A multi-dimensional feature-based energy storage battery operation and maintenance decision method and system

Abstract

The application discloses a kind of based on multi-dimensional feature's energy storage battery operation and maintenance decision method and system, it is related to energy storage battery technical field, including obtaining multi-dimensional operation time series data;ResNet unit and BiLSTM unit are included in the mixed network estimation model of construction, output SOC estimation value;SOH estimation model containing transformer encoder and data segmentation is constructed, and SOH estimation value is output;Based on SOC estimation value and SOH estimation value, design the battery equalization strategy of active and passive combination, and utilize support vector machine algorithm to generate battery operation and maintenance strategy.The application solves the problems, such as inaccurate state estimation, imperfect equalization strategy and lack of system analysis in operation and maintenance decision in the prior art.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery technology, and in particular to an energy storage battery operation and maintenance decision-making method and system based on multi-dimensional features. Background Technology

[0002] With the continuous growth of energy storage demand, energy storage batteries are increasingly widely used in electric vehicles, smart grids, distributed energy and other fields. However, there are still many problems in the operation and maintenance management of energy storage batteries. The long-term operation of energy storage batteries is easily affected by factors such as charge and discharge cycles and ambient temperature, and the performance gradually degrades. If the operation and maintenance is not done properly, it may lead to a shortened battery life and an increased safety risk. Therefore, the refined operation and maintenance of energy storage batteries is crucial.

[0003] From the perspective of battery state monitoring, most existing technologies rely on data from a single or a few dimensions for analysis. For example, they estimate the state of charge (SOC) by collecting voltage and current data and using a simple ampere-hour integration method or open-circuit voltage method. The assessment of state of harm (SOH) is also mostly based on the number of charge-discharge cycles or a single capacity decay index.

[0004] However, the main problems with the current technology are as follows: First, the accuracy of battery state estimation is insufficient. Due to the reliance on only a small amount of feature data and the simple model structure, it is difficult to accurately capture the dynamic changes of SOC and SOH, leading to deviations in the judgment of the battery's remaining capacity and health. Second, the balancing management strategy is not perfect. It mostly adopts a single passive or active balancing method, which cannot simultaneously take into account the energy balance between individual cells within the battery module and between modules. It is difficult to effectively solve the problem of declining battery pack consistency, thus affecting the overall operating efficiency and service life. In addition, the operation and maintenance decision-making lacks systematic analysis of multi-dimensional monitoring data, making it difficult to form a precise charging control, maintenance reminder, and safety early warning mechanism, and failing to meet the requirements of efficient and safe operation of energy storage batteries. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for energy storage battery operation and maintenance decision-making based on multi-dimensional features. This invention solves the problems of inaccurate state estimation, imperfect balancing strategies, and lack of systematic analysis in operation and maintenance decision-making in existing technologies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for energy storage battery operation and maintenance decision-making based on multi-dimensional features, comprising:

[0008] Acquire multidimensional runtime sequence data of the battery and perform preprocessing;

[0009] A hybrid network estimation model containing ResNet and BiLSTM units is constructed. Preprocessed multidimensional runtime data is input into the hybrid network estimation model, and the SOC estimate is output.

[0010] A SOH estimation model is constructed that includes a Transformer encoder and data segmentation. The preprocessed multidimensional runtime sequence data is segmented using data segmentation, and then the segmented data is input into the Transformer encoder to output the SOH estimate.

[0011] Based on the SOC and SOH estimates, a combined active and passive battery balancing strategy is designed, and a support vector machine model is used to generate the battery operation and maintenance strategy.

[0012] As a further technical solution, the acquired multi-dimensional operating sequence data of the battery includes the battery's individual cell voltage, current, temperature, pressure, and internal impedance data.

[0013] The preprocessing includes denoising, filtering, and normalization. The denoising uses a wavelet transform denoising algorithm, specifically decomposing the data using the db4 wavelet basis function, removing high-frequency noise, and then reconstructing the data. The filtering uses a moving average filtering method. The normalization uses a minimum-maximum normalization method.

