AI-based bbu charge-discharge state prediction and thermal management system

Abstract

The application 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: collecting basic data to form a feature vector; obtaining coupling heat release between the basic data based on energy conservation; constructing a sliding window, predicting the time in the prediction interval through the data in the sliding window to obtain the predicted values of different data, and then adjusting the loss function of the coupling heat release obtained based on the predicted values to obtain the predicted values of the data types; determining the comprehensive risk based on the predicted values, constructing a global risk assessment for each position; and distributing the total cooling resources of the battery units based on the global risk assessment to obtain the cooling resources of each space point, thereby completing thermal management. The application improves the prediction accuracy and the thermal management efficiency.

Description

Technical Field

[0001] This application relates to the field of charge and discharge control technology, specifically to an AI-based BBU charge and discharge state prediction and thermal management system. Background Technology

[0002] Battery Units (BBUs) are core components for emergency power supply in critical facilities such as data centers and communication base stations. Through built-in lithium batteries and intelligent management systems, they quickly switch power sources during mains power outages, ensuring continuous equipment operation. In the manufacturing of power transmission, distribution, and control equipment such as ground-based and underground AC charging piles, the BBU, as a power backup unit, directly affects power supply reliability and lifespan due to the charging and discharging state of its batteries. Real-time prediction of State of Charge (SOC) and charging / discharging current direction can prevent overcharging and over-discharging, optimizing energy distribution. Heat is generated during battery charging and discharging, especially in densely deployed charging pile clusters; localized overheating can pose safety risks and accelerate battery aging.

[0003] Traditional methods for BBU charge / discharge prediction and thermal management often rely on empirical rules or simple linear models. Charge / discharge state prediction is typically based on fixed-period voltage and current sampling, estimating remaining capacity through preset thresholds, ignoring the temporal correlation of load fluctuations. Thermal management employs static cooling strategies, such as allocating fixed heat dissipation resources according to the equipment's rated power, or passively initiating cooling only when the temperature at a single point exceeds a threshold, without considering the spatial distribution and dynamic conduction of heat accumulation. This not only leads to low accuracy in charge / discharge prediction, making overcharging and over-discharging prone to shortening battery life, but also results in lagging thermal management, local hotspot temperatures exceeding safety thresholds, increasing fire risk, and an imbalance in cooling resource allocation. Ground-based equipment may overheat and waste energy, while underground equipment may frequently trigger protection shutdowns due to insufficient cooling, severely impacting the stability and economy of the power transmission and distribution system. Summary of the Invention

[0004] To address the technical problems of low prediction accuracy and low thermal management efficiency, this application provides an AI-based BBU charge / discharge state prediction and thermal management system, the specific technical solution of which is as follows:

[0005] This application proposes an AI-based BBU charge / discharge state prediction and thermal management system, which includes the following modules:

[0006] The feature acquisition module collects the charge, current, and temperature at each moment and uses them to construct a feature vector;

[0007] The heat generation correlation module, based on the law of conservation of energy, treats the difference between heat release and heat absorption at each moment as the coupled heat release of charge, current, and temperature.

[0008] The prediction module constructs a sliding window for the feature vector. Based on the data within the sliding window, it uses a neural network to predict the predicted value for each time step in the prediction interval, obtaining the predicted values ​​for different data types at each time step in the prediction interval. Then, based on the predicted values, it adjusts the loss function using the coupled exothermic adjustment function to obtain the predicted values ​​for charge, current, and temperature in the prediction interval.

[0009] The risk assessment module determines the overall risk based on the differences in temperature, current, and charge predictions; the sum of the overall risks at each spatial point is used as the global risk assessment.

[0010] The thermal management module allocates the total cooling resources of the battery cells based on a global risk assessment, acquires the cooling resources for each spatial point, and completes thermal management.

