Energy storage battery management system based on Bayesian fusion and model prediction control

Abstract

The invention relates to the technical field of energy storage battery management, and provides an energy storage battery management system based on Bayesian fusion and model prediction control, and the system comprises a data collection and preprocessing module which collects the voltage, current and temperature data of an energy storage battery pack in real time; the electrochemical model prediction module is used for updating parameters by adopting a recursive least square method based on a second-order RC equivalent circuit model to obtain a capacity prediction value and prediction uncertainty; the data driving model prediction module is used for obtaining a capacity prediction value and prediction uncertainty through Monte Carlo dropout reasoning; the Bayesian fusion module is used for calculating an evidence weight based on the prediction uncertainty and carrying out Bayesian fusion on the capacity prediction value; and the model prediction control module is used for solving an optimal equalization strategy based on the fusion capacity prediction value and generating a control instruction to realize intelligent equalization scheduling of the energy storage battery pack. According to the invention, the accuracy of energy storage battery capacity prediction is improved, and adaptive battery management under complex working conditions is realized.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery management technology, and in particular to an energy storage battery management system based on Bayesian fusion and model predictive control. Background Technology

[0002] Energy storage battery systems are complex energy storage devices composed of multiple individual batteries connected in series and parallel according to a specific configuration. They are widely used in key energy fields such as electric vehicles, power storage stations, distributed energy systems, and smart grids. During long-term use, energy storage battery systems are affected by factors such as charge-discharge cycles, temperature changes, and load fluctuations, leading to various degradation problems such as inconsistent capacity decay, increased internal resistance, and thermal runaway. The battery management system (BMS), as the core control component of the energy storage battery system, ensures the safety and consistency of the battery system by monitoring battery status in real time and executing equalization scheduling strategies. However, the unique multi-scale time-varying characteristics, nonlinear dynamic behavior, and complex electrochemical environment of energy storage battery systems pose numerous challenges to accurate prediction and intelligent management of battery status.

[0003] In existing technologies, energy storage battery management mainly employs traditional single-model prediction methods and basic equalization control strategies to achieve basic battery monitoring and management functions. However, existing methods do not adequately consider the inherent physical correlation mechanism between the complex dynamic characteristics of energy storage battery systems and multimodal prediction information. This makes it difficult to organically integrate the advantages of different prediction models with the actual battery operating state characteristics, resulting in the inability to achieve high-precision and high-reliability energy storage battery state prediction and adaptive equalization control. Especially under complex operating conditions, single prediction models often suffer from insufficient adaptability, while simple weighted average fusion methods cannot effectively utilize the complementary advantages of each model, further limiting the overall performance of the battery management system. Summary of the Invention

[0004] In view of this, the present invention proposes an energy storage battery management system based on Bayesian fusion and model predictive control, which solves the problem that existing methods do not adequately consider the inherent physical correlation mechanism between the complex dynamic characteristics of energy storage battery systems and multimodal prediction information, making it difficult to organically integrate the advantages of different prediction models with the actual battery operating state characteristics, thus making it impossible to achieve high-precision and high-reliability energy storage battery state prediction and adaptive equilibrium control.

[0005] The technical solution of this invention is implemented as follows: This invention provides an energy storage battery management system based on Bayesian fusion and model predictive control, the system comprising: The data acquisition and preprocessing module is used to acquire the individual cell voltage, total current and individual cell temperature of the energy storage battery pack in real time as raw data, and to filter and normalize the raw data to obtain model input samples. The electrochemical model prediction module is used to update the model parameters based on the second-order RC equivalent circuit model using the recursive least squares method, and to make predictions using the ampere-hour integral method, thereby obtaining the electrochemical model capacity prediction value and the electrochemical model prediction uncertainty. The data-driven model prediction module is used to input the model input samples into a pre-trained Transformer model and obtain the data-driven model capacity prediction value and the data-driven model prediction uncertainty through Monte Carlo dropout inference. The Bayesian fusion module is used to calculate evidence weights based on the prediction uncertainty of the electrochemical model and the prediction uncertainty of the data-driven model, and to perform Bayesian fusion on the capacity prediction values ​​of the electrochemical model and the data-driven model according to the evidence weights to obtain a fused capacity prediction value. The model predictive control module is used to construct a state vector based on the fused capacity prediction value, solve the optimal balancing strategy using a rolling optimization algorithm, generate balancing control commands, and perform balancing scheduling of the energy storage battery pack through the balancing control commands.

[0006] Based on the above technical solutions, preferably, the electrochemical model prediction module includes: A battery equivalent model is constructed based on a second-order RC equivalent circuit model. The battery equivalent model includes ohmic internal resistance, polarization internal resistance, and polarization capacitance. The model parameter matrix is ​​updated every 5 minutes using a recursive least squares method. The weight balance between historical data and real-time data is adjusted by a forgetting factor. The influence of ambient temperature on model parameters is corrected by a temperature compensation formula. The real-time state of charge is calculated using the ampere-hour integral method. The capacity value after 5 minutes is predicted by combining the polynomial mapping relationship between open-circuit voltage and state of charge. At the same time, the electrochemical model prediction uncertainty is calculated based on the parameter covariance matrix and the capacity gradient vector to obtain the electrochemical model prediction result, which includes the electrochemical model capacity prediction value and the electrochemical model prediction uncertainty.

