A Cooling Control Method and System for Charging Piles Based on AI Predictive Reinforcement Learning

Abstract

This invention discloses a charging pile cooling control method and system based on AI predictive reinforcement learning, belonging to the field of cooling control technology. The AI ​​predictive reinforcement learning-based charging pile cooling control method and system includes: S1, collecting charging heat dissipation data from the charging pile and performing preprocessing on the data; S2, constructing a time-series feature matrix and inputting it into a temperature prediction model, outputting a temperature prediction sequence for the controlled target, assessing the risk of thermal runaway during the prediction stage, and constructing a thermal risk identification sequence; S3, constructing a cooling strategy optimization model, inputting the real-time state vector into the model, outputting adjustment commands and issuing them for execution; S4, evaluating the execution deviation of the adjustment commands, adjusting the cooling strategy based on the evaluation results, and generating cooling strategy feedback samples. This solves the problem of energy waste and cooling imbalance caused by the lack of real-time prediction and adaptive regulation capabilities in existing charging pile cooling control processes.

Description

Technical Field

[0001] This invention relates to the field of cooling control technology, specifically to a cooling control method and system for charging piles based on AI predictive reinforcement learning. Background Technology

[0002] Currently, with the continuous improvement of charging power for new energy vehicles and the development of high-rate fast charging technology, the heat generated during charging has increased significantly, placing higher demands on the intelligent control of charging pile cooling systems. Traditional cooling strategies often rely on fixed parameter settings or temperature threshold control, which are difficult to fully adapt to the complex dynamic characteristics of thermal loads under different charging conditions. This can easily lead to overcooling or undercooling, affecting equipment energy efficiency and potentially causing safety hazards and equipment aging. Therefore, how to achieve precise and efficient cooling control based on changes in operating conditions has become an important direction for research in charging pile thermal management.

[0003] For example, the invention disclosed in CN120578065A provides a method for predicting and controlling the rotational speed of a liquid cooling system for a charging gun. The method includes: acquiring performance parameters of key components in the liquid cooling system, including a radiator, water pump, fan, charging gun wire, and liquid storage tank; inputting the performance parameters into a one-dimensional simulation model to obtain the maximum coolant temperature (Tcoolant) under different operating conditions; fitting the relationship between different operating conditions and Tcoolant based on the Tcoolant under different operating conditions to obtain a rotational speed prediction model; inputting the ambient temperature and the heating power of the charging gun wire resistance under the target operating condition into the rotational speed prediction model, predicting multiple rotational speed combinations that satisfy the constraints under the Tcoolant, and selecting the rotational speed combination that minimizes the total energy consumption of the liquid cooling system from the multiple rotational speed combinations as the optimal rotational speed combination for the target operating condition; and real-time regulating the fan speed and water pump speed based on the optimal rotational speed combination during charging.

[0004] For example, the invention disclosed in CN117826615B, which relates to the field of electric vehicle thermal management technology, describes a method for determining the control parameters of the coolant in an electric vehicle power battery. This method determines the requirements of the battery pack and coolant based on the design and operating conditions of the electric vehicle power battery pack, including battery thermal characteristic requirements, coolant heat transfer performance requirements, and coolant cooling performance requirements. It then constructs a battery thermal characteristic model and a cooling-heat transfer model to simulate the thermal characteristics of the electric vehicle under different operating conditions and the heat transfer characteristics between the coolant and the battery. The battery thermal characteristic model and the cooling-heat transfer model are combined to construct a multi-model fusion thermal simulation to simulate the battery pack. Finally, experimental verification with an actual electric vehicle power battery is conducted to obtain a simulation conformity index, verify the feasibility of the multi-model fusion thermal simulation, and optimize the coolant control parameters of the electric vehicle power battery.

[0005] However, the aforementioned technologies are mostly based on static modeling or simulation prediction to establish cooling control logic, lacking the ability to dynamically perceive and adaptively control real-time data. They are unable to cope with rapid fluctuations in heat load during the charging process, especially during the constant current-constant voltage switching phase. This results in problems such as control response lag and unreasonable allocation of cooling resources, which affect the cooling efficiency of the system and the service life of the equipment.

[0006] Therefore, in order to address the above problems, there is an urgent need for a cooling control method and system for charging piles based on AI predictive reinforcement learning. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a charging pile cooling control method and system based on AI predictive reinforcement learning, which solves the problems of energy waste and cooling imbalance caused by the lack of real-time prediction and adaptive regulation capabilities in the existing charging pile cooling control process.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a charging pile cooling control method and system based on AI predictive reinforcement learning, comprising: S1, collecting charging heat dissipation data of the charging pile, and performing time alignment, anomaly removal, deburring, and normalization processing on the charging heat dissipation data to obtain preprocessed charging heat dissipation data; S2, constructing a time-series feature matrix using the preprocessed charging heat dissipation data, inputting it into a temperature prediction model, outputting a temperature prediction sequence of the controlled target, assessing the thermal runaway risk during the prediction stage, and constructing a thermal risk identification sequence; S3, generating a state vector based on the thermal risk identification sequence and the temperature prediction sequence, generating an action vector by combining the control range of the liquid cooling pump and the fan, constructing a cooling strategy optimization model, and training the cooling strategy optimization model using reinforcement learning, inputting the real-time state vector into the trained cooling strategy optimization model, outputting adjustment commands and issuing them for execution; S4, obtaining the temperature prediction sequence and thermal risk assessment value after command execution, assessing the execution deviation of the adjustment command, adjusting the cooling strategy based on the assessment results, and generating cooling strategy feedback samples to achieve adaptive closed-loop evolution of the cooling strategy.

[0011] Further, charging heat dissipation data of the charging pile is collected, and time alignment, anomaly removal, glitch removal, and normalization are performed on the charging heat dissipation data to obtain preprocessed charging heat dissipation data. The specific steps are as follows: Charging heat dissipation data during the charging process of the charging pile is collected periodically with a fixed-width sliding time window. The charging heat dissipation data includes charging power, battery temperature, IGBT temperature, charging gun temperature, coolant temperature, liquid cooling pump flow rate, liquid cooling pump speed, fan speed, ambient temperature, relative humidity, and natural wind speed. The charging heat dissipation data is time-aligned using a high-precision timestamp synchronization method to eliminate data misalignment caused by sampling frequency and communication delay between multiple sensors. A dynamic anomaly identification mechanism is constructed by combining the temperature rise change trend and sliding window volatility assessment to identify and remove pulse interference and abrupt changes in the charging heat dissipation data. A slowly varying smoothing function based on empirical response time and cooling inertia parameters is used to remove glitch and correct response delay in the charging heat dissipation data. The charging heat dissipation data is feature-compressed using a linear normalization algorithm to map inputs of different dimensions to a unified numerical range.

[0012] Further, the specific steps for constructing a time-series feature matrix from preprocessed charging heat dissipation data, inputting it into the temperature prediction model, and outputting the temperature prediction sequence of the controlled target are as follows: Based on the preprocessed charging heat dissipation data, a time-series feature matrix is ​​constructed using a fixed time window method, and the time-series feature matrix is ​​arranged according to the sampling period; the time-series feature matrix is ​​used as input, and a neural network with an associative memory structure is used to model the time-series dependencies in the charging process, constructing a temperature prediction model, and an attention mechanism is introduced to train the model; the real-time feature matrix is ​​input into the trained temperature prediction model, and the temperature prediction sequence of the controlled target at multiple future time points is output. The controlled targets include the battery, charging gun head, IGBT, and coolant. The temperature prediction sequence includes the predicted values ​​of battery temperature, charging gun head temperature, IGBT temperature, and coolant temperature.

