TWI825686B - Intelligent power management edge estimation system and construction method - Google Patents

Abstract

The present invention discloses a configured edge estimation device and a pre-computing function and prediction model for measuring voltage, current, and energy established through machine training and learning; based on a measurement established through machine learning technology The predictive model also includes an edge pre-computing function to further measure the power system. Therefore, the present invention is configured to measure one of the voltage measurement points of the voltage of a power supply system and one of the current measurement points of the current; in addition, based on measuring the energy consumption and operating temperature of the power supply system, the present invention summarizes The above parameters are predicted to understand and manage the charging status and health status of the power system, and then regulate or allocate the advanced charging and discharging behavior of the power system.

Description

Intelligent power management edge estimation system and construction method

The present invention relates to a device for determining the status of the power system, and in particular, to a device that establishes a real-time data edge computing model of the battery status to analyze, judge, learn and manage the SOC (state of charge) and battery health status of the power system. SOH (state of health) device.

In order to improve the operating efficiency of electric consumer equipment, the power management scheme (Energy management scheme) is extremely important for the battery charge state SOC (state of charge) and battery health state SOH (state of health) in electric equipment. On the one hand, in fact, the SOH and SOC of the battery itself are also important links that influence each other and determine whether the estimation is accurate. On the other hand, when further considering the consequences, when the SOH is too low and the battery can no longer be used as a battery for consumer devices, an urgent issue that must be solved will be how to deal with the increasing number of electric vehicles. To deal with the batteries used in a large number of retired electric equipment, if you want to reuse the decommissioned power sources in electric equipment or other applications, at this time, you can have an immediate and efficient estimation of SOC and SOH. Methods and systems are an important part of technology that determines safe, reliable and efficient charge and discharge control.

In terms of actual market application, among the estimation methods of lithium-ion battery SOH, the estimation methods that have been used or discussed in the past literature include "charge and discharge test method", and "battery health status estimation method" There are quite a lot of them at present, such as: electric voltage curve method, look-up table method, power slope method, battery model estimation method and data learning estimation method...etc.

In recent years, the processing of large amounts of data or big data calculations has been a very hot topic. Its applications can also be seen in battery health estimation, including artificial neural networks, support Support vector machines are common machine learning algorithms, and the data learning estimation method is based on a large amount of battery charge and discharge operating data. However, in the use of various of these methods, the estimation methods have not achieved real-time and optimization. This situation makes users who live a happy life have more worries about driving distances. .

Due to the advancement of battery manufacturing technology and materials, the cycle life has been greatly improved, or even the battery capacity and battery charge and discharge capabilities have been greatly improved, which has always been a very important topic in industry and academia. However, it is still difficult to accurately estimate the SOC and SOH of the current battery in the last mile, which makes it impossible for drivers to have varying degrees of distance anxiety due to various road conditions and estimated driving distance.

Generally speaking, the battery module of an electric consumer device can capture a variety of parameters through the physical layer of the battery management system (BMS), including, for example: battery capacity, voltage, current, operating temperature, charging FCC (Fully charged capacity), isolation membrane parameters, electrolyte parameters, positive electrode material, negative electrode material battery, SOH (State of health)... etc., whether through on-line or off-line ( Off-line) means to capture various parameters of the battery module, and determine the status of the battery module by monitoring and analyzing some of the parameters and their changes, all in order to improve the efficiency of the battery system and extend its life. Best method.

Current battery modules use the internal battery management system (Battery Management System, BMS for short) 8 to capture various parameters of the battery module 9, as shown in Figure 13. Most of them adopt an offline analysis and judgment process. Specifically, for example, when judging the battery health (SOH), it is calculated by dividing the current full charge of the battery by the full charge of a healthy battery. There are currently many different algorithms to obtain the current full charge of the battery. of full charge. This processing process basically uses a variety of parameters captured by a microcontroller unit (MCU), and then uses a specific algorithm to calculate the required battery status utilization results.

Furthermore, in the offline monitoring environment, the power management system has lost the ability to effectively grasp the operation of electric consumer equipment. It can be seen that the data obtained from its analysis is not credible enough and cannot effectively respond to the battery management system in real time. The appearance that appears.