[0011] In the above scheme, this application maps heterogeneous signals of voltage, current, temperature and their rate of change into low-dimensional embedded features. Based on these features, the GRU time series model is used to jointly predict the charge, current and temperature at the time of the prediction, while the physical correlation between charging / discharging and the thermal field is strengthened through coupling loss constraints. Finally, the prediction results are combined with the risk measurement to integrate thermal risk, charge risk and charge / discharge intensity risk, and then a soft allocation control law is used to achieve predictive cooling resource optimization. This effectively solves the problems of insufficient prediction accuracy of charge / discharge state and thermal field, lagging thermal management response and unbalanced cooling resource allocation in traditional methods. The dynamic coupling relationship between charging / discharging and thermal field is captured by the AI ​​model, which improves the prediction accuracy. The forward-looking cooling regulation based on future risks avoids local overheating and resource waste, which reduces the battery life decay and safety hazards caused by overcharging, over-discharging and high temperature, improves the utilization efficiency of cooling resources, and ensures the stability and economy of the BBU.

[0012] In one embodiment, the method for obtaining the feature vector is as follows:

[0013] Each data type at each time step and the differences in data types between adjacent time steps form a 6-dimensional original feature vector;

[0014] The original feature vectors collected are reduced in dimensionality by using an embedding layer, and the data types are normalized to obtain the feature vectors at each time step.

[0015] In one embodiment, the expression for the coupled heat release is:

[0016] , Indicates the heat dissipation coefficient. Indicates the heat dissipation area. This represents the temperature at each moment. Indicates ambient temperature. Indicates capacitance. Indicates the rate of temperature change. This represents the heat release resulting from the coupling of charge, current, and temperature.

[0017] In one embodiment, the method for obtaining the predicted values ​​of different data types at each time step within the prediction interval by using a neural network to predict the predicted values ​​based on the data within the sliding window is as follows:

[0018] A multi-layer GRU architecture is used to process temporal inputs; the first-layer GRU receives all feature vectors within a sliding window and outputs a sequence of hidden states; a stochastic gradient descent optimizer is used to minimize the loss function between the predicted and the true values.

[0019] Predicted values ​​for different data types are obtained by adjusting the hidden layer output using the weight matrix and bias generated by the network.

[0020] In one embodiment, for the predicted value of charge, an intermediate charge value is generated by the charge weight matrix, the hidden layer output at the current time, and the charge bias. The intermediate charge value is then normalized using a normalization method and used as the charge offset. The charge offset is added to the charge at the current time to obtain the predicted value of charge at the prediction time.

[0021] In one embodiment, the predicted current value is generated by using the current weight matrix, the hidden layer output at the current moment, and the current bias to generate an intermediate current value. Then, the intermediate current value is normalized using a normalization method and used as the current offset. The current offset is added to the current at the current moment to obtain the predicted current value at the prediction moment. The predicted current value is then mapped to the interval [-1, 1].

[0022] In one embodiment, the temperature prediction is for each spatial point. The battery cell is divided into multiple small cubic regions, with the center of each region serving as a spatial point. The spatial points are encoded to obtain the weight matrix and bias for each spatial point.

[0023] The expression for the predicted temperature at each spatial point is:

[0024] , This represents the temperature weight matrix for the j-th spatial point at the k-th prediction time. This represents the output of the hidden layer at time t. This represents the bias of the j-th spatial point at the k-th prediction time. This represents the predicted temperature of the j-th spatial point at time t at the k-th predicted time.

[0025] In one embodiment, the method for adjusting the coupling exothermic adjustment loss function based on the predicted value is as follows:

[0026] The coupling loss is determined by the coupled heat release of all predicted values ​​of charge, current, and temperature within the prediction interval; then, it is added to the original losses of charge, current, and temperature to obtain the combined loss.

[0027] The expression for coupling loss is:

[0028] , This represents the rate of temperature change at the predicted time. This indicates the coupled exothermic effect of the predicted values. Let represent the predicted value of the charge at time t and prediction time k. This represents the predicted value of the current at time t and prediction time k. This represents the predicted temperature of the j-th spatial point at time t at the k-th prediction time. Indicates the number of points in space. Indicates the length of the prediction interval. This represents the coupling loss.

[0029] In one embodiment, the overall risk is positively correlated with the difference between the individual temperature predictions and the absolute value of the current predictions, and negatively correlated with the charge predictions.

