BBU charge and discharge state prediction and thermal management system based on AI

Abstract

The invention relates to the technical field of charge and discharge control, in particular to a BBU charge and discharge state prediction and thermal management system based on AI. The system comprises the following steps: acquiring basic data to form a feature vector; based on energy conservation, acquiring coupling heat release among the basic data; constructing a sliding window, predicting moments in a prediction interval through data in the sliding window to obtain predicted values of different data, and adjusting a loss function based on coupling heat release obtained by the predicted values to obtain predicted values of data types; determining a comprehensive risk based on the predicted value, and constructing a global risk assessment for each position; and allocating the total cooling resource of the battery unit based on the global risk assessment to obtain the cooling resource of each spatial point, and completing thermal management. The prediction precision is improved, and the thermal management efficiency is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of charge and discharge control, in particular to an AI-based BBU charge and discharge state prediction and thermal management system. BACKGROUND

[0002] A BBU (battery unit) is a core component for emergency power supply in key facilities such as data centers and communication base stations. Through the built-in lithium battery and intelligent management system, it can quickly switch power supply when the mains power is interrupted, ensuring the continuity of equipment operation. In the manufacturing of power transmission and control equipment such as ground alternating current charging piles and underground alternating current charging piles, the BBU serves as a power guarantee unit. The charge and discharge state of its battery directly affects the power supply reliability and service life. Real-time prediction of SOC and charge and discharge current direction can avoid overcharging and overdischarging, and optimize energy distribution. Heat is generated during the battery charging and discharging process, especially in densely deployed charging pile clusters. Local overheating may cause safety risks and accelerate battery aging.

[0003] In the traditional method of BBU charge and discharge prediction and thermal management, experience rules or simple linear models are often used. The charge and discharge state prediction is usually based on fixed-period voltage and current sampling, and the remaining capacity is estimated by preset threshold values, ignoring the time sequence correlation of load fluctuations. The thermal management adopts a static cooling strategy, such as allocating fixed cooling resources according to the rated power of the equipment, or passively starting cooling only when the current single-point temperature exceeds the threshold value, without considering the spatial distribution and dynamic conduction of heat accumulation. This not only leads to low accuracy of charge and discharge prediction, making the battery life shorter due to overcharging and overdischarging, but also causes thermal management lag, local hotspot temperature exceeding the safety threshold, increasing the risk of fire, and unbalanced allocation of cooling resources, resulting in overcooling of ground equipment and frequent triggering of protection shutdown due to insufficient cooling of underground equipment, which seriously affects the stability and economy of the power transmission and distribution system. SUMMARY

[0004] To solve the problems of low prediction accuracy and low thermal management efficiency, the application provides an AI-based BBU charge and discharge state prediction and thermal management system. The technical solution is as follows: The application provides an AI-based BBU charge and discharge state prediction and thermal management system. The system includes the following modules: A feature collection module collects charge, current and temperature at each time point to form a feature vector; A heat generation correlation module, based on the conservation of energy, considers the difference between heat release and heat absorption at each time point as the coupled heat release of charge, current and temperature; The prediction module constructs a sliding window for the feature vector, predicts the prediction value of each time point in the prediction interval based on the data in the sliding window through a neural network, and obtains the prediction value of each time point in the prediction interval for different data types, and then obtains the charge prediction value, the current prediction value and the temperature prediction value in the prediction interval based on the coupling heat release adjustment loss function of the prediction value. The risk assessment module determines the comprehensive risk based on the difference of the temperature prediction value, the current prediction value and the charge prediction value, and takes the sum of the comprehensive risk of each space point as the global risk assessment. The thermal management module allocates the total cooling resource of the battery unit based on the global risk assessment to obtain the cooling resource of each space point and complete the thermal management.