In addition, even if the Internet of Things is set up, certain parameters of the battery system can be effectively collected through monitoring and analysis, as well as their changes to determine the status of the battery module. As shown in Figure 13, despite the characteristics of the Internet of Things, due to Cloud computing of the cloud server 10 cannot meet the described characteristics due to lack of location awareness, high latency, and lack of geographically distributed data centers close to IoT devices. Therefore, a novel distributed computing paradigm is needed. To be able to utilize these resources at the edge of the network, this distributed paradigm is what edge computing models.

In order for machines built on the battery management system to see, perform object detection, drive a car, understand speech, speak, walk, or otherwise imitate human skills, each of the above has execution processes that can operate through human thinking. , it is necessary to functionally replicate human intelligence.

The technical content disclosed in the present invention covers AIoT (Artificial Internet-of-Thing), artificial intelligence (AI), and edge computing (Edge computing) technologies.

Artificial intelligence uses a data structure called a deep neural network to replicate human cognition. These technologies, trained with deep neural networks (DNNs), can answer specific types of questions by showing many examples of that type of question and the correct answers. This training process, known as "deep learning," is typically run in data centers or cloud computing because training accurate models requires large amounts of data and requires data scientists to collaborate to configure the models. After training, the model becomes an "inference engine" that can answer real-world questions.

In edge AI deployments, the inference engine runs on some kind of computer or device in remote locations such as factories, hospitals, cars, satellites, and homes. When AI encounters a problem, the troublesome data is often uploaded to a cloud server to further train the original AI model, at some point replacing the inference engine at the edge. This feedback loop plays an important role in improving model performance; once edge AI models are deployed, they only get smarter and smarter.

In order to provide the above solution, the Internet of Things of the present invention combines edge computing and artificial intelligence technology to establish a neural-like learning module with a self-developed algorithm, and deploys it in the prediction and estimation of a microcontroller (MCU) with edge computing. In the measurement controller, real-time calculation results of various characteristic parameters of the battery of the power-consuming equipment are established, and a display method and device that can adapt to environmental changes and effectively and accurately display battery status and battery events are provided.

This invention aims to reduce the direct involvement of cloud servers by filtering and preprocessing data generated by an increasing number of sensors, delay-sensitive path selection, and other IoT devices connected to edge computing environments. Utilization of edge resources can lead to faster communication and task execution. On the other hand, it has been proven that the basic components of excellent edge computing will affect the computing performance of smart devices, especially for most neural networks.

The present invention also solves an interesting problem, which is to embed the framework of lightweight machine learning (TinyML) training into the intelligent power management edge estimation device and system proposed by the present invention. The system can obtain calculation results from edge computing components. Thereafter, another contribution disclosed by the present invention is to attempt to use newly designed edge computing components to evaluate the validity and reliability of the performance of intelligent power management edge estimation devices and systems.

Accordingly, the present invention proposes for the first time an intelligent power management edge estimation device and system, which is an innovative method and assembly of resource allocation model deconstruction. It not only manages the computing resources distributed in edge devices Distribution of sources and management independent of services executed on the cloud server side. Secondly, based on the configuration of the edge device model, a data transmission mechanism based on delay-sensitive selection of different paths was developed. In addition, big data can also be transferred to the cloud through path selection algorithms to change the storage method of the calculation results. Then, through extensive experiments and calculation results, we can evaluate how the previously calculated data should be transmitted and applied correctly. approach, these experiments demonstrate the effectiveness, scalability, and performance of the proposed model. The intelligent power management edge estimation device and system developed by this invention project not only embeds a framework trained by lightweight machine learning technology, but also sets many different sensors around the framework for collecting monitoring data.

Only when it can effectively utilize the energy provided by the battery system and intelligently analyze and utilize the collected parameters can it be said to be a modern high-tech electric consumption device. In order to overcome the failure to apply the collected data in an intelligent manner, the present invention uses these precious data collected over a long period of time as data samples for machine learning training, and selects the optimized neural network through experiments and measurements. The NN (Neural Network) module is used to infer the battery management system algorithm, and uses the deep learning (deep learning) training mechanism to finally achieve the highest accuracy and minimum loss rate requirements, and is first built to complete the neural-like judgment. module, complete the preliminary action of embedding this module into the power management system. Then load this valid module into a microcontroller (MCU) with edge computing capabilities, such as: SparkFun Edge, Arduino Nano BLE 33 Sense or STM32F746G... etc. MCU modules. Finally, after adjusting the algorithm, the module learned through neural network learning can be deployed in the microcontroller of the MCU, ultimately achieving the goal of real-time monitoring and power supply systems that can perform edge computing.

