parameter importance weights, Let j be the rate of change of the j-th state parameter of the i-th sample; This represents the total number of state parameters.

[0041] Step 11: Minimize the prediction error according to the following formula to optimize the loss function L of the model parameters. loss :

[0042] (11)

[0043] in, The filtered sample set, These are the model's predicted values. For the true value, For parameter importance weights, This is an L2 regularization term.

[0044] Furthermore, in step two, the specific steps of the sample decoupling and mapping process are as follows:

[0045] Step 1: Standardize the collected raw multi-dimensional feature data to obtain standardized feature values. :

[0046] (12)

[0047] in, For the i-th original feature value, The mean of this feature. The standard deviation of this feature. It is the minimum value;

[0048] Step 2: Use a two-layer neural network to capture complex feature relationships and map the standardized features to a high-dimensional feature space:

[0049] (13)

[0050] in, , This is the weight matrix. For bias terms, For activation function, These are the high-dimensional features after mapping;

[0051] Step 3: Calculate the feature importance weights using the following formula. :

[0052] (14)

[0053] in, ; for The dimension; It is a non-linear activation function;

[0054] Step 4: Obtain feature importance weights Then, the key feature vectors are extracted using the following formula. :

[0055] (15)

[0056] Step 5: Measure the key feature vector using the following formula. Dependencies between features, and calculate mutual information of features after decoupling. :

[0057] (16)

[0058] in, It is a probability distribution; This represents the summation operation;

[0059] Step 6: Calculate the features of the l-th layer node v using the following formula. :

[0060] (17)

[0061] in, Let v be the set of neighbors of node v; Aggregate weights for neighbors. For self-updating weights, For bias; Let u be the feature vector of a neighbor node u in the l-th layer; Let v be the feature vector of a neighboring node v in the l-th layer;

[0062] Step 7: Map the high-dimensional features to specific scheduling policy parameters, and calculate the mapped scheduling policy parameters using the following formula. :

[0063] (18)

[0064] in, This is the gated weight matrix; Let g be the probability distribution; g is the gate vector. For bias terms, The high-dimensional features are calculated using equation (17).

[0065] Furthermore, in step three, the specific steps of the scene vector generation process are as follows:

[0066] Step 1: Sample historical data using the following formula Medium-sized samples are used to expand the scope of scenarios and improve the generalization ability of the policy, generating new scenario data. :

[0067]

[0068] in, These are decoder parameters; It is a random noise vector; It is the sigmoid function;

[0069] Step 2: Obtain new scene data Then, the following formula is used to filter out the scene data that meets the criteria. :

[0070]

[0071] in, For node coefficients; Maximum permissible power; This represents the total allowable power capacity of the system.

[0072] Step 3: Quantify the probability of different scenarios by using the following formula to provide a basis for probabilistic decision-making in scheduling strategies, and obtain the scenario probability density function. :

[0073]

[0074] in, The mixing coefficient; and Let be the mean and covariance matrix of the k-th Gaussian distribution; Let K be the Gaussian distribution function, and K be the total number of Gaussian distributions.

[0075] Step 4: Update the probability of the scene distribution to obtain the posterior probability of the k-th distribution. :

[0076]

[0077] Step 5: Iteratively optimize the model parameters to make the scene probability model better fit the actual data distribution, and output the optimal parameter estimate using the following formula. and :

[0078]

[0079]

[0080] Step 6: Determine the set of typical scenarios, and select the mean vector of each Gaussian distribution using the following formula. As the final scene collection :

[0081]

[0082] Furthermore, in step four, the specific steps for adjusting the contactless charging of the electric vehicle are as follows:

[0083] Step 1: Obtain the final scene set Then, the objective function f for regulating the contactless charging of electric vehicles is constructed using the following formula:

[0084]

[0085] in, The weights are adaptively adjusted based on the real-time status of the power grid.

[0086] Step 2: Construct a grid stability reward based on voltage deviation and frequency fluctuation using the following formula. :

[0087]

[0088] in, Rated voltage; Node voltage; This represents the total number of power grid nodes.

[0089] Step 3: Increase the ratio of charging amount to grid loss per unit time using the following formula to obtain the charging efficiency bonus. :

[0090]

[0091] in, This refers to the charging power. For charging efficiency;

[0092] Step 4: Construct an indicator to measure the utilization rate of equipment or grid capacity using the following formula. :

[0093]

[0094] in, This represents the actual total load. Rated carrying capacity of the distribution network area;

[0095] Step 5: Suppress illegal dispatching actions using the following formula to obtain over-limit penalties when voltage exceeds limits or line overloads occur. :

[0096]

[0097] in, The number of nodes exceeding the limit, This represents the total number of nodes;

[0098] Step 6: Using voltage security constraints in the grid stability reward, construct voltage security constraints using the following formula:

[0099]

[0100] Step 7: Use line load rate constraints to constrain the degree index. Construct the line load rate constraint using the following formula:

[0101]

[0102] in, The ratio of actual load to line rated load;

[0103] Step 8: Constrain the power adjustment range using the following formula:

[0104]

[0105] in, This represents the difference in charging power between the current moment and the previous moment. The current charging power of the electric vehicle;

[0106] Step 9: Constrain the maximum load-carrying capacity of the distribution network using the following formula. Obtain the carrying capacity constraints of the distribution network:

[0107]

[0108] in, This refers to the load-bearing capacity under voltage exceeding limits. The load-bearing capacity under line overload, This refers to the transformer's overload capacity. To obtain the minimum value;

[0109] Step 10: Construct action feasibility constraints based on the switching states of the standard power grid support equipment using the following formula:

[0110]

[0111] in, This indicates whether the power grid support equipment is switched on or off; 0 indicates disconnected, 1 indicates switched on, and m is the number of equipment.