[0005] In the above scheme, the voltage, current, temperature and change rate of the heterogeneous signals are mapped into low-dimensional embedded features; based on the features, the GRU time sequence model is used to jointly predict the charge, current and temperature at the time point, and the physical correlation between charging and discharging and heat field is constrained and strengthened through coupling loss; finally, the prediction results are combined to integrate the thermal risk, power risk and charging and discharging intensity risk through risk measurement, and a soft allocation control law is used to realize predictive cooling resource optimization; effectively solving the problems of insufficient charging and discharging state and thermal field prediction accuracy, thermal management response lag and unbalanced cooling resource allocation in traditional methods; through the AI model, the dynamic coupling relationship between charging and discharging and the heat field is captured, and the prediction accuracy is improved; based on the future risk, the forward-looking cooling regulation is avoided, which not only reduces the life attenuation and safety hazards of the battery caused by overcharging and overdischarging, high temperature, but also improves the utilization efficiency of cooling resources, and guarantees the stability and economy of the BBU.

[0006] In one embodiment, the feature vector acquisition method is: Each type of data at each time point and the difference between adjacent time points of the data type form a 6-dimensional original feature vector; The original feature vector collected is reduced in dimension through an embedding layer, and the data types are normalized to obtain the feature vector at each time point.

[0007] In one embodiment, the expression of the coupling heat release is: , represents the heat dissipation coefficient, represents the heat dissipation area, represents the temperature at each time point, represents the ambient temperature, represents the capacitance, represents the temperature change rate, represents the coupling heat release of the charge, current and temperature.

[0008] In one embodiment, the prediction value of each time in the prediction interval is predicted by a neural network based on the data in the sliding window. The method for obtaining the prediction value of each time of different data types in the prediction interval is: The multi-layer GRU architecture is used to process the time series input. The first layer of GRU receives all feature vectors in the sliding window and outputs a hidden state sequence. The stochastic gradient descent optimizer is used to minimize the loss function between the predicted value and the true value. The prediction value of different data types is obtained by adjusting the hidden layer output using the weight matrix and bias generated by the network.

[0009] In one embodiment, for the prediction value of the charge, the charge intermediate value is generated by the charge weight matrix, the hidden layer output of the current time, and the charge bias. Then, the charge intermediate value is normalized by the normalization method to obtain the charge offset. The prediction value of the charge at the prediction time is obtained by adding the charge offset and the current charge.

[0010] In one embodiment, the prediction value of the current is obtained by the current weight matrix, the hidden layer output of the current time, and the current bias to generate the current intermediate value. Then, the current intermediate value is normalized by the normalization method to obtain the current offset. The prediction value of the current at the prediction time is obtained by adding the current offset and the current at the current time. The prediction value of the current is mapped to the interval [-1, 1].

[0011] In one embodiment, the prediction value of the temperature is obtained by dividing the battery cell into multiple small cubic regions, taking the center of each region as a spatial point, and encoding the spatial point to obtain the weight matrix and bias of each spatial point. The expression of the prediction value of the temperature of each spatial point is: , Tk,j represents the temperature weight matrix of the jth spatial point at the kth prediction time, ht represents the hidden layer output at the tth time, bj,k represents the bias of the jth spatial point at the kth prediction time, Tj,k represents the prediction value of the temperature of the jth spatial point at the kth prediction time.

[0012] In one embodiment, the method for obtaining the coupling heat loss adjustment loss function based on the prediction value is: The coupling loss is determined according to the prediction value of all the charges, the prediction value of the current, and the prediction value of the temperature in the prediction interval. Then, the joint loss is obtained by adding the original loss of the charge, the current, and the temperature. The expression of the coupling loss is: , represents the temperature change rate at the prediction time, represents the coupling heat release of the prediction value, represents the prediction value of the charge at the kth prediction time of the tth time, represents the prediction value of the current at the kth prediction time of the tth time, represents the prediction value of the temperature at the kth prediction time of the tth time of the jth spatial point, represents the number of spatial points, represents the length of the prediction interval, represents the coupling loss.

[0013] In one embodiment, the comprehensive risk is positively correlated with the temperature prediction value difference and the current prediction value absolute value respectively, and is negatively correlated with the charge prediction value.

[0014] In one embodiment, the method for obtaining the cooling resource of each spatial point by distributing the total cooling resource of the battery unit based on the global risk assessment is: , represents the sensitivity coefficient, represents the global risk assessment of the jth spatial point, represents the exponential function with the natural constant as the base, represents the total cooling resource of the battery unit, represents the number of spatial points, represents the cooling resource of the jth spatial point.