In order to achieve the above objectives and effects, the specific application technical means of the intelligent power management edge estimation device of the present invention is to provide a BMS system embedded or externally connected to a battery module, wherein the edge estimation device The device includes: a memory unit that pre-stores a charge and discharge algorithm instruction, a manager node demand algorithm instruction, a manager node configuration algorithm instruction, a path algorithm instruction and a lightweight machine learning instruction. Module instructions; an arithmetic logic unit electrically connected to the memory unit; an input-output unit electrically connected to the arithmetic logic unit, and the input-output unit provides several external sensors; and a communication unit electrically connected Sexually connected to the input/output unit or/and the operation logic unit; the operation logic unit performs execution and operation according to the instructions of the memory unit and the sensor signals received by the input/output unit including the voltage and current of the battery module. , transmission path selection or/and data upload, and then the execution result is outputted as a charge and discharge control signal of the BMS system through the input and output unit.

The input-output unit also receives sensor signals including the position distance and operating temperature of the battery module. The communication unit has wired communication and wireless communication.

The intelligent power management edge estimation system of the present invention provides a BMS system externally connected to several battery modules. The edge estimation system includes: a plurality of edge estimation devices, which use the aforementioned edge estimation devices, and the plurality of edge estimation devices. The estimation devices can be interconnected, wherein a first path is formed between each two edge estimation devices; and a cloud server is provided with a database containing training data and a lightweight machine learning method connected to the database. A training framework is provided, and the cloud server is interconnected with the plurality of edge estimation devices, and the training data includes voltage and current data of each battery module captured by the plurality of edge estimation devices, and the cloud server A second path is formed between each edge estimation device.

In addition, the construction method of the intelligent power management edge estimation system of the present invention includes: a step of establishing several edge estimation devices, and each edge estimation device is embedded or externally connected to The corresponding BMS system and several sensors of the battery module; a cloud server built The setting step is to build a cloud server and interconnect it with the plurality of edge estimation devices, wherein a first path is formed between each two edge estimation devices, and a first path is formed between the cloud server and each edge estimation device. a second path; a training data collection step, which collects the voltage and current data of the battery module from each edge estimation device and then uploads them to a database set up in the cloud server; an algorithm to build The configuration step is to build a charging and discharging algorithm, a manager node demand algorithm, a manager node configuration algorithm and a path algorithm in each edge estimation device; a lightweight machine learning training framework is constructed The step is to build a lightweight machine learning training framework in the cloud server, and the lightweight machine learning training framework interfaces with the training data of the database; and a lightweight machine learning module set In the setting step, a lightweight machine learning module trained by the lightweight machine learning training framework is also built in each edge estimation device.

[Invention]

A: Edge estimation device

1: Memory unit

2: Operational logic unit

3: Input and output unit

4: Communication unit

5: Sensor

B: Edge estimation system

6:Cloud server

61:Database

C: Construction method of edge estimation system

a: Several edge estimation device construction steps

b: Cloud server establishment steps

c: Training data collection steps

d: Algorithm construction steps

e: Steps to build lightweight machine learning training framework

f: Lightweight machine learning module configuration steps

[customary knowledge]

8a, 8b, 8c: Battery management system

9a, 9b, 9c: battery module

10:Cloud server

[Figure 1] An implementation diagram of the "intelligent power management edge estimation device" of the present invention.

[Figure 2] The architecture diagram of the "intelligent power management edge estimation device" of the present invention.

[Figure 3] Schematic diagram of the architecture and operation of the "intelligent power management edge estimation system" of the present invention.

[Figure 4] The step flow chart of the "charge and discharge algorithm" of the present invention.

[Figure 5] The step flow chart of the "Manager Node Requirement Algorithm" of the present invention.

[Figure 6] The step flow chart of the "Manager Node Configuration Algorithm" of the present invention.

[Figure 7] The step flow chart of the "path selection algorithm" of the present invention.

[Figure 8] The framework diagram of the present invention's "Support for vector regression through lightweight machine learning (TinyML) training".

[Figure 9] The flow chart of the steps of "lightweight machine learning training" of the present invention.