[0013] Furthermore, the specific steps for assessing the thermal runaway risk during the prediction phase and constructing a thermal risk identification sequence are as follows: Calculate the difference between the predicted temperature value and the corresponding temperature threshold at each time point in the temperature prediction sequence, obtaining the differences between the predicted battery temperature and the battery temperature threshold, the predicted charging gun temperature and the charging gun temperature threshold, the predicted IGBT temperature and the IGBT temperature threshold, and the predicted coolant temperature and the coolant temperature threshold, respectively; Square and sum these four types of temperature differences to obtain a comprehensive temperature risk accumulation value; Multiply the accumulated temperature risk value by the pre-processed charging power to obtain the heat source input factor; Calculate the heat source input factor by adjusting the liquid cooling system flow rate and fan speed... The cooling flow capacity factor is obtained by multiplying the predicted coolant temperature by one and the natural wind speed by one. The ambient temperature is then subtracted from the predicted coolant temperature and multiplied by the heat dissipation response adjustment coefficient. The negative value of the product is taken as the exponential part and subjected to a power function operation with the natural logarithm base e to obtain the heat conduction adjustment factor. The heat conduction adjustment factor is multiplied by one and then multiplied by the cooling flow capacity factor to obtain the comprehensive cooling capacity factor. The heat source input factor is divided by the comprehensive cooling capacity factor to obtain the thermal risk assessment value. The thermal risk assessment value and the thermal risk threshold are compared in real time. Time points where the thermal risk assessment value is greater than the thermal risk threshold are marked as thermal risk time points, and all thermal risk time points are arranged in chronological order to form a thermal risk identification sequence.

[0014] Furthermore, the specific steps for generating a state vector based on the thermal risk identification sequence and temperature prediction sequence, and generating an action vector by combining the control range of the liquid coolant pump and the fan, are as follows: Based on the thermal risk identification sequence and the temperature prediction sequence of the controlled target, extract the charging power, coolant temperature prediction value, coolant temperature, liquid coolant pump flow rate, liquid coolant pump speed, fan speed, ambient temperature, and natural wind speed at the corresponding time points, and combine them in the sampling order to form a state vector with temporal characteristics; according to the controllable range of the liquid coolant pump speed and fan speed, set discrete levels for selectable target values; arrange different liquid coolant pump speed adjustment commands and fan speed adjustment commands in a combination manner to generate a set of control action combinations, and organize them in vector form as action vectors.

[0015] Furthermore, a cooling strategy optimization model is constructed and trained using reinforcement learning. The specific steps for inputting the real-time state vector into the trained cooling strategy optimization model and outputting adjustment commands for execution are as follows: Using the state vector and action vector as input, a cooling strategy optimization model is constructed based on the Actor-Critic reinforcement learning architecture, and a near-end policy optimization algorithm is used to train the model. During training, the difference between the predicted battery temperature and the current battery temperature is squared, the difference between the predicted IGBT temperature and the current IGBT temperature is squared, and the difference between the predicted charging gun temperature and the current charging gun temperature is squared. The differences between the three terms are squared, and the squared differences of the three terms are added together to form the temperature control accuracy evaluation term for the cooling regulation process. The target speed of the liquid cooling pump is divided by the maximum speed of the liquid cooling pump and the square is taken, and the target speed of the fan is divided by the maximum speed of the fan and the square is taken. The two squared values ​​are added together to form the cooling resource usage term. The temperature control accuracy evaluation term and the cooling resource usage term are added together and the negative sign is taken to obtain the cooling control load index value. The cooling control load index value is used as the reinforcement learning reward corresponding to the current state and action. The real-time state vector is input into the trained cooling strategy optimization model, and the liquid cooling pump speed adjustment command and the fan speed adjustment command are output and sent to the command control terminal for execution.

[0016] Further, the specific steps for obtaining the temperature prediction sequence and thermal risk assessment value after instruction execution, and evaluating the execution deviation of the adjustment instruction, are as follows: After the adjustment instruction is executed, based on real-time charging heat dissipation data, a new thermal risk assessment and temperature prediction are performed to obtain the temperature prediction sequence and thermal risk assessment value after instruction execution; the battery temperature after instruction execution is subtracted from the battery temperature prediction value before instruction execution, and the difference is divided by the battery temperature prediction value before instruction execution to obtain the battery temperature execution deviation; the IGBT temperature after instruction execution is subtracted from the IGBT temperature prediction value before instruction execution, and the difference is divided by the IGBT temperature prediction value before instruction execution to obtain the IGBT temperature execution deviation; the instruction execution... The charging gun head temperature execution deviation is obtained by subtracting the predicted charging gun head temperature before instruction execution from the post-instruction charging gun head temperature and dividing the difference by the predicted charging gun head temperature before instruction execution. The average temperature deviation term is obtained by summing the battery temperature execution deviation, IGBT temperature execution deviation, and charging gun head temperature execution deviation. The normalized change in thermal risk is obtained by subtracting the thermal risk assessment value before instruction execution from the post-instruction thermal risk assessment value and dividing by the sum of the thermal risk assessment value before instruction execution and the minimum term. The normalized change in thermal risk is multiplied by the risk adjustment coefficient to obtain the thermal risk adjustment term. The cooling execution deviation assessment value is obtained by adding the average temperature deviation term and the thermal risk adjustment term.

[0017] Further, the specific steps for adjusting the cooling strategy based on the evaluation results are as follows: Compare the cooling execution deviation evaluation value D with the cooling deviation thresholds D1 and D2 in real time, and adjust the cooling strategy: When D ≤ D1, it is determined that the adjustment instruction has an overcooling trend. In the next round of cooling strategy training, adjust the negative feedback weight of the corresponding action vector, and control the electric valve to close the valve; When D1 < D < D2, it is determined that the adjustment instruction is effective, and the cooling strategy remains unchanged; When D ≥ D2, it is determined that the adjustment instruction has an undercooling risk. In the next round of cooling strategy training, increase the adjustment amplitude of the target speed of the liquid cooling pump and the target speed of the fan under the corresponding state variable, shorten the cooling strategy refresh interval, and increase the valve opening of the electric valve.

[0018] Further, the specific steps for generating a cooling strategy feedback sample to achieve the adaptive closed-loop evolution of the cooling strategy are as follows: Combine the charging heat dissipation data, state vector, action vector, temperature prediction sequence, and cooling execution deviation evaluation value to form a cooling strategy feedback sample. According to the cooling execution deviation evaluation value, label the overcooling, normal, and undercooling labels for the cooling strategy feedback sample, and incorporate them into the training dataset of the cooling strategy optimization model to achieve the full-process closed-loop of cooling control from prediction, decision-making, execution, feedback to strategy retraining.

[0019] The second aspect of the present invention provides a charging pile cooling control system based on AI prediction reinforcement learning, including: a heat dissipation data acquisition and preprocessing module, a temperature prediction and risk analysis module, a cooling strategy optimization and decision-making module, and a strategy feedback closed-loop training module. Among them: The heat dissipation data acquisition and preprocessing module is used to collect the charging heat dissipation data of the charging pile, and perform time alignment, anomaly elimination, de-burring, and normalization processing on the charging heat dissipation data to obtain the preprocessed charging heat dissipation data; The temperature prediction and risk analysis module is used to call the preprocessed charging heat dissipation data to construct a time series feature matrix, input it into the temperature prediction model, output the temperature prediction sequence of the controlled target, and evaluate the thermal runaway risk in the prediction stage to construct a thermal risk identification sequence; The cooling strategy optimization and decision-making module is used to generate a state vector based on the thermal risk identification sequence and the temperature prediction sequence, generate an action vector in combination with the control range of the liquid cooling pump and the fan, construct a cooling strategy optimization model, and perform reinforcement learning training on the cooling strategy optimization model. Input the real-time state vector into the trained cooling strategy optimization model, output an adjustment instruction and send it for execution; The strategy feedback closed-loop training module is used to obtain the temperature prediction sequence and the thermal risk evaluation value after the instruction is executed, evaluate the execution deviation of the adjustment instruction, adjust the cooling strategy based on the evaluation result, and generate a cooling strategy feedback sample to achieve the adaptive closed-loop evolution of the cooling strategy.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This AI-based predictive reinforcement learning-based cooling control method and system for charging piles, by introducing an LSTM-based temperature prediction model and combining it with an attention mechanism to enhance the modeling ability of key thermal variables, can accurately predict the temperature change trends of the battery, charging gun, IGBT, and coolant. Compared with the traditional passive response method that relies on real-time temperature measurement, this invention constructs a future temperature evolution sequence, transforming cooling regulation from a post-event response to an early intervention, effectively reducing the thermal hysteresis problem during charging, and dynamically adjusting the cooling strategy before the temperature rises rapidly, thus avoiding the risk of thermal runaway in advance.