[0015] The beneficial effects of the present application are: The present application maps the heterogeneous signals of voltage, current, temperature and their change rates into low-dimensional embedding features; based on the features, the GRU time sequence model is used to jointly predict the charge, current and temperature at the prediction time, and at the same time, the physical correlation between charging and discharging and the thermal field is constrained and strengthened through the coupling loss; finally, the prediction results are combined, the thermal risk, the power risk and the charging and discharging intensity risk are integrated through the risk measurement, and the predictive cooling resource optimization is realized through the soft allocation control law; the problems of insufficient charging and discharging state and thermal field prediction accuracy, thermal management response lag and unbalanced cooling resource allocation in the traditional method are effectively solved; the dynamic coupling relationship between charging and discharging and the thermal field is captured through the AI model, and the prediction accuracy is improved; based on the forward-looking cooling regulation of future risk, local overheating and resource waste are avoided, the life attenuation and safety hazards of the battery caused by overcharging and overdischarging, high temperature are reduced, the cooling resource utilization efficiency is improved, and the stability and economy of the BBU are guaranteed. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the drawings needed in the description of the embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0017] Figure 1 The AI-based BBU charging and discharging state prediction and thermal management system flowchart provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the specific embodiments, structure, features and effects of the AI-based BBU charging and discharging state prediction and thermal management system according to the present application are described in detail as follows. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0020] Embodiment of AI-based BBU charging and discharging state prediction and thermal management system: The specific scheme of the AI-based BBU charging and discharging state prediction and thermal management system provided by the present application will be specifically described below with reference to the drawings.

[0021] Please refer to Figure 1 which shows the AI-based BBU charging and discharging state prediction and thermal management system flowchart provided by an embodiment of the present application. The system includes the following modules: The feature acquisition module is connected in parallel with the BBU (battery unit) charging and discharging circuit through a voltage sensor, detects the circuit voltage in real time, calculates the charge through the voltage, and outputs the time sequence charge sequence. The current sensor is connected in series in the BBU charging and discharging circuit, and the circuit current is detected in real time. The current is positive when charging and negative when discharging, and the time sequence current sequence is output. Temperature sensors are deployed at key heat generating positions of the BBU (such as cell surface, PCB power device or heat dissipation channel inlet), and local temperature is collected in real time. If it is single-point temperature measurement, it is the temperature of this point; if it is multi-point temperature measurement, the highest temperature point is selected, and the time sequence temperature sequence is output. The collected charge, current and temperature are recorded as data types.

[0022] For the above-mentioned time sequence collected, the change amount of adjacent time is calculated to reflect the dynamic trend of the signal. The original collected data is filtered to remove high-frequency noise or burst interference. The sampling time of charge, current and temperature is strictly synchronized through hardware triggering or timestamp calibration. For data beyond the range, interpolation method is used to ensure sequence continuity. In the embodiment, the filter used is Kalman filter, the sampling frequency of the sensor is 1 kHz, and the interpolation method uses linear interpolation.

[0023] Each type of data at each time and the difference between adjacent time data types form a 6-dimensional original feature vector, which is Q(t), I(t) and T(t) represent the charge, current and temperature at time t, respectively, Q(t-1), I(t-1) and T(t-1) represent the charge, current and temperature difference between time t and time t-1, respectively, and T represents the conversion of a row vector to a column vector.

[0024] The original feature vector collected is reduced in dimension by embedding layer, and is linearly transformed, wherein the learnable parameter weight matrix is initialized by Xavier or He, the bias vector is initialized as a zero vector, and the embedding dimension is 3-6 in the embodiment; then a ReLU activation function is used to introduce nonlinearity; finally, LayerNorm is used to normalize the feature dimension of each sample. Thus, the original feature vector at each time is reduced in dimension to obtain the feature vector at each time. Different data types in the feature vector have been normalized and unified in dimension.

[0025] Linear transformation captures the correlation between different physical quantities through learnable weight matrix, ReLU introduces nonlinearity to fit complex physical relationship, and LayerNorm eliminates the training instability problem caused by the dimension difference of input features. The embedding layer not only realizes data dimension reduction and feature compression, but also retains the physical meaning in the original signal, providing a standardized, low-redundancy and information-rich input representation for subsequent time series prediction models.

