A distributed automobile liquid cooling supercharging control method and system

Abstract

The present invention relates to the field of data processing technology, and specifically provides a distributed automobile liquid-cooled supercharging control method and system, which merges the supercharging network operation information and real-time charging parameter information in the current time period to obtain merged information; performs internal weight focusing and allocation on the merged information to obtain a temporary vector representation, and performs multi-objective decision prediction based on the temporary vector representation to obtain target charging scheduling strategy information at one or more levels; based on the target charging parameter information determined by the target charging scheduling strategy information, initiates charging scheduling decisions for supercharging charging demands in the next time period when supercharging of charging piles is managed. Because different types of charging parameter information and supercharging network operation information are prioritized based on the internal weight focusing strategy, the effect of advanced features is expanded, so that the output target charging scheduling strategy information can increase the scheduling effect when used for charging scheduling decisions in the next time period.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a distributed automobile liquid cooling supercharging control method and system. Background Art

[0002] With the vigorous development of the electric vehicle industry, charging issues have increasingly become a key factor restricting the further development of the industry. At present, traditional automobile charging technology has gradually exposed many limitations in the face of the rapidly growing demand for electric vehicles. On the one hand, slow charging speed is a prominent problem. The long waiting time for charging seriously affects the user experience and reduces the convenience and efficiency of electric vehicles. On the other hand, the stability and safety of the charging process are also difficult to be effectively guaranteed. For example, the instability of charging power and abnormal changes in battery temperature may not only shorten the battery life, but also pose certain safety hazards. Furthermore, the existing charging control method lacks effective scheduling strategies when dealing with large-scale distributed charging scenarios, making it difficult to achieve reasonable allocation and optimal utilization of charging resources, which can easily cause excessive load on local power grids and affect the stable operation of the power grid. At the same time, the existing charging technology has poor adaptability to different battery states and charging requirements, and cannot be flexibly adjusted according to real-time charging parameters, resulting in poor charging results. These problems urgently need a more efficient, stable and intelligent distributed automobile charging control technology to solve them in order to meet the needs of the rapid development of the electric vehicle industry, improve the user's charging experience, and ensure the stable operation of the power grid and the safe use of batteries. Summary of the invention

[0003] In view of this, the present invention provides a distributed automobile liquid cooling supercharging control method and system. The technical solution of the present invention is implemented as follows:

[0004] On the one hand, the present invention provides a distributed automobile liquid-cooled supercharging control method, the method comprising: obtaining real-time charging parameter information in a current time period, and supercharging network operation information caused by charging scheduling decisions based on the real-time charging parameter information; the real-time charging parameter information includes charging power, battery status and charging rate; the real-time charging parameter information is determined based on past charging scheduling strategy information in a previous time period; the supercharging network operation information and the real-time charging parameter information in the current time period are merged to obtain merged information; internal weight focusing and allocation are performed on the merged information to obtain a temporary vector representation, and multi-objective decision prediction is performed based on the temporary vector representation to obtain target charging scheduling strategy information at one or more levels; based on the target charging parameter information determined using the target charging scheduling strategy information, a charging scheduling decision is made for the supercharging charging demand in the next time period during the supercharging management of the charging pile.

[0005] On the other hand, the present invention provides a computer system, comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0006] In the distributed automobile liquid-cooled supercharging control method provided by the present invention, the real-time charging parameter information in the current time period and the supercharging network operation information caused by the charging scheduling decision based on the real-time charging parameter information are merged to obtain the merged information that merges multiple levels of information. Among them, the real-time charging parameter information and the supercharging network operation information can globally characterize the charging state, so after merging the real-time charging parameter information and the supercharging network operation information, it can more comprehensively and accurately guide the reasoning of the charging scheduling strategy information. Then, after the internal weight focusing allocation is performed on the merged information, a multi-objective decision prediction is performed to obtain the target charging scheduling strategy information at one or more levels. In this way, different types of charging parameter information and supercharging network operation information can be prioritized based on the internal weight focusing strategy, expanding the effect of advanced features, so that the output target charging scheduling strategy information can be used for charging scheduling decisions in the next time period. When it can increase the scheduling effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of the implementation flow of a distributed automobile liquid cooling supercharging control method provided in an embodiment of the present invention.

[0008] Figure 2 A hardware entity schematic diagram of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0009] The embodiment of the present invention provides a distributed automobile liquid cooling supercharging control method, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, a desktop computer, a mobile device, etc.

[0010] Figure 1 A schematic diagram of a distributed automobile liquid cooling supercharging control method according to an embodiment of the present invention is provided. Figure 1 As shown, the method includes:

[0011] Step 100: Acquire real-time charging parameter information in the current time period, and supercharging network operation information caused by charging scheduling decisions based on the real-time charging parameter information; the real-time charging parameter information includes charging power, battery status and charging rate; the real-time charging parameter information is determined based on past charging scheduling strategy information in the previous time period.

[0012] In step 100, the real-time charging parameter information reflects the current charging state of the vehicle battery, while the supercharging network operation information reflects the operation status of the entire charging network after scheduling decisions are made based on the current charging parameters. Through the comprehensive analysis of these two aspects of information, the computer system can have a deeper understanding of the various factors in the charging process, thereby providing strong support for subsequent decision-making.

[0013] In specific implementation, the computer system can obtain real-time charging parameter information with the help of a variety of sensors and monitoring equipment. For charging power, the computer system can measure the power value output by the charging pile in real time through a power sensor. For example, when an electric car is performing liquid cooling supercharging, the power sensor can accurately record the changes in power during the charging process. The computer system can read the data of the power sensor every certain period of time (such as 1 second) to obtain the current charging power. The battery status can be obtained by relying on the battery management system (BMS). The BMS can monitor the battery voltage, current, temperature and other parameters in real time. The computer system can communicate with the BMS to obtain these parameters and comprehensively evaluate the battery status, such as the remaining battery power and health status.

[0014] The real-time charging parameter information is determined based on the past charging scheduling strategy information of the previous time period. This means that the scheduling strategy of the previous time period will have an impact on the current charging status. For example, if the charging scheduling strategy of the previous time period is to use high-power fast charging, then in the current time period, the battery may be at a higher temperature, the charging power may be subject to certain restrictions, and the battery status and charging rate will also change accordingly.

[0015] The operation information of the supercharging network is caused by the charging scheduling decision based on the real-time charging parameter information. After obtaining the real-time charging parameter information, the computer system makes charging scheduling decisions according to the preset scheduling algorithm. For example, when multiple electric vehicles are charging in a supercharging network at the same time, the computer system will reasonably allocate the power of the charging piles according to the real-time charging parameter information of each vehicle to ensure the stable operation of the entire charging network. In this process, the operation status of the supercharging network will change, such as the load of the power grid, the use of the charging piles, etc. The computer system can obtain the operation information of the supercharging network in real time through network monitoring equipment, such as smart meters, network flow sensors, etc. Smart meters can record the total power consumption of the power grid and the power consumption of each charging pile, and the computer system can calculate the load rate of the power grid based on these data. The network flow sensor can monitor the data transmission between the charging pile and the computer system to ensure the stability of communication and the timeliness and accuracy of data.

[0016] In practical applications, the computer system can use a data acquisition module to obtain real-time charging parameter information and supercharging network operation information. The data acquisition module can be connected to various sensors and monitoring devices to read data through interface protocols (such as Modbus, CAN bus, etc.). The collected data can be stored in a database for subsequent analysis and processing. At the same time, the computer system can also use cloud computing and big data technology to analyze and mine a large amount of historical data and real-time data to further optimize the charging scheduling strategy.

[0017] Step 200: Merge the supercharging network operation information and the real-time charging parameter information in the current time period to obtain merged information.

[0018] In step 200, in order to merge the supercharging network operation information and the real-time charging parameter information, the computer system needs to first extract features from these two types of information to convert the original information into a feature vector representation that is easier to process and analyze. For the supercharging network operation information, it may include data on multiple aspects such as the load rate of the power grid, the usage of charging piles, and the stability of power transmission. The computer system can use a dimensionality reduction algorithm such as principal component analysis (PCA) to process these data and extract the main features that can represent the operation status of the supercharging network. For example, in a supercharging network containing multiple charging piles, the computer system can collect data such as the real-time power and the number of connected vehicles of each charging pile, and convert these high-dimensional data into several low-dimensional principal components through the PCA algorithm. These principal components are the feature vector representations of the supercharging network operation information. For the real-time charging parameter information, since it includes parameters such as charging power, battery status, and charging rate, the computer system can normalize these parameters so that they have the same scale range, and then combine the processed data into a vector as a feature vector representation of the real-time charging parameter information. For example, the charging power may range from 0 to 500 kW, the battery status can be represented by a value between 0 and 1 as the percentage of remaining power, and the charging rate can be represented by the percentage of power charged per minute. The computer system normalizes these parameters and combines them into a vector, such as [0.8, 0.6, 0.2], which represent the normalized charging power, battery status, and charging rate, respectively.

[0019] After the feature extraction is completed, the computer system merges the feature vector representations of the supercharging network operation information and the real-time charging parameter information. The merging method can be splicing, that is, connecting the two feature vectors in sequence into a longer vector.

[0020] Step 200 provides a more comprehensive and accurate information basis for the subsequent steps of the distributed vehicle liquid-cooled supercharging control method by extracting and merging the features of the supercharging network operation information and the real-time charging parameter information.

[0021] Step 300: Perform internal weight focusing allocation on the merged information to obtain a temporary vector representation, and perform multi-objective decision prediction based on the temporary vector representation to obtain target charging scheduling strategy information at one or more levels.

[0022] In step 300, when performing internal weight focusing allocation, the computer system can embed and map the merged information based on one or more internal weight focusing networks to obtain a temporary vector representation. The role of the internal weight focusing network is to perform a nonlinear transformation on the input information, map it into a new vector space, and emphasize the importance of different features through the internal weight allocation mechanism. For example, in a super charging network containing multiple charging piles, the merged information may include multiple features such as the real-time power of each charging pile, the charging demand of the vehicle, and the load condition of the power grid. The internal weight focusing network can assign different weights to each feature based on the importance of these features to the charging scheduling decision. If the current power grid load rate is high, the weight of the power grid load condition feature may be assigned a higher value to highlight its importance in the decision.

[0023] In the multi-objective decision prediction stage, the computer system determines one or more levels of target charging scheduling strategy information based on the temporary vector representation. Multi-objective decision-making refers to the process of making the best decision while considering multiple interrelated but potentially conflicting objectives. In the distributed vehicle liquid-cooled supercharging scenario, these objectives may include improving charging efficiency, ensuring charging stability, and reducing grid load. The computer system comprehensively considers these objectives, extracts information related to each objective from the temporary vector representation, and makes decisions based on preset decision rules. For example, when the temporary vector representation shows that a charging pile has a high charging efficiency but causes excessive grid load, the computer system weighs the two objectives of improving charging efficiency and reducing grid load, and ultimately determines a suitable charging scheduling strategy.

[0024] In order to achieve internal weight focus allocation and multi-objective decision prediction, the computer system can adopt a variety of technical means. In terms of internal weight focus allocation, the attention mechanism in deep learning can be used. The attention mechanism can automatically learn the importance of different parts of the input information and assign corresponding weights to them. For example, the multi-head attention mechanism in the Transformer architecture is a feasible attention mechanism, which can process the input information through multiple parallel attention heads to capture feature information at different levels. In terms of multi-objective decision prediction, the computer system can use reinforcement learning algorithms. Reinforcement learning is a method of continuously learning the optimal strategy through interaction between the agent and the environment. In the distributed automobile liquid cooling supercharging scenario, the agent can be a computer system, and the environment can be the supercharging network and the charging status of the vehicle. By constantly trying different charging scheduling strategies and adjusting the strategy according to the feedback from the environment (such as charging efficiency, grid load, etc.), the optimal charging scheduling strategy is finally learned.

[0025] Step 300 provides key charging scheduling strategy information for the distributed vehicle liquid-cooled supercharging system through internal weight focus allocation and multi-objective decision prediction.

[0026] Step 400: Based on the target charging parameter information determined by using the target charging scheduling strategy information, a charging scheduling decision is made for the supercharging demand in the next time period during the supercharging management of the charging pile.

[0027] In step 400, the computer system determines the target charging parameter information based on the target charging scheduling strategy information. The target charging scheduling strategy information may include multiple levels of content, such as giving priority to meeting the charging needs of certain vehicles, balancing the grid load, etc. The computer system calculates specific target charging parameters based on these strategies and the actual situation of the current supercharging network. For example, if the target charging scheduling strategy is to give priority to charging vehicles with high power requirements, and the current grid load allows, then the computer system can allocate higher charging power to these vehicles. Assuming that the battery power of an electric car is low and the user sets a fast charging requirement, the computer system determines to provide 300kW of charging power as the target charging parameter for the vehicle based on the target charging scheduling strategy.