[0014] As a further technical solution, the ResNet unit contains two asymmetric convolutional modules, each consisting of a 1×3 convolutional layer and a 3×1 convolutional layer, and each convolutional layer is followed by a BatchNorm layer and a ReLU activation function; the BiLSTM unit contains two bidirectional LSTM layers, namely a forward LSTM and a backward LSTM, and also contains a hidden layer and an output layer, wherein the output layer is a fully connected layer.

[0015] As a further technical solution, the data segmentation is specifically as follows: first, the number of battery charge-discharge cycles is set and the cycle value is determined, and then the preprocessed multidimensional runtime sequence data is divided into data segments of different lengths according to the cycle value.

[0016] As a further technical solution, when the length of the last data segment is insufficient for the corresponding period value, the last value is repeatedly used to fill the standard length, and then the data segment is normalized and aligned according to the maximum and minimum values ​​within the data segment.

[0017] As a further technical solution, the Transformer encoder includes a four-layer structure, each layer consisting of a multi-head attention mechanism and a feedforward network.

[0018] As a further technical solution, the design incorporates a combined active and passive battery balancing strategy. Specifically, the passive balancing strategy involves connecting a resistor calculated based on the battery's rated voltage and current in parallel to each individual battery cell. The active balancing strategy uses an inductor as an energy transfer element and employs PWM control to control the inductor to transfer energy between high and low voltage modules.

[0019] Secondly, the present invention provides an energy storage battery operation and maintenance decision-making system based on multi-dimensional features, comprising the following modules:

[0020] The data acquisition module is configured to acquire multidimensional runtime sequence data of the battery and perform preprocessing.

[0021] The State of Charge (SOC) estimation module is configured to: construct a hybrid network estimation model containing ResNet and BiLSTM units, input preprocessed multidimensional runtime data into the hybrid network estimation model, and output SOC estimates.

[0022] The health status estimation module is configured to: construct a SOH estimation model that includes a Transformer encoder and data segmentation; segment the preprocessed multidimensional runtime sequence data using data segmentation; input the segmented data into the Transformer encoder; and output the SOH estimate.

[0023] The strategy output module is configured to: design a combined active and passive battery balancing strategy based on the SOC estimate and SOH estimate, and generate a battery operation and maintenance strategy using a support vector machine model.

[0024] One or more technical solutions of the present invention have the following beneficial effects:

[0025] This invention avoids the limitations of single data by acquiring five core data types: voltage, current, temperature, pressure, and internal impedance. It also constructs a hybrid network estimation model containing ResNet and BiLSTM units to estimate the SOC value, and constructs an SOH estimation model containing a Transformer encoder and data segmentation to estimate the SOH value. This enables accurate judgment of the battery's remaining power and health status, and solves the "estimation bias" problem in the prior art.

[0026] This invention designs a battery balancing strategy that combines active and passive methods. The combination of the two balancing methods covers both the "cell-module" and "module" scenarios, effectively solving the problem of decreased consistency and extending the overall lifespan of the battery pack.

[0027] This invention is based on SOC and SOH estimates and uses clear thresholds to form a precise strategy for the entire process of "charging-maintenance-early warning-shutdown" to meet the requirements of efficient and safe battery operation. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0029] Figure 1 This is a flowchart of an energy storage battery operation and maintenance decision-making method based on multi-dimensional features in this invention. Detailed Implementation

[0030] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0031] Example 1

[0032] This embodiment provides a multi-dimensional feature-based energy storage battery operation and maintenance decision-making method, such as... Figure 1 The method flowchart is shown below, and the specific method steps are as follows:

[0033] S1: Obtain multi-dimensional runtime sequence data of the battery and perform preprocessing.