[0023] (2) The AI-based predictive reinforcement learning-based cooling control method and system for charging piles utilizes a reinforcement learning mechanism to construct a cooling strategy optimization model. This model can dynamically adjust the control strategy based on the thermal risk performance and historical execution results during the charging process, without relying on specific physical models or manual rules. The model supports self-learning and strategy updates, thereby achieving precise and efficient coordinated control of the fan and liquid cooling pump, overcoming the problems of poor adaptability and lag in traditional fixed threshold control methods.

[0024] (3) This AI-based predictive reinforcement learning-based charging pile cooling control method and system incorporates the cost of cooling resource usage into the objective function during the cooling strategy optimization process, constructing a control strategy that simultaneously considers temperature control accuracy and energy consumption. During the model training phase, energy consumption indicators are measured by the squared values ​​of normalized fan and liquid-cooled pump speeds, achieving flexible control of cooling resources. Under the premise of ensuring temperature control objectives, redundant energy consumption is minimized, improving the economic efficiency of the coordinated operation of the air-cooled and liquid-cooled systems, and reducing the operating cost of the entire charging thermal management system.

[0025] (4) The AI-based predictive reinforcement learning-based charging pile cooling control method and system, by establishing a thermal risk assessment mechanism and a temperature execution deviation feedback mechanism, can identify insufficient or overcooling trends and dynamically adjust the training direction of the control strategy, thereby reducing thermal fatigue caused by temperature stress on the equipment. The refined and dynamic adjustment capability of temperature control reduces mechanical losses and energy waste caused by high-frequency overcooling and undercooling operations, slows down the aging rate of liquid cooling pumps, fans and other actuators, and thus improves the overall reliability and long-term service life of the system. Attached Figure Description

[0026] Figure 1 This is a flowchart of a charging pile cooling control method based on AI predictive reinforcement learning.

[0027] Figure 2 This is a structural diagram of a charging pile cooling control system based on AI predictive reinforcement learning.

[0028] Figure 3 A diagram illustrating the effect of cooling execution deviation assessment values ​​and adjustment instructions;

[0029] Figure 4 This is a structural diagram of the charging pile cooling control system.

[0030] In the diagram, 1. Charging pile; 2. AI intelligent algorithm controller; 3. Coolant reservoir; 4. Condenser; 5. Liquid cooling pump; 6. Electric valve; 7. Sensor group. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Please see Figures 1-4 This invention provides a technical solution: a charging pile cooling control method and system based on AI predictive reinforcement learning, comprising: S1, collecting charging heat dissipation data of charging pile 1, and performing time alignment, anomaly removal, deburring and normalization processing on the charging heat dissipation data to obtain preprocessed charging heat dissipation data; S2, calling the preprocessed charging heat dissipation data to construct a time-series feature matrix, inputting it into a temperature prediction model, outputting a temperature prediction sequence of the controlled target, and assessing the thermal runaway risk in the prediction stage to construct a thermal risk identification sequence; S3, generating a state vector based on the thermal risk identification sequence and the temperature prediction sequence, generating an action vector by combining the control range of liquid cooling pump 5 and fan, constructing a cooling strategy optimization model, and performing reinforcement learning training on the cooling strategy optimization model, inputting the real-time state vector into the trained cooling strategy optimization model, outputting adjustment commands and issuing them for execution; S4, obtaining the temperature prediction sequence and thermal risk assessment value after command execution, assessing the execution deviation of the adjustment command, adjusting the cooling strategy based on the assessment results, and generating cooling strategy feedback samples to achieve adaptive closed-loop evolution of the cooling strategy.

[0033] Specifically, the charging heat dissipation data of charging pile 1 is collected, and time alignment, anomaly removal, de-glitching, and normalization processing are performed on the charging heat dissipation data to obtain the preprocessed charging heat dissipation data. The specific steps are as follows: The operation process of charging pile 1 is periodically sampled with a fixed-width sliding time window to obtain the charging heat dissipation data in real time. The charging heat dissipation data includes charging power, battery temperature, IGBT temperature, charging gun head temperature, coolant temperature, liquid cooling pump flow rate, liquid cooling pump speed, fan speed, ambient temperature, ambient humidity, and natural wind speed. The charging heat dissipation data is collected by sensor group 7, which includes temperature, flow rate, and power sensors. Then, the charging heat dissipation data is preprocessed by AI intelligent algorithm controller 2: First, time alignment is performed based on a high-precision timestamp synchronization method, with the main AI intelligent algorithm controller 2... Using the control system clock as a reference, the charging heat dissipation data is calibrated to eliminate the multi-channel data misalignment problem caused by differences in sampling frequencies and communication delays between sensors, ensuring that different signal channels are processed synchronously under a unified timeline. Subsequently, combining the battery thermal behavior model and historical charging operation data, a dynamic anomaly identification mechanism is constructed with temperature rise trend and sliding window volatility as joint criteria to remove pulse anomalies and discontinuous abrupt changes caused by external electromagnetic interference, sensor jitter, etc. After anomaly removal, a slowly varying smoothing function reflecting the thermal inertia and response hysteresis of the cooling system is further constructed to correct and smooth high-frequency glitches and hysteresis responses in the data, enhancing the physical consistency of the data. Finally, a linear normalization algorithm is used to normalize and compress the collected data of different physical quantity dimensions, mapping the charging heat dissipation data of different dimensions to a standardized range.

[0034] In this implementation plan, by collecting charging heat dissipation data in real time and performing time alignment, anomaly removal, de-glitching, and normalization on the charging heat dissipation data, the temporal consistency, stability, and input standardization of the original multi-source data are significantly improved. This ensures that the charging heat dissipation data has the characteristics of strong continuity, realistic response, and unified dimensions, providing accurate and reliable data input support for subsequent temperature prediction models and cooling strategy optimization models. This effectively improves the perception accuracy and decision-making quality of the entire cooling control method in dynamic thermal environments.