[0026] At this point, the feature vector at each time is obtained.

[0027] ​​​​​​The heat generation correlation module, in the BBU, the charging and discharging state and the thermal state have a strong coupling relationship, and large current charging and discharging can exacerbate battery heat generation, and temperature change can affect the electrochemical performance and life of the battery. The traditional method usually models the two independently, ignoring the dynamic interaction, resulting in insufficient prediction accuracy. In addition, the thermal response of the battery system has hysteresis (temperature change lags behind current change), and if thermal management is based only on the current state, it will lead to control lag and be difficult to cope with fast charging and discharging scenarios. Therefore, a joint model that can predict the future charging and discharging state and thermal field evolution at the same time is needed to provide a basis for proactive thermal management.

[0028] In the battery system, the charging and discharging current, charge and temperature have a strong physical coupling relationship, and the traditional multi-task learning only optimizes each prediction task independently, ignoring this internal relationship, resulting in prediction results that may violate physical laws. Based on the energy conservation law design, the deviation of the prediction result from the physical law is quantified. The core is to build a heat generation function, estimate the heat generation per unit time through historical data, and establish a constraint relationship with the temperature change rate. By minimizing the difference between the predicted temperature change and the theoretical heat generation based on current and charge, the model is forced to learn the coupled relationship that conforms to the physical law, enhancing the consistency of cross-variable prediction.

[0029] Therefore, based on energy conservation, the difference between heat release and heat absorption at each time is the coupled heat release of charge, current and temperature, i.e. the heat generation function, where the heat is the heat generated by the influence of charge and current on temperature.

[0030] The expression of the coupled heat release of charge, current and temperature is: , represents the heat dissipation coefficient, represents the heat dissipation area, represents the temperature at each time, represents the ambient temperature, represents the capacitance, represents the temperature change rate, represents the coupled heat release of charge, current and temperature.

[0031] At this point, the coupled heat release of charge, current and temperature is obtained.

[0032] The prediction module, for all the collected feature vectors, constructs a sliding window, and takes a continuous segment of a predetermined length as a feature input. For each sliding window, the goal is to output the state variables in the prediction interval, including the charge sequence, the current sequence and the temperature field sequence. In this embodiment, the predetermined length is the time constant of the system, and the value is 10; the length of the prediction interval is the adjustment period of the system, and the trend is 5.

[0033] The multi-layer GRU architecture is used to process the time series input. The first layer of GRU receives all feature vectors within the sliding window and outputs a hidden state sequence. To capture the dependence of different time scales, multiple layers of GRU (at least three layers) can be stacked, and the hidden state of each layer is passed to the next layer. The last value of the hidden state sequence is recorded as the hidden layer output at the current time t. During training, a stochastic gradient descent optimizer is used to minimize the loss function between the predicted value and the true value. In this embodiment, the gradient descent optimizer is the Adam optimizer. To prevent overfitting, a Dropout layer is added between the GRU layers, and L2 regularization is applied to constrain the model parameters.

[0034] For different data types, different data type prediction heads are constructed to predict each time in the prediction interval. The time in the prediction interval is referred to as the prediction time.

[0035] The charge prediction head is constructed. For each prediction time, the charge intermediate value is generated by the charge weight matrix, the hidden layer output at the current time, and the charge bias. Then, the charge intermediate value is normalized by a normalization method to obtain the charge offset. The charge offset and the charge at the current time are added to obtain the charge prediction value at the prediction time. When the network is first trained, the first charge weight matrix and the charge bias are randomly generated. The network generates new weight matrix and bias for iterative training. In this embodiment, the normalization method is the sigmoid function.

[0036] In this embodiment, the expressions of the charge intermediate value and the charge prediction value are: , represents the charge weight matrix at the kth prediction time, represents the hidden layer output at the tth time, represents the bias at the kth prediction time, represents the charge intermediate value at the kth prediction time.

[0037] , represents the charge at the tth time, represents the normalization function, represents the charge prediction value at the kth prediction time at the tth time.

[0038] The method of constructing the current prediction head is the same as that of the charge prediction head. The current intermediate value is generated by the current weight matrix, the hidden layer output at the current time, and the current bias, and then the current prediction value at the prediction time is obtained by combining the current at the current time. It is worth noting that, since the current has positive and negative directions, the tanh function is used to map the intermediate value to the interval [-1, 1].