[0034] In step S1, voltage sensors, current sensors, temperature sensors, pressure sensors, and internal impedance measuring instruments are used to acquire real-time data on the individual cell voltage, current, temperature, pressure, and internal impedance of the battery.

[0035] In this embodiment, the voltage sensor has an accuracy of ±0.01V, the current sensor has an accuracy of ±0.1A, the temperature sensor has an accuracy of ±0.5℃, the pressure sensor has an accuracy of ±0.05MPa, and the internal impedance measuring instrument has an accuracy of ±0.001Ω; the industrial camera has a resolution of 1920×1080 and features autofocus and low-light shooting capabilities; all sensors use the CAN bus protocol for data transmission with a communication rate of 500kbps.

[0036] In step S1, the acquired multi-dimensional battery runtime sequence data is preprocessed, including denoising, filtering, and normalization. Denoising employs a wavelet transform denoising algorithm, selecting the db4 wavelet basis function for three-level decomposition. After decomposition, high-frequency and low-frequency coefficients are obtained. A hard threshold function is used to process the high-frequency coefficients, with the threshold set to 0.02 times the maximum high-frequency coefficient to remove noise components. The denoised data is then reconstructed using the processed high-frequency and low-frequency coefficients. Filtering uses a moving average filtering method, setting the filtering window to 5, and taking the average of the previous and next 5 data points as the result. Normalization uses the minimum-maximum normalization method, with the following formula:

[0037] ,in The original data, The minimum value of this dimension. The maximum value of this dimension is given; the normalized data is output in the form of a time series window of length 30, with each window containing five dimensions of features: voltage, current, temperature, pressure, and internal impedance.

[0038] S2: Construct a hybrid network estimation model containing ResNet units and BiLSTM units, input the preprocessed multidimensional runtime data into the hybrid network estimation model, and output the SOC estimate.

[0039] In step S2, the ResNet unit contains two asymmetric convolutional modules. Each asymmetric convolutional module consists of a 1×3 convolutional layer and a 3×1 convolutional layer, replacing the traditional 3×3 convolution to reduce parameters and enhance spatial feature extraction capabilities. Specifically, the asymmetric convolutional module contains three convolutional layers: the first layer has a 1×3 kernel, a stride of 1, and padding of 1; the second layer has a 3×1 kernel, a stride of 1, and padding of 1; and the third layer has a 1×3 kernel, a stride of 1, and padding of 1. The pooling layer is average pooling with a 2×2 window and a stride of 2. Each convolutional layer is followed by a BatchNorm layer and a ReLU activation function, and finally, feature fusion is completed through a 1×1 convolutional layer.

[0040] The BiLSTM unit contains two bidirectional LSTM layers, each with 64 neurons: a forward LSTM and a backward LSTM. The forward LSTM processes forward temporal information, while the backward LSTM processes reverse temporal information. It also includes hidden layers and an output layer. The hidden layer is initialized to 0 with a dropout rate of 0.2 to prevent overfitting and enhance the model's generalization ability. The output layer is a fully connected layer with one neuron, which directly outputs the SOC estimate using a linear activation function.

[0041] In this embodiment, the pre-training stage of the hybrid network estimation model is as follows: the model is trained on a public dataset with a batch size of 64, a learning rate of 0.001, and the optimizer Adam, where β1 is 0.9, β2 is 0.999, and the loss function is mean squared error; the model is iterated for 50 rounds, and a validation is performed every 5 rounds. The model parameters with the highest accuracy on the validation set are saved to ensure that the model fully learns the general features.

[0042] Transfer learning phase: Freeze the parameters of ResNet and BiLSTM units, and fine-tune only the fully connected layers; train on the self-test dataset, set the batch size to 32, reduce the learning rate to 0.0001, and iterate for 30 rounds to solve the overfitting problem of small datasets and quickly adapt to new battery types.

[0043] Experimental verification

[0044] (1) Experimental scheme

[0045] Hardware environment: CPU is AMD Ryzen 9 5900X, GPU is NVIDIA GeForce RTX 3080Ti, memory is 64GB, and operating system is Ubuntu 18.04.