[0035] Specifically, the steps for constructing a time-series feature matrix from preprocessed charging heat dissipation data, inputting it into a temperature prediction model, and outputting a temperature prediction sequence for the controlled target are as follows: Based on the preprocessed charging heat dissipation data, a fixed time window is divided according to the sampling timestamp. The window length is limited to five to fifteen consecutive sampling periods, and the sliding step size is limited to one sampling period. Within each time window, charging power, coolant temperature, liquid cooling pump flow rate, liquid cooling pump speed, fan speed, ambient temperature, natural wind speed, battery temperature, charging gun head temperature, and IGBT temperature are written in chronological order to form a time-series feature matrix arranged by time steps. The time-series feature matrix is ​​then arranged in the sampling period index order to ensure the continuity and consistency of the input sequence. Using the time-series feature matrix as input, a two-layer stacked long short-term memory neural network is used to model the temporal dependencies of the charging process. The number of nodes in the first hidden layer is set to sixty-four, and the number of nodes in the second hidden layer is set to thirty-two. The full-time hidden state output by the second hidden layer is input into the time attention layer. The attention weights of each time step are calculated, and an attention convergence feature vector is generated. Based on the attention convergence feature... The vector model employs a four-way parallel output structure. Each path uses a fully connected mapping to output a future temperature prediction vector for the corresponding controlled target. The first path outputs a sequence of predicted battery temperatures, the second a sequence of predicted charging gun temperatures, the third a sequence of predicted IGBT temperatures, and the fourth a sequence of predicted coolant temperatures. All four outputs are arranged sequentially according to multiple future sampling time points to form the temperature prediction sequence. During the training phase, historical window samples are used as input, and the battery temperature, charging gun temperature, IGBT temperature, and coolant temperature at the corresponding future time points are used as supervision labels. The squared prediction errors of each of the four paths are calculated and summed item by item to obtain the training loss. After parameter iteration, the real-time feature matrix is ​​input into the trained temperature prediction model, outputting a sequence of predicted temperatures for the controlled target at multiple future time points. The controlled target includes the battery, charging gun, IGBT module, and coolant. The temperature prediction sequences output the predicted battery temperature, charging gun temperature, IGBT temperature, and coolant temperature, respectively, providing forward-looking input support for subsequent thermal risk assessment and cooling strategy optimization.

[0036] In this implementation scheme, by constructing a temporal feature matrix and combining it with a neural network with associative memory structure and an attention mechanism for training and prediction, the prediction accuracy and stability of the charging pile 1 for battery temperature, charging gun head temperature, IGBT temperature, and coolant temperature under complex thermal load environments are effectively improved. This achieves dynamic perception and accurate prediction of the thermal evolution trend during the charging process, providing continuous and high-quality predictive input for subsequent thermal risk identification and cooling strategy optimization, thus contributing to improved timeliness, accuracy, and intelligence of cooling control.

[0037] Specifically, the steps for assessing the thermal runaway risk during the prediction phase and constructing a thermal risk identification sequence are as follows: The difference between the predicted temperature value and the corresponding temperature threshold at each time point in the temperature prediction sequence is calculated, obtaining the differences between the predicted battery temperature and the battery temperature threshold, the predicted charging gun temperature and the charging gun temperature threshold, the predicted IGBT temperature and the IGBT temperature threshold, and the predicted coolant temperature and the coolant temperature threshold, respectively. These four types of temperature differences are squared and summed to obtain a comprehensive temperature risk accumulation value, which reflects the coupling degree of multi-target temperature exceedances. The accumulated temperature risk value is multiplied by the pre-processed charging power to construct a heat source input factor, used to quantitatively characterize the heat source. The driving effect of intensity on the system temperature rise trend is investigated. The liquid cooling pump flow rate data, fan speed data, and natural wind speed data are processed by adding one and then multiplying them sequentially to obtain the cooling flow capacity factor, which characterizes the cooling medium circulation capacity of the liquid cooling system under the current control state. The predicted coolant temperature minus the ambient temperature is multiplied by the heat dissipation response adjustment coefficient, and the negative value is used as the exponential part in a power function operation with the natural logarithm base e to construct a heat conduction adjustment factor, reflecting the heat dissipation sensitivity of the liquid cooling system to temperature difference changes. The heat dissipation response adjustment coefficient is calculated based on known historical charging power, coolant temperature, and ambient temperature time-series data, using the least squares method to fit the response relationship between coolant temperature rise and heat dissipation capacity. The heat conduction adjustment factor is multiplied by one and then multiplied by the cooling flow capacity factor to form a comprehensive cooling capacity factor, which is used to uniformly measure the overall performance of cooling flux and conduction efficiency under the current state. The heat source input factor is used as the numerator and the comprehensive cooling capacity factor is used as the denominator to perform a division operation to obtain the thermal risk assessment value at the current time point. The thermal risk assessment value is compared with the thermal risk threshold in real time, and all time points with thermal risk assessment values ​​higher than the thermal risk threshold are marked as thermal risk time points. They are arranged in chronological order according to the prediction time axis to construct a thermal risk identification sequence for cooling strategy response.

[0038] The specific formula for calculating the thermal risk assessment value is as follows:

[0039] ;

[0040] In the formula, Indicates the thermal risk assessment value. Indicates the battery temperature difference. This indicates the temperature difference at the nozzle tip. This indicates the temperature difference between the IGBTs. This indicates the temperature difference of the coolant. Indicates charging power. Indicates the flow rate of the liquid cooling pump. Indicates fan speed. Indicates natural wind speed. This indicates the predicted coolant temperature. Indicates ambient temperature. This represents the heat dissipation response adjustment coefficient.

[0041] In this implementation scheme, a thermal risk assessment calculation path based on the coupling of multi-objective temperature prediction sequences and charging power is constructed. This path comprehensively considers the differences between predicted battery temperature, charging gun temperature, IGBT temperature, and coolant temperature and their corresponding temperature thresholds. During the temperature rise risk quantification process, a cumulative temperature risk value and a heat source input factor are introduced, effectively improving the forward-looking identification capability of thermal runaway trends. Simultaneously, a cooling flow capacity factor is jointly constructed using liquid-cooled pump flow rate, fan speed, and natural wind speed. This factor, combined with the predicted coolant temperature and ambient temperature, calculates a heat conduction adjustment factor, forming a comprehensive cooling capacity factor oriented towards real-time changes in operating conditions. This enables dynamic calculation of thermal risk assessment values ​​and accurate identification of thermal risk time points. This method significantly improves the accuracy and response efficiency of thermal risk assessment, providing a stable and reliable data foundation for subsequent cooling strategy optimization.

[0042] Specifically, the steps for generating state vectors based on thermal risk identification sequences and temperature prediction sequences, and generating action vectors by combining the control ranges of liquid-cooled pump 5 and fan, are as follows: Based on the thermal risk identification sequence and the temperature prediction sequence of the controlled target, the charging power, coolant temperature prediction value, coolant temperature, liquid-cooled pump flow rate, liquid-cooled pump speed, fan speed, ambient temperature, and natural wind speed at corresponding time points are extracted and combined in the sampling order to form a state vector with temporal characteristics; the state vector is standardized and encoded in the AI ​​intelligent algorithm controller 2 and then used as the input of the cooling strategy optimization model to ensure the uniformity of the variables on the numerical scale and the continuity of the temporal arrangement; according to the adjustable range of liquid-cooled pump speed and fan speed, discrete classification of selectable target values ​​is set, and the classification rules are set in combination with the physical performance curves and response stability of liquid-cooled pump 5 and fan; different liquid-cooled pump speed adjustment commands and fan speed adjustment commands are arranged in combination to generate a set of control action combinations, which covers all feasible cooling strategies and is organized in vector form as action vectors for use as the basis for action selection in the reinforcement learning training stage of the subsequent cooling strategy optimization model.

[0043] In this implementation scheme, by constructing a state vector with temporal characteristics and an action vector covering the entire control range, a fine modeling of the heat dissipation control strategy of charging pile 1 between input features and output actions is achieved. This enables the liquid cooling pump speed, fan speed, and environmental thermal disturbances to be embedded in the cooling strategy optimization model in a unified structure, thereby improving the cooling control strategy's responsiveness to complex thermal coupling dynamics and the accuracy of multi-variable linkage adjustment. This provides a stable and efficient input-output data structure foundation for subsequent reinforcement learning-based cooling strategy training.