[Figure 10] Experimental comparison chart between the discharge curve of the present invention and the original discharge curve.

[Figure 11] Schematic diagram of the mean square error MSE of the entire experimental results of the present invention.

[Figure 12] The step flow chart of the "Construction Method of Intelligent Power Management Edge Estimation System" of the present invention.

[Figure 13] The architecture diagram of a general traditional battery management system (BMS).

In order to correctly demonstrate the practical feasibility of the innovative method of the intelligent power management edge estimation device and system of the present invention, please refer to Figures 1 and 2 for demonstration examples. The intelligent power management edge estimation device of the present invention mainly provides built-in or external A battery management system 8a of a battery module 9a, in which the edge estimation device A includes: a memory unit 1 that pre-stores a charge and discharge algorithm command, a manager node demand algorithm command, and a manager node Algorithm instructions, a path algorithm instruction and a lightweight machine learning module instruction are configured; an arithmetic logic unit 2 is electrically connected to the memory unit 1; an input-output unit 3 is electrically connected to the arithmetic logic Unit 2, and the input and output unit 3 provides several external sensors 5; and a communication unit 4 is electrically connected to the input and output unit 3 or/and the operation logic unit 2; through the operation logic unit 2 according to the memory The instructions of the unit 1 and the sensor 5 signals received by the input and output unit 3 including the voltage, current, position distance, operating temperature, etc. of the battery management system 8a of the battery module 9a are used for execution, calculation, transmission path selection, or/ and data upload, and then the execution result is outputted as a charge and discharge control signal of the BMS system through the input and output unit.

As for the intelligent power management edge estimation system of the present invention, as shown in Figure 3, a BMS system externally connected to several battery modules is provided. The edge estimation system B includes: a plurality of edge estimation devices A, which use the aforementioned An edge estimation device A, and the plurality of edge estimation devices A can be horizontally interconnected, wherein a first path is formed between every two edge estimation devices A; and a cloud server 6 is provided with a data including training data Database 61 and a lightweight machine learning training framework connected to the database 61, and the cloud server 6 is vertically interconnected with the plurality of edge estimation devices A, and the training data It includes the voltage and current data of each battery module captured by the plurality of edge estimation devices A, and a second path is formed between the cloud server 6 and each edge estimation device A.

Therefore, for the edge estimation device A of intelligent power management disclosed in the present invention, the environment can be regarded as connecting the Internet of Things via the wireless AP (access point) plus the edge service layer (IoT+Edge Computing layer) and the cloud service layer. (loud layer) devices, thus forming an interconnected environment that can communicate within the set network topology. The data transmission is based on the manager node and the local server transmitting big data through different paths, thereby entering the cloud server 6, thereby completing the intelligent power management edge estimator system to transmit the data obtained by the power system as a whole. In the case deployment in Figure 3, there are two worker nodes and a manager node, which are run as two hosts with the same performance consumption capabilities. In order to set up and evaluate the concept and implementation of the intelligent power management edge estimation proposed by the present invention, in the implementation work, an Arduino Nano 33 BLE microcontroller with built-in temperature sensing, humidity sensing and sound sensor is used. Controller (MCU) board. As shown in Figure 3, since the CPU speed of the Arduino Nano 33 BLE board is quite low and the RAM storage resources are limited, this lightweight-based operating system can perform intelligent power management edge estimation of the values captured by the edge, including, electric The voltage, current, operating temperature, and power consumption of energy-consuming equipment...etc.

As for how the data is stored and transmitted after the data is retrieved and processed by the edge computing algorithm, a more detailed description is as follows. Regarding an important beginning for operating the intelligent power management edge estimation device and system of the present invention, first The first setup step is to register the local access network (LAN) in the network file after refreshing the operating system to the memory card. After the operating system is successfully flashed, the upstream allocation of the network address (internet protocol, IP) space is 192.168.2.0/24. The complete actual deployment implemented in the system is shown in Figure 3, which involves voltage, current, operating temperature, and power consumption and corresponding effective sensors 5. These signal sensing devices are regarded as necessary tools for the Internet of Things layer. do. A node container runtime version is installed and run in the manager node and worker node. The edge layer shown in Figure 3 implies both the worker node and the manager node. In addition, the first working and manager node, the second working and manager node and the cloud server are also deployed in the cloud server 6 . In the scenario of implementation of intelligent power management edge estimation, modular self-propelled construction by Arduino MCU plays the roles of manager node and manager node respectively. These two paths are clearly shown in Figure 3 of the disclosure. Therefore, the first path and the second path correspond to two paths, one of which can be designated as a wired path, and the other can take the format of a wireless path.