[0030] In one embodiment, the method for allocating the total cooling resources of the battery cells based on a global risk assessment to obtain the cooling resources for each spatial point is as follows:

[0031] , Represents the sensitivity coefficient. This represents the global risk assessment of the j-th spatial point. This represents an exponential function with the natural constant as its base. This indicates the total cooling resources of the battery cell. Indicates the number of points in space. This represents the cooling resources at the j-th spatial point.

[0032] The beneficial effects of this application are as follows:

[0033] This application maps heterogeneous signals of voltage, current, temperature, and their rates of change into low-dimensional embedded features. Based on these features, a GRU time-series model is used to jointly predict the charge, current, and temperature at a given moment. Simultaneously, coupling loss constraints are used to strengthen the physical correlation between charging / discharging and the thermal field. Finally, combining the prediction results, thermal risk, charge risk, and charge / discharge intensity risk are integrated through risk measurement, and a soft allocation control law is used to achieve predictive cooling resource optimization. This effectively solves the problems of insufficient accuracy in predicting charge / discharge states and thermal fields, delayed thermal management response, and unbalanced cooling resource allocation in traditional methods. By jointly capturing the dynamic coupling relationship between charging / discharging and the thermal field using an AI model, prediction accuracy is improved. Proactive cooling regulation based on future risks avoids localized overheating and resource waste, reducing battery life degradation and safety hazards caused by overcharging, over-discharging, and high temperatures, while also improving cooling resource utilization efficiency and ensuring the stability and economy of the BBU. Attached Figure Description

[0034] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart of an AI-based BBU charge / discharge state prediction and thermal management system provided in one embodiment of this application. Detailed Implementation

[0036] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the AI-based BBU charge / discharge state prediction and thermal management system proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0037] 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 this application pertains.

[0038] Example of an AI-based BBU charge / discharge state prediction and thermal management system:

[0039] The following description, in conjunction with the accompanying drawings, details the specific solution of the AI-based BBU charge / discharge state prediction and thermal management system provided in this application.

[0040] Please see Figure 1 The diagram illustrates a flowchart of an AI-based BBU charge / discharge state prediction and thermal management system according to an embodiment of this application. The system includes the following modules:

[0041] The feature acquisition module uses a voltage sensor connected in parallel to the charging and discharging circuit of the BBU (battery unit) to monitor the circuit voltage in real time, calculate the charge based on the voltage, and output a time-series charge sequence. A current sensor is connected in series to the BBU's charging and discharging circuit to monitor the circuit current in real time. The current is defined as positive during charging and negative during discharging, and a time-series current sequence is output. Temperature sensors are deployed at key heat-generating locations in the BBU (such as the cell surface, PCB power devices, or heat dissipation channel inlets) to collect local temperature data in real time. For single-point temperature measurement, the temperature at that point is recorded; for multi-point temperature measurement, the highest temperature point is selected, and a time-series temperature sequence is output. The collected charge, current, and temperature are recorded as data types.

[0042] For the time series acquired above, the changes between adjacent time points are calculated to reflect the dynamic trend of the signal. The raw acquired data is filtered to remove high-frequency noise or sudden interference. Strict synchronization of the sampling times for charge, current, and temperature is ensured through hardware triggering or timestamp calibration. For data exceeding... For the data within the range, interpolation is used to ensure sequence continuity. In this embodiment, Kalman filtering is used; the sensor sampling frequency is 1kHz; and linear interpolation is used.

[0043] Each data type at each time step and the differences in data types between adjacent time steps are used to construct a 6-dimensional original feature vector. This original feature vector is... ,in , , Let represent the charge, current, and temperature at time t, respectively. , and Let T represent the differences in charge, current, and temperature between time t and time t-1, respectively, and T denotes the transpose of the row vector into a column vector.

[0044] The original feature vectors are reduced in dimensionality through an embedding layer. A linear transformation is performed, where the learnable parameter weight matrix is ​​initialized using Xavier or He, the bias vector is initialized to zero, and the embedding dimension is set to 3-6 in this embodiment. Next, the ReLU activation function is used to introduce non-linearity. Finally, LayerNorm is used to normalize the feature dimension of each sample. This reduces the dimensionality of the original feature vector at each time step to obtain the feature vector at that time step. Different data types in the feature vector have been normalized to a unified dimension.