[0028] Next, the computer system will use these target charging parameter information to make charging scheduling decisions for the supercharging demand in the next time period. This involves the rational allocation and management of charging pile resources. The computer system considers multiple factors, such as the number of charging piles, the available power of each charging pile, the charging needs of the vehicle, etc. For example, in a supercharging station with 10 charging piles, there are 5 electric vehicles that need to be charged. The computer system assigns a suitable charging pile to each vehicle based on the target charging parameter information and adjusts the output power of the charging pile. If the target charging power of one of the vehicles is 300kW, and the maximum output power of a charging pile is 400kW, the computer system can assign the vehicle to the charging pile and adjust the output power of the charging pile to 300kW.

[0029] In order to achieve this process, the computer system can use a variety of technical means. In determining the target charging parameter information, optimization algorithms such as linear programming and genetic algorithms can be used. These algorithms can solve the optimal target charging parameters based on the target charging scheduling strategy information and the constraints of the super charging network. For example, the linear programming algorithm can maximize the charging efficiency or minimize the charging cost while meeting the constraints such as the power limit of the power grid and the power limit of the charging pile. When making charging scheduling decisions, the computer system can use real-time monitoring and feedback mechanisms. By monitoring the operating status of the charging pile and the charging status of the vehicle in real time, the computer system can adjust the charging scheduling strategy in time. For example, if a charging pile fails, the computer system can immediately transfer the vehicle assigned to the charging pile to other available charging piles.

[0030] Step 400 realizes the practical application of the distributed automobile liquid-cooled supercharging system by converting the target charging scheduling strategy information into target charging parameter information and making charging scheduling decisions for the supercharging demand in the next time period.

[0031] As an implementation mode, step 200, combining the supercharging network operation information and the real-time charging parameter information in the current time period to obtain combined information, includes:

[0032] Step 210: extracting features from the real-time charging parameter information and the supercharging network operation information to obtain corresponding feature vector representations;

[0033] Step 220: Merge the feature vector representations to obtain merge information.

[0034] In step 210, feature extraction is the process of converting the original real-time charging parameter information and supercharging network operation information into a more representative and computable feature vector representation. For the real-time charging parameter information, which includes charging power, battery status, and charging rate, the computer system can use methods such as principal component analysis (PCA) to extract features.

[0035] For the supercharging network operation information, since it reflects the operation status of the entire supercharging network under the current charging scheduling decision, it may contain data on multiple aspects such as the load rate of the power grid, the usage of charging piles, and the stability of power transmission. The computer system can use independent component analysis (ICA) for feature extraction. Independent component analysis is a technology used for blind source separation. It assumes that the observed data is a linear mixture of multiple independent source signals, and the feature extraction of data is achieved by finding these independent source signals. For example, in a supercharging network, the load rate of the power grid may be affected by multiple factors, such as electricity demand in different time periods, power distribution of charging piles, etc. The computer system can use the ICA algorithm to separate these independent factors from the supercharging network operation information and represent them as feature vectors of the supercharging network operation information.

[0036] After obtaining the feature vector representations of the real-time charging parameter information and the supercharging network operation information, in step 220, the computer system merges these feature vector representations to obtain merged information. The purpose of the merger is to integrate two different aspects of information together for subsequent unified analysis and processing. The computer system can merge by splicing, that is, connecting the feature vectors of the real-time charging parameter information and the feature vectors of the supercharging network operation information in sequence into a longer vector.

[0037] As an implementation mode, step 300, performing internal weight focusing allocation on the merged information to obtain a temporary vector representation, includes:

[0038] Step 310: Based on one or more internal weight focusing networks, the merged information is embedded and mapped to obtain a temporary vector representation; if there are multiple internal weight focusing networks, the multiple internal weight focusing networks are cascaded, and the execution data of the first internal weight focusing network is the merged information, and the execution data of the internal weight focusing networks other than the first one is the execution result after the previous internal weight focusing network connected to it is embedded and mapped, and the execution result of the last internal weight focusing network is the temporary vector representation.

[0039] Embedding mapping refers to mapping the original data space into a new low-dimensional or high-dimensional vector space. In this new space, the characteristics of the data can be more clearly displayed, which is convenient for subsequent analysis and processing. The embedding mapping process can be implemented using an encoder.

[0040] If there are multiple internal weight focusing networks in cascade, this structure can be regarded as a deep feature extractor. Each internal weight focusing network will further process and refine the output of the previous layer, just like a multi-layer filter, gradually filtering out unimportant information and retaining key features. For example, the first internal weight focusing network may extract some basic features from the merged information, such as the approximate range of charging power, the fluctuation trend of battery temperature, etc.; the second internal weight focusing network will further explore deeper information based on these basic features, such as the potential relationship between charging power and battery temperature. Through this cascading method, the computer system can obtain a more comprehensive and accurate temporary vector representation.

[0041] As an implementation mode, the internal weight focusing network includes a multi-channel joint focusing component and a feedforward neural network component; for any internal weight focusing network, the embedding mapping process of any internal weight focusing network includes:

[0042] Step 311: Based on the multi-channel joint focusing component in the internal weighted focusing network, a multi-channel joint focusing operation is performed on the corresponding execution data to obtain a multi-channel joint focusing result;

[0043] Step 312: Based on the feedforward neural network component in the internal weighted focusing network, feedforward processing is performed on the multi-channel joint focusing result to obtain the execution result of the corresponding internal weighted focusing network.

[0044] In step 311, the multi-channel joint focusing component is the core part of implementing the multi-channel joint focusing operation. It can be a multi-head attention network that can focus and analyze the execution data from multiple different channels or angles to highlight the important feature information in the data. The execution data can be the merged information or the execution result of the previous internal weight focusing network. The purpose of the multi-channel joint focusing operation is to convert the execution data into a more representative and discriminative multi-channel joint focusing result through a series of calculations and transformations.

[0045] Taking the distributed automobile liquid cooling supercharging scenario as an example, assume that the execution data is a vector containing supercharging network operation information and real-time charging parameter information, such as [0.2, 0.3, 0.4, 0.5, 0.6, 0.7]. The multi-channel joint focusing component will first perform an affine transformation on this execution data, which is a full connection operation. The mathematical formula of the affine transformation is y=Wx+b, where x is the input vector, W is the weight matrix, b is the bias vector, and y is the output vector. Assume that the weight matrix W is a 6×6 matrix , the bias vector b is [0.1, 0.2, 0.3, 0.4, 0.5, 0.6], then after the affine transformation, each element y of the output vector yi The calculation method is The calculated output vector is the affine transformation vector representation.

[0046] Next, the multi-channel joint focusing component decomposes the affine transformation vector representation to obtain multiple affine transformation sub-vectors. This step is similar to splitting a whole piece of information into multiple parts for independent analysis. For example, the affine transformation vector representation [0.8, 0.9, 1.0, 1.1, 1.2, 1.3] is decomposed into three affine transformation sub-vectors [0.8, 0.9], [1.0, 1.1], and [1.2, 1.3]. Then, a normalized dot product focusing operation is performed on each affine transformation sub-vector to obtain multiple multi-channel focused sub-vectors. The normalized dot product focusing operation enables the computer system to pay attention to the importance of different elements in each affine transformation sub-vector.

[0047] Finally, multiple multi-channel focusing sub-vectors are combined and then an affine transformation is performed to obtain the multi-channel joint focusing result. Through the multi-channel joint focusing operation, the computer system can analyze and process the execution data from multiple dimensions, capture the key features in the data, and provide more valuable information for subsequent feedforward processing.

[0048] In order to realize the multi-channel joint focusing operation, the computer system can use the relevant functions and modules in the deep learning framework. For example, in frameworks such as PyTorch or TensorFlow, functions for implementing operations such as affine transformation, vector decomposition, and normalization are provided, and the computer system can use these functions to build a multi-channel joint focusing component.

[0049] After obtaining the multi-channel joint focusing result, in step 312, the computer system performs feedforward processing on it based on the feedforward neural network component in the internal weight focusing network. The feedforward neural network is a feasible neural network structure, which consists of an input layer, a hidden layer and an output layer. Information is passed forward from the input layer to the hidden layer and the output layer in sequence without feedback connection. The process of feedforward processing is to input the multi-channel joint focusing result into the feedforward neural network, and finally obtain the execution result of the corresponding internal weight focusing network after nonlinear transformation and weight calculation of the hidden layer.

[0050] Assume that the result of multi-channel joint focusing is a vector [1.5, 1.6, 1.7], the hidden layer of the feedforward neural network has two neurons, and the output layer has one neuron. Each neuron in the hidden layer performs a weighted summation on the input vector and performs a nonlinear transformation through an activation function (such as the ReLU function f(x)=max(0, x)). Assume that the weight vector of the first neuron in the hidden layer is [0.2, 0.3, 0.4] and the bias is 0.1, then the output of the neuron is f(0.2×1.5+0.3×1.6+0.4×1.7+0.1)=f(1.66)=1.66. Similarly, the output of the second neuron in the hidden layer is calculated. Then, the neurons in the output layer perform a weighted summation on the output of the hidden layer to obtain the final execution result.

[0051] The feedforward neural network component can further extract features and integrate information from the multi-channel joint focusing results through its internal weight matrix and activation function, and convert them into a form that is more suitable for subsequent multi-objective decision-making predictions. In the distributed automobile liquid cooling supercharging control method, the execution results after feedforward processing can reflect the comprehensive characteristics of the current charging state and supercharging network operation information, providing an important basis for determining the target charging scheduling strategy information.

[0052] In summary, step 311 and step 312 realize the embedding mapping process of any internal weight focusing network through multi-channel joint focusing operation and feedforward processing. The computer system uses the deep learning framework and related algorithms to continuously optimize the component parameters to improve the processing capability and feature extraction effect of input information, and provide more accurate and effective decision support for the distributed automobile liquid cooling supercharging control method.

[0053] As an implementation mode, step 311, based on the multi-channel joint focusing component in the internal weighted focusing network, performs a multi-channel joint focusing operation on the corresponding execution data to obtain a multi-channel joint focusing result, including:

[0054] Step 3111: Based on the multi-channel joint focusing component in the internal weighted focusing network, perform affine transformation on the corresponding execution data to obtain an affine transformation vector representation;

[0055] Step 3112: Decompose the affine transformation vector representation to obtain multiple affine transformation sub-vectors;

[0056] Step 3113: performing normalized dot product focusing operations on each affine transformation sub-vector respectively to obtain a plurality of multi-channel focused sub-vectors;

[0057] Step 3114: combine multiple multi-channel focusing sub-vectors to obtain a multi-channel focusing vector representation;

[0058] Step 3115: Perform affine transformation on the multi-channel focusing vector representation to obtain a multi-channel joint focusing result.

[0059] In step 3111, the computer system performs an affine transformation on the corresponding execution data based on the multi-channel joint focusing component in the internal weight focusing network to obtain an affine transformation vector representation. Affine transformation is a linear transformation plus a translation operation, and its mathematical formula is y=Wx+b, where x is the input vector, that is, the execution data; W is the weight matrix; b is the bias vector; y is the output vector, that is, the affine transformation vector representation. In the distributed automobile liquid-cooled supercharging scenario, the execution data may contain information in multiple dimensions such as the load condition of the supercharging network, the charging requirements of the vehicle, and the battery status. For example, the execution data x is a vector containing 5 elements [0.2, 0.3, 0.4, 0.5, 0.6], and the weight matrix W is a 5×5 matrix , the bias vector b is [0.1, 0.2, 0.3, 0.4, 0.5]. Then, the affine transformation vector represents the first element y 1 The calculation process of y is: 1 =0.1×0.2+0.2×0.3+0.3×0.4+0.4×0.5+0.5×0.6+0.1=0.1×0.2+(0.2×0.3+0.3×0.4+0.4×0.5+0.5×0.6)+ 0.1=0.1×0.2+0.62+ 0.1=0.722, and the other elements of y can be calculated in the same way. Through affine transformation, the computer system can perform linear combination and offset adjustment on the execution data, map it to a new vector space, highlight certain features in the data, and prepare for subsequent processing.

[0060] In step 3112, the computer system decomposes the affine transformation vector representation to obtain multiple affine transformation sub-vectors. The purpose of decomposition is to split the affine transformation vector representation into multiple parts so that each part can be processed and analyzed independently later. For example, assuming that the affine transformation vector representation y is [0.722, 0.833, 0.944, 1.055, 1.166], the computer system can decompose it into multiple affine transformation sub-vectors according to certain rules. One feasible decomposition method is to divide it evenly according to the order of elements, such as decomposing it into two affine transformation sub-vectors [0.722, 0.833] and [0.944, 1.055, 1.166]. Another decomposition method can be divided according to the semantic information of the data, for example, the elements related to the super charging network load are divided into one sub-vector, and the elements related to the vehicle charging demand are divided into another sub-vector. Through decomposition, the computer system can process the data more carefully and mine the characteristics of different parts of the data.