[0112] Step 11: Obtain the complete state information required for scheduling decisions using equation (25). Construct the state space of the seamless charging scheduling strategy for electric vehicles. :

[0113]

[0114] Step 12: Select the charging pile power control, equipment switching control, and charging priority adjustment control scheme, and construct the action space of the electric vehicle seamless charging adjustment model using the following formula. :

[0115]

[0116] in, Prioritize charging scheduling;

[0117] Step 13: Design actions for the agent The reward and punishment mechanism is constructed by using the following formula to establish the reward function of the electric vehicle contactless charging regulation model. :

[0118]

[0119] in, For action-based penalties;

[0120] Step 14: Use a hierarchical penalty function to obtain the penalty for actions that violate voltage / load rate constraints. :

[0121]

[0122] in, This is the penalty coefficient;

[0123] Step 15: Replace fully connected layers with depthwise separable convolutions to obtain a reduction in the number of parameters. :

[0124]

[0125] in, The kernel size is [size]. Number of input / output channels;

[0126] Step 16: Concatenate the state with the basic features to generate the feature weight vector using the following formula. :

[0127]

[0128] in, This involves concatenating the state with the basic features. For gating parameters, g-weight bias term;

[0129] Step 17: Introduce entropy regularization using the following formula to obtain the loss function of the near-end policy optimization algorithm. :

[0130]

[0131] in, For deep reinforcement learning loss, For entropy regularization, The regularization coefficient is used.

[0132] Step 18: Use gating vectors to measure feature diversity and indirectly encourage action exploration. Construct the entropy regularization term using the following formula. :

[0133]

[0134] in, This is the gate vector;

[0135] Step 19: Limit the update range of network parameters by using clip loss, and update the network parameters using the following formula. :

[0136]

[0137] in, The strategy ratio, For clip loss, The learning rate; This is the clipping threshold;

[0138] Step 20: Use a greedy strategy to select the optimal action and obtain the optimal action. :

[0139]

[0140] in, To explore probability, probability To utilize probability; For policy functions; For the action space, It is a random function. To retrieve the value corresponding to the maximum index.

[0141] According to another aspect of the present invention, a contactless charging control system for electric vehicles based on lightweight artificial intelligence technology is provided, characterized in that it comprises the following modules:

[0142] Data processing module: used to perform pre-training operations on the prediction data;

[0143] Sample decoupling module: Used to decouple stored samples and map them to scheduling policies through lightweight functions;

[0144] Scene generation module: used to construct multi-dimensional scene vectors, model and generate typical scenes, and cluster historical scenes to generate multi-scene data;

[0145] Seamless control module: Used to design a seamless charging scheduling and control model for lightweight electric vehicles, and to perform seamless charging control for electric vehicles.

[0146] Furthermore, the specific steps of the data processing module for performing pre-training operations on the prediction data are as follows:

[0147] Step 1: Compress the storage space of the sample using the following formula to obtain the quantized INT8 integer q:

[0148]

[0149] in, These are the original floating-point weights; Let be the weighted mean and standard deviation, respectively. This is the scaling factor;

[0150] Step 2: Calculate the quantization error of the predicted data by evaluating the errors generated during the quantization process using the following formula. :

[0151]

[0152] Where N is the total number of parameters; This is the quantization step size;

[0153] Step 3: Calculate the similarity between each sample and existing centers using the following formula. Up to recently, there are already centers Distance:

[0154]

[0155] Where M is the feature dimension; These are the k-th feature values ​​of the sample and the center, respectively;

[0156] Step 4: Determine the probability distribution of the initial center using the following formula, and calculate the initial center probability. :

[0157]

[0158] Where Y is the sample set;

[0159] Step 5: Based on the following formula, through the time factor Dynamically eliminate outdated samples:

[0160]

[0161] in, As the initial time factor, The time decay coefficient, Generate intervals for the scene. For parameter importance weights, Let j be the rate of change of the j-th state parameter of the i-th sample;

[0162] Step 6: Minimize the prediction error according to the following formula to optimize the loss function L of the model parameters. loss :

[0163]

[0164] in, The filtered sample set, These are the model's predicted values. For the true value, For parameter importance weights, This is an L2 regularization term.

[0165] Furthermore, the specific steps of the sample decoupling module in the sample decoupling and mapping process are as follows:

[0166] Step 1: Standardize the above sample data to obtain standardized feature values. :

[0167]

[0168] in, For the i-th original feature value, The mean of this feature. The standard deviation of this feature. It is the minimum value;

[0169] Step 2: Use a two-layer neural network to capture complex feature relationships and map the standardized features to a high-dimensional feature space:

[0170]

[0171] in, , This is the weight matrix. For bias terms, For activation function, These are the high-dimensional features after mapping;

[0172] Step 3: Calculate the feature importance weights using the following formula. :

[0173]

[0174] in, ; for The dimension;

[0175] Step 4: Obtain feature importance weights Then, the key feature vectors are extracted using the following formula. :

[0176]

[0177] Step 5: Measure the key feature vector using the following formula. Dependencies between features, and calculate mutual information of features after decoupling. :

[0178]

[0179] in, It is a probability distribution;

[0180] Step 6: Calculate the features of the l-th layer node v using the following formula. :

[0181]

[0182] in, Let v be the set of neighbors of node v; Aggregate weights for neighbors. For self-updating weights, For bias;

[0183] Step 7: Map the high-dimensional features to specific scheduling policy parameters, and calculate the mapped scheduling policy parameters using the following formula. :

[0184]

[0185] in, This is the gated weight matrix; Let g be the probability distribution; g is the gate vector. This is a bias term.

[0186] Furthermore, the specific steps of the scene generation module in the scene vector generation process are as follows:

[0187] Step 1: Integrate multi-dimensional charging scenario information using the following formula to construct a scenario vector containing parameters such as power grid, vehicle, market, and time. :

[0188]

[0189] Step 2: Quantify the power grid operating status (reflecting power grid stability and load conditions) using the following formula to construct power grid operating parameters. :

[0190]

[0191] in, , , and These are node voltage deviation, line load rate, transformer capacity utilization rate, and reactive power deficit, respectively.

[0192] Step 3: Construct electric vehicle behavior parameters by describing the charging behavior characteristics of electric vehicles (affecting charging demand and scheduling priority) using the following formula. :

[0193]

[0194] in, , , and These are the current charging power, state of charge, estimated charging time, and vehicle-to-charging station connection topology matrix, respectively.

[0195] Step 4: Construct market trading parameters using the following formula To obtain the market-level load fluctuation forecast status:

[0196]

[0197] in, and This represents the predicted charging load fluctuation values ​​for the next 15 minutes and 1 hour;

[0198] Step 5: Construct time feature parameters by capturing the impact of time factors on charging scenarios (such as peak hours, seasonal differences, etc.) using the following formula. :

[0199]

[0200] in, , , These are time, date type, and season, respectively.