[0044] Specifically, a cooling strategy optimization model is constructed and trained using reinforcement learning. The real-time state vector is input into the trained cooling strategy optimization model, and adjustment commands are output and executed. The specific steps are as follows: The state vector, composed of sampled parameters such as charging power, predicted coolant temperature, coolant temperature, liquid pump flow rate, liquid pump speed, fan speed, ambient temperature, and natural wind speed, and the action vector, composed of liquid pump speed adjustment commands and fan speed adjustment commands, are used as input. A cooling strategy optimization model is constructed based on the Actor-Critic reinforcement learning architecture, and a near-end strategy optimization algorithm with advantages in sample efficiency and strategy stability is used to iteratively train the cooling strategy optimization model. During the training process, the difference between the predicted battery temperature and the current battery temperature is squared, the difference between the predicted IGBT temperature and the current IGBT temperature is squared, and the difference between the predicted charging gun temperature and the current charging gun temperature is squared. The temperature difference is squared, and the squared differences of the above three terms are added together to form the temperature control accuracy evaluation term of the cooling regulation process. Then, the target speed of the liquid cooling pump is divided by the maximum speed of the liquid cooling pump and the square is taken, and the target speed of the fan is divided by the maximum speed of the fan and the square is taken. The two results are added together to form the cooling resource usage term of the liquid cooling pump 5 and the fan. Then, the temperature control accuracy evaluation term and the cooling resource usage term are added together and the negative sign is taken to obtain the cooling control load index value characterizing the comprehensive regulation performance. The cooling control load index value is used as the reinforcement learning immediate reward corresponding to the current state and action combination and input to the policy optimization network to drive the policy iteration direction. After the model training is completed, the real-time state vector is input to the trained cooling policy optimization model, and the liquid cooling pump speed adjustment command and fan speed adjustment command are output in real time. They are sent to the execution end through the control interface module to regulate the operating state of the liquid cooling pump 5 and the fan, thereby realizing the closed-loop temperature regulation of the battery, IGBT and charging gun.

[0045] The specific formula for calculating the cooling control load index value is as follows:

[0046] ;

[0047] In the formula, This indicates the cooling control load index value. This represents the difference between the predicted battery temperature and the current battery temperature. This represents the difference between the predicted temperature of the charging gun head and the current temperature of the charging gun head. This represents the difference between the predicted IGBT temperature and the current IGBT temperature. Indicates the target fan speed. Indicates the target speed of the liquid cooling pump. This indicates the maximum speed of the liquid cooling pump. This indicates the maximum fan speed.

[0048] In this implementation scheme, a cooling strategy optimization model is constructed using state vectors and action vectors as inputs. An Actor-Critic reinforcement learning architecture and a near-end policy optimization algorithm are introduced to achieve adaptive evolution of the cooling strategy under complex thermal environments. During training, the differences between predicted battery temperature, IGBT temperature, and charging gun head temperature and their corresponding current temperatures are integrated to construct a temperature control accuracy evaluation term. This, combined with a cooling resource usage term normalized to the target speeds of the liquid-cooled pump and fan, forms a cooling control load index value, which serves as a reinforcement learning reward signal. This significantly improves the comprehensive balance between temperature control accuracy and energy consumption control of the cooling strategy optimization model. This method can output finely adjustable liquid-cooled pump speed adjustment commands and fan speed adjustment commands based on real-time state vectors, ensuring the temperature safety and cooling resource allocation efficiency of charging pile 1 under multi-source heat interference, effectively enhancing the system's intelligence, economy, and responsiveness.

[0049] Specifically, the specific steps for obtaining the temperature prediction sequence and thermal risk assessment value after command execution, and evaluating the execution deviation of the adjustment command, are as follows: After the adjustment command is executed, based on the real-time charging heat dissipation data collected by sensor group 7, the AI ​​intelligent algorithm controller 2 is invoked again to perform thermal risk assessment and temperature prediction on the operating status of charging pile 1, and the temperature prediction sequence and thermal risk assessment value after command execution are obtained; the battery temperature value after command execution is subtracted from the battery temperature prediction value output by the temperature prediction model before command execution, and the resulting temperature difference is divided by the battery temperature prediction value before command execution to calculate the battery temperature execution deviation; the IGBT temperature value after command execution is subtracted from the IGBT temperature prediction value generated by the temperature prediction model before command execution, and the difference is divided by the IGBT temperature prediction value before command execution to obtain the IGBT temperature execution deviation; the charging gun head temperature value after command execution is subtracted from the charging gun head temperature prediction value obtained based on time-series feature matrix inference before command execution, and the temperature difference is divided by the prediction value itself to calculate the battery temperature execution deviation. The execution deviation of the charging gun head temperature is obtained. The arithmetic mean of the execution deviations of the battery temperature, IGBT temperature, and charging gun head temperature is defined as the average temperature deviation term, which is used to characterize the degree of deviation of the temperature control of multiple heat sources. The thermal risk assessment value after instruction execution is subtracted from the thermal risk assessment value output by the thermal risk identification sequence before instruction execution, and the difference is divided by the sum of the thermal risk assessment value before instruction execution and the minimum term to obtain the normalized change of thermal risk. The minimum term is a small positive real constant that is not zero, which is used to avoid the problem of zero denominator when the thermal risk assessment value is zero, and to improve numerical stability and computational robustness. The normalized change is multiplied by the risk adjustment coefficient to obtain the thermal risk adjustment term, which is used to reflect the weighted impact of risk changes on the overall deviation assessment. The risk adjustment coefficient is obtained by using Bayesian optimization algorithm to train parameters through long-term labeled samples of historical thermal risk assessment value sequences and corresponding cooling execution deviation assessment values, and is used to quantify the degree of influence of thermal risk changes on the adjustment range of cooling strategy. Finally, the average temperature deviation term and the thermal risk adjustment term are added together to output the cooling execution deviation assessment value, which provides deviation feedback signal support for the adaptive training of the subsequent cooling strategy optimization model.

[0050] The specific formula for calculating the cooling performance deviation assessment value is as follows:

[0051] ;

[0052] In the formula, This indicates the evaluation value of cooling performance deviation. This indicates the battery temperature after the command is executed. This indicates the predicted battery temperature before the instruction is executed. This indicates the IGBT temperature after the instruction is executed. This indicates the predicted IGBT temperature before the instruction is executed. This indicates the nozzle temperature after the command is executed. This indicates the predicted temperature of the charging gun head before the command is executed. This indicates the thermal risk assessment value before the instruction is executed. This indicates the thermal risk assessment value after the command is executed. This represents the risk adjustment coefficient. Indicates a minus term.

[0053] In this embodiment, Table 1 is a data table of cooling execution deviation assessment values, which lists the key variables and cooling execution deviation assessment values ​​of 5 adjustment commands under the same operating environment. The key variables include: battery temperature, IGBT temperature, nozzle temperature and thermal risk assessment value before and after command execution. The specific data is as follows: In instruction 1, the current battery temperature is 42°C, the predicted battery temperature is 40°C, the current IGBT temperature is 75°C, the predicted IGBT temperature is 70°C, the current temperature of the cooling gun is 58°C, and the predicted temperature of the cooling gun is 55°C. The thermal risk assessment value before execution is 0.95, the thermal risk assessment value after execution is 1.00, the risk adjustment coefficient is 0.5, and the corresponding cooling execution deviation assessment value is 0.03. In instruction 2, the current battery temperature is 41°C, the predicted battery temperature is 39°C, the current IGBT temperature is 77°C, the predicted IGBT temperature is 72°C, the current temperature of the cooling gun is 57°C, and the predicted temperature of the cooling gun is 54°C. The thermal risk assessment value before execution is 1.10, the thermal risk assessment value after execution is 1.00, the risk adjustment coefficient is 0.5, and the corresponding cooling execution deviation assessment value is 0.11. In instruction 3, the current battery temperature is 45°C, the predicted battery temperature is 42°C, the current IGBT temperature is 80°C, the predicted IGBT temperature is 76°C, the current temperature of the cooling gun is 60°C, and the predicted temperature of the cooling gun is 76°C. The predicted temperature of the cooling gun is 56°C. The thermal risk assessment value before execution is 1.30, and the thermal risk assessment value after execution is 1.00. The risk adjustment coefficient is 0.5, and the corresponding cooling execution deviation assessment value is 0.21. In instruction 4, the current battery temperature is 43°C, the predicted battery temperature is 41°C, the current IGBT temperature is 78°C, the predicted IGBT temperature is 74°C, the current temperature of the cooling gun is 59°C, and the predicted temperature of the cooling gun is 55°C. The thermal risk assessment value before execution is 1.05, and the thermal risk assessment value after execution is 1.00. The risk adjustment coefficient is 0.5, and the corresponding cooling execution deviation assessment value is 0.08. In instruction 5, the current battery temperature is 44°C, the predicted battery temperature is 42°C, the current IGBT temperature is 79°C, the predicted IGBT temperature is 75°C, the current temperature of the cooling gun is 61°C, and the predicted temperature of the cooling gun is 56°C. The thermal risk assessment value before execution is 1.25, the thermal risk assessment value after execution is 1.00, the risk adjustment coefficient is 0.5, and the corresponding cooling execution deviation assessment value is 0.19.