[0061] In step 3113, the computer system performs a normalized dot product focusing operation on each affine transformation sub-vector to obtain multiple multi-channel focused sub-vectors. The normalized dot product focusing operation is an attention mechanism that allows the computer system to pay attention to the importance of different elements in each affine transformation sub-vector. Specifically, for each affine transformation sub-vector, the dot product is first calculated with itself to obtain the dot product calculation result. The formula for the dot product calculation is , where a is the affine transformation subvector and n is the length of the subvector. For example, for the affine transformation subvector a=[0.722, 0.833], the dot product calculation result is 0.722×0.722+ 0.833×0.833=0.521284+0.693889=1.215173. Then, the dot product calculation result is normalized, for example, using the softmax function for normalization. The formula of the softmax function is , where x is the result of the dot product calculation. Finally, the influence coefficient vector is multiplied with the corresponding affine transformation subvector to obtain the multi-channel focused subvector corresponding to the corresponding affine transformation subvector. Through the normalized dot product focusing operation, the computer system can highlight the influence of important elements in the affine transformation subvector and suppress the interference of unimportant elements, thereby obtaining a more representative multi-channel focused subvector. In implementation, the computer system can use the dot product function, softmax function, and nonlinear activation function in the deep learning framework to complete these operations.

[0062] In step 3114, the computer system combines the multiple multi-channel focusing sub-vectors to obtain a multi-channel focusing vector representation. The combination may be by concatenation, that is, connecting the multiple multi-channel focusing sub-vectors in sequence into a longer vector.

[0063] In step 3115, the computer system performs an affine transformation on the multi-channel focus vector representation to obtain a multi-channel joint focusing result. This step is similar to step 3111, and also uses the affine transformation formula y=Wx+b, where x is the multi-channel focus vector representation, W is the weight matrix, b is the bias vector, and y is the multi-channel joint focusing result. Through this affine transformation, the computer system further performs linear combination and offset adjustment on the multi-channel focus vector representation to convert it into a multi-channel joint focusing result that is more suitable for subsequent processing. In implementation, the computer system can also use matrix multiplication and addition operations in the deep learning framework to complete this step.

[0064] Overall, steps 3111 - 3115 implement a multi-channel joint focusing operation on the execution data through a series of operations, including affine transformation, vector decomposition, normalized dot product focusing, vector combination, and affine transformation again. This process can help the computer system extract key features from the execution data, integrate and process information of different dimensions, highlight the impact of important information, and suppress the interference of unimportant information. In the distributed automobile liquid-cooled supercharging control method, the multi-channel joint focusing results can more accurately reflect the operating status of the supercharging network and the charging needs of the vehicle, and provide a more reliable basis for subsequent multi-objective decision-making predictions. For example, when formulating a charging scheduling strategy, the multi-channel joint focusing results can help the computer system more accurately determine which vehicles need to be charged first, and how to reasonably allocate the power of the charging piles, thereby improving the charging efficiency and ensuring the stable operation of the supercharging network.

[0065] In summary, steps 3111 - 3115 play an important role in the distributed automobile liquid cooling supercharging control method. They realize multi-channel joint focusing of execution data through a series of complex operations, providing key support for subsequent decision-making.

[0066] As an implementation mode, step S3113 performs a normalized dot product focusing operation on each affine transformation sub-vector to obtain a plurality of multi-channel focused sub-vectors, including:

[0067] Step S31131: for any affine transformation sub-vector, perform dot product calculation on the corresponding affine transformation sub-vector and itself to obtain a dot product calculation result;

[0068] Step S31132: normalizing the dot product calculation result to obtain a normalized result;

[0069] Step S31133: based on the first nonlinear transformation operator, transform the normalized processing result into an influence coefficient vector corresponding to the corresponding affine transformation sub-vector;

[0070] Step S31134: Multiply the influence coefficient vector with the corresponding affine transformation sub-vector to obtain a multi-channel focusing sub-vector corresponding to the corresponding affine transformation sub-vector.

[0071] In step 31131, the computer system performs a dot product calculation on any affine transformation subvector and itself to obtain a dot product calculation result. Through the dot product calculation, the computer system can measure the "length" or "energy" of the affine transformation subvector, and can also highlight the sum of squares of elements in the subvector, in preparation for subsequent normalization processing.

[0072] In step 31132, the computer system normalizes the dot product calculation result obtained in step 31131 to obtain a normalized result. The purpose of normalization is to map the dot product calculation result to a specific range, usually the interval [0,1], to facilitate subsequent processing and comparison. A feasible normalization method is to use the softmax function. Through normalization, the computer system can unify the dot product calculation results of different affine transformation subvectors into a comparable range, highlighting the relative importance of important elements.

[0073] In step 31133, the computer system transforms the normalized processing result obtained in step 31132 into an influence coefficient vector corresponding to the corresponding affine transformation subvector based on the first nonlinear transformation operator. The first nonlinear transformation operator can enhance the expressiveness of the model and enable the model to learn more complex patterns and relationships. The feasible first nonlinear transformation operator includes the ReLU (Rectified Linear Unit) function, which is defined as f(x)=max(0, x). For example, assuming that the normalized processing result obtained in step 31132 is [0.304, 0.326, 0.371], the ReLU function is used for transformation. Since these values ​​are all greater than 0, the influence coefficient vector is still [0.304, 0.326, 0.371]. If there are negative numbers in the normalized processing result, after being processed by the ReLU function, the negative numbers will be set to 0. Through the processing of the first nonlinear transformation operator, the computer system can introduce nonlinear factors so that the influence coefficient vector can better reflect the importance of elements in the affine transformation subvector. In terms of implementation, the computer system can use conditional statements in the programming language to implement the ReLU function, or it can use the ReLU activation function provided in the deep learning framework.

[0074] In step 31134, the computer system multiplies the influence coefficient vector obtained in step 31133 with the corresponding affine transformation subvector to obtain a multi-channel focusing subvector corresponding to the corresponding affine transformation subvector. The product operation is to multiply each element in the influence coefficient vector by the element at the corresponding position in the affine transformation subvector. For example, assuming that the affine transformation subvector is [0.2, 0.3, 0.4] and the influence coefficient vector is [0.304, 0.326, 0.371], then the multi-channel focusing subvector is [0.2×0.304, 0.3×0.326, 0.4×0.371]=[0.0608, 0.0978, 0.1484]. Through the product operation, the computer system can weight the affine transformation subvector according to the influence coefficient vector, highlight the influence of important elements, and suppress the interference of unimportant elements, so as to obtain a more representative multi-channel focusing subvector.

[0075] Steps 31131-31134 implement the normalized dot product focusing operation on the affine transformation subvector through dot product calculation, normalization, nonlinear transformation and product operation. This process can help the computer system extract key information from the affine transformation subvector, quantify and adjust the importance of different elements, and provide more valuable input for subsequent multi-channel joint focusing and multi-target decision prediction.

[0076] Steps 31131 - 31134 play an important role in the distributed automobile liquid cooling supercharging control method. They realize the normalized dot product focusing of the affine transformation subvector through a series of complex operations, providing key support for subsequent decision-making.

[0077] As an implementation mode, in step 300, a multi-objective decision prediction is performed based on the temporary vector representation to obtain target charging scheduling strategy information at one or more levels, including:

[0078] Step 320: For each of the one or more levels, perform identity mapping on the temporary vector representation to obtain an identity mapping result corresponding to each level;

[0079] Step 330: performing affine transformation on each identity mapping result to obtain a corresponding affine transformation result;

[0080] Step 340: Based on the second nonlinear transformation operator, the affine transformation results are transformed into target charging scheduling strategy information corresponding to the corresponding levels.

[0081] In step 320, the computer system performs identity mapping on the temporary vector representation for each level in one or more levels to obtain the identity mapping results corresponding to each level. Identity mapping is a simple and basic mapping operation, which keeps the original information of the input vector unchanged and outputs it directly. In the distributed automobile liquid cooling supercharging scenario, the temporary vector representation is obtained after the internal weight focus allocation of the merged information in the previous step, which integrates the key features of the supercharging network operation information and the real-time charging parameter information. The "level" here can be understood as different decision-making considerations, such as charging efficiency level, power grid stability level, user satisfaction level, etc. The significance of identity mapping is to ensure that each level can be subsequently processed based on the complete temporary vector representation information, avoid information loss or distortion, and prepare for subsequent more in-depth analysis and decision-making. In actual implementation, the computer system can complete the identity mapping through a simple assignment operation, such as directly assigning the temporary vector representation to a new variable in the programming language, and this new variable is the identity mapping result.

[0082] In step 330, the computer system performs affine transformation on each identity mapping result to obtain the corresponding affine transformation result. Different weight matrices and bias vectors are used at different levels to reflect the different decision-making emphases at each level. For example, at the charging efficiency level, the weight matrix W 1 It may focus more on elements related to charging power and charging rate, while at the grid stability level, the weight matrix W 2 They may pay more attention to the load conditions of the supercharging grid, power fluctuations, and other elements. Through affine transformation, the computer system can linearly combine and offset the identity mapping results, highlight the features of concern at different levels, and transform the temporary vector representation into an intermediate result that better meets the decision-making needs at all levels.

[0083] In step 340, the computer system transforms the affine transformation results obtained in step 330 into target charging scheduling strategy information corresponding to the corresponding level based on the second nonlinear transformation operator. The role of the second nonlinear transformation operator is to introduce nonlinear factors, enhance the expressiveness of the model, and enable the model to learn more complex decision-making patterns. The feasible second nonlinear transformation operators include Sigmoid function, Tanh function, etc. Different second nonlinear transformation operators are suitable for different decision-making scenarios. For example, the Sigmoid function is feasible to output probability values, which can be used to determine the possibility of adopting a certain charging scheduling strategy; the Tanh function maps the input value to the (-1, 1) interval, which can be used to represent the intensity or direction of a certain scheduling strategy. Through the processing of the second nonlinear transformation operator, the computer system can convert the affine transformation results into specific target charging scheduling strategy information that can be used for actual scheduling decisions, such as the charging power allocation ratio of a charging pile, the charging priority of a vehicle, etc.

[0084] Steps 320 to 340 gradually transform the temporary vector representation into one or more levels of target charging scheduling strategy information through three steps: identity mapping, affine transformation, and nonlinear transformation. Identity mapping ensures that each level can obtain complete temporary vector representation information, affine transformation linearly combines and adjusts the information according to the characteristics of each level, and nonlinear transformation introduces nonlinear factors to make the decision results more flexible and accurate. In the distributed vehicle liquid cooling supercharging control method, these steps can comprehensively consider multiple goals, such as improving charging efficiency, ensuring grid stability, meeting user needs, etc., to formulate a reasonable charging scheduling strategy.

[0085] In the embodiment of the present invention, the above-mentioned distributed automobile liquid cooling supercharging control method can be implemented based on a pre-trained charging scheduling neural network. Then, the distributed automobile liquid cooling supercharging control method provided by the present invention also includes a training process of the charging scheduling neural network, specifically including:

[0086] Step 10: Obtain a pre-charging scheduling neural network obtained by pre-debugging using a plurality of basic training example sets;

[0087] Step 20: For each non-real-time time period in the non-real-time debugging, a non-real-time training example for the next non-real-time time period is constructed based on the non-real-time training example for the current non-real-time time period and the non-real-time charging scheduling inference result output by the front-end charging scheduling neural network; the non-real-time training example includes non-real-time charging parameter information and non-real-time super charging network operation information;

[0088] Step 30: determining a non-real-time overall performance indicator for each non-real-time time period based on the non-real-time training examples for each non-real-time time period;

[0089] Step 40: progressively debug the front-end charging scheduling neural network in non-real-time time periods according to the non-real-time training examples and non-real-time overall performance indicators corresponding to the multiple non-real-time time periods, terminate the debugging when the non-real-time debugging convergence requirements are met, and obtain the transition charging scheduling neural network;

[0090] Step 50: for each real-time time period in the real-time debugging, based on the real-time training examples of the current real-time time period, determine the real-time overall performance index of the current real-time time period;

[0091] Step 60: Based on the real-time training examples and real-time overall performance indicators corresponding to the multiple real-time time periods, the transition charging scheduling neural network is progressively debugged in real-time time periods, and the debugging is terminated when the real-time debugging convergence requirements are met to obtain a trained charging scheduling neural network.

[0092] In step 10, the computer system obtains a pre-charging scheduling neural network obtained by pre-debugging using multiple basic training example sets. The basic training example set is a sample set for pre-training, which contains past supercharging network operation information, past charging parameter information, past scheduling comparison information and past overall performance indicators corresponding to the same past time period. Pre-debugging can be understood as pre-training, and its purpose is to allow the neural network to initially learn the basic patterns and laws of charging scheduling. For example, the computer system collects supercharging network operation data (such as grid load rate, charging pile usage, etc.) and vehicle charging parameters (such as charging power, battery status, etc.) at different time periods every day in the past week, and organizes these data according to time periods to construct multiple basic training example sets. Then, the computer system uses these basic training example sets to pre-debug the original charging scheduling neural network to be debugged. During the pre-debugging process, the neural network will continuously adjust its own parameters to minimize the difference between the prediction results and the past scheduling comparison information. Finally, when the pre-debugging convergence requirements are met (for example, the value of the loss function no longer decreases significantly), the computer system obtains the pre-charging scheduling neural network.