[0046] Software environment: Based on PyTorch 1.9.1 framework, CUDA 11.2 acceleration, and programmed using Python 3.8.

[0047] Evaluation metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Maximum Error (ME), calculated using the following formulas:

[0048] ;

[0049] in, For the sample size, For predicted values, This is the actual value.

[0050] Comparison models: LSTM, ResNet, and Transformer, all employing the same training strategy.

[0051] (2) The experimental data are as follows:

[0052]

[0053] The experimental data above show that on the public dataset, the ME of ResNet-BiLSTM is only 1.49%, significantly lower than other models, proving that it has a better ability to fuse temporal and spatial features. After transfer learning, the ME of the self-test dataset decreased from 7.78% to 2.43%, solving the overfitting problem of small datasets and verifying the model's generalization ability. Under different initial SOC conditions, the ME does not exceed 3%, meeting the stringent requirements of energy storage devices for SOC estimation accuracy (usually allowing an error ≤5%).

[0054] S3: Construct a SOH estimation model that includes a Transformer encoder and data segmentation. Use data segmentation to segment the preprocessed multidimensional runtime sequence data, and then input the segmented data into the Transformer encoder to output the SOH estimate.

[0055] In step S3, the data segmentation is performed according to the battery cycle parameter variation law. Specifically, the number of battery charge and discharge cycles is first set and the cycle value is determined. Then, the preprocessed multidimensional runtime sequence data is divided into data segments of different lengths according to the cycle value. Among them, the long cycle Ta is set to 5 cycles, the medium cycle Tb is set to 3 cycles, and the short cycle Tc is set to 1 cycle.

[0056] In this embodiment, when the length of the last data segment is insufficient for the corresponding period value, it is repeatedly padded to the standard length with the last value, and then normalized and aligned according to the maximum and minimum values ​​within the data segment to enhance the ability to capture local time-series features.

[0057] In step S3, the Transformer encoder contains a four-layer structure, each consisting of a multi-head attention mechanism and a feedforward network. The multi-head attention mechanism has 8 heads, each with a dimension of 32, and uses Scaled Dot-Product Attention to calculate attention scores, thereby enhancing the capture of long-term temporal associations. The feedforward network has 128 neurons in its hidden layer and uses GELU as the activation function to improve the non-linear transformation capability of features.

[0058] Position coding uses the sine and cosine formulas, the specific formulas are as follows:

[0059]

[0060]

[0061] in This refers to the time step position in the time series data; Group indexes by dimensions; The model feature dimension is 256; this ensures that the model is aware of the temporal position of the data.

[0062] In this embodiment, the training parameters of the SOH estimation model are set as follows: batch size is 32, learning rate is 0.0005, optimizer is AdamW, weight decay is 0.01 to suppress overfitting, and loss function is mean squared error; iterates for 60 rounds, and the learning rate is multiplied by 0.5 every 10 rounds to decay, balancing training efficiency and accuracy.

[0063] Experimental verification

[0064] (1) Experimental scheme

[0065] Hardware environment: CPU is AMD Ryzen 9 5900X, GPU is NVIDIA GeForce RTX 3080Ti, memory is 64GB, data set is stored on 1TB SSD, and operating system is Ubuntu 18.04;

[0066] Software environment: Based on TensorFlow 2.4 framework, using AdamW optimizer, learning rate decay strategy (0.5 every 10 rounds).

[0067] Evaluation metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Maximum Error (ME), calculation formula:

[0068] ;

[0069] in, For the sample size, For predicted values, This is the actual value.

[0070] Comparison models: Crossformer, Transformer, FEDformer, comparing the performance differences before and after data segmentation.