[0201] Step 6: Sample historical data using the following formula Medium-sized samples are used to expand the scope of scenarios and improve the generalization ability of the policy, generating new scenario data. :

[0202]

[0203] in, These are decoder parameters; It is a random noise vector; It is the sigmoid function;

[0204] Step 7: Obtain new scene data Then, the following formula is used to filter out the scene data that meets the criteria. :

[0205]

[0206] in, For node coefficients; Maximum permissible power; This represents the total allowable power capacity of the system.

[0207] Step 8: Quantify the probability of different scenarios by using the following formula to provide a basis for probabilistic decision-making in scheduling strategies, and obtain the scenario probability density function. :

[0208]

[0209] in, The mixing coefficient; and Let be the mean and covariance matrix of the k-th Gaussian distribution;

[0210] Step 9: Update the probability of the scene distribution to obtain the posterior probability of the k-th distribution. :

[0211]

[0212] Step 10: Iteratively optimize the model parameters to make the scene probability model better fit the actual data distribution, and output the optimal solution of the parameter estimation using the following formula. :

[0213]

[0214]

[0215] Step 11: Determine the set of typical scenarios, and select the mean vector of each Gaussian distribution using the following formula. As the final scene collection :

[0216]

[0217] Furthermore, the specific steps for the contactless control module to adjust the contactless charging of electric vehicles are as follows:

[0218] Step 1: Obtain the final scene set Then, the objective function f for regulating the contactless charging of electric vehicles is constructed using the following formula:

[0219]

[0220] in, The weights are adaptively adjusted based on the real-time status of the power grid.

[0221] Step 2: Construct a grid stability reward based on voltage deviation and frequency fluctuation using the following formula. :

[0222]

[0223] in, Rated voltage; Node voltage;

[0224] Step 3: Increase the ratio of charging amount to grid loss per unit time using the following formula to obtain the charging efficiency bonus. :

[0225]

[0226] in, This refers to the charging power. For charging efficiency;

[0227] Step 4: Construct an indicator to measure the utilization rate of equipment or grid capacity using the following formula. :

[0228]

[0229] in, This represents the actual total load. Rated carrying capacity of the distribution network area;

[0230] Step 5: Suppress illegal dispatching actions using the following formula to obtain over-limit penalties when voltage exceeds limits or line overloads occur. :

[0231]

[0232] in, The number of nodes exceeding the limit, This represents the total number of nodes;

[0233] Step 6: Using voltage security constraints in the grid stability reward, construct voltage security constraints using the following formula:

[0234]

[0235] Step 7: Use line load rate constraints to constrain the degree index. Construct the line load rate constraint using the following formula:

[0236]

[0237] in, The ratio of actual load to line rated load;

[0238] Step 8: Constrain the power adjustment range using the following formula:

[0239]

[0240] in, This represents the difference in charging power between the current moment and the previous moment. The current charging power of the electric vehicle;

[0241] Step 9: Constrain the maximum load-carrying capacity of the distribution network using the following formula. Obtain the carrying capacity constraints of the distribution network:

[0242]

[0243] in, This refers to the load-bearing capacity under voltage exceeding limits. The load-bearing capacity under line overload, This refers to the transformer's overload capacity.

[0244] Step 10: Construct action feasibility constraints based on the switching states of the standard power grid support equipment using the following formula:

[0245]

[0246] in, This indicates whether the power grid support equipment is switched on or off; 0 indicates disconnected, 1 indicates switched on, and m is the number of equipment.

[0247] Step 11: Construct the state space of the electric vehicle seamless charging scheduling strategy by defining the complete state information required for scheduling decisions (including factors affecting scheduling such as power grid, vehicles, and environment) using the following formula. :

[0248]

[0249] Step 12: Select control schemes such as charging pile power control, equipment switching control, and charging priority adjustment, and construct the action space of the electric vehicle seamless charging adjustment model through the following formula. :

[0250]

[0251] in, , and These are respectively the charging pile power adjustment, the switching status of grid support equipment, and the charging priority scheduling;

[0252] Step 13: Design actions for the agent The reward and punishment mechanism is constructed by using the following formula to establish the reward function of the electric vehicle contactless charging regulation model. :

[0253]

[0254] in, For action-based penalties;

[0255] Step 14: Use a hierarchical penalty function to obtain the penalty for actions that violate voltage / load rate constraints. :

[0256]

[0257] in, This is the penalty coefficient;

[0258] Step 15: Replace fully connected layers with depthwise separable convolutions to obtain a reduction in the number of parameters. :

[0259]

[0260] in, The kernel size is [size]. Number of input / output channels;

[0261] Step 16: Concatenate the state with the basic features to generate the feature weight vector using the following formula. :

[0262]

[0263] in, This involves concatenating the state with the basic features. For gating parameters, g-weight bias term;

[0264] Step 17: Introduce entropy regularization using the following formula to obtain the loss function for the near-end policy optimization algorithm:

[0265]

[0266] in, For deep reinforcement learning loss, For entropy regularization, The regularization coefficient is used.

[0267] Step 18: Use gating vectors to measure feature diversity and indirectly encourage action exploration. Construct the entropy regularization term using the following formula. :

[0268]

[0269] in, This is the gate vector;

[0270] Step 19: Limit the update range of network parameters by using clip loss, and update the network parameters using the following formula. :

[0271]

[0272] in, The strategy ratio, For clip loss, The learning rate;

[0273] Step 20: Use a greedy strategy to select the optimal action and obtain the optimal action. :

[0274]

[0275] in, To explore probability, probability To utilize probability; For policy functions; For movement space.