[0039] In constructing the temperature prediction head, a space-time separated architecture is adopted, and for a battery cell, it is divided into multiple small cubic regions, and the center of each region is taken as a space point. The space points are encoded to obtain the weight matrix and bias of each space point, and then combined with the output of the current time hidden layer to obtain the predicted value of the temperature.

[0040] The expression of the predicted value of the temperature is: , represents the temperature weight matrix of the jth space point at the kth prediction time, represents the hidden layer output at the tth time, represents the bias of the jth space point at the kth prediction time, represents the predicted value of the temperature at the kth prediction time at the tth time.

[0041] Through the above steps, the charge prediction value, the current prediction value and the temperature prediction value of each space point at each time in the prediction interval are obtained; the coupling loss is determined according to the coupling heat release of all the charge prediction values, the current prediction values and the temperature prediction values in the prediction interval; and then the joint loss is obtained by adding the original loss of the charge, the current and the temperature.

[0042] The expression of the coupling loss is: , represents the temperature change rate at the prediction time, represents the coupling heat release of the charge prediction value, the current prediction value and the temperature prediction value, represents the number of space points, represents the length of the prediction interval, represents the coupling loss.

[0043] The coupling loss forces the model to learn consistent physical behavior by explicitly modeling physical constraints, improving the physical rationality of the prediction and the system-level collaborative optimization ability.

[0044] The original loss in the neural network is replaced by the joint loss, and the neural network is performed again to generate the charge prediction value, the current prediction value and the temperature prediction value of the prediction interval according to the prediction result.

[0045] The present application integrates the charge, the charge and discharge current and the temperature into a unified prediction framework through the GRU time sequence model, realizes the joint dynamic prediction of multiple physical fields, can not only output the time sequence prediction curve of the prediction interval, but also can capture the coupling effect of the charge and discharge behavior and the heat conduction.

[0046] At this point, the charge prediction value, the current prediction value and the temperature prediction value in the prediction interval are obtained.

[0047] The risk assessment module, the conventional battery thermal management system is usually based on the current temperature state passive cooling control, there is a response lag, prone to local hot spot accumulation and further affect the battery life and safety. And the battery charging and discharging process, temperature change and charge, there is a strong coupling relationship between the charging and discharging current, and the heat conduction itself has the space-time lag characteristics. Therefore, only rely on the current temperature feedback can not deal with the potential thermal risk in advance, need to combine the future charging and discharging state prediction for forward-looking thermal management.

[0048] The risk degree total has three modules, respectively, thermal risk, power risk and charging intensity risk.

[0049] Among them, the thermal risk is represented by the difference between the predicted values of different spatial points, the power risk is represented by the predicted value of the charge, and the charging intensity risk is represented by the predicted value of the current. Based on the three, the comprehensive risk is obtained.

[0050] The comprehensive risk is positively correlated with the difference between the predicted values of the temperature and the absolute value of the predicted value of the current, and is negatively correlated with the predicted value of the charge.

[0051] It should be noted that the positive correlation means that the increase of one variable leads to the increase of the other variable, the two variables change in the same direction, and when one variable changes from large to small or from small to large, the other variable also changes from large to small or from small to large. The specific relationship is determined by actual application, and the present application does not make special limitation.

[0052] It should be noted that the negative correlation means that the increase of one variable leads to the decrease of the other variable, the two variables change in opposite directions, and when one variable changes from large to small or from small to large, the other variable also changes from small to large or from large to small. The specific relationship is determined by actual application, and the present application does not make special limitation.

[0053] The expression of the comprehensive risk is: , represents the predicted value of the charge at the kth predicted moment at the tth moment, represents the predicted value of the temperature at the kth predicted moment at the tth moment, represents the predicted value of the current at the kth predicted moment at the tth moment, represents the temperature threshold value, which reflects the risk condition after exceeding the temperature threshold value, represents the comprehensive risk of the jth spatial point at the kth predicted moment at the tth moment, represents a normalization function. represents a positive value function, that is, x is greater than zero, x is taken, otherwise zero is taken.