[0071] (2) The experimental data are as follows:

[0072]

[0073] The experimental data above show that the ME is as low as 0.52% on the public dataset, proving that the data segmentation method is better than the traditional point-by-point input in capturing local temporal features. On the self-test dataset, the proposed method has an ME of 0.88%, which is 64% lower than Transformer (2.47%), verifying the encoder's ability to mine long-term temporal associations. When the amount of training data is reduced to 30%, the ME is still ≤1.55%, indicating that the model has strong adaptability to small sample data and is suitable for scenarios where data acquisition of energy storage devices is limited.

[0074] S4: Based on the SOC and SOH estimates, design a combined active and passive battery balancing strategy, and use a support vector machine model to generate a battery operation and maintenance strategy.

[0075] In step S4, the battery balancing strategy combining active and passive approaches is designed as follows: a multi-layer balancing circuit combining active and passive approaches is designed to balance balancing efficiency and stability. The specific scheme is as follows: the battery pack is divided into 3 modules, each containing 8 lithium iron phosphate batteries with a specification of 3.2V 2800mAh. A passive balancing strategy is adopted within the module, with a resistor calculated based on the battery's rated voltage and current connected in parallel to each battery cell. When the voltage difference between the individual cell and the average voltage of the module is greater than 50mV, the resistor is connected via a MOSFET switch. Specifically, each battery cell is connected in parallel with a 10Ω 2W resistor and an IRF540 MOSFET. The MOSFET gate is directly connected to the GPIO port of the FPGA, and the conduction state is controlled by a PWM signal to dissipate excess energy.

[0076] In the passive balancing strategy within the module, the resistor power is calculated using the following formula: ;in R is the maximum charge / discharge current of a single battery cell, and R is the resistance value.

[0077] The MOSFET switch is an N-channel enhancement type with an on-resistance of ≤10mΩ. It is controlled by an STM32 microcontroller outputting a PWM signal. When the voltage difference ΔV = 50mV, the duty cycle is 50%. For every 10mV increase in voltage difference, the duty cycle increases by 10%, with a maximum duty cycle of 80%.

[0078] An active balancing strategy is adopted between modules, using a 10mH inductor as an energy transfer element. When the voltage difference between modules is greater than 100mV, PWM control is used to control the inductor to transfer energy between high and low voltage modules. Specifically, a 220μF electrolytic capacitor and a 10mH inductor are connected in series as energy transfer elements, and the energy flow is controlled by a G6K-2P-Y relay. The relay coil is driven by the PWM signal of the FPGA, with an operating voltage of 5V and an operating current of 10mA, to achieve efficient energy transfer.

[0079] The control circuit for the active balancing strategy between modules is based on an STM32H743 microcontroller, which collects voltage data between modules every 10ms. When the voltage difference ΔV = 100mV, the duty cycle of the PWM signal is 30%. For every 50mV increase in voltage difference, the duty cycle increases by 15%, with a maximum duty cycle of 70%. The charging and discharging of the inductor is controlled by the IR2110 driver chip to achieve an energy transfer efficiency of ≥85%.

[0080] Equilibrium strategy and control logic:

[0081] Intra-module equalization: Individual cell voltage is acquired in real time via BQ76940 at a sampling frequency of 10Hz.

[0082] When the maximum voltage of the single cell and minimum value When the difference is greater than 50mV, the passive equalization strategy is activated to calculate the average voltage. .

[0083] For voltage higher The single unit, with duty cycle D equal to 0.5 multiplied by the current voltage minus Divide the difference by 50mV to output a 50kHz PWM signal until... and The difference is less than 25mV.

[0084] Inter-module balancing: The total voltage of the modules is collected by a combination of 100kΩ and 10kΩ voltage divider resistors, with an accuracy controlled within ±0.05V.

[0085] When the difference between the maximum and minimum module voltages is greater than 100mV, active balancing is initiated. The high-voltage module charges the capacitor through the inductor, and the PWM duty cycle D is equal to 0.3 multiplied by the difference between the high voltage and the low voltage divided by 100mV. Then, the capacitor discharges to the low-voltage module until the difference is less than 50mV.