[0045] Linear transformation captures the correlation between different physical quantities through a learnable weight matrix, ReLU introduces nonlinearity to fit complex physical relationships, and LayerNorm eliminates the training instability problem caused by differences in the dimensions of input features. This embedding layer achieves both data dimensionality reduction and feature compression while preserving the physical meaning of the original signal, providing a standardized, low-redundancy, and information-rich input representation for subsequent time series prediction models.

[0046] At this point, the feature vector for each time step has been obtained.

[0047] The thermal generation correlation module in the Battery Burner (BBU) exhibits a strong coupling relationship between charge / discharge states and thermal states. High-current charge / discharge exacerbates battery heat generation, while temperature changes affect the battery's electrochemical performance and lifespan. Traditional methods typically model these two aspects independently, neglecting this dynamic interaction and resulting in insufficient prediction accuracy. Furthermore, the thermal response of the battery system exhibits hysteresis (temperature changes lag behind current changes). If thermal management is based solely on the current state, control lag will occur, making it difficult to cope with rapid charge / discharge scenarios. Therefore, a joint model capable of simultaneously predicting future charge / discharge states and thermal field evolution is needed to provide a basis for proactive thermal management.

[0048] In battery systems, charging and discharging current, charge, and temperature exhibit a strong physical coupling. Traditional multi-task learning optimizes each prediction task independently, ignoring this inherent correlation, which can lead to predictions that violate physical laws. This paper addresses this issue by designing a system based on the law of conservation of energy to quantify the deviation between predictions and physical laws. The core of this approach is to construct a heat generation function, estimating the heat generated per unit time using historical data and establishing a constraint relationship with the rate of temperature change. By minimizing the difference between the predicted temperature change and the theoretical heat generation calculated based on current and charge, the model is forced to learn a coupling relationship that conforms to physical laws, enhancing the consistency of cross-variable predictions.

[0049] Therefore, based on the law of conservation of energy, the difference between heat release and heat absorption at each moment is taken as 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.

[0050] The expression for the heat release through the coupling of charge, current, and temperature is:

[0051] , Indicates the heat dissipation coefficient. Indicates the heat dissipation area. This represents the temperature at each moment. Indicates ambient temperature. Indicates capacitance. Indicates the rate of temperature change. This represents the heat release resulting from the coupling of charge, current, and temperature.

[0052] Thus, the coupled exothermic reaction of charge, current, and temperature was obtained.

[0053] The prediction module constructs a sliding window for all collected feature vectors, extracting continuous segments of a preset length as feature inputs. The objective of each sliding window is to output the state variables of the prediction interval, including charge, current, and temperature field sequences. In this embodiment, the preset length is the system's time constant, with a value of 10; the length of the prediction interval is the system's adjustment period, with a trend of 5.

[0054] A multi-layer GRU architecture is used to process temporal inputs. The first-layer GRU receives all feature vectors within a sliding window and outputs a sequence of hidden states. To capture dependencies at different time scales, multiple GRU layers (at least three layers) can be stacked, with the hidden states of each layer passed to the next. 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 and true values; in this embodiment, the gradient descent optimizer is the Adam optimizer. To prevent overfitting, Dropout layers are added between the GRU layers, and L2 regularization is applied to constrain the model parameters.

[0055] For different data types, different prediction heads are constructed respectively. Predictions are then made for each time point within the prediction interval, and these time points are denoted as prediction times.

[0056] A charge prediction head is constructed. For each prediction time step, an intermediate charge value is generated using the charge weight matrix, the hidden layer output at the current time step, and the charge bias. This intermediate charge value is then normalized using a normalization method and used as the charge bias. The charge bias is added to the charge at the current time step to obtain the predicted charge value for that prediction time step. Initially, the charge weight matrix and charge bias are randomly generated during network training. Subsequent training iterates using new weight matrices and biases. In this embodiment, the normalization method is the sigmoid function.