[0054] Table 1. Cooling Execution Deviation Evaluation Values ​​Data Table

[0055]

[0056] like Figure 3 The figure shows the cooling execution deviation assessment values ​​and effect judgment results corresponding to five sets of adjustment commands, used to evaluate the execution deviation level and cooling adequacy of the cooling strategy under different control commands. The figure uses a color bar chart to represent the cooling execution deviation assessment value of each command, where green indicates an overcooling trend, blue indicates normal cooling effect, and red indicates a risk of undercooling, visually clearly distinguishing the cooling control performance of each adjustment command. Two dashed threshold lines are set in the figure as references for classifying cooling deviation levels: the orange dashed line represents the first-level cooling deviation threshold D1, used to determine whether there is an overcooling trend; the purple dashed line represents the second-level cooling deviation threshold D2, used to identify whether there is a risk of undercooling. As can be seen from the figure, the cooling execution deviation assessment values ​​of commands 1 and 4 are 0.03 and 0.08 respectively, both lower than D1, and are judged as overcooled; the deviation values ​​of commands 2 and 5 are 0.11 and 0.19 respectively, between D1 and D2, indicating reasonable cooling control and are judged as normal; while the deviation value of command 3 is 0.21, exceeding D2, and is already in the undercooling risk range. Figure 3 It effectively reflects the control error performance under different cooling strategies, providing intuitive data for the training feedback and strategy optimization of the cooling strategy optimization model.

[0057] In this implementation scheme, by dynamically calculating and multidimensionally decomposing the cooling execution deviation assessment value, and combining the fusion processing of the average temperature deviation term and the thermal risk adjustment term, an accurate assessment of the cooling control effect after command execution is achieved. This method can comprehensively reflect the execution deviation of the adjustment command in terms of battery temperature, IGBT temperature, and charging gun head temperature control, integrate the changing trend of the thermal risk assessment value, and introduce a risk adjustment coefficient to enhance the sensitivity to potential thermal runaway risks. This provides a quantitative feedback basis for the cooling strategy optimization model, improving the convergence efficiency and adaptability of the cooling strategy training.

[0058] Specifically, the steps for adjusting the cooling strategy based on the evaluation results are as follows: First, compare the cooling execution deviation evaluation value D with the first-level cooling deviation threshold D1 and the second-level cooling deviation threshold D2 in real time. Adjust the parameter update mechanism of the cooling strategy optimization model in the cooling control feedback path according to the deviation level. When the cooling execution deviation evaluation value D is less than or equal to the threshold D1, it is determined that the currently issued liquid cooling pump speed adjustment command, fan speed adjustment command, and electric valve 6 valve opening adjustment command have an overcooling trend. In the next round of cooling strategy training based on the Actor-Critic architecture, the expected reward value of the corresponding action vector in the current state is reduced by increasing the weight of the negative feedback gradient, while simultaneously controlling electric valve 6 to close the valve, reducing the actual flow capacity of the coolant. When the cooling execution... When the deviation evaluation value D is between the threshold D1 and the threshold D2, it is determined that the adjustment command is effective in the current cooling task, and the cooling control strategy has achieved a relatively good balance under the current state vector input conditions. Therefore, the structure and parameters of the cooling strategy optimization model are not adjusted, and the cooling control strategy remains unchanged. When the cooling execution deviation evaluation value D is greater than or equal to the threshold D2, it is determined that there is a risk of insufficient cooling in the current command. In the next round of cooling strategy training, the adjustment range of the target speed of the liquid cooling pump and the target speed of the fan under the corresponding state variables is increased to enhance the model's response to cooling capacity. At the same time, the cooling strategy refresh interval is shortened, the control update frequency is increased, and the valve opening is increased in the electric valve 6 control path to increase the coolant flow rate, thereby achieving a multi-parameter synergistic enhancement of the cooling regulation effect.

[0059] In this implementation scheme, by comparing the cooling execution deviation assessment value with the cooling deviation threshold in real time, and based on the feedback results of the cooling control load index, the training mechanism of the cooling strategy optimization model is dynamically adjusted. This effectively realizes the refined updating of three control commands—the target speed of the liquid cooling pump, the target speed of the fan, and the opening degree of the electric valve 6—under different deviation conditions, improving the adaptability of the cooling strategy to changes in the state of the controlled target. By applying negative feedback and closing the electric valve 6 under the overcooling trend, and increasing the adjustment amplitude and opening degree under the undercooling risk, precise allocation of cooling resources and risk-oriented dynamic response are achieved, improving the efficiency and stability of cooling control and providing a feasible parameter update path for the closed-loop optimization of the intelligent cooling system.

[0060] Specifically, the steps for generating cooling strategy feedback samples and realizing the adaptive closed-loop evolution of the cooling strategy are as follows: The charging heat dissipation data processed by the AI ​​intelligent algorithm controller 2, the state vector extracted from the corresponding time point in the thermal risk identification sequence, the action vector constructed by combining the target speed of the liquid cooling pump and the target speed of the fan, the temperature prediction sequence output by the temperature prediction model, and the cooling execution deviation evaluation value are systematically integrated to form a complete and variable-defined cooling strategy feedback sample. Based on the comparison results between the cooling execution deviation evaluation value D and the cooling deviation thresholds D1 and D2, each cooling strategy feedback sample is labeled as overcooled, normal, or undercooled to ensure that the training dataset has a clear performance orientation. The labeled cooling strategy feedback samples are uniformly incorporated into the training dataset of the cooling strategy optimization model to continuously update the strategy network parameters and strategy control boundaries in the cooling strategy optimization model, constructing a dynamic control feedback mechanism that runs through the entire charging heat dissipation process. This enables the adaptive closed-loop evolution of cooling control from temperature prediction, strategy decision-making, command execution, effect evaluation to strategy retraining, improving the self-learning ability and generalization stability of the cooling strategy.

[0061] In this implementation scheme, by structurally combining charging heat dissipation data, state vectors, action vectors, temperature prediction sequences, and cooling execution deviation evaluation values, a cooling strategy feedback sample containing complete control elements is formed. Based on the cooling execution deviation evaluation values, the feedback sample is accurately labeled, enabling continuous dynamic updates to the training dataset of the cooling strategy optimization model. This effectively constructs a closed-loop path for strategy optimization with cooling control performance as the core indicator, significantly enhancing the adaptability and control accuracy of the cooling strategy optimization model under different thermal load conditions, and improving the response robustness and energy-saving coordination of the cooling control system under complex operating conditions.