[0007] Based on the above technical solutions, preferably, the step of updating the model parameter matrix every 5 minutes using the recursive least squares method includes: The voltage and current time series data of the last 10 seconds are collected to construct an observation vector. The observation vector includes the current current value and an integral term reflecting the charging and discharging process of the polarized capacitor. The Kalman gain at the current moment is calculated through the parameter update matrix. The model parameter vector is updated based on the prediction error and the Kalman gain. At the same time, the parameter covariance matrix is ​​updated to reflect the uncertainty of parameter estimation. The partial derivatives of capacity with respect to each model parameter are calculated using the central difference method to obtain the capacity gradient vector. The capacity gradient vector is then multiplied by the parameter covariance matrix to obtain the initial electrochemical model prediction uncertainty. When the battery temperature deviates from the standard temperature, the initial electrochemical model prediction uncertainty is corrected, and the electrochemical model capacity prediction value and electrochemical model prediction uncertainty are output.

[0008] Based on the above technical solutions, preferably, the data-driven model prediction module includes: A Transformer model is constructed, which includes a 2-layer encoder and a 4-head attention mechanism. An 8-dimensional feature sequence from the input sample of the model is input to the Transformer encoder. The 8-dimensional feature sequence includes voltage, current, temperature and the rate of change of voltage, the rate of change of current and the rate of change of temperature. Temporal location information is added through position encoding, and after processing by a 64-dimensional embedding layer and a feedforward neural network, a normalized state of charge (SOC) is output. Monte Carlo dropout technique is used to perform 50 random sampling inferences to calculate the prediction mean and prediction variance. The normalized SOC is then back-inferred to obtain the capacity prediction value of the data-driven model. The prediction uncertainty of the data-driven model is calculated based on the prediction variance. When the deviation between the capacity prediction value and the electrochemical model prediction value exceeds 5 amps, re-inference is performed to eliminate abnormal data interference.

[0009] Based on the above technical solutions, preferably, the step of using Monte Carlo dropout technology to perform 50 random sampling inferences includes: During the model inference phase, the dropout layer is kept active, and 50 forward propagation calculations are performed on the same input sample. Different neuron connections are randomly dropped during each propagation to obtain 50 different normalized state of charge prediction values. The arithmetic mean of 50 predicted values ​​is calculated as the prediction mean, and the variance of the 50 predicted values ​​is calculated as the original estimate of the prediction uncertainty. The normalized uncertainty is mapped to the actual state of charge range by variance scaling. The capacity prediction uncertainty is calculated in combination with the rated capacity. When the prediction uncertainty exceeds a set threshold, 500 sets of recent sample data are collected for incremental training to optimize the model performance. The complete data-driven model prediction result is output, which includes the capacity prediction value, prediction uncertainty, and prediction distribution.

[0010] Based on the above technical solutions, preferably, the Bayesian fusion module includes: The prediction uncertainties of the electrochemical model and the data-driven model are converted into confidence indices. Evidence weights for each model are calculated based on the confidence indices. The sum of the evidence weights is normalized to 1. Bayesian evidence theory is used to weight and fuse the capacity predictions of the electrochemical model and the data-driven model according to their evidence weights. Simultaneously, the fused prediction uncertainty is calculated by combining the prediction uncertainties of each model and the evidence weights. When the difference between the two model predictions exceeds 10 amperes, a conflict detection mechanism is activated to adjust the evidence weight allocation. When the fused prediction uncertainty exceeds twice the historical average, the weight of the electrochemical model is increased, and the fused capacity prediction value and fused prediction uncertainty are output.

[0011] Based on the above technical solutions, preferably, the calculation of the evidence weights of each model based on the trust index includes: The electrochemical model prediction uncertainty is transformed by its reciprocal to obtain the electrochemical model confidence level. The data-driven model prediction uncertainty is transformed by its reciprocal to obtain the data-driven model confidence level. The confidence levels under low temperature and high temperature conditions are adjusted by temperature correction factors. The proportion of each model confidence level to the total confidence level is calculated to obtain the initial evidence weight. The prediction errors of each model in the most recent period are statistically analyzed by sliding window. The initial evidence weights are dynamically adjusted based on the prediction errors. When the prediction error of the electrochemical model is less than that of the data-driven model for three consecutive times, the evidence weight of the electrochemical model is increased. When the prediction uncertainty of the data-driven model decreases continuously, the evidence weight of the data-driven model is increased. The evidence weights are avoided from fluctuating drastically by using a weight smoothing algorithm. The final evidence weights are output after dynamic adjustment.

[0012] Based on the above technical solutions, preferably, the formulas for calculating the Bayesian evidence weight and the fusion prediction uncertainty are as follows: ; ; in, Predicting uncertainties for electrochemical models; To predict uncertainty for data-driven models; and These are the normalized weights of evidence; This represents the prediction uncertainty after fusion.