[0276] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0277] This invention presents a seamless charging control scheme for electric vehicles constructed using lightweight artificial intelligence technology. It compresses sample storage space through INT8 quantization and replaces traditional structures with deep separable convolutions to significantly reduce the number of model parameters and lower computational resource consumption. Secondly, it utilizes mutual information theory to achieve deep decoupling of feature vectors and combines this with the neighbor aggregation mechanism of graph neural networks to accurately extract spatial correlation features between the power grid and traffic flow, improving the accuracy of control decisions. Regarding environmental adaptability, this invention uses a Gaussian mixture model to characterize the probability distribution of multi-dimensional scenarios and enhances the generalization ability of the scheduling strategy by simulating historical scenarios and generating random noise. Finally, through an improved near-end policy optimization algorithm combined with entropy regularization and a hierarchical penalty mechanism, it achieves a dynamic balance between charging efficiency, power grid stability, and equipment utilization, suppressing illegal scheduling behaviors such as voltage exceeding limits and ensuring the safe and efficient operation of the seamless charging process in complex power-traffic coupled networks. Attached Figure Description

[0278] Figure 1 This is a schematic diagram of a contactless charging control method for electric vehicles based on lightweight artificial intelligence technology provided in an embodiment of the present invention;

[0279] Figure 2 This is a detailed step diagram of a pre-training operation on prediction data provided by an embodiment of the present invention;

[0280] Figure 3 This is a detailed step diagram of a sample decoupling and mapping process provided in an embodiment of the present invention;

[0281] Figure 4 This is a detailed step diagram of a scene vector generation process provided by an embodiment of the present invention;

[0282] Figure 5 This is a detailed step diagram illustrating the adjustment process for contactless charging of an electric vehicle according to an embodiment of the present invention. Detailed Description

[0283] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0284] like Figure 1 As shown, a method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology is provided, including the following steps:

[0285] Step 1: Collect raw data from the power-transportation coupled network, compress sample storage space using INT8 quantization technology and evaluate quantization error to complete the pre-training preparation of the data;

[0286] Step 2: Perform mutual information decoupling and scheduling policy mapping operations on the pre-trained stored samples. Through a lightweight gating mapping function, the decoupled high-dimensional features are converted into specific scheduling policy parameters.

[0287] Step 3: Construct multi-dimensional scene vectors, model and generate typical scenes, and cluster historical scenes to generate multi-scene data;

[0288] Step 4: Design a lightweight electric vehicle seamless charging scheduling and control model, and implement seamless charging control for the electric vehicle.

[0289] like Figure 2 The diagram illustrates the specific steps involved in collecting and pre-training raw data from a power-transportation coupled network, as provided in an embodiment of the present invention. The specific steps are as follows:

[0290] Step A1: Collect data samples from the power-transportation coupled network;

[0291] Through monitoring equipment and dispatching systems, real-time data including power grid operation parameters, electric vehicle behavior parameters, market transaction parameters, and time characteristic parameters are acquired. The collected power grid operation parameters... Electric vehicle behavior parameters Market trading parameters Time characteristic parameters Scene vectors Format the samples and store them uniformly as a set of original samples to be processed. :

[0292] (1)

[0293] Step A2: Quantify the power grid operating status (reflecting power grid stability and load conditions) using the following formula to construct power grid operating parameters. :

[0294] (2)

[0295] in, , , and These are node voltage deviation, line load rate, transformer capacity utilization rate, and reactive power deficit, respectively.

[0296] Step A3: Construct electric vehicle behavior parameters by describing the charging behavior characteristics of electric vehicles (affecting charging demand and scheduling priority) using the following formula. :

[0297] (3)

[0298] in, , , and These are the current charging power, state of charge, estimated charging time, and vehicle-to-charging station connection topology matrix, respectively.

[0299] Step A4: Construct market trading parameters using the following formula To obtain the market-level load fluctuation forecast status:

[0300] (4)

[0301] in, and This represents the predicted charging load fluctuation values ​​for the next 15 minutes and 1 hour;

[0302] Step A5: Capture the impact of time factors on charging scenarios (such as peak hours, seasonal differences, etc.) using the following formula to construct time feature parameters. :

[0303] (5)

[0304] in, , , These are time, date type, and season, respectively.

[0305] Step A6: Compress the storage space of the sample data in the original sample set using the following formula to obtain the quantized INT8 integer q:

[0306] (6)

[0307] in, This represents the clipping function. This represents the floor function. These are the original floating-point weights; Let be the weighted mean and standard deviation, respectively. This is the scaling factor;

[0308] Step A7: Calculate the quantization error of the original data sample by evaluating the error generated during the quantization process using the following formula. :

[0309] (7)

[0310] Where N is the total number of parameters; This is the quantization step size; This represents the i-th original floating-point weight parameter in the model; It is the i-th INT8 integer after quantization;

[0311] Step A8: Calculate the similarity of each sample to existing centers using the following formula. Up to recently, there are already centers distance :

[0312] (8)

[0313] Where M is the feature dimension; These are the k-th feature values ​​of the sample and the center, respectively;

[0314] Step A9: Determine the probability distribution of the initial center using the following formula, and calculate the initial center probability. :

[0315] (9)

[0316] Where Y is the sample set;

[0317] Step A10: Based on the following formula, through the time factor Dynamically eliminate outdated samples:

[0318] (10)

[0319] in, As the initial time factor, The time decay coefficient, Generate intervals for the scene. For parameter importance weights, Let j be the rate of change of the j-th state parameter of the i-th sample; This represents the total number of state parameters.

[0320] Step A11: Minimize the prediction error according to the following formula to optimize the loss function L of the model parameters. loss :

[0321] (11)

[0322] in, The filtered sample set, These are the model's predicted values. For the true value, For parameter importance weights, This is an L2 regularization term.

[0323] like Figure 3The diagram illustrates the specific steps of a sample decoupling and mapping process provided in an embodiment of the present invention. The specific steps are as follows:

[0324] Step B1: Standardize the collected raw multi-dimensional feature data to obtain standardized feature values. :

[0325] (12)

[0326] in, For the i-th original feature value, The mean of this feature. The standard deviation of this feature. It is the minimum value;

[0327] Step B2: Use a two-layer neural network to capture complex feature relationships and map the standardized features to a high-dimensional feature space:

[0328] (13)

[0329] in, , This is the weight matrix. For bias terms, For activation function, These are the high-dimensional features after mapping; These are the standardized eigenvalues.

[0330] Step B3: Calculate the feature importance weights using the following formula. :

[0331] (14)

[0332] in, ; for The dimension; It is a non-linear activation function;

[0333] Step B4: Obtain feature importance weights Then, the key feature vectors are extracted using the following formula. :

[0334] (15)

[0335] Step B5: Measure the key feature vector using the following formula. Dependencies between features, and calculate mutual information of features after decoupling. :

[0336] (16)

[0337] in, It is a probability distribution; This represents the summation operation;

[0338] Step B6: Calculate the features of the l-th layer node v using the following formula. :

[0339] (17)

[0340] in, Let v be the set of neighbors of node v; Aggregate weights for neighbors. For self-updating weights, For bias; Let u be the feature vector of a neighbor node u in the l-th layer; Let v be the feature vector of a neighboring node v in the l-th layer;

[0341] Step B7: Map the high-dimensional features to specific scheduling policy parameters, and calculate the mapped scheduling policy parameters using the following formula. :

[0342] (18)

[0343] in, This is the gated weight matrix; Let g be the probability distribution; g is the gate vector. For bias terms, The high-dimensional features are calculated using equation (17).