[0057] In this embodiment, the expressions for the intermediate charge value and the predicted charge value are:

[0058] , This represents the charge weight matrix at the k-th prediction time. This represents the output of the hidden layer at time t. This represents the bias at the k-th prediction time. This represents the intermediate charge value at the k-th prediction time.

[0059] , Let represent the charge at time t. Represents the normalization function. This represents the predicted value of the charge at time t and the predicted value at time k.

[0060] The method for constructing the current prediction head is the same as that for the charge prediction head. After generating an intermediate current value using the current weight matrix, the output of the hidden layer at the current moment, and the current bias, the predicted current value at the prediction moment is obtained by combining it with the current at the current moment. 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].

[0061] When constructing the temperature prediction head, a spatial and temporal separation architecture is adopted. For a single battery cell, it is divided into multiple small cubic regions, with the center of each region serving as a spatial point. The spatial points are encoded to obtain the weight matrix and bias of each point, and then combined with the output of the hidden layer at the current time step to obtain the predicted temperature value.

[0062] The expression for the predicted temperature is:

[0063] , This represents the temperature weight matrix for the j-th spatial point at the k-th prediction time. This represents the output of the hidden layer at time t. This represents the bias of the j-th spatial point at the k-th prediction time. This represents the predicted temperature at time t, which is the predicted value at time k.

[0064] The above steps are used to obtain the predicted values ​​of charge, current, and temperature at each spatial point at each moment within the prediction interval. The coupling loss is determined based on the coupled heat release of all the predicted values ​​of charge, current, and temperature within the prediction interval. The combined loss is then obtained by adding these values ​​to the original losses of charge, current, and temperature.

[0065] The expression for coupling loss is:

[0066] , This represents the rate of temperature change at the predicted time. This indicates the coupled exothermic reaction of the predicted charge, predicted current, and predicted temperature values. Indicates the number of points in space. Indicates the length of the prediction interval. This represents the coupling loss.

[0067] Coupling loss, by explicitly modeling physical constraints, forces the model to learn consistent physical behavior, thereby improving the physical rationality of predictions and the system-level collaborative optimization capability.

[0068] The original loss in the neural network is replaced with a joint loss, and the neural network is run again. Based on the prediction results, the predicted values ​​of charge, current and temperature for the prediction interval are generated.

[0069] This application incorporates charge, charging / discharging current, and temperature into a unified prediction framework using the GRU time series model, enabling joint dynamic prediction of multiple physics fields. It can not only output the time series prediction curve of the prediction interval, but also capture the coupling effect between charging / discharging behavior and heat conduction.

[0070] Thus, the predicted values ​​of charge, current, and temperature within the prediction interval have been obtained.

[0071] In the risk assessment module, traditional battery thermal management systems typically rely on passive cooling control based on the current temperature state, which suffers from response lag and can easily lead to the accumulation of localized hot spots, thus affecting battery life and safety. However, during battery charging and discharging, temperature changes are strongly coupled with charge and charging / discharging current, and heat conduction itself exhibits spatiotemporal lag characteristics. Therefore, relying solely on current temperature feedback is insufficient to proactively address potential thermal risks; proactive thermal management combining predictions of future charge / discharge states is necessary.

[0072] The risk level is divided into three modules: thermal risk, power risk, and charging intensity risk.

[0073] Thermal risk is represented by the difference in predicted temperature at different spatial points, electrical risk by predicted charge, and charging intensity risk by predicted current. A comprehensive risk is derived based on these three factors.

[0074] The overall risk is positively correlated with the difference between the individual temperature prediction values ​​and the absolute value of the current prediction values, and negatively correlated with the charge prediction values.

[0075] It should be noted that positive correlation means that when one variable increases, the other variable also increases, and the two variables change in the same direction. 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 the actual application, and this application does not impose any special restrictions.

[0076] It should be noted that negative correlation means that when one variable increases, the other variable decreases accordingly, and the two variables change in opposite directions. 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 practical application, and this application does not impose any special restrictions.