[0013] Based on the above technical solutions, preferably, the data acquisition and preprocessing module includes: The individual cell voltage, total current, and individual cell temperature of the energy storage battery pack are collected in real time as raw data. The raw data are then processed by moving average filtering and Kalman filtering to remove noise. Outliers are removed using the three-standard-deviation criterion to obtain filtered data. The individual unit voltage, total current, and individual unit temperature in the filtered data are respectively processed by minimum and maximum value normalization. The normalized continuous time step data are constructed into a multi-dimensional feature sequence containing voltage, current, temperature and their rate of change using the fixed-length sequence method to obtain the model input sample.

[0014] Based on the above technical solutions, preferably, the model prediction control module includes: A 6-dimensional state vector is constructed based on the fused capacity prediction value. The 6-dimensional state vector includes the 5-minute capacity prediction value, the 24-hour capacity trend, the current capacity difference, the capacity change rate, the real-time temperature, and the prediction uncertainty. At the same time, the real-time voltage and temperature data of 100 cells in the energy storage battery pack are collected, and the difference value of the capacity of each cell and the average capacity of the battery pack and the capacity change rate are calculated. A battery state evolution model is constructed, which includes capacity evolution equation, temperature evolution equation, and voltage evolution equation. A multi-objective optimization function is established with the goal of minimizing the battery pack capacity variance, temperature penalty term, and energy conversion loss in the next hour. Multiple physical constraints are set, including the equalization current range, the upper limit of single cell temperature, the single cell voltage range, and the conversion efficiency. A sequential quadratic programming algorithm is used to perform rolling optimization every 5 minutes to generate equalization control instructions. The equalization control instructions include pairing combinations, equalization current values, execution time, and priority order.

[0015] The energy storage battery management system based on Bayesian fusion and model predictive control of the present invention has the following advantages over the prior art: (1) By integrating the dual-modal prediction technology of electrochemical model and data-driven model, the second-order RC equivalent circuit model and Transformer deep learning model are used to perform capacity prediction and uncertainty quantification respectively. The evidence weight is dynamically calculated based on the prediction uncertainty in combination with Bayesian evidence theory. The model prediction control optimization is carried out by constructing a multi-dimensional state vector based on the fused capacity prediction value. This improves the accuracy and reliability of energy storage battery capacity prediction. At the same time, adaptive battery management under complex working conditions is realized through explicit uncertainty quantification and multi-objective optimization equilibrium control. (2) By integrating the recursive least squares method with the second-order RC equivalent circuit model, the adaptive parameter update mechanism is used to adjust the weight balance of historical data and real-time data by using the forgetting factor, and the influence of environmental factors on model parameters is dynamically corrected by combining the temperature compensation formula and the ampere-hour integral method. The prediction uncertainty is quantified based on the parameter covariance matrix and capacity gradient vector, thereby improving the adaptability of energy storage battery capacity prediction. (3) By integrating the Transformer deep learning model with Monte Carlo dropout uncertainty quantification technology, 8-dimensional time-series feature analysis and capacity prediction calculation are performed using a 2-layer encoder and a 4-head attention mechanism. The temporal dependency of battery state is dynamically captured by combining position encoding and a 64-dimensional embedding layer. The uncertainty of the model output is evaluated based on 50 random sampling inferences and prediction variance calculation, which improves the generalization ability of energy storage battery capacity prediction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a structural diagram of an energy storage battery management system based on Bayesian fusion and model predictive control according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides an energy storage battery management system based on Bayesian fusion and model predictive control, the system comprising: The data acquisition and preprocessing module is used to acquire the individual cell voltage, total current and individual cell temperature of the energy storage battery pack in real time as raw data, and to filter and normalize the raw data to obtain model input samples. The electrochemical model prediction module is used to update the model parameters based on the second-order RC equivalent circuit model using the recursive least squares method, and to make predictions using the ampere-hour integral method, thereby obtaining the electrochemical model capacity prediction value and the electrochemical model prediction uncertainty. The data-driven model prediction module is used to input the model input samples into a pre-trained Transformer model and obtain the data-driven model capacity prediction value and the data-driven model prediction uncertainty through Monte Carlo dropout inference. The Bayesian fusion module is used to calculate evidence weights based on the prediction uncertainty of the electrochemical model and the prediction uncertainty of the data-driven model, and to perform Bayesian fusion on the capacity prediction values ​​of the electrochemical model and the data-driven model according to the evidence weights to obtain a fused capacity prediction value. The model predictive control module is used to construct a state vector based on the fused capacity prediction value, solve the optimal balancing strategy using a rolling optimization algorithm, generate balancing control commands, and perform balancing scheduling of the energy storage battery pack through the balancing control commands.