[0093] In step 20, the computer system constructs a non-real-time training example for the next non-real-time time period for each non-real-time time period in the non-real-time debugging based on the non-real-time training example of the current non-real-time time period and the non-real-time charging scheduling reasoning result output by the front charging scheduling neural network. The non-real-time training example includes non-real-time charging parameter information and non-real-time super charging network operation information. Non-real-time debugging is the process of debugging the network using historical data or simulated data in a non-real-time environment. For example, the computer system selects historical data from the past month as the data source for non-real-time debugging, and divides this period into multiple non-real-time time periods. For each non-real-time time period, the computer system obtains a non-real-time training example for the time period, including non-real-time charging parameter information (such as simulated charging power, battery power, etc.) and non-real-time super charging network operation information (such as simulated power grid load conditions). Then, these non-real-time training examples are input into the front charging scheduling neural network to obtain non-real-time charging scheduling reasoning results. Based on this information, the computer system can construct a non-real-time training example for the next non-real-time time period. For example, if the non-real-time charging scheduling inference result of the current time period shows that the usage frequency of a certain charging pile will increase, then in the non-real-time super charging network operation information of the next time period, the load rate of the charging pile can be increased accordingly. Constructing the non-real-time training examples for the next non-real-time time period can be completed based on the generation component in the generative adversarial network. The generative adversarial network can learn the distribution law of the data, thereby generating training examples that are more in line with the actual situation.

[0094] In step 30, the computer system determines the non-real-time overall performance index of each non-real-time time period based on the non-real-time training examples of each non-real-time time period. The non-real-time overall performance index is used to evaluate the effect of charging scheduling in each non-real-time time period. Specifically, the computer system first determines the charging efficiency based on the charging current of the current non-real-time time period. The charging efficiency can be obtained by calculating the ratio of the actual charging power to the theoretical charging power. For example, in a non-real-time time period, the battery should theoretically be charged from 30% to 80%, but it is actually only charged to 70%, then the charging efficiency is (70% - 30%) / (80% - 30%) = 80%. Then, the charging stability index is determined based on the charging fluctuation information of the time period (such as power instability, charging efficiency loss, number of charging interruptions, battery temperature changes, etc.). The charging stability index can be calculated by comprehensively considering these fluctuation factors, for example, a weight can be assigned to each fluctuation factor, and then the weighted sum is calculated. Next, the floating index is determined based on the charging current of the current non-real-time time period and the charging current of the next non-real-time time period. The floating index can reflect the change of the charging current. For example, the absolute value of the difference between the charging currents of two time periods can be calculated. Finally, based on the charging efficiency, charging stability index, and floating index, the non-real-time performance index of the current non-real-time time period is calculated. The non-real-time performance index can be the weighted sum of these three indicators, for example, the non-real-time performance index = 0.5 × charging efficiency + 0.3 × charging stability index + 0.2 × floating index. Then, based on the non-real-time overall performance index of the next non-real-time time period and the non-real-time performance index of the current non-real-time time period, determine the non-real-time overall performance index of the current non-real-time time period. For example, the exponential weighted average method can be used, and the non-real-time overall performance index of the current non-real-time time period = 0.8 × the non-real-time performance index of the current non-real-time time period + 0.2 × the non-real-time overall performance index of the next non-real-time time period.

[0095] In step 40, the computer system performs progressive debugging of the front-end charging scheduling neural network in non-real-time time periods according to the non-real-time training examples and non-real-time overall performance indicators corresponding to the multiple non-real-time time periods, and terminates the debugging when the non-real-time debugging convergence requirements are met to obtain the transition charging scheduling neural network. Progressive debugging in non-real-time time periods is a rolling training process, and the computer system continuously optimizes the parameters of the neural network in each non-real-time time period. Specifically, the computer system first processes the non-real-time training examples of the current non-real-time time period based on the decision module (i.e., Actor) of the current non-real-time time period to obtain the non-real-time charging scheduling reasoning result of the current non-real-time time period. Then, the second log-likelihood cost value is determined based on the non-real-time charging scheduling reasoning result of the current non-real-time time period, and the second log-likelihood cost value is used to measure the degree of matching between the reasoning result and the actual situation. The first non-real-time evaluation function is determined based on the second log-likelihood cost value, and the first non-real-time evaluation function is used to evaluate the performance of the decision module. At the same time, the computer system processes the non-real-time training examples of the current non-real-time time period based on the feedback module (i.e., Critic) of the current non-real-time time period to obtain the non-real-time overall performance indicator reasoning result of the current time period. According to the inference result of the non-real-time overall performance index of the current non-real-time time period and the non-real-time overall performance index of the current non-real-time time period, the second performance index cost value is determined, and the second performance index cost value is used to measure the prediction error of the feedback module. A second non-real-time evaluation function is generated based on the second performance index cost value, and the second non-real-time evaluation function is used to evaluate the performance of the feedback module. Finally, the decision module is debugged based on the first non-real-time evaluation function, and the feedback module is debugged based on the second non-real-time evaluation function. During the debugging process, the computer system uses an optimization algorithm (such as a gradient descent algorithm) to update the parameters of the neural network so that the value of the evaluation function continues to decrease. When the convergence requirements of the non-real-time debugging are met (for example, the value of the evaluation function fluctuates within a certain range and no longer changes), the debugging is ended and the transition charging scheduling neural network is obtained.

[0096] In step 50, the computer system determines the real-time overall performance index of the real-time time period for each real-time time period in the real-time debugging based on the real-time training examples of the real-time time period. Real-time debugging is a debugging process performed in the actual operating environment, and the real-time training examples include real-time charging parameter information and real-time supercharging network operation information. The computer system obtains these real-time training examples in each real-time time period, and then determines the real-time overall performance index in a similar manner to step 30. For example, the computer system monitors the charging current, battery status, grid load and other information of the charging pile in real time, calculates the charging efficiency, charging stability index and floating index, and then obtains the real-time performance index. Combined with the real-time overall performance indicators of the subsequent time period, the real-time overall performance indicators of the current real-time time period are determined by methods such as exponential weighted average.

[0097] In step 60, the computer system performs real-time time period progressive debugging on the transition charging scheduling neural network according to the real-time training examples and real-time overall performance indicators corresponding to multiple real-time time periods, and ends the debugging when the real-time debugging convergence requirements are met to obtain a trained charging scheduling neural network. The real-time time period progressive debugging is similar to the non-real-time time period progressive debugging. The computer system optimizes the transition charging scheduling neural network in each real-time time period. The real-time training examples are processed based on the decision module to obtain the real-time charging scheduling reasoning result, and the second log-likelihood cost value and the first real-time evaluation function are calculated; the real-time training examples are processed based on the feedback module to obtain the real-time overall performance indicator reasoning result, and the second performance indicator cost value and the second real-time evaluation function are calculated. Then, the optimization algorithm is used to update the parameters of the neural network according to the first real-time evaluation function and the second real-time evaluation function. During the real-time debugging process, since the data is updated in real time, the computer system responds and adjusts the parameters of the neural network in a timely manner. When the real-time debugging convergence requirements are met (for example, the value of the real-time evaluation function is stable in a small range), the debugging is ended. At this time, the neural network obtained is a trained charging scheduling neural network, which can accurately formulate a reasonable charging scheduling strategy based on real-time charging parameter information and super charging network operation information.

[0098] During the training process of the entire charging scheduling neural network, the computer system continuously adjusts the parameters of the neural network to improve its prediction accuracy of the charging scheduling strategy and the reliability of performance evaluation. At the same time, the computer system also needs to pay attention to the convergence and stability during the training process to avoid overfitting or underfitting. In order to improve the training efficiency, the computer system can adopt technical means such as batch training and parallel computing. In addition, the computer system can also pre-process the training data, such as normalization, noise reduction and other operations, to improve the quality of the data, thereby improving the training effect of the neural network. Through the training of the above steps, the charging scheduling neural network can better adapt to the needs of the distributed vehicle liquid cooling supercharging scenario and provide strong support for the actual charging scheduling decision. For example, in a large electric vehicle charging station, the trained charging scheduling neural network can reasonably allocate the power of the charging pile according to the real-time vehicle charging demand and power grid conditions, improve the charging efficiency, and ensure the stable operation of the power grid.

[0099] In summary, the training process of the charging scheduling neural network gradually optimizes the parameters of the neural network through three stages: pre-debugging, non-real-time debugging, and real-time debugging, so that it can accurately make charging scheduling decisions. Each step has its specific role and significance, and together constitutes a complete training system, which provides a guarantee for the effective implementation of the distributed vehicle liquid cooling supercharging control method.

[0100] As an implementation method, the pre-charging scheduling neural network is obtained based on pre-debugging, and the process of pre-debugging is as follows:

[0101] Step 101: constructing past training examples from past supercharging network operation information and past charging parameter information corresponding to the same past time period in the collected monitoring information;

[0102] Step 102: for the real-time past time period, determine the past scheduling comparison information corresponding to the real-time past time period according to the past charging parameter information of the next past time period of the real-time past time period, and determine the past overall performance index corresponding to the real-time past time period based on the past training examples of the real-time past time period;

[0103] Step 103: taking the past training examples, past scheduling comparison information and past overall performance indicators corresponding to the same past time period as a basic training example set;

[0104] Step 104: pre-debugging the original charging scheduling neural network to be debugged according to a plurality of basic training example sets, and ending the debugging when the pre-debugging convergence requirements are met to obtain the pre-charging scheduling neural network.

[0105] In step 101, the computer system constructs a past training example by collecting the past supercharging network operation information and past charging parameter information corresponding to the same past time period in the collected monitoring information. The collected monitoring information refers to the data obtained by the computer system monitoring the operation of the supercharging network and the charging status of the vehicle in the past period of time. The past supercharging network operation information may include the load rate of the power grid, the usage of the charging piles, the stability of power transmission, etc., and the past charging parameter information covers the charging power, battery status and charging rate, etc. For example, the computer system records the real-time power of each charging pile in the supercharging network, the number of connected vehicles, and the charging power, battery remaining power percentage and charging rate of each vehicle during the time period from 9 to 10 a.m. on a certain day. Integrating these past supercharging network operation information and past charging parameter information corresponding to the same time period together constitutes a past training example. The computer system can store this information in a data structure (such as an array, a list, or a dictionary) for subsequent processing.

[0106] In step 102, the computer system determines the past scheduling comparison information corresponding to the real-time past time period based on the past charging parameter information of the next past time period of the real-time past time period, and determines the past overall performance index corresponding to the real-time past time period based on the past training examples of the real-time past time period. The real-time past time period refers to the past time period currently being processed. The past scheduling comparison information is the specific information of the charging scheduling strategy adopted for the past time period in the actual situation, which can be used as a reference standard for measuring the prediction accuracy of the neural network in subsequent training. For example, assuming that the real-time past time period is from 9 am to 10 am, and the next past time period is from 10 am to 11 am. The computer system can infer the reasonable charging scheduling strategy that should be adopted in the time period from 9 am to 10 am based on the past charging parameter information from 10 am to 11 am (such as the change in the charging demand of the vehicle, the battery status, etc.), such as which vehicles are allocated more charging power, whether the output of certain charging piles needs to be adjusted, etc., and this information constitutes the past scheduling comparison information corresponding to 9 am to 10 am. The past overall performance index is used to evaluate the comprehensive effect of charging scheduling in the past time period, and multiple factors such as charging efficiency, charging stability, and grid load can be comprehensively considered. The computer system can calculate the values ​​of these factors based on the information in the past training examples of the real-time past time period, and obtain the past overall performance index through a certain weighted summation method. For example, the charging efficiency can be calculated by the ratio of the actual charging power to the theoretical charging power, the charging stability can be evaluated based on the fluctuation of the charging power, and the grid load can be measured by the ratio of the total power of the supercharging network to the rated power. Assuming that the weight of the charging efficiency is 0.5, the weight of the charging stability is 0.3, and the weight of the grid load is 0.2, then the past overall performance index = 0.5×charging efficiency+0.3×charging stability+0.2×grid load.

[0107] In step 103, the computer system takes the past training examples, past scheduling control information and past overall performance indicators corresponding to the same past time period as a basic training example set. Each basic training example set contains complete information within the past time period, including input data (past training examples), expected output (past scheduling control information) and evaluation indicators (past overall performance indicators). For example, for the time period from 9 to 10 a.m., the computer system combines the previously constructed past training examples (including supercharging network operation information and charging parameter information), the determined past scheduling control information (such as charging pile power allocation strategy) and the calculated past overall performance indicators (such as comprehensive evaluation values) to form a basic training example set. By combining multiple such basic training example sets, the computer system obtains a sample set for pre-debugging.