[0344] like Figure 4 The diagram shows the specific steps of a scene vector generation process provided by an embodiment of the present invention. The specific steps are as follows:

[0345] Step C1: Sample historical data using the following formula Medium-sized samples are used to expand the scope of scenarios and improve the generalization ability of the policy, generating new scenario data. :

[0346]

[0347] Step C2: Obtain new scene data Then, the following formula is used to filter out the scene data that meets the criteria. :

[0348]

[0349] in, For node coefficients; Maximum permissible power; This represents the total allowable power capacity of the system.

[0350] Step C3: Quantify the probability of different scenarios by using the following formula to provide a basis for probabilistic decision-making in scheduling strategies, and obtain the scenario probability density function. :

[0351]

[0352] in, The mixing coefficient; and Let be the mean and covariance matrix of the k-th Gaussian distribution; Let K be the Gaussian distribution function, and K be the total number of Gaussian distributions.

[0353] Step C4: Update the probability of the scene distribution to obtain the posterior probability of the k-th distribution. :

[0354]

[0355] Step C5: Iteratively optimize the model parameters to make the scene probability model better fit the actual data distribution, and output the optimal parameter estimate using the following formula. and :

[0356]

[0357]

[0358] Step C6: Determine the set of typical scenarios, and select the mean vector of each Gaussian distribution using the following formula. As the final scene collection :

[0359]

[0360] like Figure 5 The diagram shows the specific steps of a contactless charging adjustment method for an electric vehicle according to an embodiment of the present invention. The specific steps are as follows:

[0361] Step D1: Obtain the final scene set Then, the objective function f for regulating the contactless charging of electric vehicles is constructed using the following formula:

[0362]

[0363] in, The weights are adaptively adjusted based on the real-time status of the power grid.

[0364] Step D2: Construct a grid stability reward based on the degree of voltage deviation and frequency fluctuation using the following formula. :

[0365]

[0366] in, Rated voltage; Node voltage; This represents the total number of power grid nodes.

[0367] Step D3: Increase the ratio of charging amount to grid loss per unit time using the following formula to obtain the charging efficiency bonus. :

[0368]

[0369] in, This refers to the charging power. For charging efficiency;

[0370] Step D4: Construct an indicator to measure the utilization rate of equipment or grid capacity using the following formula. :

[0371]

[0372] in, This represents the actual total load. Rated carrying capacity of the distribution network area;

[0373] Step D5: Suppress illegal dispatching actions using the following formula to obtain over-limit penalties when voltage exceeds limits or line overloads occur. :

[0374]

[0375] in, The number of nodes exceeding the limit, This represents the total number of nodes;

[0376] Step D6: Using voltage security constraints in the grid stability reward, construct voltage security constraints using the following formula:

[0377] (31)

[0378] Step D7: Use line load rate constraints to constrain the degree index. Construct the line load rate constraint using the following formula:

[0379]

[0380] in, The ratio of actual load to line rated load;

[0381] Step D8: Constrain the power adjustment range using the following formula:

[0382] (32)

[0383] in, This represents the difference in charging power between the current moment and the previous moment. The current charging power of the electric vehicle;

[0384] Step D9: Constrain the maximum load-carrying capacity of the distribution network using the following formula. Obtain the carrying capacity constraints of the distribution network:

[0385] (33)

[0386] in, This refers to the load-bearing capacity under voltage exceeding limits. The load-bearing capacity under line overload, This refers to the transformer's overload capacity. To obtain the minimum value;

[0387] Step D10: Construct action feasibility constraints based on the switching states of the standard power grid support equipment using the following formula:

[0388]

[0389] in, This indicates whether the power grid support equipment is switched on or off; 0 indicates disconnected, 1 indicates switched on, and m is the number of equipment.

[0390] Step D11: Obtain the complete state information required for scheduling decisions using equation (25). Construct the state space of the seamless charging scheduling strategy for electric vehicles. :

[0391]

[0392] Step D12: Select the charging pile power control, equipment switching control, and charging priority adjustment control scheme, and construct the action space of the electric vehicle seamless charging adjustment model through the following formula. :

[0393]

[0394] in, Prioritize charging scheduling;

[0395] Step D13: Design actions for the agent The reward and punishment mechanism is constructed by using the following formula to establish the reward function of the electric vehicle contactless charging regulation model. :

[0396]

[0397] in, f represents the action penalty term; f is the objective function for adjusting the inductive charging of electric vehicles.

[0398] Step D14: Use a hierarchical penalty function to obtain the penalty for actions that violate voltage / load rate constraints. :

[0399]

[0400] in, This is the penalty coefficient;

[0401] Step D15: Replace fully connected layers with depthwise separable convolutions to obtain a reduction in the number of parameters. :

[0402]

[0403] in, The kernel size is [size]. Number of input / output channels;

[0404] Step D16: Concatenate the state with the basic features to generate the feature weight vector using the following formula. :

[0405]

[0406] in, This involves concatenating the state with the basic features. For gating parameters, g-weight bias term;

[0407] Step D17: Introduce entropy regularization using the following formula to obtain the loss function of the near-end policy optimization algorithm. :

[0408]

[0409] in, For deep reinforcement learning loss, For entropy regularization, The regularization coefficient is used.

[0410] Step D18: Use gating vectors to measure feature diversity and indirectly encourage action exploration. Construct the entropy regularization term using the following formula. :

[0411]

[0412] in, This is the gate vector;

[0413] Step D19: Limit the update range of network parameters by using clip loss, and update the network parameters using the following formula. :

[0414]

[0415] in, The strategy ratio, For clip loss, The learning rate; This is the clipping threshold;

[0416] Step D20: Use a greedy strategy to select the optimal action and obtain the optimal action. :

[0417]

[0418] in, To explore probability, probability To utilize probability; For policy functions; For the action space, It is a random function. To retrieve the value corresponding to the maximum index.