[0020] Specifically, this embodiment employs a dual-modal prediction technique that integrates an electrochemical model and a data-driven model. It utilizes a second-order RC equivalent circuit model and a Transformer deep learning model for capacity prediction and uncertainty quantification, respectively. Bayesian evidence theory is combined to dynamically calculate evidence weights based on prediction uncertainty, and a multi-dimensional state vector is constructed based on the fused capacity prediction values ​​for model predictive control optimization. This embodiment addresses the shortcomings of single prediction models in terms of adaptability and the inability of simple fusion methods to fully utilize the complementary advantages of models through a rolling optimization algorithm and a dynamic weight adjustment mechanism. This improves the accuracy and reliability of energy storage battery capacity prediction. Furthermore, it achieves adaptive battery management under complex operating conditions through explicit uncertainty quantification and multi-objective optimization equilibrium control.

[0021] The data acquisition and preprocessing module includes: The raw data consists of real-time acquisition of individual cell voltage, total current, and individual cell temperature of the energy storage battery pack. The raw data is then processed by moving average filtering and Kalman filtering to remove noise. Outliers are removed using the three-standard-deviation criterion to obtain filtered data.

[0022] In one specific embodiment, a moving average filter with a window size of 10 is used to denoise the individual cell voltage and total current. A Kalman filter with a process noise covariance of 0.01 and an observation noise covariance of 0.1 is used to further filter the total current. Outliers in the original data are identified and removed by setting a three-standard-deviation criterion for the individual cell voltage range of 2.5V to 3.65V, the total current range of -200A to +200A, and the individual cell temperature range of -10°C to 60°C.

[0023] The individual unit voltage, total current, and individual unit temperature in the filtered data are respectively processed by minimum and maximum value normalization. The normalized continuous time step data are constructed into a multi-dimensional feature sequence containing voltage, current, temperature and their rate of change using the fixed-length sequence method to obtain the model input sample.

[0024] In one specific embodiment, the individual cell voltage is normalized to the range of 0 to 1 according to the range of 2.5 volts to 3.65 volts, the total current is normalized to the range of -200 amperes to 200 amperes, and the individual cell temperature is normalized to the range of -10 degrees Celsius to 45 degrees Celsius. The rate of change of voltage, current and temperature is calculated by numerical differentiation. A fixed-length sequence is constructed using 20 consecutive time steps to form a model input sample containing 8-dimensional features.

[0025] Specifically, this embodiment employs a dual denoising mechanism combining moving average filtering and Kalman filtering, utilizes a three-standard-deviation criterion for outlier detection and removal, dynamically identifies data anomalies based on specific voltage, current, and temperature range constraints, and constructs a multidimensional feature sequence containing the original signal and its rate of change using minimum-maximum normalization and numerical differentiation methods. This embodiment addresses the issues of noise interference and insufficient feature representation in the original sensor data through a fixed-length time window and 8-dimensional feature vector construction, improving the quality and reliability of energy storage battery state data. Furthermore, the standardized data preprocessing workflow and multidimensional feature engineering meet the high-quality input data requirements of electrochemical and deep learning models.

[0026] The electrochemical model prediction module includes: A battery equivalent model is constructed based on a second-order RC equivalent circuit model. The battery equivalent model includes ohmic internal resistance, polarization internal resistance, and polarization capacitance. The model parameter matrix is ​​updated every 5 minutes using a recursive least squares method. The weight balance between historical data and real-time data is adjusted by a forgetting factor. The influence of ambient temperature on model parameters is corrected by a temperature compensation formula. The real-time state of charge is calculated using the ampere-hour integral method. The capacity value after 5 minutes is predicted by combining the polynomial mapping relationship between open-circuit voltage and state of charge. At the same time, the electrochemical model prediction uncertainty is calculated based on the parameter covariance matrix and the capacity gradient vector to obtain the electrochemical model prediction result, which includes the electrochemical model capacity prediction value and the electrochemical model prediction uncertainty.

[0027] The method of updating the model parameter matrix every 5 minutes using recursive least squares includes: The voltage and current time series data of the last 10 seconds are collected to construct an observation vector. The observation vector includes the current current value and an integral term reflecting the charging and discharging process of the polarized capacitor. The Kalman gain at the current moment is calculated through the parameter update matrix. The model parameter vector is updated based on the prediction error and the Kalman gain. At the same time, the parameter covariance matrix is ​​updated to reflect the uncertainty of parameter estimation. The partial derivatives of capacity with respect to each model parameter are calculated using the central difference method to obtain the capacity gradient vector. The capacity gradient vector is then multiplied by the parameter covariance matrix to obtain the initial electrochemical model prediction uncertainty. When the battery temperature deviates from the standard temperature, the initial electrochemical model prediction uncertainty is corrected, and the electrochemical model capacity prediction value and electrochemical model prediction uncertainty are output.

[0028] Specifically, this embodiment employs an adaptive parameter update mechanism that integrates recursive least squares and a second-order RC equivalent circuit model. It utilizes a forgetting factor to adjust the weight balance between historical and real-time data, combines a temperature compensation formula and an ampere-hour integral method to dynamically correct the impact of environmental factors on model parameters, and quantifies prediction uncertainty based on the parameter covariance matrix and capacity gradient vector. This embodiment addresses the problems of traditional electrochemical models with fixed parameters being unable to adapt to battery aging and environmental changes, and lacking prediction reliability assessment, by using the central difference method and temperature correction mechanism, thus improving the adaptability of energy storage battery capacity prediction.