[0054] When the temperature prediction value exceeds the temperature threshold, the greater the difference, the higher the thermal risk term value, meaning the greater the possibility of performance degradation, shortened life, and safety hazards due to high temperature at the location; if the predicted temperature does not exceed the threshold, the thermal risk term is 0, avoiding excessive allocation of cooling resources to normal temperature areas. When the predicted charge value is closer to 0, the risk of battery life attenuation due to excessive discharge is greater. When the absolute value of the predicted current value is greater, the current load is greater, and large current charging and discharging will exacerbate heat generation and accelerate battery aging, and the risk is greater.

[0055] The global risk assessment of each spatial point is obtained by adding all the comprehensive risks of each spatial point in the prediction interval. It not only retains the time lag characteristic of heat conduction, but also captures the spatial heterogeneity of risk. Compared with the traditional thermal management which only focuses on the current temperature peak, the present application captures the risk evolution trend through time weighting and realizes preventive control.

[0056] At this point, the global risk assessment of each spatial point is obtained.

[0057] The thermal management module has inherent errors based on AI-based charging and discharging state prediction, especially long-term prediction, and the traditional hard allocation strategy will amplify the influence of prediction errors. Soft allocation reduces the dependence on single-point prediction accuracy through exponential smoothing, making the system more robust to local prediction deviations. The traditional thermal management system only focuses on the current high temperature point, while the present application needs to deal with potential hot spots in the future. Soft allocation automatically directs resources to areas with low current risk but potential future deterioration through risk negative exponential mapping, achieving preventive cooling.

[0058] When a certain area is already in a serious over-temperature state, continuing to invest in cooling resources may have limited effect, and more attention should be paid to manual processing. Soft allocation automatically reduces the resource allocation of such areas and transfers the capacity to more valuable preventive cooling, improving the overall energy efficiency ratio.

[0059] The total cooling resource of the battery unit is allocated through global risk assessment, and different cooling resources are allocated to different spatial points.

[0060] The expression of the cooling resource is: , represents the sensitivity coefficient, represents the global risk assessment of the jth spatial point, represents the exponential function with the natural constant as the base, represents the total cooling resource of the battery unit, represents the number of spatial points, The cooling resource of the jth spatial point is represented. The smaller the sensitivity coefficient is, the smoother the allocation is, the anti-interference ability is enhanced, but the response to the real risk is lagged; on the contrary, the response speed to the predicted risk is accelerated, but the resource may fluctuate sharply due to local prediction error. In the embodiment, the value is 5.

[0061] Based on this, the cooling resource at the current time is obtained. The cooling resource at each time is obtained by the above formula, the low-risk but possibly fast-warming area is preferentially protected, and the cooling investment in the area already in high risk is appropriately reduced. When the global risk assessment of the spatial point is greater than the preset overheating threshold, the overheating state may be irreversible, at which time an alarm is given to the artificial processing. In the embodiment, the overheating threshold is 0.7. This design not only meets the total resource constraint, but also smoothes the influence of prediction error through probabilistic allocation, and achieves a dynamic balance between avoiding thermal runaway risk and reducing cooling energy consumption, which embodies the core advantage of predictive thermal management.

[0062] It should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features thereof; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

[0063] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments.

Claims

1. An AI-based BBU state of charge prediction and thermal management system, characterized in that, The system comprises the following modules: a feature acquisition module, which acquires the charge, current and temperature at each time point, and forms a feature vector therefrom; a heat generation correlation module, which, based on the law of conservation of energy, takes the difference between heat release and heat absorption at each time point as the coupled heat release of the charge, current and temperature; a prediction module, which constructs a sliding window for the feature vector, and, based on the data in the sliding window, predicts the prediction value at each time point in the prediction interval by means of a neural network, to obtain the prediction value of each type of data at each time point in the prediction interval, and then adjusts the loss function based on the coupled heat release obtained from the prediction value, to obtain the charge prediction value, current prediction value and temperature prediction value in the prediction interval; a risk assessment module, which determines the comprehensive risk based on the difference in the temperature prediction value, the current prediction value and the charge prediction value; the sum of the comprehensive risks of each spatial point is taken as the global risk assessment; a thermal management module, which, based on the global risk assessment, allocates the total cooling resources of the battery unit to obtain the cooling resources of each spatial point, to complete thermal management.