[0108] In step 104, the computer system performs pre-debugging on the original charging scheduling neural network to be debugged according to multiple basic training example sets, and ends the debugging when the pre-debugging convergence requirements are met to obtain the pre-charging scheduling neural network. Specifically, the computer system first processes the past training examples in the basic training example set according to the decision module in the original charging scheduling neural network to be debugged to obtain the past scheduling reasoning results. The function of the decision module is to predict the charging scheduling strategy that should be adopted in the time period based on the input past training examples. For example, the past training examples from 9 am to 10 am are input into the decision module, and the decision module will output a past scheduling reasoning result, such as recommending that a specific charging power be allocated to certain vehicles. Then, the computer system determines the first log-likelihood cost value based on the past scheduling reasoning results and past scheduling control information corresponding to each basic training example set. The first log-likelihood cost value is used to measure the degree of match between the past scheduling reasoning results and the past scheduling control information, and its calculation formula can be based on the log-likelihood function. Assuming that the past scheduling control information is a probability distribution and the past scheduling reasoning result is also a probability distribution, the first log-likelihood cost value can be obtained by calculating the log-likelihood difference between the two distributions. Next, the computer system processes the past training examples in the basic training example set according to the feedback module in the original charging scheduling neural network to be debugged to obtain the past overall performance index reasoning result. The function of the feedback module is to predict the past overall performance index within the time period based on the input past training examples. For example, the past training examples from 9 am to 10 am are input into the feedback module, and the feedback module will output a past overall performance index reasoning result, such as the predicted comprehensive evaluation value. Then, based on the error between the past overall performance index reasoning result and the corresponding past overall performance indicator, the first performance indicator cost value is determined. The first performance indicator cost value is used to measure the prediction error of the feedback module. A feasible calculation method is the mean square error (MSE), that is, the first performance indicator cost value = ,in is the actual overall past performance indicator, is the inference result of the past overall performance index, and n is the number of samples. Afterwards, the computer system determines the pre-debugging cost value based on the first log-likelihood cost value and the first performance index cost value. The pre-debugging cost value can be the weighted sum of these two cost values. Finally, based on the pre-debugging cost value, the computer system uses an optimization algorithm (such as stochastic gradient descent, Adam algorithm, etc.) to pre-debug the original charging scheduling neural network, and continuously adjusts the parameters of the neural network to gradually reduce the pre-debugging cost value. When the pre-debugging convergence requirements are met (for example, the pre-debugging cost value fluctuates within a certain range and no longer decreases significantly), the debugging is ended, and the neural network obtained at this time is the pre-charging scheduling neural network.

[0109] During the entire pre-debugging process, the computer system processes a large amount of historical data and ensures the accuracy of calculations in each step. In order to improve the debugging efficiency, the computer system can use batch training to input multiple basic training example sets into the neural network for processing. At the same time, the computer system can also pre-process the training data, such as normalization and standardization, to accelerate the convergence speed of the neural network. In addition, the computer system also needs to continuously monitor the changes in the pre-debugging cost value, adjust the parameters of the optimization algorithm according to its changing trend, and ensure the stability and effectiveness of the debugging process. Through pre-debugging, the original charging scheduling neural network can initially learn the basic mode and law of charging scheduling, laying a good foundation for subsequent non-real-time debugging and real-time debugging. For example, in a large electric vehicle charging station, the pre-debugging charging scheduling neural network can preliminarily predict the charging demand and reasonable scheduling strategy in different time periods based on historical data, thereby improving the charging efficiency and the stability of the power grid.

[0110] As an implementation mode, step 104, pre-debugging the original charging scheduling neural network to be debugged is performed according to multiple basic training example sets, and debugging is terminated when the pre-debugging convergence requirements are met to obtain the pre-charging scheduling neural network, including:

[0111] Step 1041: Processing past training examples in the basic training example set according to the decision module in the original charging scheduling neural network to be debugged to obtain past scheduling inference results;

[0112] Step 1042: Determine a first log-likelihood cost value based on the past scheduling inference results and past scheduling comparison information corresponding to each basic training example set;

[0113] Step 1043: Processing past training examples in the basic training example set according to the feedback module in the original charging scheduling neural network to be debugged, to obtain past overall performance index inference results;

[0114] Step 1044: determining a first performance indicator cost value based on the past overall performance indicator reasoning result and the error between the corresponding past overall performance indicators;

[0115] Step 1045: Determine a pre-debugging cost value according to the first log-likelihood cost value and the first performance indicator cost value;

[0116] Step 1046: Based on the pre-debugging cost value, pre-debugging is performed on the original charging scheduling neural network, and the debugging is terminated when the pre-debugging convergence requirement is met to obtain the pre-charging scheduling neural network.

[0117] In step 1041, the computer system processes the past training examples in the basic training example set according to the decision module in the original charging scheduling neural network to be debugged to obtain the past scheduling reasoning result. The main function of the decision module is to predict the charging scheduling strategy that should be adopted in the past time period based on the input past training examples. The past training examples contain the past super charging network operation information and past charging parameter information corresponding to the same past time period.

[0118] In step 1042, the computer system determines a first log-likelihood cost value based on the past scheduling reasoning results and past scheduling control information corresponding to each basic training example set. The past scheduling control information is the specific information of the charging scheduling strategy adopted for the past time period in actual conditions. It can be used as a reference standard for measuring the prediction accuracy of the decision-making module. The first log-likelihood cost value is used to measure the degree of match between the past scheduling reasoning results and the past scheduling control information. Assuming that the past scheduling control information can be expressed as a probability distribution P, and the past scheduling reasoning results can also be expressed as a probability distribution Q, then the first log-likelihood cost value can be obtained by calculating the log-likelihood difference between the two distributions. Specifically, for discrete probability distributions, the calculation formula for the first log-likelihood cost value L is , where i represents different scheduling strategies.

[0119] In step 1043, the computer system processes the past training examples in the basic training example set according to the feedback module in the original charging scheduling neural network to be debugged, and obtains the inference result of the past overall performance index. The function of the feedback module is to predict the past overall performance index in the past time period based on the input past training examples. The past overall performance index is used to evaluate the comprehensive effect of charging scheduling in the past time period, which can comprehensively consider multiple factors such as charging efficiency, charging stability, and grid load.

[0120] In step 1044, the computer system determines a first performance indicator cost value based on the error between the past overall performance indicator reasoning result and the corresponding past overall performance indicator. The first performance indicator cost value is used to measure the prediction error of the feedback module, and a feasible calculation method is mean square error (MSE).

[0121] In step 1045, the computer system determines the pre-debugging cost value based on the first log-likelihood cost value and the first performance index cost value. The pre-debugging cost value is a comprehensive indicator used to measure the performance of the entire original charging scheduling neural network under the current parameter settings. It can be a weighted sum of the first log-likelihood cost value and the first performance index cost value.

[0122] In step 1046, the computer system performs pre-debugging on the original charging scheduling neural network based on the pre-debugging cost value, and ends the debugging when the pre-debugging convergence requirements are met to obtain the pre-charging scheduling neural network. The goal of pre-debugging is to gradually reduce the pre-debugging cost value by adjusting the parameters of the original charging scheduling neural network, thereby improving the performance of the network. The computer system can use optimization algorithms (such as stochastic gradient descent, Adam algorithm, etc.) to achieve this goal. In each iteration, the computer system calculates the gradient of the pre-debugging cost value relative to the network parameters, and then updates the parameters of the network according to the direction and size of the gradient. As the iteration proceeds, the pre-debugging cost value will gradually decrease. When the pre-debugging convergence requirements are met (for example, the pre-debugging cost value fluctuates within a certain range and no longer decreases significantly), the computer system ends the debugging, and the neural network obtained at this time is the pre-charging scheduling neural network.

[0123] During the entire pre-debugging process, the computer system processes a large number of basic training example sets and ensures the calculation accuracy of each step. In order to improve the debugging efficiency, the computer system can use batch training to input multiple basic training example sets into the neural network for processing. At the same time, the computer system can also pre-process the training data, such as normalization and standardization, to accelerate the convergence speed of the neural network. In addition, the computer system also needs to continuously monitor the changes in the pre-debugging cost value, adjust the parameters of the optimization algorithm (such as learning rate) according to its change trend, and ensure the stability and effectiveness of the debugging process. Through pre-debugging, the original charging scheduling neural network can initially learn the basic mode and law of charging scheduling, laying a good foundation for subsequent non-real-time debugging and real-time debugging. For example, in a large electric vehicle charging station, the pre-debugging charging scheduling neural network can preliminarily predict the charging demand and reasonable scheduling strategy in different time periods based on historical data, thereby improving the charging efficiency and the stability of the power grid.

[0124] As an implementation method, the non-real-time training examples include charging current and charging fluctuation information. Based on this, step 30, based on the non-real-time training examples of each non-real-time time period, determines the non-real-time overall performance index of each non-real-time time period, including:

[0125] Step 30A: for the current non-real-time time period, determining the charging efficiency of the current non-real-time time period based on the charging current of the current non-real-time time period;

[0126] Step 30B: determining a charging stability index for the current non-real-time time period based on the charging fluctuation information for the current non-real-time time period;

[0127] Step 30C: determining a floating index of the current non-real-time time period based on the charging current of the current non-real-time time period and the charging current of the next non-real-time time period of the current non-real-time time period;

[0128] Step 30D: Calculate the non-real-time performance index of the current non-real-time time period based on the charging efficiency, charging stability index, and floating index of the current non-real-time time period;

[0129] Step 30E: Determine the non-real-time overall performance indicator of the current non-real-time time period according to the non-real-time overall performance indicator of the next non-real-time time period and the non-real-time performance indicator of the current non-real-time time period.

[0130] In step 30A, the computer system determines the charging efficiency of the current non-real-time time period based on the charging current of the time period. Charging efficiency refers to the ratio of the actual amount of electricity charged into the battery to the theoretically chargeable amount of electricity within a certain period of time. In actual calculations, the computer system can estimate the actual amount of electricity charged by measuring the charging current and charging time, and the theoretically chargeable amount of electricity can be determined based on the rated capacity and initial capacity of the battery. For example, assuming that the current non-real-time time period is from 9 am to 10 am, during this time period, the computer system monitors that the average charging current of a vehicle is I=50A, the charging time is t=1 hour, and the battery voltage is V=400V, then the actual amount of electricity charged is Q 实际 =I× V× t=50×400×1=20000Wh=20kWh. If the rated capacity of the battery is C=100kWh, the initial charge is Q 初始 =30kWh, theoretically, the amount of electricity that can be charged in this time period is Q 理论 = min(CQ 初始 , P max × t), where P max is the maximum charging power of the charging device. Assume that P max = 30kW, then Q 理论 = min(100-30, 30×1)= 30kWh. Then the charging efficiency in this period is To achieve this step, the computer system can use sensors to monitor the charging current and battery voltage in real time, and calculate the power in combination with the time information, while calculating the theoretical chargeable power based on the rated parameters and initial power of the battery.

[0131] In step 30B, the computer system determines the charging stability index of the current non-real-time time period based on the charging fluctuation information of the current non-real-time time period. The charging fluctuation information includes multiple aspects such as power instability, charging efficiency loss, number of charging interruptions, battery temperature changes, etc. These factors will affect the stability of charging, and thus affect the battery life and charging effect. The computer system can assign a weight to each fluctuation factor, and then calculate the charging stability index by weighted summation. For example, assuming the weight of power instability is w 1 =0.4, the weight of charging efficiency loss is w 2 =0.3, the weight of the number of charging interruptions is w 3 =0.2, the weight of battery temperature change is w 4 =0.1. The power instability can be measured by calculating the standard deviation of the charging power. Assuming that in the current non-real-time time period, the standard deviation of the charging power is , it can be normalized to the interval [0, 1]. Assume that the normalized power instability is x 1 =0.3. The charging efficiency loss can be calculated by comparing the actual charging efficiency with the ideal charging efficiency, assuming that the actual charging efficiency is , the ideal charging efficiency is = 0.8, then the charging efficiency loss is The number of charging interruptions can be used as an indicator directly. Assuming that the charging is interrupted n=1 times during this time period, it is normalized to the interval [0, 1]. Assuming that the normalized number of charging interruptions is x 3 =0.2. The battery temperature change can be measured by measuring the difference between the highest and lowest temperatures of the battery during this period of time. Assuming that the temperature change is , the normalized battery temperature change is x 4 =0.1. Then the charging stability index To achieve this step, the computer system monitors the charging power, charging efficiency, number of charging interruptions and battery temperature in real time, and performs corresponding calculations and normalization processing.