[0029] The data-driven model prediction module includes: A Transformer model is constructed, which includes a 2-layer encoder and a 4-head attention mechanism. An 8-dimensional feature sequence from the input sample of the model is input to the Transformer encoder. The 8-dimensional feature sequence includes voltage, current, temperature and the rate of change of voltage, the rate of change of current and the rate of change of temperature. Temporal location information is added through position encoding, and after processing by a 64-dimensional embedding layer and a feedforward neural network, a normalized state of charge (SOC) is output. Monte Carlo dropout technique is used to perform 50 random sampling inferences to calculate the prediction mean and prediction variance. The normalized SOC is then back-inferred to obtain the capacity prediction value of the data-driven model. The prediction uncertainty of the data-driven model is calculated based on the prediction variance. When the deviation between the capacity prediction value and the electrochemical model prediction value exceeds 5 amps, re-inference is performed to eliminate abnormal data interference.

[0030] The method of using Monte Carlo dropout to perform 50 random sampling inferences includes: During the model inference phase, the dropout layer is kept active, and 50 forward propagation calculations are performed on the same input sample. Different neuron connections are randomly dropped during each propagation to obtain 50 different normalized state of charge prediction values. The arithmetic mean of 50 predicted values ​​is calculated as the prediction mean, and the variance of the 50 predicted values ​​is calculated as the original estimate of the prediction uncertainty. The normalized uncertainty is mapped to the actual state of charge range by variance scaling. The capacity prediction uncertainty is calculated in combination with the rated capacity. When the prediction uncertainty exceeds a set threshold, 500 sets of recent sample data are collected for incremental training to optimize the model performance. The complete data-driven model prediction result is output, which includes the capacity prediction value, prediction uncertainty, and prediction distribution.

[0031] Specifically, this embodiment integrates the Transformer deep learning model with Monte Carlo dropout uncertainty quantification technology. It utilizes a 2-layer encoder and a 4-head attention mechanism for 8-dimensional temporal feature analysis and capacity prediction calculation. It combines position encoding and a 64-dimensional embedding layer to dynamically capture the temporal dependencies of battery states, and evaluates the uncertainty of the model output based on 50 random sampling inferences and prediction variance calculation. This embodiment addresses the problems of traditional deep learning models lacking prediction reliability evaluation and unable to adaptively update through anomaly detection mechanisms and incremental training strategies, thereby improving the generalization ability of energy storage battery capacity prediction.

[0032] The Bayesian fusion module includes: The prediction uncertainties of the electrochemical model and the data-driven model are converted into confidence indices. Evidence weights for each model are calculated based on the confidence indices. The sum of the evidence weights is normalized to 1. Bayesian evidence theory is used to weight and fuse the capacity predictions of the electrochemical model and the data-driven model according to their evidence weights. Simultaneously, the fused prediction uncertainty is calculated by combining the prediction uncertainties of each model and the evidence weights. When the difference between the two model predictions exceeds 10 amperes, a conflict detection mechanism is activated to adjust the evidence weight allocation. When the fused prediction uncertainty exceeds twice the historical average, the weight of the electrochemical model is increased to ensure prediction stability. The fused capacity prediction value and the fused prediction uncertainty are then output.

[0033] The calculation of evidence weights for each model based on the trust index includes: The electrochemical model prediction uncertainty is transformed by its reciprocal to obtain the electrochemical model confidence level. The data-driven model prediction uncertainty is transformed by its reciprocal to obtain the data-driven model confidence level. The confidence levels under low temperature and high temperature conditions are adjusted by temperature correction factors. The proportion of each model confidence level to the total confidence level is calculated to obtain the initial evidence weight. The prediction errors of each model in the most recent period are statistically analyzed by sliding window. The initial evidence weights are dynamically adjusted based on the prediction errors. When the prediction error of the electrochemical model is less than that of the data-driven model for three consecutive times, the evidence weight of the electrochemical model is increased. When the prediction uncertainty of the data-driven model decreases continuously, the evidence weight of the data-driven model is increased. The evidence weights are avoided from fluctuating drastically by using a weight smoothing algorithm. The final evidence weights are output after dynamic adjustment.

[0034] In one specific embodiment, the formulas for calculating the Bayesian evidence weight and the fusion prediction uncertainty are as follows: ; ; in, Predicting uncertainties for electrochemical models; To predict uncertainty for data-driven models; and These are the normalized weights of evidence; To address the uncertainty of the fused predictions, this implementation utilizes the principle of adding the inverse variances of independent information sources in Bayesian theory to achieve optimal fusion of the uncertainties of the two models.