[0419] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0420] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology, characterized in that, The method includes the following steps: Step 1: Collect raw data from the power-transportation coupled network, compress sample storage space using INT8 quantization technology and evaluate quantization error to complete the pre-training preparation of the data; Step 2: Perform mutual information decoupling and scheduling policy mapping operations on the pre-trained stored samples. Through a lightweight gating mapping function, the decoupled high-dimensional features are converted into specific scheduling policy parameters. Step 3: Construct multi-dimensional scene vectors, model and generate typical scenes, and cluster historical scenes to generate multi-scene data; Step 4: Design a lightweight electric vehicle seamless charging scheduling and control model, and implement seamless charging control for the electric vehicle.

2. The method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology according to claim 1, characterized in that, In step one, the specific steps for collecting and pre-training the raw data in the power-transportation coupled network are as follows: Step 1-1: Collect data samples from the power-transportation coupled network; Through monitoring equipment and dispatching systems, real-time data including power grid operation parameters, electric vehicle behavior parameters, market transaction parameters, and time characteristic parameters are acquired. The collected power grid operation parameters... Electric vehicle behavior parameters Market trading parameters Time characteristic parameters Scene vectors Format the samples and store them uniformly as a set of original samples to be processed. : (1) Steps 1-2: Quantify the power grid operating status (reflecting power grid stability and load conditions) using the following formula to construct power grid operating parameters. : (2) in, , , and These are node voltage deviation, line load rate, transformer capacity utilization rate, and reactive power deficit, respectively. Steps 1-3: Construct electric vehicle behavior parameters by describing the charging behavior characteristics of electric vehicles (affecting charging demand and scheduling priority) using the following formula. : (3) in, , , and These are the current charging power, state of charge, estimated charging time, and vehicle-to-charging station connection topology matrix, respectively. Steps 1-4: Construct market trading parameters using the following formula To obtain the market-level load fluctuation forecast status: (4) in, and This represents the predicted charging load fluctuation values ​​for the next 15 minutes and 1 hour; Steps 1-5: Capture the impact of time factors on charging scenarios (such as peak hours, seasonal differences, etc.) using the following formula to construct time feature parameters. : (5) in, , , These are time, date type, and season, respectively. Steps 1-6: Compress the storage space of the sample data in the original sample set using the following formula to obtain the quantized INT8 integer q: (6) in, This represents the clipping function. This represents the floor function. These are the original floating-point weights; Let be the weighted mean and standard deviation, respectively. This is the scaling factor; Steps 1-7: Evaluate the errors generated during the quantization process using the following formula, and calculate the quantization error of the original data sample. : (7) Where N is the total number of parameters; This is the quantization step size; This represents the i-th original floating-point weight parameter in the model; It is the i-th INT8 integer after quantization; Steps 1-8: Calculate the similarity between each sample and existing centers using the following formula. Up to recently, there are already centers distance : (8) Where M is the feature dimension; These are the k-th feature values ​​of the sample and the center, respectively; Steps 1-9: Determine the probability distribution of the initial center using the following formula, and calculate the initial center probability. : (9) Where Y is the sample set; Steps 1-10: Based on the following formula, using the time factor... Dynamically eliminate outdated samples: (10) in, As the initial time factor, The time decay coefficient, Generate intervals for the scene. For parameter importance weights, Let j be the rate of change of the j-th state parameter of the i-th sample; This represents the total number of state parameters. Steps 1-11: Minimize the prediction error and optimize the loss function L of the model parameters according to the following formula. loss : (11) in, The filtered sample set, These are the model's predicted values. For the true value, For parameter importance weights, This is an L2 regularization term.

3. The method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology according to claim 1, characterized in that, In step two, the specific steps of the sample decoupling and mapping process are as follows: Step 2-1: Standardize the collected raw multi-dimensional feature data to obtain standardized feature values. : (12) in, For the i-th original feature value, The mean of this feature. The standard deviation of this feature. It is the minimum value; Step 2-2: Use a two-layer neural network to capture complex feature relationships and map the standardized features to a high-dimensional feature space: (13) in, , This is the weight matrix. For bias terms, For activation function, These are the high-dimensional features after mapping; Steps 2-3: Calculate the feature importance weights using the following formula. : (14) in, ; for The dimension; It is a non-linear activation function; Steps 2-4: Obtain feature importance weights Then, the key feature vector is extracted using the following formula. : (15) Steps 2-5: Measure the key feature vector using the following formula. Dependencies between features, and calculate mutual information of features after decoupling. : (16) in, It is a probability distribution; This represents the summation operation; Steps 2-6: Calculate the features of the l-th layer node v using the following formula. : (17) in, Let v be the set of neighbors of node v; Aggregate weights for neighbors. For self-updating weights, For bias; Let u be the feature vector of a neighbor node u in the l-th layer; Let v be the feature vector of a neighboring node v in the l-th layer; Steps 2-7: Map the high-dimensional features to specific scheduling policy parameters, and calculate the mapped scheduling policy parameters using the following formula. : (18) in, This is the gated weight matrix; Let g be the probability distribution; g is the gate vector. For bias terms, The high-dimensional features are calculated using equation (17).