[0132] In step 30C, the computer system determines the floating index of the time period based on the charging current of the current non-real-time time period and the charging current of the next non-real-time time period. The floating index reflects the change of the charging current between adjacent time periods, and it can measure the smoothness and stability of the charging process. The computer system can obtain the floating index by calculating the absolute value of the difference between the charging currents of the two time periods and normalizing it to the interval [0, 1]. For example, assuming that the average charging current of the current non-real-time time period is I 1 =50A, the average charging current in the next non-real-time time period is I 2 =60A, then the difference in charging current is Assuming that the charging current range is [0,100A], the normalized floating index is To achieve this step, the computer system records the charging current information of adjacent time periods and performs difference calculation and normalization processing.

[0133] In step 30D, the computer system calculates the non-real-time performance index of the current non-real-time time period based on the charging efficiency, charging stability index, and floating index of the current non-real-time time period. The non-real-time performance index is a comprehensive index that takes into account the charging efficiency, charging stability, and floating conditions of the charging current. The computer system can assign different weights to these three indicators and then calculate the non-real-time performance index by weighted summation. For example, assuming the weight of the charging efficiency is , the weight of the charging stability index is , the weight of the floating indicator is . Non-real-time performance indicators To achieve this step, the computer system only needs to perform weighted summation of the charging efficiency, charging stability index and floating index calculated previously according to the corresponding weights.

[0134] In step 30E, the computer system determines the non-real-time overall performance index of the current non-real-time time period based on the non-real-time overall performance index of the next non-real-time time period and the non-real-time performance index of the current non-real-time time period. This step takes into account the time series factor and more accurately evaluates the overall performance of the current time period by combining the information of the next time period. The computer system can use the exponential weighted average method to calculate the non-real-time overall performance index of the current non-real-time time period. The formula for exponential weighted average is , where G t is the non-real-time overall performance index of the current non-real-time time period t, P_t is the non-real-time performance index of the current non-real-time time period t, G t+1 is the non-real-time overall performance indicator for the next non-real-time time period t+1, is the weighting coefficient, which usually takes a value between [0, 1]. For example, suppose =0.8, the non-real-time performance index P in the current non-real-time time period t =0.423625, the non-real-time overall performance index G for the next non-real-time time period t+1 = 0.45, then the non-real-time overall performance index G in the current non-real-time time period t =0.8×0.423625+(1-0.8)×0.45=0.3389+ 0.09=0.4289. To achieve this step, the computer system records the non-real-time performance indicators and the non-real-time overall performance indicators of adjacent time periods and calculates them according to the formula of exponential weighted average.

[0135] During the entire process of determining the non-real-time overall performance indicators, the computer system monitors and records a large amount of charging data in real time, including charging current, charging power, battery temperature and other information. At the same time, the computer system also needs to process and analyze these data, including power calculation, normalization, weighted summation and other operations. The computer system adjusts the weights and weighting coefficients of each indicator according to the actual situation to ensure that the non-real-time overall performance indicators can accurately reflect the effect of charging scheduling. By accurately calculating the non-real-time overall performance indicators, the computer system can better evaluate the pros and cons of the charging scheduling strategy and provide a basis for subsequent optimization and adjustment. For example, in an electric vehicle charging station, if the non-real-time overall performance indicators of a certain time period are low, it means that there may be problems with the charging scheduling of this time period. The computer system can analyze the specific indicators to find out the problem, such as low charging efficiency may be caused by charging equipment failure or unreasonable scheduling, and unstable charging may be caused by power grid fluctuations or battery aging, and then make targeted improvements and optimizations.

[0136] In summary, steps 30A - 30E calculate the charging efficiency, charging stability index, floating index and non-real-time performance index, and combine the non-real-time overall performance index of the next time period to finally determine the non-real-time overall performance index of the current non-real-time time period. This process comprehensively considers multiple factors in the charging process and provides an important reference for the optimization of the charging scheduling strategy.

[0137] As an implementation method, the front charging scheduling neural network includes a decision module and a feedback module. Based on this, step 40, according to the non-real-time training examples and non-real-time overall performance indicators corresponding to the multiple non-real-time time periods, the front charging scheduling neural network is progressively debugged in the non-real-time time period, and the debugging is terminated when the non-real-time debugging convergence requirements are met to obtain the transition charging scheduling neural network, including:

[0138] Step 40A: Based on the decision module of the current non-real-time time period, the non-real-time training examples of the current non-real-time time period are processed to obtain the non-real-time charging scheduling inference result of the current non-real-time time period;

[0139] Step 40B: determining a second log-likelihood cost value according to the non-real-time charging scheduling inference result of the current non-real-time time period, and determining a first non-real-time evaluation function based on the second log-likelihood cost value;

[0140] Step 40C: Based on the feedback module of the current non-real-time time period, the non-real-time training examples of the current non-real-time time period are processed to obtain the non-real-time overall performance indicator inference result of the current time period;

[0141] Step 40D: determining a second performance indicator cost value according to the non-real-time overall performance indicator inference result of the current non-real-time time period and the non-real-time overall performance indicator of the current non-real-time time period, and generating a second non-real-time evaluation function according to the second performance indicator cost value;

[0142] Step 40E: Debug the decision module based on the first non-real-time evaluation function, and debug the feedback module based on the second non-real-time evaluation function, and end the debugging when the non-real-time debugging convergence requirements are met to obtain a transition charging scheduling neural network.

[0143] In step 40A, the computer system processes the non-real-time training examples of the current non-real-time time period based on the decision module of the current non-real-time time period to obtain the non-real-time charging scheduling inference result of the current non-real-time time period. The decision module is used to predict the charging scheduling strategy that should be adopted in the time period based on the input non-real-time training examples. The non-real-time training examples include the non-real-time charging parameter information and the non-real-time super charging network operation information of the current non-real-time time period.

[0144] In step 40B, the computer system determines a second log-likelihood cost value based on the non-real-time charging scheduling reasoning result of the current non-real-time time period, and determines the first non-real-time evaluation function based on the second log-likelihood cost value. The second log-likelihood cost value is used to measure the degree of match between the non-real-time charging scheduling reasoning result and the actual situation. Assuming that the actual charging scheduling strategy can be expressed as a probability distribution P, and the non-real-time charging scheduling reasoning result can also be expressed as a probability distribution Q, then the second log-likelihood cost value L 2 The calculation formula is , where i represents different scheduling strategy situations. For example, suppose that in the current non-real-time time period, the actual charging scheduling strategy indicates that the probability of allocating 80% of the charging power to vehicle A and 20% of the charging power to vehicle B is 0.8, while the non-real-time charging scheduling inference result predicts that the probability of allocating 80% of the charging power to vehicle A and 20% of the charging power to vehicle B is 0.6. Then, according to the above formula, the log-likelihood value in this case is -0.8log 0.6. The computer system sums all possible scheduling strategy situations to obtain the final second log-likelihood cost value. The first non-real-time evaluation function is constructed based on the second log-likelihood cost value, which is used to evaluate the performance of the decision module. The feasible first non-real-time evaluation function can directly use the second log-likelihood cost value, or it can make some adjustments on this basis. For example, the first non-real-time evaluation function ,in is a regularization coefficient, and R is a regularization term used to prevent the model from overfitting. In order to calculate the second log-likelihood cost value and the first non-real-time evaluation function, the computer system can use a mathematical library in a programming language to implement logarithmic operations and summation operations.

[0145] In step 40C, the computer system processes the non-real-time training examples of the current non-real-time time period based on the feedback module of the current non-real-time time period to obtain the non-real-time overall performance index reasoning result of the current time period. The feedback module predicts the non-real-time overall performance index within the time period based on the input non-real-time training examples. The non-real-time overall performance index is used to evaluate the comprehensive effect of charging scheduling within the time period, which comprehensively considers multiple factors such as charging efficiency, charging stability, and grid load. For example, the computer system inputs the non-real-time training examples from 9 am to 10 am into the feedback module, and the feedback module analyzes and processes this information based on its internal neural network structure and parameters, thereby outputting a non-real-time overall performance index reasoning result, such as a predicted comprehensive evaluation value. Similar to the decision module, the computer system can use a deep learning framework to build a feedback module and use the forward propagation function to obtain the non-real-time overall performance index reasoning result.

[0146] In step 40D, the computer system determines the second performance indicator cost value based on the non-real-time overall performance indicator reasoning result of the current non-real-time time period and the non-real-time overall performance indicator of the current non-real-time time period, and generates a second non-real-time evaluation function based on the second performance indicator cost value. The second performance indicator cost value is used to measure the prediction error of the feedback module, and a feasible calculation method is the mean square error (MSE). For example, assuming that in the current non-real-time time period, the actual non-real-time overall performance indicator (calculated by comprehensively considering factors such as charging efficiency, charging stability, and grid load) is 0.8, and the non-real-time overall performance indicator reasoning result output by the feedback module is 0.7. Then, the error in this sample is (0.8 -0.7) 2 =0.01. The computer system sums the samples of all non-real-time time periods and divides it by the number of samples n to obtain the final second performance indicator cost value. The second non-real-time evaluation function is constructed based on the second performance indicator cost value, which is used to evaluate the performance of the feedback module. The feasible second non-real-time evaluation function can directly use the second performance indicator cost value or make some adjustments on this basis. For example, the second non-real-time evaluation function ,in is a regularization coefficient and S is a regularization term used to prevent the model from overfitting.

[0147] In step 40E, the computer system debugs the decision module based on the first non-real-time evaluation function, and debugs the feedback module based on the second non-real-time evaluation function. The debugging is terminated when the non-real-time debugging convergence requirements are met to obtain the transition charging scheduling neural network. The goal of debugging is to gradually reduce the values ​​of the first non-real-time evaluation function and the second non-real-time evaluation function by adjusting the parameters of the decision module and the feedback module, thereby improving the performance of the network. The computer system can use optimization algorithms (such as stochastic gradient descent, Adam algorithm, etc.) to achieve this goal. For the decision module, the optimization algorithm will update the parameters of the decision module according to the gradient of the first non-real-time evaluation function relative to the parameters of the decision module. For example, the update formula of the stochastic gradient descent algorithm is ,in is the parameter of the decision module at the current moment, is the learning rate, is the first non-real-time evaluation function J 1 Relative to the decision module parameters Gradient. For the feedback module, the optimization algorithm updates the parameters of the feedback module according to the gradient of the second non-real-time evaluation function relative to the parameters of the feedback module. In each iteration, the computer system calculates the gradient of the evaluation function relative to the network parameters, and then updates the parameters of the network according to the direction and magnitude of the gradient. As the iteration proceeds, the values ​​of the first non-real-time evaluation function and the second non-real-time evaluation function will gradually decrease. When the non-real-time debugging convergence requirements are met (for example, the value of the evaluation function fluctuates within a certain range and no longer decreases significantly), the computer system ends debugging, and the neural network obtained at this time is the transition charging scheduling neural network.

[0148] During the progressive debugging process of the entire non-real-time time period, the computer system processes a large number of non-real-time training examples and non-real-time overall performance index data, and ensures the calculation accuracy of each step. In order to improve the debugging efficiency, the computer system can use batch training to input multiple training examples of non-real-time time periods into the neural network for processing. At the same time, the computer system can also pre-process the training data, such as normalization and standardization, to accelerate the convergence speed of the neural network. In addition, the computer system also needs to continuously monitor the changes of the first non-real-time evaluation function and the second non-real-time evaluation function, and adjust the parameters of the optimization algorithm (such as learning rate) according to their changing trends to ensure the stability and effectiveness of the debugging process. Through progressive debugging of non-real-time time periods, the front-end charging scheduling neural network can further learn the mode and law of charging scheduling, improve its prediction accuracy of charging scheduling strategy and the reliability of evaluation of non-real-time overall performance indicators, and lay a more solid foundation for subsequent real-time debugging and practical application.

[0149] As an implementation mode, the network training process provided by the present invention further includes:

[0150] Step 40A1: Based on the previous generation of decision modules, non-real-time training examples in the current non-real-time time period are processed to obtain non-real-time decision support information in the current non-real-time time period;

[0151] Step 40A2: determining a difference metric cost according to the non-real-time decision support information corresponding to the current non-real-time time period, the density function of the non-real-time decision support information, and the density function of the non-real-time charging scheduling reasoning result;

[0152] Step 40A3: Determine the decision correction cost based on the difference measurement cost and the decision gain; the decision gain is the difference between the non-real-time overall performance index and the non-real-time overall performance index reasoning result in each non-real-time time period;

[0153] Step 40A4: According to the incentive and punishment strategy, the decision correction cost is adjusted to obtain the adjusted decision correction cost.

[0154] Based on this, step 40B, determining a first non-real-time evaluation function based on the second log-likelihood cost value, includes:

[0155] Step 40B1: Generate a first non-real-time evaluation function based on the second log-likelihood cost value and the adjusted decision correction cost.