[0035] Specifically, this embodiment integrates Bayesian evidence theory with a dynamic weight adjustment mechanism. It utilizes the inverse transformation of prediction uncertainty to calculate the confidence index and allocate evidence weights. It dynamically adjusts the model weight ratios by combining a temperature correction factor and sliding window prediction error statistics. Furthermore, it adaptively handles anomalies using a conflict detection mechanism and a weight smoothing algorithm. This embodiment addresses the shortcomings of simple weighted averaging in fully utilizing the complementary advantages of dual-modal models and lacking adaptive adjustment capabilities through multi-level weight optimization and stability assurance mechanisms, thereby improving the accuracy of the fused prediction results for energy storage battery capacity.

[0036] The model prediction control module includes: A 6-dimensional state vector is constructed based on the fused capacity prediction value. The 6-dimensional state vector includes the 5-minute capacity prediction value, the 24-hour capacity trend, the current capacity difference, the capacity change rate, the real-time temperature, and the prediction uncertainty. At the same time, real-time voltage and temperature data of 100 cells in the energy storage battery pack are collected, and the difference value of the capacity of each cell and the average capacity of the battery pack and the capacity change rate are calculated.

[0037] In one specific embodiment, the 5-minute capacity prediction value and prediction uncertainty of 100 individual cells are acquired through real-time data acquisition. Real-time voltage and temperature data of each individual cell are acquired simultaneously. The capacity decay trend for the next 24 hours is obtained based on the remaining lifespan prediction network. The difference between the capacity of each individual cell and the average capacity of the battery pack is calculated. The capacity change rate is calculated using a numerical differentiation method. The 5-minute capacity prediction value, 24-hour capacity trend, current capacity difference, capacity change rate, real-time temperature, and prediction uncertainty are combined to form a 6-dimensional state vector. When the capacity difference exceeds 2 amp-hours, it is marked as a high-priority balancing object. When the prediction uncertainty exceeds the historical average value, the weight of the state vector is corrected.

[0038] A battery state evolution model is constructed, which includes capacity evolution equation, temperature evolution equation, and voltage evolution equation. A multi-objective optimization function is established with the goal of minimizing the battery pack capacity variance, temperature penalty term, and energy conversion loss in the next hour. Multiple physical constraints are set, including the equalization current range, the upper limit of single cell temperature, the single cell voltage range, and the conversion efficiency. A sequential quadratic programming algorithm is used to perform rolling optimization every 5 minutes to generate equalization control instructions. The equalization control instructions include pairing combinations, equalization current values, execution time, and priority order.

[0039] In one specific embodiment, a capacity evolution equation considering the equalization current effect and natural aging degradation, a temperature evolution equation calculating the temperature rise based on the square relationship of the equalization current, and a voltage evolution equation reflecting the voltage change during the equalization process are established. A multi-objective optimization function is constructed, including a battery pack capacity variance minimization term, a temperature penalty term, and an energy conversion loss term, with weights set to 0.6, 0.2, and 0.2, respectively. Physical constraints are set, including an equalization current range of 0.5 amps to 2.0 amps, an upper limit of single-cell temperature of 45 degrees Celsius, a single-cell voltage range of 2.7 volts to 3.5 volts, and a DC-DC converter conversion efficiency of not less than 90%. A sequential quadratic programming algorithm is used to solve for the optimal equalization strategy. The optimal equalization command for the first 15 minutes is extracted and a safety check is performed to determine the pairing combination of high-capacity cells and low-capacity cells and the equalization current value based on multi-objective optimization.

[0040] Specifically, this embodiment integrates multi-dimensional state vector construction and sequential quadratic programming optimization algorithms, utilizes 6-dimensional state features and battery state evolution models to calculate multi-objective optimization functions and solve equalization strategies, combines multiple physical constraints and dynamic weight correction mechanisms to adjust control parameters in real time, and dynamically generates and verifies equalization commands based on rolling optimization and safety verification mechanisms. Through multi-objective coordinated optimization and physical constraint guarantee mechanisms, it solves the problems of traditional equalization control methods being single-objective-oriented, lacking physical constraint considerations, and unable to adaptively adjust, thereby improving the accuracy and safety of energy storage battery equalization control. At the same time, through explicit multi-dimensional state perception and rolling optimization strategies, it achieves adaptive intelligent equalization management under complex operating conditions.

[0041] In one specific embodiment, this embodiment also provides an energy storage battery management system based on Bayesian fusion and model predictive control, including a battery pack, a data acquisition module, a dual-modal capacity prediction module, a system control module, and a communication module; The battery pack includes: 2 parallel 100 series lithium iron phosphate battery packs, with a single cell capacity of 100Ah, a nominal voltage of 3.2V, and a total capacity of 200Ah (640Wh); each 10 series is equipped with one NTC temperature sensor (accuracy ±0.5℃) to monitor the battery temperature distribution in real time.