[0077] The expression for comprehensive risk is:

[0078] , This represents the predicted charge value at the k-th prediction time, at time t. This represents the predicted temperature value at time t and the predicted temperature at time k. This represents the predicted current value at the k-th prediction time, which is the t-th time. This indicates the temperature threshold, reflecting the risk situation after exceeding the temperature threshold. This represents the comprehensive risk of the spatial point at time t, the k-th predicted time, and the j-th spatial point. This represents the normalization function. This indicates a function that takes positive values, meaning it takes x when x is greater than zero, and zero otherwise.

[0079] When the predicted temperature exceeds the temperature threshold, the larger the difference, the higher the thermal risk value, meaning that the location is more likely to experience performance degradation, shortened lifespan, and safety hazards due to high temperatures. If the predicted temperature does not exceed the threshold, the thermal risk value is 0, avoiding excessive allocation of cooling resources to the normal temperature range. The closer the predicted charge value is to 0, the greater the risk of battery life degradation due to over-discharge. The larger the absolute value of the predicted current, the greater the current load; high-current charging and discharging will exacerbate heat generation and accelerate battery aging, thus increasing the risk.

[0080] The overall risk assessment is obtained by summing all the comprehensive risks at each spatial point within the prediction interval. This method retains the time lag characteristic of heat conduction while capturing the spatial heterogeneity of risk. Compared to traditional thermal management that only focuses on the current temperature peak, this application captures the risk evolution trend through time weighting, thereby achieving preventative control.

[0081] At this point, a global risk assessment for each spatial point has been obtained.

[0082] The thermal management module, based on AI, suffers from inherent errors in its charge / discharge state prediction, especially for long-term forecasts. Traditional hard allocation strategies amplify these errors. Soft allocation, through exponential smoothing, reduces reliance on the accuracy of single-point predictions, making the system more robust to localized prediction deviations. Traditional thermal management systems only focus on current high-temperature points, while this application addresses potential future hotspots. Soft allocation, through risk negative exponential mapping, automatically directs resources to areas with low current risk but potentially worsening future conditions, achieving preventative cooling.

[0083] When an area is already severely overheated, continuing to allocate cooling resources may have limited effect, and manual intervention should be initiated. Soft allocation automatically reduces resource allocation to such areas, diverting capacity to more valuable preventative cooling, thereby improving the overall energy efficiency ratio.

[0084] The total cooling resources of the battery cells are allocated through a global risk assessment, and different cooling resources are assigned to different spatial points.

[0085] The expression for cooling resources is:

[0086] , Represents the sensitivity coefficient. This represents the global risk assessment of the j-th spatial point. This represents an exponential function with the natural constant as its base. This indicates the total cooling resources of the battery cell. Indicates the number of points in space. This represents the cooling resources at the j-th spatial point. A smaller sensitivity coefficient results in smoother allocation and enhanced anti-interference capability, but a delayed response to actual risks; conversely, a larger sensitivity coefficient results in a faster response to predicted risks, but may cause drastic fluctuations in resources due to local prediction errors. In this embodiment, the value is set to 5.

[0087] Based on this, the cooling resources at the current moment are obtained. The cooling resources for each moment are obtained using the above formula, prioritizing the protection of low-risk areas that may experience rapid temperature rise, while appropriately reducing cooling input for areas already at high risk. When the global risk assessment at a spatial point exceeds the preset overheating threshold, the overheating state may be irreversible, at which point an alarm is triggered to notify manual intervention. In this embodiment, the overheating threshold is 0.7. This design satisfies both the total resource constraints and smooths out the impact of prediction errors through probabilistic allocation, achieving a dynamic balance between avoiding thermal runaway risks and reducing cooling energy consumption, demonstrating the core advantages of predictive thermal management.