[0156] In the training process of the charging scheduling neural network, steps 40A1 - 40A4 and sub-step 40B1 of step 40B are a series of closely related and important operations, which work together to progressively debug the non-real-time time period of the front charging scheduling neural network to obtain a more optimized transition charging scheduling neural network. These steps will be explained in detail below.

[0157] Step 40A1 describes that the computer system processes the non-real-time training examples of the current non-real-time time period based on the previous generation of decision modules, thereby obtaining non-real-time decision support information for the current non-real-time time period. The previous generation decision module here refers to the decision module that has been obtained in the previous training stage, which is gradually learned and optimized during the continuous training process. The non-real-time training examples include non-real-time charging parameter information and non-real-time supercharging network operation information, etc., which are important inputs in the training process. The computer system uses the learning ability and decision-making mechanism of the previous generation of decision modules to analyze and process these non-real-time training examples and output non-real-time decision support information.

[0158] Step 40A2 involves the computer system determining the difference measurement cost based on the non-real-time decision support information corresponding to the current non-real-time time period, the density function of the non-real-time decision support information, and the density function of the non-real-time charging scheduling reasoning result. The density function of the non-real-time decision support information describes the probability distribution of the non-real-time decision support information under different values, while the density function of the non-real-time charging scheduling reasoning result describes the probability distribution of the reasoning results obtained by processing the non-real-time training examples through the current decision module. The computer system determines the difference measurement cost by comparing the difference between the two density functions. A feasible technical means is to use Kullback-Leibler (KL) divergence to measure the difference between the two probability distributions. The formula for KL divergence is: , where P(x) is the density function of non-real-time decision support information, and Q(x) is the density function of non-real-time charging scheduling reasoning results.

[0159] In step 40A3, the computer system determines the decision correction cost based on the difference metric cost and the decision gain. The decision gain refers to the difference between the non-real-time overall performance index and the inference result of the non-real-time overall performance index in each non-real-time time period. It reflects the gap between the performance of the current decision module and the actual performance, such as the advantage function value. The computer system determines the decision correction cost by combining the difference metric cost with the decision gain. Specifically, first calculate the first multiplication result between the difference metric cost and the decision gain, then calculate the second multiplication result between the cutoff value of the difference metric cost and the decision gain, and then select the smaller value from the two multiplication results as the decision correction cost. The setting of the cutoff value is to avoid excessive influence of the difference metric cost on the decision correction. For example, assuming that the difference metric cost is D, the decision gain is A, and the cutoff value is C, then the first multiplication result is D×A, the second multiplication result is min(D,C)×A, and the decision correction cost L corr =min(D×A,min(D,C)×A). If the difference measurement cost is large but the decision gain is small, then by selecting a smaller multiplication result as the decision correction cost, it is possible to avoid over-adjustment of the decision module and make the decision correction more reasonable.

[0160] Step 40A4 describes that the computer system adjusts the decision correction cost according to the incentive and punishment strategy to obtain the adjusted decision correction cost. The incentive and punishment strategy can be, for example, a positive / negative incentive mechanism to encourage the decision module to make better decisions while punishing unreasonable decisions. Specifically, if the current decision correction cost is negative, the computer system uses the larger of the current decision correction cost and the second performance indicator cost value multiplied by the set positive integer as the adjusted decision correction cost; if the current decision correction cost is 0 or a positive value, the current decision correction cost is not adjusted. The second performance indicator cost value is determined based on the error between the non-real-time overall performance indicator inference result of the current non-real-time time period and the non-real-time overall performance indicator of the current non-real-time time period. For example, set the positive integer to k, and the second performance indicator cost value is , the current decision correction cost is , then the adjusted decision correction cost For: When <0, ;when hour, This adjustment method can ensure that when the decision correction cost is negative, appropriate adjustments can be made according to the actual performance indicator error to avoid excessive punishment or reward; and when the decision correction cost is positive or 0, the decision correction cost remains unchanged to maintain the stability of the decision.

[0161] Sub-step 40B1 of step 40B describes that the computer system generates a first non-real-time evaluation function based on the second log-likelihood cost value and the adjusted decision correction cost. The second log-likelihood cost value is determined based on the non-real-time charging scheduling inference result of the current non-real-time time period, which reflects the degree of match between the output of the current decision module and the actual situation. The computer system combines the second log-likelihood cost value and the adjusted decision correction cost to generate a first non-real-time evaluation function. A feasible combination method is to add the two. The first non-real-time evaluation function comprehensively considers the output accuracy of the decision module and the impact of decision correction, and is used to debug the decision module. By minimizing the first non-real-time evaluation function, the computer system can continuously adjust the parameters of the decision module to make the non-real-time charging scheduling inference results output by it more accurate and reasonable.

[0162] Throughout the process, the computer system gradually optimizes the decision module of the front charging scheduling neural network through these steps. From using the previous generation decision module to obtain empirical data, to determining the difference measurement cost by comparing the density function, and then combining the decision gain and incentive penalty strategy to determine the adjusted decision correction cost, and finally generating the first non-real-time evaluation function to debug the decision module, each step is closely linked and affects each other. These operations help improve the performance of the decision module, enabling it to better cope with different non-real-time training examples, and lay the foundation for finally obtaining the transition charging scheduling neural network. At the same time, these steps also reflect the application of reinforcement learning in the training of charging scheduling neural networks. Through continuous feedback and adjustment, the network can learn better decision strategies, thereby improving the effect and efficiency of charging scheduling.

[0163] For the acquisition of non-real-time decision support information, the computer system can adopt a variety of technical means. For example, sequence models such as recurrent neural networks (RNN) or long short-term memory networks (LSTM) can be used to process time series information in non-real-time training examples to better capture the dynamic changes of charging parameters and supercharging network operation information. When determining the difference measurement cost, in addition to KL divergence, other measurement methods can be used, such as Jensen-Shannon (JS) divergence or Wasserstein distance, which may provide more accurate difference measurement results in some cases. In the setting of the incentive and penalty strategy, the value of the positive integer k can be adjusted according to the specific application scenario and experimental results to achieve the best adjustment effect. For the generation of the first non-real-time evaluation function, in addition to the addition method, the weighted summation method can also be considered to assign different weights to the second log-likelihood cost value and the adjusted decision correction cost to better balance the output accuracy of the decision module and the impact of decision correction. At the same time, the computer system can dynamically adjust these weights during the training process, optimize the composition of the first non-real-time evaluation function according to the training conditions and performance indicators at different stages, and further improve the performance of the decision module.

[0164] In the entire charging scheduling neural network training process, steps 40A1 to 40A4 and sub-step 40B1 of step 40B constitute a complex and sophisticated decision correction and optimization mechanism. Through these steps, the computer system can make full use of historical experience and real-time feedback to effectively debug and improve the decision module, so that it can make more accurate and reasonable charging scheduling decisions in a non-real-time environment. These decisions not only help improve charging efficiency and stability, but also better adapt to the operating requirements of the supercharging network, providing strong support for distributed vehicle liquid cooling supercharging control.

[0165] As an implementation mode, step 40A4, adjusting the decision correction cost according to the incentive and penalty strategy to obtain the adjusted decision correction cost, includes:

[0166] Step 40A41: if the current decision correction cost is negative, the current decision correction cost and the second performance indicator cost value multiplied by the set positive integer are used as the adjusted decision correction cost.

[0167] Step 40A42: If the current decision correction cost is 0 or a positive value, the current decision correction cost is not adjusted.

[0168] When the computer system executes step 40A4 in the charging scheduling neural network training process, it will adjust the decision correction cost according to the incentive and punishment strategy to obtain the adjusted decision correction cost, wherein steps 40A41 and 40A42 respectively process different situations of the decision correction cost, and sub-step 40B1 of step 40B generates the first non-real-time evaluation function based on the relevant cost. Step 40A41 stipulates that when the current decision correction cost is negative, the computer system will take the current decision correction cost and the second performance indicator cost value multiplied by the set positive integer as the adjusted decision correction cost. The decision correction cost reflects the difference between the current decision and the ideal decision. A negative decision means that the current decision may deviate from the expected direction, but there may be accidental factors that cause it to be over-punished; the second performance indicator cost value is determined based on the error between the non-real-time overall performance indicator reasoning result of the current non-real-time time period and the non-real-time overall performance indicator of the current non-real-time time period, which measures the accuracy of the decision module's prediction of the performance indicator. The introduction of setting a positive integer is to make a certain degree of correction with the second performance indicator cost value as a reference when the decision correction cost is unreasonably negative. For example, assuming that the positive integer is set to 3, the current decision correction cost is -5, and the second performance indicator cost value is 2, then the second performance indicator cost value multiplied by the set positive integer is 6, and the computer system selects 6 as the adjusted decision correction cost. The purpose of this is to avoid excessive punishment caused by accidental factors and make the adjustment of the decision module more reasonable. In implementation, the computer system can implement this step through comparison operation, that is, compare the decision correction cost and the second performance indicator cost value multiplied by the set positive integer, and select the larger value. If the current decision correction cost is 0 or a positive value, the computer system does not adjust the current decision correction cost. When the decision correction cost is 0, it means that the degree of correction of the current decision is just right and no further adjustment is required; when the decision correction cost is a positive value, it means that the current decision has been corrected in the right direction. At this time, keeping the cost unchanged can maintain the stability of the decision correction. For example, if the decision correction cost is 3, the computer system will directly use 3 as the adjusted decision correction cost. This step is mainly implemented through conditional judgment. The computer system only needs to judge the positive or negative nature of the decision correction cost. If it is 0 or positive, the decision correction cost is directly used as the adjusted result. Sub-step 40B1 of step 40B describes that the computer system generates a first non-real-time evaluation function based on the second log-likelihood cost value and the adjusted decision correction cost. The second log-likelihood cost value reflects the degree of match between the output of the current decision module and the actual situation, and it measures the accuracy of the decision module's inference results for non-real-time charging scheduling; the adjusted decision correction cost takes into account factors such as the difference between the decision and the empirical data and the decision gain, and optimizes the correction of the decision module.The computer system combines the two costs to generate a first non-real-time evaluation function, for example, by adding them together. The first non-real-time evaluation function comprehensively considers the output accuracy of the decision module and the impact of decision correction, and is used to debug the decision module. The computer system continuously adjusts the parameters of the decision module by minimizing the first non-real-time evaluation function, so that the non-real-time charging scheduling reasoning results output by it are more accurate and reasonable. During implementation, the computer system can use optimization algorithms, such as stochastic gradient descent and its variant algorithms, to update the parameters of the decision module according to the gradient information of the first non-real-time evaluation function, thereby optimizing the decision module.

[0169] As an implementation method, in step 20, the step of constructing the non-real-time training example for the next non-real-time time period based on the non-real-time training example for the current non-real-time time period and the non-real-time charging scheduling inference result output by the front-end charging scheduling neural network is completed based on the generation component in the generative adversarial network. The generative adversarial network is trained by the following steps:

[0170] Step 20A: construct a charging state training example using the supercharging network operation information and charging parameter information corresponding to the same debugging time period in the real-time monitoring information;

[0171] Step 20B: for the real-time debugging time period, according to the charging parameter information of the next debugging time period of the real-time debugging time period, determine the state charging scheduling strategy information corresponding to the real-time debugging time period;

[0172] Step 20C: Based on the generation component in the generative adversarial network to be debugged, the charging state training examples and the state charging scheduling strategy information of the real-time debugging time period are processed to construct the charging state reasoning information of the next debugging time period;

[0173] Step 20D: Determine the state of charge cost based on the error between the state of charge reasoning information of each debugging time period and the state of charge training example of each debugging time period;

[0174] Step 20E: determining the cost of the generation component based on the authenticity of the construction when the generation component constructs the charging state reasoning information of each debugging time period, and determining the cost of the judgment component based on the judgment accuracy when judging the charging state reasoning information and the charging state training examples of each debugging time period respectively in the judgment component in the generative adversarial network to be debugged;

[0175] Step 20F: Generate a target cost value of the generative adversarial network to be debugged based on the charging state cost, the generation component cost, and the judgment component cost, and repeatedly debug the generative adversarial network to be debugged based on the target cost value, and terminate the debugging when the adversarial generation convergence requirements are met to obtain a debugged generative adversarial network.

[0176] In step 20A, the computer system constructs the supercharging network operation information and charging parameter information corresponding to the same debugging time period in the real-time monitoring information into a charging state training example. The real-time monitoring information is various data continuously collected during the actual charging process. The supercharging network operation information includes the voltage, frequency, power and other parameters of the power grid, and the charging parameter information covers the charging current, battery temperature, charging rate, etc. The computer system integrates this information within the same debugging time period to form a complete charging state training example.