[0062] like Figure 2As shown, the second aspect of this invention provides a charging pile cooling control system based on AI predictive reinforcement learning, including: a heat dissipation data acquisition and preprocessing module, a temperature prediction risk analysis module, a cooling strategy optimization decision module, and a strategy feedback closed-loop training module. The heat dissipation data acquisition and preprocessing module is used to acquire charging heat dissipation data from charging pile 1, and perform time alignment, anomaly removal, deburring, and normalization processing on the charging heat dissipation data to obtain preprocessed charging heat dissipation data. The temperature prediction risk analysis module is used to construct a time-series feature matrix from the preprocessed charging heat dissipation data, input a temperature prediction model, output a temperature prediction sequence of the controlled target, and evaluate the heat dissipation during the prediction stage. To address runaway risk, a thermal risk identification sequence is constructed. A cooling strategy optimization decision module generates a state vector based on the thermal risk identification sequence and temperature prediction sequence, and generates an action vector by combining the control range of the liquid-cooled pump 5 and the fan. This module constructs a cooling strategy optimization model and trains it using reinforcement learning. The real-time state vector is input into the trained cooling strategy optimization model, which then outputs adjustment commands and issues them for execution. A strategy feedback closed-loop training module obtains the temperature prediction sequence and thermal risk assessment value after command execution, evaluates the execution deviation of the adjustment commands, adjusts the cooling strategy based on the evaluation results, and generates cooling strategy feedback samples to achieve adaptive closed-loop evolution of the cooling strategy.

[0063] like Figure 4 The diagram illustrates the structural components of a charging pile cooling control system. This system includes a charging pile 1, an AI intelligent algorithm controller 2, a coolant reservoir 3, a condenser 4, a liquid-cooled pump 5, an electric valve 6, and a sensor group 7, forming a closed-loop cooling circuit. The charging pile 1, as the controlled object, generates heat during charging. The AI ​​intelligent algorithm controller 2 receives charging heat dissipation data collected by the sensor group 7 and performs thermal risk prediction and cooling strategy optimization. The coolant reservoir 3 provides and recycles circulating coolant. The condenser 4, located in the cooling circuit, reduces the coolant temperature. The liquid-cooled pump 5 provides the driving force for coolant circulation. The electric valve 6 regulates the flow path of the coolant. The sensor group 7 includes temperature, flow, and power sensors to collect real-time operating data from the battery, IGBT module, charging gun, and coolant. Through this structural configuration, the system can achieve real-time perception and response control of the thermal load during charging, improving the intelligence and energy efficiency of the cooling system and effectively mitigating the risk of thermal runaway.

[0064] In this implementation plan, a charging pile cooling control system is constructed, which includes a heat dissipation data acquisition and preprocessing module, a temperature prediction and risk analysis module, a cooling strategy optimization and decision-making module, and a strategy feedback closed-loop training module. This system achieves high-quality acquisition and standardized processing of charging heat dissipation data, forward-looking prediction of temperature changes of the controlled target, reinforcement learning optimization and command output of the cooling strategy, and real-time evaluation and closed-loop feedback training of the cooling control effect. It constructs a complete intelligent control process from data-driven to strategy iteration, which significantly improves the dynamic adaptability and control accuracy of the cooling strategy and effectively ensures the thermal safety of charging pile 1 and the energy efficiency synergy of the cooling system under complex operating conditions.