4. The method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology according to claim 1, characterized in that, In step three, the specific steps of the scene vector generation process are as follows: Step 3-1: Sample historical data using the following formula Medium-sized samples are used to expand the scope of scenarios and improve the generalization ability of the policy, generating new scenario data. : in, These are decoder parameters; It is a random noise vector; It is the sigmoid function; Step 3-2: Obtain new scene data Then, the following formula is used to filter out the scene data that meets the criteria. : in, For node coefficients; Maximum permissible power; This represents the total allowable power capacity of the system. Step 3-3: Quantify the probability of different scenarios by using the following formula to provide a basis for probabilistic decision-making in scheduling strategies, and obtain the scenario probability density function. : in, The mixing coefficient; and Let be the mean and covariance matrix of the k-th Gaussian distribution; Let K be the Gaussian distribution function, and K be the total number of Gaussian distributions. Steps 3-4: Update the probability of the scene distribution to obtain the posterior probability of the k-th distribution. : Steps 3-5: Iteratively optimize the model parameters to make the scene probability model better fit the actual data distribution, and output the optimal parameter estimate using the following formula. and : Steps 3-6: Determine the set of typical scenarios, and select the mean vector of each Gaussian distribution using the following formula. As the final scene collection : 。 5. The method for contactless charging control of electric vehicles based on lightweight artificial intelligence technology according to claim 1, characterized in that, In step four, the specific steps for adjusting the contactless charging of electric vehicles are as follows: Step 4-1: Obtain the final scene set Then, the objective function f for regulating the contactless charging of electric vehicles is constructed using the following formula: in, The weights are adaptively adjusted based on the real-time status of the power grid. Step 4-2: Construct a grid stability reward based on voltage deviation and frequency fluctuation using the following formula. : in, Rated voltage; Node voltage; This represents the total number of power grid nodes. Step 4-3: Increase the ratio of charging amount to grid loss per unit time using the following formula to obtain the charging efficiency bonus. : in, This refers to the charging power. For charging efficiency; Step 4-4: Construct an indicator to measure the utilization rate of equipment or power grid capacity using the following formula. : in, This represents the actual total load. Rated carrying capacity of the distribution network area; Steps 4-5: Suppress illegal dispatching actions using the following formula to obtain over-limit penalties when voltage exceeds limits or line overload occurs. : in, The number of nodes exceeding the limit, This represents the total number of nodes. Steps 4-6: Using voltage security constraints in the grid stability reward, construct voltage security constraints using the following formula: Steps 4-7: Use line load rate constraints to constrain the degree index. The line load rate constraint is constructed using the following formula: in, The ratio of actual load to line rated load; Steps 4-8: Constrain the power adjustment range using the following formula: in, This represents the difference in charging power between the current moment and the previous moment. The current charging power of the electric vehicle; Steps 4-9: Constrain the maximum load-carrying capacity of the distribution network using the following formula. Obtain the carrying capacity constraints of the distribution network: in, This refers to the load-bearing capacity under voltage exceeding limits. The load-bearing capacity under line overload, This refers to the transformer's overload capacity. To obtain the minimum value; Step 4-10: Construct action feasibility constraints based on the switching states of the standard power grid support equipment using the following formula: in, This indicates whether the power grid support equipment is switched on or off; 0 indicates disconnected, 1 indicates switched on, and m is the number of equipment. Step 4-11: Obtain the complete state information required for scheduling decisions using equation (25). Construct the state space of the seamless charging scheduling strategy for electric vehicles. : Steps 4-12: Select the charging pile power control, equipment switching control, and charging priority adjustment control schemes, and construct the action space of the electric vehicle seamless charging adjustment model using the following formula. : in, Prioritize charging scheduling; Step 4-13: Design actions for the agent The reward and punishment mechanism is constructed by using the following formula to establish the reward function of the electric vehicle contactless charging regulation model. : in, For action-based penalties; Step 4-14: Use a hierarchical penalty function to obtain the penalty for actions that violate voltage / load rate constraints. : in, This is the penalty coefficient; Step 4-15: Replace fully connected layers with depthwise separable convolutions to obtain a reduction in the number of parameters. : in, The kernel size is [size]. Number of input / output channels; Step 4-16: Concatenate the state with the basic features to generate the feature weight vector using the following formula. : in, This involves concatenating the state with the basic features. For gating parameters, g-weight bias term; Step 4-17: Introduce entropy regularization using the following formula to obtain the loss function of the near-end policy optimization algorithm. : in, For deep reinforcement learning loss, For entropy regularization, The regularization coefficient is used. Step 4-18: Use gating vectors to measure feature diversity and indirectly encourage action exploration. Construct the entropy regularization term using the following formula. : in, This is the gate vector; Step 4-19: Limit the update range of network parameters by using clip loss, and update the network parameters using the following formula. : in, The strategy ratio, For clip loss, The learning rate; This is the clipping threshold; Step 4-20: Use a greedy strategy to select the optimal action and obtain the optimal action. : in, To explore probability, probability To utilize probability; For policy functions; For the action space, It is a random function. To retrieve the value corresponding to the maximum index.

6. A contactless charging control system for electric vehicles based on lightweight artificial intelligence technology, characterized in that, For implementing the contactless charging control method for electric vehicles based on lightweight artificial intelligence technology as described in any one of claims 1-5, the system comprises the following modules: Data processing module: used to perform pre-training operations on the prediction data; Sample decoupling module: Used to decouple stored samples and map them to scheduling policies through lightweight functions; Scene generation module: used to construct multi-dimensional scene vectors, model and generate typical scenes, and cluster historical scenes to generate multi-scene data; Seamless control module: Used to design a seamless charging scheduling and control model for lightweight electric vehicles, and to perform seamless charging control for electric vehicles.

7. The electric vehicle contactless charging control system based on lightweight artificial intelligence technology according to claim 6, characterized in that, The specific steps of the data processing module to perform pre-training operations on the prediction data are as follows: Step 1: Compress the storage space of the sample using the following formula to obtain the quantized INT8 integer q: in, These are the original floating-point weights; Let be the weighted mean and standard deviation, respectively. This is the scaling factor; Step 2: Calculate the quantization error of the predicted data by evaluating the errors generated during the quantization process using the following formula. : Where N is the total number of parameters; This is the quantization step size; Step 3: Calculate the similarity between each sample and existing centers using the following formula. Up to recently, there are already centers Distance: Where M is the feature dimension; These are the k-th feature values ​​of the sample and the center, respectively; Step 4: Determine the probability distribution of the initial center using the following formula, and calculate the initial center probability. : Where Y is the sample set; Step 5: Based on the following formula, through the time factor Dynamically eliminate outdated samples: in, As the initial time factor, The time decay coefficient, Generate intervals for the scene. For parameter importance weights, Let j be the rate of change of the j-th state parameter of the i-th sample; Step 6: Minimize the prediction error according to the following formula to optimize the loss function L of the model parameters. loss : in, The filtered sample set, These are the model's predicted values. For the true value, For parameter importance weights, This is an L2 regularization term.

8. The electric vehicle contactless charging control system based on lightweight artificial intelligence technology according to claim 6, characterized in that, The specific steps of the sample decoupling module in the sample decoupling and mapping process are as follows: Step 1: Standardize the above sample data to obtain standardized feature values. : in, For the i-th original feature value, The mean of this feature. The standard deviation of this feature. It is the minimum value; Step 2: Use a two-layer neural network to capture complex feature relationships and map the standardized features to a high-dimensional feature space: in, , This is the weight matrix. For bias terms, For activation function, These are the high-dimensional features after mapping; Step 3: Calculate the feature importance weights using the following formula. : in, ; for The dimension; Step 4: Obtain feature importance weights Then, the key feature vector is extracted using the following formula. : Step 5: Measure the key feature vector using the following formula. Dependencies between features, and calculate mutual information of features after decoupling. : in, It is a probability distribution; Step 6: Calculate the features of the l-th layer node v using the following formula. : in, Let v be the set of neighbors of node v; Aggregate weights for neighbors. For self-updating weights, For bias; Step 7: Map the high-dimensional features to specific scheduling policy parameters, and calculate the mapped scheduling policy parameters using the following formula. : in, This is the gated weight matrix; Let g be the probability distribution; g is the gate vector. This is a bias term.