[0086] (1) Simulation verification

[0087] Simulation environment: Matlab / Simulink2021a, using a supercapacitor (12F) to simulate battery characteristics, with the equalization capacitor set to 220μF, inductor to 10mH, and MOSFET switching frequency to 50kHz.

[0088] Evaluation metrics: Equalization time (time to reach a voltage difference ≤ 20mV), final voltage difference;

[0089] Experimental groups: Three sets of initial voltage conditions to simulate different inconsistent scenarios.

[0090] (2) The experimental data are as follows:

[0091]

[0092] The experimental data above shows that all three sets of experiments completed balancing within 5 seconds, with a final voltage difference of ≤20mV, which meets the requirements of energy storage devices for balancing speed (≤10s) and accuracy. The active-passive combination design takes into account both efficiency and stability: passive balancing within the module quickly dissipates small differences in energy, while active balancing between modules reduces energy loss, making it suitable for low-maintenance scenarios.

[0093] In step S4, a battery operation and maintenance strategy is generated using a support vector machine model. Specifically, the kernel function of the support vector machine algorithm is the radial basis function (RBF), the penalty parameter C is 10, and the kernel function parameter γ is 0.1.

[0094] When constructing the support vector machine model, the labeled historical charge and discharge data of the battery, historical fault records, real-time collected voltage data, real-time collected current data, real-time collected temperature data, SOC estimation results, and SOH estimation results are divided into training set and test set in an 8:2 ratio.

[0095] During training, parameters are adjusted using 5-fold cross-validation. Specifically, the training set is randomly divided into 5 subsets. One subset is selected as the validation set, and the other 4 subsets are selected as the training subsets for model training and validation. This process is repeated 5 times, and the average performance index is used to determine the optimal parameters.

[0096] The evaluation metrics for the support vector machine model on the test set must meet the following requirements: accuracy ≥ 95%, recall ≥ 90%, and F1 score ≥ 92%.

[0097] The generated battery operation and maintenance strategy is to trigger charging when the SOC value is below 20%, remind maintenance when the SOH value is below 80%, and issue a warning through an audible and visual alarm device when the model detects a voltage fluctuation of ±10%. The alarm volume of the audible and visual alarm device is 100 decibels and the warning light flashes 3 times per second.

[0098] When a voltage fluctuation of ±15% is detected, the emergency shutdown mechanism of the battery pack is triggered, and the connection between the battery pack and the main circuit is cut off through a solid-state relay. The response time of the solid-state relay is ≤1 millisecond. After the cut-off action is completed, a signal containing the fault type and timestamp is sent to the operation and maintenance terminal.

[0099] Example 2

[0100] This embodiment provides an energy storage battery operation and maintenance decision-making system based on multi-dimensional features, including the following modules:

[0101] The data acquisition module is configured to acquire multidimensional runtime sequence data of the battery and perform preprocessing.

[0102] The State of Charge (SOC) estimation module is configured to: construct a hybrid network estimation model containing ResNet and BiLSTM units, input preprocessed multidimensional runtime data into the hybrid network estimation model, and output SOC estimates.

[0103] The health status estimation module is configured to: construct a SOH estimation model that includes a Transformer encoder and data segmentation; use data segmentation to segment the preprocessed multidimensional runtime sequence data; then input the segmented data into the Transformer encoder and output the SOH estimate;

[0104] The strategy output module is configured to: design a combined active and passive battery balancing strategy based on the SOC estimate and SOH estimate, and generate a battery operation and maintenance strategy using the support vector machine algorithm.

[0105] In this embodiment, an energy storage battery operation and maintenance decision-making system is implemented based on FPGA, integrating state estimation and equalization control functions, as specifically implemented as follows:

[0106] Hardware Design and Connection

[0107] The main control core uses a ZCU104 FPGA, equipped with an XCZU7EV MPSoC, 4GB DDR4 memory, 16GB eMMC flash memory, and connects to the BQ76940 via an I2C interface with an interface rate of 400kHz to ensure efficient data transmission.