[0177] In step 20B, the computer system determines the state charging scheduling strategy information corresponding to the real-time debugging time period based on the charging parameter information of the next debugging time period of the real-time debugging time period. The state charging scheduling strategy information is a scheduling plan formulated according to the charging demand and actual situation of the next debugging time period, which will guide the charging process. For example, if the battery power in the next debugging time period is low and needs to be charged quickly, the state charging scheduling strategy information may indicate to increase the charging current and power. The computer system can establish a mapping relationship between charging parameters and scheduling strategies by analyzing historical data and charging models, thereby determining the corresponding state charging scheduling strategy information according to the charging parameter information of the next debugging time period.

[0178] Step 20C describes that the computer system processes the charging state training examples and the state charging scheduling strategy information of the real-time debugging time period based on the generation component in the generative adversarial network to be debugged to construct the charging state reasoning information of the next debugging time period. The generation component is one of the core parts of the generative adversarial network, which has the ability to learn and generate data. The computer system inputs the charging state training examples and the state charging scheduling strategy information into the generation component, and the generation component analyzes and processes this information through the internal neural network structure and algorithm to generate the charging state reasoning information of the next debugging time period. For example, the generation component can predict the charging current, battery temperature and other parameters of the next debugging time period based on the current charging state and scheduling strategy.

[0179] Step 20D involves the computer system determining a state of charge cost based on the error between the state of charge reasoning information for each debugging period and the state of charge training examples for each debugging period. The state of charge cost measures the degree of difference between the state of charge reasoning information generated by the generation component and the true state of charge training examples. The computer system can use a variety of error metrics, such as mean square error (MSE), mean absolute error (MAE), etc. The computer system calculates the error for each debugging period and accumulates or averages it to obtain a final state of charge cost.

[0180] Step 20E requires the computer system to determine the cost of the generated component and the cost of the judged component. The cost of the generated component is based on the construction authenticity of the generated component when constructing the charging state inference information for each debugging time period, that is, the similarity between the generated data and the real data. The computer system can use some measurement methods, such as the kernel maximum mean difference (KMMD) to measure the distribution difference between the generated data and the real data. The formula for the kernel maximum mean difference is: , where P is the distribution of real data, Q is the distribution of generated data, k is the kernel function, is the reproducing kernel Hilbert space. The smaller the cost of the generation component, the closer the generated data is to the real data. The cost of the determination component is based on the determination accuracy of the determination component in the generative adversarial network to be debugged when determining the charging state inference information and charging state training examples for each debugging time period. The task of the determination component is to distinguish between generated data and real data. The higher the determination accuracy, the more accurately the determination component can identify the source of the data. The computer system can use the cross entropy loss function to calculate the cost of the determination component. The formula of the cross entropy loss function is: , where y i is the true label (0 or 1), is the predicted probability of the decision component.

[0181] In step 20F, the computer system generates the target cost value of the generative adversarial network to be debugged based on the charging state cost, the generation component cost and the determination component cost, and repeatedly debugs the generative adversarial network to be debugged based on the target cost value, and ends the debugging when the adversarial generation convergence requirement is met to obtain the debugged generative adversarial network. The target cost value is a comprehensive cost indicator that comprehensively considers the accuracy and authenticity of the data generated by the generation component and the determination accuracy of the determination component. The computer system can add the charging state cost, the generation component cost and the determination component cost according to certain weights to obtain the target cost value. The computer system gradually reduces the target cost value by continuously adjusting the parameters of the generation component and the determination component until the adversarial generation convergence requirement is met.

[0182] The computer system continuously optimizes the performance of the generation component and the determination component through steps 20A-20F. From constructing charging state training examples and determining state charging scheduling strategy information, to generating charging state reasoning information and calculating various costs, to synthesizing these costs to generate target cost values ​​and debugging, each step is interrelated and jointly promotes the training of the generative adversarial network. In this way, the generative adversarial network can learn the real distribution and change rules of the charging state, generate high-quality charging state reasoning information, and provide strong support for subsequent charging scheduling decisions.

[0183] Figure 2A hardware entity diagram of a computer system provided by an embodiment of the present invention is as follows Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

Claims

1. A distributed automobile liquid cooling supercharging control method, characterized in that: The method comprises: Acquire real-time charging parameter information in the current time period, and supercharging network operation information caused by charging scheduling decisions based on the real-time charging parameter information; the real-time charging parameter information includes charging power, battery status and charging rate; the real-time charging parameter information is determined based on past charging scheduling strategy information in the previous time period; Extracting features of the real-time charging parameter information and the supercharging network operation information respectively to obtain corresponding feature vector representations; Merging the feature vectors to obtain merged information; Based on one or more internal weight focusing networks, the merged information is embedded and mapped to obtain a temporary vector representation; if there are multiple internal weight focusing networks, the multiple internal weight focusing networks are cascaded, and the execution data of the first internal weight focusing network is the merged information, and the execution data of the internal weight focusing networks other than the first one is the execution result after the previous internal weight focusing network connected to it is embedded and mapped, and the execution result of the last internal weight focusing network is the temporary vector representation; wherein the internal weight focusing network includes a multi-channel joint focusing component and a feedforward neural network component; for any internal weight focusing network, the embedding mapping process of any internal weight focusing network includes: based on the multi-channel joint focusing component in the internal weight focusing network, a multi-channel joint focusing operation is performed on the corresponding execution data to obtain a multi-channel joint focusing result; based on the feedforward neural network component in the internal weight focusing network, the multi-channel joint focusing result is feedforward processed to obtain the execution result of the corresponding internal weight focusing network; multi-objective decision prediction is performed based on the temporary vector representation to obtain target charging scheduling strategy information at one or more levels; According to the target charging parameter information determined by using the target charging scheduling strategy information, a charging scheduling decision is made for the supercharging demand in a subsequent time period during the supercharging management of the charging pile.

2. The method according to claim 1, characterized in that The multi-channel joint focusing component based on the internal weighted focusing network performs a multi-channel joint focusing operation on the corresponding execution data to obtain a multi-channel joint focusing result, including: Based on the multi-channel joint focusing component in the internal weighted focusing network, performing an affine transformation on the corresponding execution data to obtain an affine transformation vector representation; Decomposing the affine transformation vector representation to obtain a plurality of affine transformation sub-vectors; Performing a normalized dot product focusing operation on each of the affine transformation sub-vectors to obtain a plurality of multi-channel focused sub-vectors; Combining a plurality of the multi-channel focusing sub-vectors to obtain a multi-channel focusing vector representation; An affine transformation is performed on the multi-channel focusing vector representation to obtain a multi-channel joint focusing result.

3. The method according to claim 2, characterized in that The step of performing a normalized dot product focusing operation on each of the affine transformation sub-vectors to obtain a plurality of multi-channel focused sub-vectors comprises: For any affine transformation subvector, perform dot product calculation on the corresponding affine transformation subvector and itself to obtain a dot product calculation result; Normalizing the dot product calculation result to obtain a normalized result; Based on a first nonlinear transformation operator, transforming the normalized processing result into an influence coefficient vector corresponding to a corresponding affine transformation subvector; The influence coefficient vector is multiplied with the corresponding affine transformation sub-vector to obtain a multi-channel focusing sub-vector corresponding to the corresponding affine transformation sub-vector.

4. The method according to claim 1, characterized in that: The multi-objective decision prediction is performed based on the temporary vector representation to obtain target charging scheduling strategy information at one or more levels, including: For each of the one or more levels, the temporary vector representation is subjected to identity mapping to obtain an identity mapping result corresponding to each level; Performing affine transformation on each of the identity mapping results to obtain a corresponding affine transformation result; Based on the second nonlinear transformation operator, the affine transformation results are transformed into target charging scheduling strategy information corresponding to the corresponding levels.

5. The method according to any one of claims 1 to 4, characterized in that: The method is implemented based on a pre-trained charging scheduling neural network, and the method also includes a training process of the charging scheduling neural network, including: Obtaining a pre-charging scheduling neural network obtained by pre-debugging using a plurality of basic training example sets; For each non-real-time time period in the non-real-time debugging, a non-real-time training example for the next non-real-time time period is constructed based on the non-real-time training example for the current non-real-time time period and the non-real-time charging scheduling inference result output by the front-end charging scheduling neural network; the non-real-time training example includes non-real-time charging parameter information and non-real-time super charging network operation information; Determine a non-real-time overall performance indicator for each non-real-time time period based on the non-real-time training examples for each non-real-time time period; According to the non-real-time training examples and the non-real-time overall performance indicators respectively corresponding to the plurality of non-real-time time periods, the front-end charging scheduling neural network is progressively debugged in the non-real-time time period, and the debugging is terminated when the non-real-time debugging convergence requirements are met to obtain a transition charging scheduling neural network; For each real-time time period in real-time debugging, the real-time overall performance index of the current real-time time period is determined based on the real-time training examples of the current real-time time period; According to the real-time training examples and real-time overall performance indicators corresponding to multiple real-time time periods, the transition charging scheduling neural network is progressively debugged in real-time time periods, and the debugging is ended when the real-time debugging convergence requirements are met to obtain a trained charging scheduling neural network.

6. The method according to claim 5, characterized in that The pre-charging scheduling neural network is obtained based on pre-debugging, and the process of pre-debugging is as follows: Construct past training examples by using the past supercharging network operation information and past charging parameter information corresponding to the same past time period in the collected monitoring information; For the real-time past time period, determine the past scheduling comparison information corresponding to the real-time past time period according to the past charging parameter information of the next past time period of the real-time past time period, and determine the past overall performance index corresponding to the real-time past time period based on the past training examples of the real-time past time period; The past training examples, past scheduling comparison information and past overall performance indicators corresponding to the same past time period are regarded as a basic training example set; The original charging scheduling neural network to be debugged is pre-debugged according to a plurality of basic training example sets, and the debugging is terminated when the pre-debugging convergence requirements are met to obtain the pre-charging scheduling neural network.

7. The method according to claim 6, characterized in that The method of pre-debugging the original charging scheduling neural network to be debugged according to the plurality of basic training example sets, and ending the debugging when the pre-debugging convergence requirements are met to obtain the pre-charging scheduling neural network includes: According to the decision module in the original charging scheduling neural network to be debugged, the past training examples in the basic training example set are processed to obtain the past scheduling reasoning results; Determine a first log-likelihood cost value based on the past scheduling reasoning results and the past scheduling comparison information corresponding to each basic training example set; Processing past training examples in the basic training example set according to a feedback module in the original charging scheduling neural network to be debugged to obtain past overall performance indicator inference results; Determining a first performance indicator cost value based on the error between the past overall performance indicator reasoning result and the corresponding past overall performance indicator; Determining a pre-debugging cost value according to the first log-likelihood cost value and the first performance indicator cost value; Based on the pre-debugging cost value, pre-debugging is performed on the original charging scheduling neural network, and debugging is ended when the pre-debugging convergence requirement is met to obtain the pre-charge scheduling neural network.

8. The method according to claim 5, characterized in that The non-real-time training example includes charging current and charging fluctuation information, and the non-real-time overall performance index of each non-real-time time period is determined based on the non-real-time training example of each non-real-time time period, including: For a current non-real-time time period, determining a charging efficiency of the current non-real-time time period based on a charging current of the current non-real-time time period; Determining a charging stability index for the current non-real-time time period based on the charging fluctuation information for the current non-real-time time period; Determine a floating index of the current non-real-time time period based on the charging current of the current non-real-time time period and the charging current of the next non-real-time time period of the current non-real-time time period; Calculate the non-real-time performance index of the current non-real-time time period based on the charging efficiency, charging stability index, and floating index of the current non-real-time time period; Determine the non-real-time overall performance indicator of the current non-real-time time period according to the non-real-time overall performance indicator of the next non-real-time time period and the non-real-time performance indicator of the current non-real-time time period; The front-end charging scheduling neural network includes a decision module and a feedback module. The front-end charging scheduling neural network is progressively debugged in non-real-time time periods according to non-real-time training examples and non-real-time overall performance indicators corresponding to a plurality of non-real-time time periods, and the debugging is terminated when the non-real-time debugging convergence requirements are met to obtain a transition charging scheduling neural network, including: Based on the decision module of the current non-real-time time period, the non-real-time training examples of the current non-real-time time period are processed to obtain the non-real-time charging scheduling inference result of the current non-real-time time period; Determine a second log-likelihood cost value according to the non-real-time charging scheduling inference result of the current non-real-time time period, and determine a first non-real-time evaluation function based on the second log-likelihood cost value; Based on the feedback module of the current non-real-time time period, the non-real-time training examples of the current non-real-time time period are processed to obtain the inference result of the non-real-time overall performance index of the current time period; Determine a second performance indicator cost value according to the non-real-time overall performance indicator inference result of the current non-real-time time period and the non-real-time overall performance indicator of the current non-real-time time period, and generate a second non-real-time evaluation function according to the second performance indicator cost value; The decision module is debugged based on the first non-real-time evaluation function, and the feedback module is debugged based on the second non-real-time evaluation function. The debugging is terminated when the non-real-time debugging convergence requirement is met to obtain a transition charging scheduling neural network.

9. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 8 are implemented.

Images