9. As described in claim 6, characterized in that, The specific steps of the scene generation module in the scene vector generation process are as follows: Step 1: Integrate multi-dimensional charging scenario information using the following formula to construct a scenario vector containing parameters such as power grid, vehicle, market, and time. : Step 2: Quantify the power grid operating status (reflecting power grid stability and load conditions) using the following formula to construct power grid operating parameters. : in, , , and These are node voltage deviation, line load rate, transformer capacity utilization rate, and reactive power deficit, respectively. Step 3: Construct electric vehicle behavior parameters by describing the charging behavior characteristics of electric vehicles (affecting charging demand and scheduling priority) using the following formula. : in, , , and These are the current charging power, state of charge, estimated charging time, and vehicle-to-charging station connection topology matrix, respectively. Step 4: Construct market trading parameters using the following formula To obtain the market-level load fluctuation forecast status: in, and This represents the predicted charging load fluctuation values ​​for the next 15 minutes and 1 hour; Step 5: Construct time feature parameters by capturing the impact of time factors on charging scenarios (such as peak hours, seasonal differences, etc.) using the following formula. : in, , , These are time, date type, and season, respectively. Step 6: Sample historical data using the following formula Medium-sized samples are used to expand the scope of scenarios and improve the generalization ability of the policy, generating new scenario data. : in, These are decoder parameters; It is a random noise vector; It is the sigmoid function; Step 7: Obtain new scene data Then, the following formula is used to filter out the scene data that meets the criteria. : in, For node coefficients; Maximum permissible power; This represents the total allowable power capacity of the system. Step 8: Quantify the probability of different scenarios by using the following formula to provide a basis for probabilistic decision-making in scheduling strategies, and obtain the scenario probability density function. : in, The mixing coefficient; and Let be the mean and covariance matrix of the k-th Gaussian distribution; Step 9: Update the probability of the scene distribution to obtain the posterior probability of the k-th distribution. : Step 10: Iteratively optimize the model parameters to make the scene probability model better fit the actual data distribution, and output the optimal solution of the parameter estimation using the following formula. : Step 11: Determine the set of typical scenarios, and select the mean vector of each Gaussian distribution using the following formula. As the final scene collection : 。 10. As described in claim 6, characterized in that, The specific steps for the contactless control module to adjust the contactless charging of electric vehicles are as follows: Step 1: Obtain the final scene set Then, the objective function f for regulating the contactless charging of electric vehicles is constructed using the following formula: in, The weights are adaptively adjusted based on the real-time status of the power grid. Step 2: Construct a grid stability reward based on voltage deviation and frequency fluctuation using the following formula. : in, Rated voltage; Node voltage; Step 3: Increase the ratio of charging amount to grid loss per unit time using the following formula to obtain the charging efficiency bonus. : in, This refers to the charging power. For charging efficiency; Step 4: Construct an indicator to measure the utilization rate of equipment or grid capacity using the following formula. : in, This represents the actual total load. Rated carrying capacity of the distribution network area; Step 5: Suppress illegal dispatching actions using the following formula to obtain over-limit penalties when voltage exceeds limits or line overloads occur. : in, The number of nodes exceeding the limit, This represents the total number of nodes. Step 6: Using voltage security constraints in the grid stability reward, construct voltage security constraints using the following formula: Step 7: Use line load rate constraints to constrain the degree index. Construct the line load rate constraint using the following formula: in, The ratio of actual load to line rated load; Step 8: Constrain the power adjustment range using the following formula: in, This represents the difference in charging power between the current moment and the previous moment. The current charging power of the electric vehicle; Step 9: Constrain the maximum load-carrying capacity of the distribution network using the following formula. Obtain the carrying capacity constraints of the distribution network: in, This refers to the load-bearing capacity under voltage exceeding limits. The load-bearing capacity under line overload, This refers to the transformer's overload capacity. Step 10: Construct action feasibility constraints based on the switching states of the standard power grid support equipment using the following formula: in, This indicates whether the power grid support equipment is switched on or off; 0 indicates disconnected, 1 indicates switched on, and m is the number of equipment. Step 11: Construct the state space of the electric vehicle seamless charging scheduling strategy by defining the complete state information required for scheduling decisions (including factors affecting scheduling such as power grid, vehicles, and environment) using the following formula. : Step 12: Select control schemes such as charging pile power control, equipment switching control, and charging priority adjustment, and construct the action space of the electric vehicle seamless charging adjustment model through the following formula. : in, , and These are respectively the charging pile power adjustment, the switching status of grid support equipment, and the charging priority scheduling; Step 13: Design actions for the agent The reward and punishment mechanism is constructed by using the following formula to establish the reward function of the electric vehicle contactless charging regulation model. : in, For action-based penalties; Step 14: Use a hierarchical penalty function to obtain the penalty for actions that violate voltage / load rate constraints. : in, This is the penalty coefficient; Step 15: Replace fully connected layers with depthwise separable convolutions to obtain a reduction in the number of parameters. : in, The kernel size is [size]. Number of input / output channels; Step 16: Concatenate the state with the basic features to generate the feature weight vector using the following formula. : in, This involves concatenating the state with the basic features. For gating parameters, g-weight bias term; Step 17: Introduce entropy regularization using the following formula to obtain the loss function for the near-end policy optimization algorithm: in, For deep reinforcement learning loss, For entropy regularization, The regularization coefficient is used. Step 18: Use gating vectors to measure feature diversity and indirectly encourage action exploration. Construct the entropy regularization term using the following formula. : in, This is the gate vector; Step 19: Limit the update range of network parameters by using clip loss, and update the network parameters using the following formula. : in, The strategy ratio, For clip loss, The learning rate; Step 20: Use a greedy strategy to select the optimal action and obtain the optimal action. : in, To explore probability, probability To utilize probability; For policy functions; For movement space.

Images