[0042] The data acquisition module includes: 1. Voltage acquisition: ADS1281 chip (16-bit precision, 1kHz sampling frequency), differential input for anti-interference, 1kΩ current-limiting resistor and 50V Zener diode in series for each channel; 2. Current acquisition: ACS758 Hall sensor (±200A range, response time <1μs), the output is amplified 5 times by OPA2340 op-amp and then connected to ADC; 3. Temperature acquisition: DS18B20 sensor (resolution 0.0625℃), cascaded via 1-Wire bus; 4. Data storage: 32GB industrial-grade SD card, 1 minute / data storage cycle, historical data retention for ≥1 year.

[0043] The dual-modal capacity prediction module includes: hardware carrier: NVIDIA Jetson Nano (4-core ARMA57, 128-core GPU), supporting edge computing; model integration: electrochemical model (second-order RC equivalent circuit) + data-driven model (Transformer network); deployment method: lightweight deployment based on TensorFlowLite, model memory usage <50MB, inference latency <100ms.

[0044] The system control module includes: main control chip: STM32H743 (Cortex-M7 core, 480MHz main frequency); software environment: running FreeRTOS real-time operating system, task scheduling cycle 1ms; core functions: task coordination, charging and discharging strategy optimization, and safety protection instruction generation.

[0045] The communication modules include: internal communication: CAN2.0B (baud rate 500kbps, data exchange between modules, latency <10ms); external communication: Ethernet (100Mbps, connection to host computer / energy storage converter, static IP 192.168.1.100).

[0046] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery management system for energy storage based on Bayesian fusion and model predictive control, characterized in that, The system comprises: A data acquisition preprocessing module for acquiring the single cell voltage, total current and single cell temperature of the energy storage battery pack as raw data in real time, filtering and normalizing the raw data to obtain model input samples; An electrochemical model prediction module for updating model parameters based on a second-order RC equivalent circuit model using a recursive least squares method, combining an ampere-hour integral method for prediction to obtain an electrochemical model capacity prediction value and an electrochemical model prediction uncertainty; A data-driven model prediction module for inputting the model input samples into a pre-trained Transformer model to obtain a data-driven model capacity prediction value and a data-driven model prediction uncertainty through Monte Carlo dropout reasoning; A Bayesian fusion module for calculating an evidence weight based on the electrochemical model prediction uncertainty and the data-driven model prediction uncertainty, and performing Bayesian fusion on the electrochemical model capacity prediction value and the data-driven model capacity prediction value according to the evidence weight to obtain a fusion capacity prediction value; A model prediction control module for constructing a state vector based on the fusion capacity prediction value, solving an optimal equilibrium strategy using a rolling optimization algorithm, generating an equilibrium control instruction, and performing equilibrium scheduling on the energy storage battery pack through the equilibrium control instruction.

2. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 1, wherein, The electrochemical model prediction module comprises: A battery equivalent model is constructed based on a second-order RC equivalent circuit model, the battery equivalent model includes an ohmic internal resistance, a polarization internal resistance and a polarization capacitance, a model parameter matrix is updated every 5 minutes using a recursive least squares method, the weight balance of historical data and real-time data is adjusted through a forgetting factor, the influence of environmental temperature on the model parameters is corrected through a temperature compensation formula, the real-time state of charge is calculated through an ampere-hour integral method, the capacity value after 5 minutes is predicted based on the polynomial mapping relationship between the open circuit voltage and the state of charge, and the electrochemical model prediction uncertainty is calculated based on the parameter covariance matrix and the capacity gradient vector to obtain the electrochemical model prediction result, the electrochemical model prediction result includes the electrochemical model capacity prediction value and the electrochemical model prediction uncertainty.

3. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 2, wherein, The recursive least squares method is used to update the model parameter matrix every 5 minutes, which comprises: The voltage and current time series data of the last 10 seconds are collected to construct an observation vector, the observation vector includes the current current value and the integral term reflecting the polarization capacitance charging and discharging process, the Kalman gain at the current time is calculated through the parameter update matrix, the model parameter vector is updated based on the prediction error and the Kalman gain, and the parameter covariance matrix is updated to reflect the uncertainty of parameter estimation; The partial derivative of the capacity with respect to each model parameter is calculated by the central difference method to obtain a capacity gradient vector, the capacity gradient vector is multiplied by the parameter covariance matrix to obtain an initial electrochemical model prediction uncertainty, the initial electrochemical model prediction uncertainty is corrected when the battery temperature deviates from the standard temperature, and the electrochemical model capacity prediction value and the electrochemical model prediction uncertainty are output.

4. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 1, wherein, The data-driven model prediction module comprises: A Transformer model is constructed, the Transformer model comprising 2 layers of encoder and 4 heads of attention mechanism, an 8-dimensional feature sequence in the sample is input into the Transformer encoder, the 8-dimensional feature sequence comprising voltage, current, temperature, voltage rate of change, current rate of change, temperature rate of change; Time sequence position information is added through position coding, normalized state of charge is output after being processed by a 64-dimensional embedding layer and a feedforward neural network, Monte Carlo dropout technology is used to perform 50 times of random sampling reasoning, a prediction mean and a prediction variance are calculated, the normalized state of charge is back calculated to obtain a data-driven model capacity prediction value, a data-driven model prediction uncertainty is calculated based on the prediction variance, and when a capacity prediction value and an electrochemical model prediction value deviate by more than 5 ampere-hours, re-reasoning is performed to exclude abnormal data interference.

5. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 4, wherein, The 50 times of random sampling reasoning by using the Monte Carlo dropout technology comprises: In the model reasoning stage, the dropout layer is activated, 50 times of forward propagation calculation are performed on the same input sample, different neuron connections are randomly discarded in each propagation process, and 50 different normalized state of charge prediction values are obtained; An arithmetic mean of the 50 prediction values is calculated as the prediction mean, a variance of the 50 prediction values is calculated as an original estimation of the prediction uncertainty, the normalized uncertainty is mapped to the actual state of charge range through variance scaling, and the capacity prediction uncertainty is calculated in combination with the rated capacity, when the prediction uncertainty exceeds a set threshold, 500 groups of recent sample data are collected for incremental training to optimize the model performance, and a complete data-driven model prediction result is output, the complete data-driven model prediction result comprising the capacity prediction value, the prediction uncertainty and the prediction distribution.

6. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 1, wherein, The Bayesian fusion module comprises: The electrochemical model prediction uncertainty and the data-driven model prediction uncertainty are respectively converted into trustworthiness indexes, the evidence weight of each model is calculated based on the trustworthiness indexes, the sum of the evidence weights is 1 through normalization processing, the electrochemical model capacity prediction value and the data-driven model capacity prediction value are weighted and fused according to the evidence weight by using the Bayesian evidence theory, and the fusion prediction uncertainty is calculated in combination with the prediction uncertainty of each model and the evidence weight, when the prediction values of the two models deviate by more than 10 ampere-hours, a conflict detection mechanism is started to adjust the evidence weight distribution, when the fusion prediction uncertainty exceeds twice the historical average value, the electrochemical model weight is increased, and a fusion capacity prediction value and a fusion prediction uncertainty are output.

7. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 6, wherein, The calculation of the evidence weight of each model based on the trustworthiness indexes comprises: The electrochemical model prediction uncertainty is transformed into the electrochemical model trustworthiness through reciprocal transformation, the data-driven model prediction uncertainty is transformed into the data-driven model trustworthiness through reciprocal transformation, the trustworthiness under low-temperature and high-temperature working conditions is adjusted through a temperature correction factor, the proportion of the trustworthiness of each model in the total trustworthiness is calculated, and the initial evidence weight is obtained. The prediction errors of each model in the recent period are counted through a sliding window, the initial evidence weight is dynamically corrected based on the prediction errors, when the prediction error of the electrochemical model is less than that of the data-driven model for three times in succession, the evidence weight of the electrochemical model is increased, when the prediction uncertainty of the data-driven model is continuously reduced, the evidence weight of the data-driven model is increased, the weight smoothing algorithm is used to avoid the evidence weight from fluctuating sharply, and the final evidence weight after dynamic adjustment is output.

8. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 7, wherein, The calculation formula of the Bayesian evidence weight and the fusion prediction uncertainty is: ； ； wherein, is the electrochemical model prediction uncertainty; is the data-driven model prediction uncertainty; and are the normalized evidence weights, respectively; is the fused prediction uncertainty.

9. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 1, wherein, The data acquisition and preprocessing module comprises: The single cell voltage, total current and single cell temperature of the energy storage battery pack are collected in real time as raw data, the raw data are subjected to moving average filtering and Kalman filtering to remove noise, the abnormal values are removed by using the three standard deviation criterion, and the filtered data are obtained; The single cell voltage, total current and single cell temperature in the filtered data are subjected to minimum maximum value normalization respectively, the normalized continuous time step data are constructed into a multi-dimensional feature sequence containing voltage, current, temperature and change rate by using the fixed length sequence method, and the model input sample is obtained.

10. The energy storage battery management system based on Bayesian fusion and model predictive control of claim 1, wherein, The model prediction control module comprises: A 6-dimensional state vector is constructed based on the fusion capacity prediction value, the 6-dimensional state vector comprises 5-minute capacity prediction value, 24-hour capacity trend, current capacity difference, capacity change rate, real-time temperature and prediction uncertainty, the real-time voltage and temperature data of 100 single cells in the energy storage battery pack are collected, and the difference value and capacity change rate of each single cell and the average capacity of the battery pack are calculated; A battery state evolution model is constructed, the battery state evolution model comprises a capacity evolution equation, a temperature evolution equation and a voltage evolution equation, a multi-objective optimization function is established with the optimization objective of minimizing the battery pack capacity variance within 1 hour in the future, a temperature penalty term and an energy conversion loss, multiple physical constraints are set, the multiple physical constraints comprise a balanced current range, a single cell temperature upper limit, a single cell voltage range and a conversion efficiency, a sequence quadratic programming algorithm is used to perform rolling optimization and solving every 5 minutes, a balancing control instruction is generated, and the balancing control instruction comprises a pairing combination, a balanced current value, an execution time and a priority order.

Images