2. The AI-based BBU charge-discharge state prediction and thermal management system of claim 1, wherein, The method for obtaining the feature vector is: each type of data at each time point and the difference between adjacent time points are taken to form a 6-dimensional original feature vector; the original feature vector acquired is reduced in dimension by means of an embedding layer, and the data types are normalized to obtain the feature vector at each time point.

3. The AI-based BBU charge-discharge state prediction and thermal management system of claim 1, wherein, The expression of the coupled heat release is: , represents a heat dissipation coefficient, represents a heat dissipation area, represents a temperature at each time, represents an ambient temperature, represents a capacitance, represents a temperature change rate, represents a coupling heat release of the three of charge, current, and temperature.

4. The AI-based BBU state of charge prediction and thermal management system of claim 1, wherein, The method for predicting the prediction value at each time point in the prediction interval by means of a neural network based on the data in the sliding window is: a multi-layer GRU architecture is used to process the time series input; the first layer of GRU receives all the feature vectors in the sliding window and outputs a hidden state sequence; a stochastic gradient descent optimizer is used to minimize the loss function between the prediction value and the true value; the prediction value of each type of data is obtained by adjusting the hidden layer output using the weight matrix and bias generated by the network.

5. The AI-based BBU state of charge prediction and thermal management system of claim 4, wherein, For the prediction value of the charge, a charge intermediate value is generated by means of the charge weight matrix, the hidden layer output at the current time point and the charge bias, and then the charge intermediate value is normalized by means of a normalization method to obtain a charge offset, and the charge offset and the current charge are added to obtain the prediction value of the charge at the prediction time point.

6. The AI-based BBU state of charge prediction and thermal management system of claim 5, wherein, For the prediction value of the current, a current intermediate value is generated by means of the current weight matrix, the hidden layer output at the current time point and the current bias, and then the current intermediate value is normalized by means of a normalization method to obtain a current offset, and the current offset and the current at the current time point are added to obtain the prediction value of the current at the prediction time point; the prediction value of the current is mapped to the interval [-1, 1].

7. The AI-based BBU state of charge prediction and thermal management system of claim 4, wherein, The prediction value of the temperature is obtained for each spatial point by dividing the battery unit into a plurality of small cubic regions, taking the center of each region as a spatial point, encoding the spatial point and obtaining the weight matrix and bias of each spatial point. The expression of the prediction value of the temperature of each spatial point is: , represents the temperature weight matrix of the jth spatial point at the kth prediction time, represents the hidden layer output at the tth time, represents the bias of the jth spatial point at the kth prediction time, represents the predicted value of the temperature of the jth spatial point at the kth prediction time at the tth time.

8. The AI-based BBU state of charge prediction and thermal management system of claim 1, wherein, The method for adjusting the loss function based on the coupled heat release obtained from the prediction value is: the coupled loss is determined according to the coupled heat release of all the charge prediction values, current prediction values and temperature prediction values in the prediction interval; and then the joint loss is obtained by adding the original loss of the charge, current and temperature; The expression of the coupling loss is: , represents a temperature rate of change at a prediction time, represents a coupled heat release of a prediction value, represents a prediction value of a charge at a kth prediction time at a tth time, represents a prediction value of a current at a kth prediction time at a tth time, represents a prediction value of a temperature at a kth prediction time at a tth time at a jth spatial point, represents a number of spatial points, represents a length of a prediction interval, represents a coupling loss.

9. The AI-based BBU state of charge prediction and thermal management system of claim 1, wherein, The comprehensive risk is positively correlated with the temperature prediction value difference and the current prediction value absolute value, and is negatively correlated with the charge prediction value.

10. The AI-based BBU state of charge prediction and thermal management system of claim 1, wherein, The method for obtaining the cooling resource of each space point by distributing the total cooling resource of the battery unit based on the global risk assessment is: , denotes a sensitivity coefficient, denotes a global risk assessment of the jth spatial point, denotes an exponential function with a natural constant as base, denotes the total cooling resource of the battery cell, denotes the number of spatial points, denotes the cooling resource of the jth spatial point.

Images