Furthermore, the algorithm flow for operating the intelligent power management edge estimation device and system of the present invention is explained. This algorithm flow includes the pre-process of data collection, then the pre-processing process after data collection, and then the data training process. The framework description of the neural network finally reveals the calculation process of path selection for data transmission in the edge computing and cloud service layers of the present invention. These are disclosed in the demonstration expressions or processes of the present invention and are shown in Figures 4 to 4 respectively. Among 7.

Further, regarding the "variables and parameters" of the algorithm of the present invention: the monitoring system flow chart of the intelligent power management edge estimation device and system of the battery management system disclosed in the present invention is shown in Figure 4. The variables and parameters used are explained in Table 1 below:

That is, the real-time value of the remaining power is represented by SOC, the real-time value of the discharge power is represented by ID, IC represents the real-time value of the charging current, SOCH represents the high value of the remaining power, the low value of the remaining power is represented by SOCL, IDH is the high value of the discharge current limit, and the high value of the charging current limit. Value indicates ICH. When the entire battery management system starts to operate, the initial parameters of SOCH, SOCL, IDH and ICH are first set, and then the SOC, ID and IC data are measured and calculated in real time, and then the SOC value is determined. That is, the instant value of the residual power is compared with the low value of the residual power of the SOC. When the low value of the residual power is compared, when the SOC value is less than the SOCL value, the charging mechanism is started. Otherwise, the SOC value is compared with the SOCH value. When the SOC value is higher than the residual power value, the charging mechanism is started. When the power reaches a high value, the charging mechanism is turned off; otherwise, the instant value of the discharge amount is compared with the high value of the discharge current limit. If the conditions are met, the charging mechanism is started; otherwise, the IC, that is, the instant value of the charging current is compared with the IDH charging current limit. The high value is compared, and if the comparison result is established, the charging mechanism is turned off; otherwise, the supervision of judging whether the battery management system continues to be executed. If the execution is not continued, the supervision operation of the battery management system will be stopped, and the entire battery management system will be terminated, completing the entire process; if the execution command of the battery management system has not been reached, then continue to measure and calculate the SOC real-time Data calculation.

Further, the manager node demand calculation process of the intelligent power management edge estimation device and system is explained. As shown in Figure 5, after first registering with the manager node in the managed management area, it depends on whether the guidance of the manager node is obtained. If so, the passive manager node mark is maintained and waits for T seconds before sending the request signal; if no guidance from the manager node is obtained, the manager node mark is updated to the active manager node mark and waits for T seconds before sending the request signal.

In addition to the above-mentioned manager node demand process, the calculation process of the intelligent power management edge estimation device and system of the present invention also includes two configurations in order to allow the manager node to operate in a high availability mode. The algorithm for selecting the leader is shown in Figure 6. Worker nodes are first checked, and then the previous worker node is checked to see if there is an active manager node. If there is an active manager node, the new The manager node changes its status to inactive mode. If there is no active manager node, update the node's status to active. The edge computing environment has two manager nodes, and their status is checked through the signals sensed by the sensors. The algorithm for assigning combinations is shown in Figure 6. After that, the allocation differences are classified, check if the manager node is active, if it is active, the allocation is sent, the manager node is responsible for checking the work of the load node, when the load is no longer working, the allocation will be carried out by worker node to execute. If another task is to be further computed, it will be asked to check whether there are potential worker nodes. In the previously mentioned case, it is checked whether it is loaded and then the allocation is performed by the worker node. The research algorithm for resource allocation presupposes that the network is enhanced through signaling and resource reservation mechanisms that are designed to allocate different paths with some new streams.

Furthermore, during the calculation process of the intelligent power management edge estimation device and system, the path selection algorithm designed by it selects the highest minimum remaining bandwidth among the candidate paths of P seconds. Figure 7 shows the flow chart of the path selection algorithm. The path selection algorithm first determines whether the minimum excess bandwidth of path P is greater than the selected path. If not, the routing process is carried out along the selected path. If the candidate path P is selected as the selected path, then enter the (P+1) nest. After formulating the candidate path, it is judged whether the P path is greater than the number of the candidate path. If not, the routing process is carried out along the selected path. If so, the aforementioned judgment of "whether the minimum excess bandwidth of path P is greater than the selected path" is repeated.