[0108] The data acquisition board uses the BQ76940 chip, with a built-in 14-bit ADC for voltage acquisition (0.625mV resolution) and a 16-bit ADC for current acquisition (0.01mA resolution). It also integrates a temperature sensor with an accuracy of ±0.5℃. The board is connected to the battery pack via DuPont wires, with each battery cell connected in series with a 10kΩ current-limiting resistor to protect the acquisition circuit.

[0109] Communication and display section: The FPGA connects to the server via a USB 3.0 interface with a transmission rate of 480Mbps; the server is equipped with a high-performance processor, 64GB of memory, and runs the Ubuntu system to realize data storage and remote monitoring.

[0110] Software Design and Development

[0111] The front end is developed using QT5.12.10 and includes a battery status panel that displays SOC, SOH, voltage, current, and temperature in real time; a data curve window that displays time-voltage and cycle-SOH curves; and a threshold setting interface that allows users to customize parameters such as a minimum SOC of 20% and a maximum temperature of 60℃.

[0112] The backend is written in Python 3.8 and uses PyQt5 for user interface interaction. The SQLite database stores one record every 10 seconds to ensure data integrity. The ONNX format SOC / SOH model is used for real-time estimation with a response time of less than 1 second.

[0113] FPGA algorithm implementation: Use Vivado 2021.2 to generate hardware circuit, add ZYNQIP core, and set pl_clk0 to 100MHz; use VitisAI 2.5 to quantize the model to INT8 precision, generate xmodel file, and boot from SD card to meet low power consumption requirements.

[0114] System testing

[0115] (1) Experimental scheme

[0116] Test environment: Laboratory environment (temperature 25±2℃, humidity 60±5%), continuous operation for 72 hours;

[0117] Evaluation metrics: parameter refresh rate, SOC / SOH estimation error, equalization response time, and data integrity (no data loss).

[0118] Test content: Functional testing (parameter display, curve plotting, threshold alarm) and stability testing (long-term operating performance).

[0119] (2) The experimental data are as follows:

[0120]

[0121] The experimental data above show that all system indicators meet the design requirements, and the SOC / SOH estimation accuracy is consistent with the algorithm verification results, proving the effectiveness of hardware and software integration. During long-term operation, there was no data loss and the FPGA temperature remained stable, verifying the reliability of the system in long-term operation and laying the foundation for actual deployment.

[0122] Various modifications and variations of this invention will be apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for operation and maintenance decision-making of energy storage batteries based on multi-dimensional features, characterized in that, include: The multidimensional runtime timing data of the battery is acquired and preprocessed. The acquired multidimensional runtime timing data of the battery includes the individual cell voltage, current, temperature, pressure and internal impedance data. A hybrid network estimation model containing ResNet and BiLSTM units is constructed. Preprocessed multidimensional runtime data is input into the hybrid network estimation model, and the SOC estimate is output. A SOH estimation model is constructed that includes a Transformer encoder and data segmentation. The preprocessed multidimensional runtime sequence data is segmented using data segmentation, and then the segmented data is input into the Transformer encoder to output the SOH estimate. A combined active and passive battery balancing strategy is designed based on the voltage difference threshold, and a battery operation and maintenance strategy is generated using a support vector machine model. Specifically, the battery balancing strategy combining active and passive approaches is as follows: the passive balancing strategy involves connecting a resistor calculated based on the battery's rated voltage and current in parallel to each battery cell; the active balancing strategy involves using an inductor as an energy transfer element and employing PWM control to control the inductor to transfer energy between high and low voltage modules. Among them, the battery operation and maintenance strategy is generated using the support vector machine model. Specifically, the kernel function of the support vector machine algorithm is the radial basis kernel function, the penalty parameter C is 10, and the kernel function parameter γ is 0.