[0065] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0066] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A charging pile cooling control method based on AI prediction reinforcement learning, characterized in that, The method comprises the following steps: S1, collecting charging heat dissipation data of the charging pile (1), and performing time alignment, abnormality elimination, deburring and normalization processing on the charging heat dissipation data to obtain preprocessed charging heat dissipation data; S2, calling the preprocessed charging heat dissipation data to construct a time sequence feature matrix, inputting the time sequence feature matrix into a temperature prediction model, outputting a temperature prediction sequence of the controlled target, and evaluating a thermal runaway risk in a prediction stage to construct a thermal risk identification sequence; S3, generating a state vector based on the thermal risk identification sequence and the temperature prediction sequence, generating an action vector in combination with a regulation range of the liquid cooling pump (5) and the fan, constructing a cooling strategy optimization model, and performing reinforcement learning training on the cooling strategy optimization model; inputting a real-time state vector into the trained cooling strategy optimization model, outputting an adjustment instruction, and executing the adjustment instruction; The specific steps of constructing the cooling strategy optimization model, performing reinforcement learning training on the cooling strategy optimization model, inputting the real-time state vector into the trained cooling strategy optimization model, outputting the adjustment instruction, and executing the adjustment instruction are as follows: The state vector and the action vector are taken as inputs, the cooling strategy optimization model is constructed based on an Actor-Critic reinforcement learning architecture, and a proximal policy optimization algorithm is used to train the cooling strategy optimization model; during the training process, the square of the difference between the battery temperature prediction value and the current battery temperature, the square of the difference between the IGBT temperature prediction value and the current IGBT temperature, and the square of the difference between the charging gun head temperature prediction value and the current charging gun head temperature are calculated, and the three square differences are added together to form a temperature control accuracy evaluation item of the cooling regulation process; the liquid cooling pump target speed is divided by the maximum liquid cooling pump speed, and the square of the result is taken, the fan target speed is divided by the maximum fan speed, and the square of the result is taken, and the two square values are added together to form a cooling resource usage item; the temperature control accuracy evaluation item and the cooling resource usage item are added together, and the negative of the result is taken to obtain a cooling control load index value; the cooling control load index value is taken as the reinforcement learning reward corresponding to the current state and the action; The real-time state vector is input into the trained cooling strategy optimization model to output liquid cooling pump speed adjustment instructions and fan speed adjustment instructions, which are executed by the instruction control end; S4, obtaining the temperature prediction sequence and the thermal risk evaluation value after executing the instructions, evaluating the execution deviation of the adjustment instructions, adjusting the cooling strategy based on the evaluation result, and generating a cooling strategy feedback sample to realize adaptive closed-loop evolution of the cooling strategy. 2.The AI prediction reinforcement learning-based charging pile cooling control method of claim 1, wherein: The specific steps of collecting the charging heat dissipation data of the charging pile (1) and performing time alignment, abnormality elimination, deburring and normalization processing on the charging heat dissipation data to obtain preprocessed charging heat dissipation data are as follows: The charging heat dissipation data of the charging pile (1) during the charging process is periodically collected by a fixed-width sliding time window, and the charging heat dissipation data includes charging power, battery temperature, IGBT temperature, charging gun head temperature, cooling liquid temperature, liquid cooling pump flow, liquid cooling pump speed, fan speed, ambient temperature, relative humidity, and natural wind speed; The high-precision timestamp synchronization method is used to time-align the charging heat dissipation data, so as to eliminate the data misalignment problem caused by the sampling frequency and communication delay among multiple sensors; a dynamic anomaly identification mechanism is constructed by combining the temperature rise trend and the sliding window volatility evaluation, so as to identify and eliminate the pulse interference and sudden jump in the charging heat dissipation data; a slow change smoothing function is constructed based on the empirical response time and cooling inertia parameters, so as to perform burr removal and response delay correction on the charging heat dissipation data; and a linear normalization algorithm is used to compress the features of the charging heat dissipation data, so as to map the input of different dimensions to a unified numerical interval. 3.The AI prediction reinforcement learning-based charging pile cooling control method of claim 1, wherein: The specific steps of calling the preprocessed charging heat dissipation data to construct a time sequence feature matrix, inputting the temperature prediction model, and outputting the temperature prediction sequence of the controlled target are as follows: Based on the preprocessed charging heat dissipation data, a time sequence feature matrix is constructed in a fixed time window manner, and the time sequence feature matrix is arranged according to the sampling period; the time sequence feature matrix is taken as the input, a neural network with an associative memory structure is used to model the time sequence dependence relationship in the charging process, a temperature prediction model is constructed, and an attention mechanism is introduced to train the model; The real-time feature matrix is input into the trained temperature prediction model, and the temperature prediction sequence of the controlled target at multiple future time points is output, the controlled target including a battery, a charging gun head, an IGBT, and a cooling liquid, and the temperature prediction sequence including a battery temperature prediction value, a charging gun head temperature prediction value, an IGBT temperature prediction value, and a cooling liquid temperature prediction value. 4.The AI prediction reinforcement learning-based charging pile cooling control method of claim 1, wherein: The specific steps of evaluating the thermal runaway risk in the prediction stage and constructing a thermal risk identification sequence are as follows: The temperature prediction values at each time point in the temperature prediction sequence are subtracted from the corresponding temperature thresholds to obtain the difference between the battery temperature prediction value and the battery temperature threshold, the difference between the charging gun head temperature prediction value and the charging gun head temperature threshold, the difference between the IGBT temperature prediction value and the IGBT temperature threshold, and the difference between the cooling liquid temperature prediction value and the cooling liquid temperature threshold, respectively; The above four types of temperature differences are squared and summed to obtain a comprehensive temperature risk accumulation value; The temperature risk accumulation value is multiplied by the preprocessed charging power to obtain a heat source input factor; the liquid cooling system flow rate, fan speed, and natural wind speed are multiplied in sequence to obtain a cooling flow capacity factor; the cooling liquid temperature prediction value is subtracted from the ambient temperature and multiplied by a heat dissipation response adjustment coefficient, and the negative value of the product is taken as the exponential part to perform a power function operation with the natural logarithm base e as the base to obtain a heat conduction adjustment factor; the heat conduction adjustment factor is multiplied by one to obtain a cooling capacity comprehensive factor; The heat source input factor is divided by the cooling capacity comprehensive factor to obtain a thermal risk evaluation value; The thermal risk evaluation value and the thermal risk threshold are compared in real time, time points at which the thermal risk evaluation value is greater than the thermal risk threshold are marked as thermal risk time points, and all the thermal risk time points are arranged in chronological order to form a thermal risk identification sequence. 5.The AI prediction reinforcement learning based charging pile cooling control method of claim 1, wherein: The specific steps of generating a state vector based on the thermal risk identification sequence and the temperature prediction sequence, and generating an action vector combining the control range of the liquid cooling pump (5) and the fan are as follows: Based on the thermal risk identification sequence and the temperature prediction sequence of the controlled target, the charging power, the cooling liquid temperature prediction value, the cooling liquid temperature, the liquid cooling pump flow, the liquid cooling pump speed, the fan speed, the ambient temperature and the natural wind speed at the corresponding time point are extracted, and combined in the sampling order to form a state vector with time sequence characteristics; According to the adjustable range of the liquid cooling pump speed and the fan speed, the discrete classification of the optional target value is set; different liquid cooling pump speed adjustment instructions and fan speed adjustment instructions are arranged in combination to generate a control action combination set, and organized in the form of a vector as an action vector. 6.The AI prediction reinforcement learning based charging pile cooling control method of claim 1, wherein: The specific steps of the temperature prediction sequence and the thermal risk evaluation value after the execution of the acquisition instruction are as follows: After the execution of the adjustment instruction, based on the real-time charging and cooling data, the thermal risk evaluation and temperature prediction are re-performed to obtain the temperature prediction sequence and the thermal risk evaluation value after the execution of the instruction; The battery temperature after the execution of the instruction is subtracted from the battery temperature prediction value before the execution of the instruction, and the difference is divided by the battery temperature prediction value before the execution of the instruction to obtain the battery temperature execution deviation; the IGBT temperature after the execution of the instruction is subtracted from the IGBT temperature prediction value before the execution of the instruction, and the difference is divided by the IGBT temperature prediction value before the execution of the instruction to obtain the IGBT temperature execution deviation; the charging gun head temperature after the execution of the instruction is subtracted from the charging gun head temperature prediction value before the execution of the instruction, and the difference is divided by the charging gun head temperature prediction value before the execution of the instruction to obtain the charging gun head temperature execution deviation; the battery temperature execution deviation, the IGBT temperature execution deviation and the charging gun head temperature execution deviation are added and averaged to obtain the temperature average deviation term; the thermal risk evaluation value after the execution of the instruction is subtracted from the thermal risk evaluation value before the execution of the instruction, and divided by the sum of the thermal risk evaluation value before the execution of the instruction and the minimum term to obtain the thermal risk normalized change; The thermal risk normalized change is multiplied by the risk adjustment coefficient to obtain the thermal risk adjustment term; The temperature average deviation term and the thermal risk adjustment term are added to obtain the cooling execution deviation evaluation value. 7.The charging pile cooling control method based on AI prediction reinforcement learning according to claim 1, characterized in that: The specific steps of adjusting the cooling strategy based on the evaluation result are as follows: Real-time comparison of the cooling execution deviation evaluation value D and the cooling deviation threshold D1 and D2 to adjust the cooling strategy: When D≤D1, it is determined that the adjustment instruction has an overcooling trend, and the negative feedback weight of the corresponding action vector is adjusted in the next round of cooling strategy training, and the electric valve (6) is controlled to close the valve; When D1<D<D2, it is determined that the adjustment instruction is effective, and the cooling strategy remains unchanged; When D≥D2, it is determined that the adjustment instruction has an undercooling risk, and the adjustment range of the liquid cooling pump target speed and the fan target speed of the corresponding state variable is increased in the next round of cooling strategy training, the cooling strategy refresh interval is shortened, and the valve opening of the electric valve (6) is increased. 8.The charging pile cooling control method based on AI prediction reinforcement learning according to claim 1, characterized in that: The specific steps of generating a cooling strategy feedback sample to realize adaptive closed-loop evolution of the cooling strategy are as follows: The charging heat dissipation data, the state vector, the action vector, the temperature prediction sequence and the cooling execution deviation evaluation value are combined to form a cooling strategy feedback sample, and according to the cooling execution deviation evaluation value, the cooling strategy feedback sample is labeled as over-cooling, normal and under-cooling labels, and is included in the training data set of the cooling strategy optimization model, so as to realize the whole process closed loop of cooling control from prediction, decision, execution, feedback to strategy retraining.

9. The charging pile cooling control system based on AI prediction reinforcement learning, used for executing the charging pile cooling control method based on AI prediction reinforcement learning in any one of claims 1-8, characterized in that: It comprises: a heat dissipation data acquisition and preprocessing module, a temperature prediction risk analysis module, a cooling strategy optimization and decision module, and a strategy feedback closed loop training module, wherein: The heat dissipation data acquisition and preprocessing module is used to acquire the charging heat dissipation data of the charging pile (1), and perform time alignment, abnormal elimination, deburring and normalization processing on the charging heat dissipation data to obtain preprocessed charging heat dissipation data; The temperature prediction risk analysis module is used to call the preprocessed charging heat dissipation data to construct a time sequence feature matrix, input a temperature prediction model, output a temperature prediction sequence of the controlled target, and evaluate the thermal runaway risk in the prediction stage to construct a thermal risk identification sequence; The cooling strategy optimization and decision module is used to generate a state vector based on the thermal risk identification sequence and the temperature prediction sequence, generate an action vector by combining the regulation and control range of the liquid cooling pump (5) and the fan, construct a cooling strategy optimization model, and perform reinforcement learning training on the cooling strategy optimization model; input the real-time state vector into the trained cooling strategy optimization model, output an adjustment instruction and execute it; The strategy feedback closed loop training module is used to obtain the temperature prediction sequence and the thermal risk evaluation value after the instruction execution, evaluate the execution deviation of the adjustment instruction, adjust the cooling strategy based on the evaluation result, generate a cooling strategy feedback sample, and realize the adaptive closed loop evolution of the cooling strategy.

Images