[0088] It should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

[0089] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. An AI-based BBU charge / discharge state prediction and thermal management system, characterized in that, The system includes the following modules: The feature acquisition module collects the charge, current, and temperature at each moment and uses them to construct a feature vector; The heat generation correlation module, based on the law of conservation of energy, treats the difference between heat release and heat absorption at each moment as the coupled heat release of charge, current, and temperature. The prediction module constructs a sliding window for the feature vector. Based on the data within the sliding window, it uses a neural network to predict the predicted value for each time step in the prediction interval, obtaining the predicted values ​​for different data types at each time step in the prediction interval. Then, based on the predicted values, it adjusts the loss function using the coupled exothermic adjustment function to obtain the predicted values ​​for charge, current, and temperature in the prediction interval. The risk assessment module determines the overall risk based on the differences in temperature predictions, current predictions, and charge predictions. The sum of the comprehensive risks of each spatial point is used as the global risk assessment; The thermal management module allocates the total cooling resources of the battery cells based on a global risk assessment, acquires the cooling resources of each spatial point, and completes thermal management. Among them, the temperature prediction value is for each spatial point. The battery cell is divided into multiple small cubic regions, and the center of each region is a spatial point. The spatial points are encoded to obtain the weight matrix and bias of each spatial point. The expression for the predicted temperature value at each spatial point is: , This represents the temperature weight matrix for the j-th spatial point at the k-th prediction time. This represents the output of the hidden layer at time t. This represents the bias of the j-th spatial point at the k-th prediction time. This represents the predicted temperature value of the j-th spatial point at the k-th prediction time at the t-th time. The method for adjusting the coupling exothermic adjustment loss function based on the predicted value is as follows: The coupling loss is determined by the coupled heat release of all predicted charge, current, and temperature values ​​within the prediction interval; then, it is added to the original losses of charge, current, and temperature to obtain the combined loss. The expression for coupling loss is: , This represents the rate of temperature change at the predicted time. This indicates the coupled exothermic effect of the predicted values. This represents the predicted charge value at the k-th prediction time, at time t. This represents the predicted current value at the k-th prediction time, which is the t-th time. This represents the predicted temperature value of the j-th spatial point at time t at the k-th prediction time. Indicates the number of points in space. Indicates the length of the prediction interval. This represents the coupling loss.

2. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 1, characterized in that, The method for obtaining the feature vector is as follows: Each data type at each time step and the differences in data types between adjacent time steps form a 6-dimensional original feature vector; The original feature vectors collected are reduced in dimensionality by using an embedding layer, and the data types are normalized to obtain the feature vectors at each time step.

3. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 1, characterized in that, The expression for the coupled heat release is: , Indicates the heat dissipation coefficient. Indicates the heat dissipation area. This represents the temperature at each moment. Indicates ambient temperature. Indicates capacitance. Indicates the rate of temperature change. This represents the heat release resulting from the coupling of charge, current, and temperature.

4. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 1, characterized in that, The method for obtaining the predicted values ​​of different data types at each time step within the prediction interval based on the data within the sliding window using a neural network is as follows: A multi-layer GRU architecture is used to process temporal inputs; the first-layer GRU receives all feature vectors within a sliding window and outputs a sequence of hidden states; a stochastic gradient descent optimizer is used to minimize the loss function between the predicted and the true values. Predicted values ​​for different data types are obtained by adjusting the hidden layer output using the weight matrix and bias generated by the network.

5. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 4, characterized in that, The intermediate charge value is generated by the charge weight matrix, the hidden layer output at the current time, and the charge bias. Then, the intermediate charge value is normalized by the normalization method and used as the charge offset. The charge offset is added to the charge at the current time to obtain the predicted charge value at the prediction time.

6. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 5, characterized in that, The intermediate current value is generated by the current weight matrix, the hidden layer output at the current time and the current bias. The intermediate current value is then normalized by the normalization method and used as the current offset. The current offset is added to the current at the current time to obtain the current prediction value at the prediction time. The predicted current values ​​are mapped to the interval [-1, 1].

7. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 1, characterized in that, The overall risk is positively correlated with the difference between the individual temperature prediction values ​​and the absolute value of the current prediction values, and negatively correlated with the charge prediction values.

8. The AI-based BBU charge / discharge state prediction and thermal management system as described in claim 1, characterized in that, The method for allocating the total cooling resources of the battery cells and obtaining the cooling resources for each spatial point based on global risk assessment is as follows: , Represents the sensitivity coefficient. This represents the global risk assessment of the j-th spatial point. This represents an exponential function with the natural constant as its base. This indicates the total cooling resources of the battery cell. Indicates the number of points in space. This represents the cooling resources at the j-th spatial point.

Images