1. When constructing the support vector machine model, the labeled historical charge and discharge data of the battery, historical fault records, real-time collected voltage data, real-time collected current data, real-time collected temperature data, SOC estimation results, and SOH estimation results are divided into training set and test set in an 8:2 ratio. During training, parameters are adjusted using 5-fold cross-validation. Specifically, the training set is randomly divided into 5 subsets. One subset is selected as the validation set and the other 4 subsets are selected as the training subsets for model training and validation. This process is repeated 5 times, and the average performance index is used to determine the optimal parameters. The evaluation metrics for the support vector machine model on the test set must meet the following requirements: accuracy ≥ 95%, recall ≥ 90%, and F1 score ≥ 92%. The generated battery operation and maintenance strategy is to trigger charging when the SOC value is below 20%, remind maintenance when the SOH value is below 80%, and issue a warning through an audible and visual alarm device when the model detects a voltage fluctuation of ±10%. The alarm volume of the audible and visual alarm device is 100 decibels and the warning light flashes 3 times per second. When a voltage fluctuation of ±15% is detected, the emergency shutdown mechanism of the battery pack is triggered, and the connection between the battery pack and the main circuit is cut off through a solid-state relay. The response time of the solid-state relay is ≤1 millisecond. After the cut-off action is completed, a signal containing the fault type and timestamp is sent to the operation and maintenance terminal.

2. The energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in claim 1, characterized in that, The preprocessing includes denoising, filtering, and normalization. The denoising uses a wavelet transform denoising algorithm, specifically decomposing the data using the db4 wavelet basis function, removing high-frequency noise, and then reconstructing the data. The filtering uses a moving average filtering method. The normalization uses a minimum-maximum normalization method.

3. The energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in claim 1, characterized in that, The ResNet unit contains two asymmetric convolutional modules, each consisting of a 1×3 convolutional layer and a 3×1 convolutional layer, with each convolutional layer followed by a BatchNorm layer and a ReLU activation function; the BiLSTM unit contains two bidirectional LSTM layers, namely a forward LSTM and a backward LSTM, and also contains hidden layers and an output layer, wherein the output layer is a fully connected layer.

4. The energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in claim 1, characterized in that, The data segmentation is specifically as follows: First, the number of battery charge-discharge cycles is set and the cycle value is determined. Then, the preprocessed multidimensional runtime sequence data is divided into data segments of different lengths according to the cycle value.

5. The energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in claim 4, characterized in that, When the length of the last data segment is insufficient for the corresponding period value, it is padded with the last value until the standard length is reached, and then normalized and aligned according to the maximum and minimum values ​​within the data segment.

6. The energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in claim 1, characterized in that, The Transformer encoder comprises a four-layer structure, with each layer consisting of a multi-head attention mechanism and a feedforward network.

7. A multi-dimensional feature-based energy storage battery operation and maintenance decision-making system employing the multi-dimensional feature-based energy storage battery operation and maintenance decision-making method as described in claim 1, characterized in that, Includes the following modules: The data acquisition module is configured to acquire multidimensional runtime sequence data of the battery and perform preprocessing. The State of Charge (SOC) estimation module is configured to: construct a hybrid network estimation model containing ResNet and BiLSTM units, input preprocessed multidimensional runtime data into the hybrid network estimation model, and output SOC estimates. The health status estimation module is configured to: construct a SOH estimation model that includes a Transformer encoder and data segmentation; use data segmentation to segment the preprocessed multidimensional runtime sequence data; then input the segmented data into the Transformer encoder and output the SOH estimate; The strategy output module is configured to design a combined active and passive battery balancing strategy based on the voltage difference threshold, and to generate a battery operation and maintenance strategy using a support vector machine model.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the energy storage battery operation and maintenance decision-making method based on multi-dimensional features as described in any one of claims 1-6.

Images