Regarding the framework training and deployment of lightweight machine learning (TinyML) training for the intelligent power management edge estimation device and system of the present invention. The training and deployment work proceeds as shown in Figures 8 and 9. Figure 8 illustrates that the present invention selects the Support Vector Regression SVR (Support Vector Regression) neural algorithm to conduct framework training of intelligent power management edge estimation devices and systems. The data collected in this framework training comes from offline experimental data, supporting Vector regression generally fits a broad range of predictions and predictions, with a selection of linear and nonlinear regression types of curves. SVR is based on support vector machine SVM (Support vector machine) element, basically, where the support vector is a hyperplane in the feature space that is closer to the generated n-point dimension, and it significantly isolates the data points with respect to the hyperplane. The SVR model performs the equations broadly, and the equation for the hyperplane can be expressed as y=wX+b, where w is the weight and b is the intercept at X=0. The tolerance margin is represented by ε. You can use the SVM class of the Sklearn python library to import the SVR regression model.

Among the methods for SOC estimation, the most traditional method is to use the Kalman filter. Among the general SOC estimation methods, there are also DC impedance methods and Coulomb integration method, which is the most commonly used method. In addition, there is also a method using voltage measurement, because it can be achieved at an extremely low price and is accurate. The advantage of high degree. In addition, it also includes a linear model method, which can estimate SOC based on the battery status calculated from real-time data. Another one is the so-called impedance tracking method, which uses a set of models in the battery pack to correct the chemical capacitance for estimation. Until recently, neural network-like technology has made neural network-like technology faster due to hardware acceleration of computing speed. Lu's technology can be applied to the field of battery state estimation SOC, and there are many and successive papers or methods published in this field. In the intelligent power management edge estimation device and system for energy management disclosed in the present invention, the design is based on a neural network, in which the statistical method of supporting vector regression SVR is used to target the neural network model trained therein. The SOC of the battery is estimated. Before estimation, of course, the data must be pre-processed, and the SVR-type neural network is used to estimate the SOC of the battery according to the flow method and process in Figure 9.

When conducting SVR statistical analysis to estimate the SOC of the BMS, the BMS data is collected from the lithium iron battery through the human-machine interface through the CAN BUS or RS485 communication protocol on the BMS circuit board, in which the entire PACK is the data collected from the load generated by each cell unit (CELL). They include battery capacity, voltage, current, operating temperature, health Degree (State of health, SOH)…. These data are becoming the best background data for us to estimate SOC for BMS. Before developing a BMS-like neural network strategy, we must preprocess the data collected by BMS. In the pre-processing data processing, of course, we must first visually inspect the rationalization performance of the entire data, and then normalize it. After our BMS believes that the data obtained by the normalized processing is useful, efficient, and reasonable. Yes, then proceed to the training data of the data. Before training, the amount of data to be trained and tested on the neural network should be reasonably divided. In the neural network training in this report, 80% of the amount is used as training data, and other test data accounts for 20%. %. After determining the size of the training and testing data in this way, you can determine the type of kernel function to be used and then divide the data into N equal parts for cross-validation. The SVR used in this training uses the radial basis function RBF (Radial Basis Function) [https://www.ycc.idv.tw/ml-course-techniques_7.html] as the basis of the kernel function.

Divide the collected data into N equal parts for cross-validation, then set the parameters and use N - 1 equal parts of the data to train SVR, and then use the trained SVR model to predict the remaining data. When After the data is predicted, it is necessary to compare the accuracy of the model and obtain the output result of the best model. After that, it is judged whether the obtained result is optimal and has the expected accuracy; if it does not reach the accuracy The expected target of the degree of accuracy is repeated, and the N training times are repeated, and the expected result of accuracy is repeatedly judged whether it is the output of the best model. If it is time to reach the expected accuracy, of course it is to predict that the random test data can be completed, and then randomly take out 30 pieces of data for testing. In this continuous cycle of testing and training of training data, the expected neural network-like prediction model is finally achieved.

SVR model training and test results: The experimental battery charging method of the present invention adopts the constant-current (CC for short)/constant-voltage (CV for short) charging method. First, charge at 0.2C Charging is carried out at a current rate, and when the battery voltage reaches 54V, it switches to constant voltage charging. When charging at constant voltage, charging until the battery current reaches 0.01A is considered to be fully charged. After charging, let it sit for another hour before performing the battery discharge experiment. When performing a discharge experiment, the greater the discharge current, the greater the overvoltage of the battery, so the discharged power will be limited. Therefore, the maximum available capacity calculated by the present invention is the total amount of electricity released when the battery terminal voltage is discharged from 54V to the cut-off voltage of 2V at the latest 0.2C discharge rate.

Figure 10 shows the program writing process planned by the present invention to support the vector regression statistical model. The plan emphasizes that the entire data is used to develop the SVR model during the regression calculation process. In such data, the entire accumulated time is Training and testing are conducted through discharge. The discharge time is 3000 seconds, and the obtained SOC results are shown in Figure 10. It can be clearly seen in the figure that the original data uses a pink presentation method to express the discharge curve, while the green one is the result where the test data accounts for 30%. The two curves can be very close to each other, forming a SOC state that discharges first. Therefore, it can be found that the results trained by this SVR can indeed be used to test the prediction results of SOC.

In order to understand the SOC state prediction and discharge power of the battery No. 32700 lithium iron battery used in this invention, it is used as the prediction process of proposing the SVR model neural network statistical regression. Therefore, this project is based on the calculation of the so-called discharge power. It is to compare the multiplication of discharge voltage and current. It refers to the different ways of presenting the reference voltage and the trained SVR statistical regression model. It is found that the same curve properties can be achieved, but the way of presentation is indeed different. This patent is designed according to the program. The horizontal axis represents the time performance of the discharge for 3000 seconds. The left side of the central axis is the performance of the prediction results using SOC, and the left and right sides are the performance of the discharge power watts, and are presented in Figure 11 Among them, the pink one is the discharge process of the original data, and the black dots show the radiation results of the test data. After the two are consistent, the blue curve is used to express the discharge power process. The results presented are in a decreasing manner. Therefore, it is judged that the discharge pattern of the entire energy discharge after the load is added is quite consistent. However, as for the details, a small amount of correction must be made to understand the chemical change process of the entire lithium iron battery. whether it is functioning properly.

After training in the neural network SVR regression statistical model, 80% of the training data and 20% of the test data are used. In the process of program simulation, the calculation process of mean square error (MSE) is also used. The mean square error between testing and training is displayed in Figure 11 in the form of time on the horizontal axis and MSE on the vertical axis. It can be clearly seen that the maximum result of the entire mean square error MSE will not exceed 0.1. If the MSE shown in the results is compared with past research results, it can be seen that past research also used lithium iron batteries for SOC state estimation, and what they used was obtained by deep learning neural network DNN. The MSE result is 0.6. Therefore, it can be concluded that the results of the SVR statistical regression with the MSE developed by the institute show that the intelligent power management edge estimation device and system of the present invention have excellent advantages.

Therefore, the present invention summarizes the aforementioned intelligent power management edge estimation device and system, and can obtain the construction method C of the intelligent power management edge estimation system of the present invention, as shown in Figure 2, Figure 3 and Figure 12, which includes: one or more The edge estimation device construction step a is to build several edge estimation devices A, and each edge estimation device A is embedded or externally connected to the BMS system of the corresponding battery module and several sensors 5; a cloud server builds Setup step b is to build a cloud server 6 and interconnect it with the plurality of edge estimation devices A, wherein a first path is formed between each two edge estimation devices A, and the cloud server 6 is connected to each edge estimation device A. A second path is formed between the estimating devices A; a training data collection step c is to capture the voltage and current data of the battery module from each edge estimating device A, and then upload it to the cloud server 6 Assuming a database 61; an algorithm building step d, a charging and discharging algorithm, a charging and discharging algorithm, and an algorithm are built in each edge estimation device. A manager node demand algorithm, a manager node configuration algorithm and a path algorithm; a lightweight machine learning training framework building step e is to build a lightweight machine learning training framework on the cloud server In the server 6, the lightweight machine learning training framework interfaces with the training data of the database 61; and a lightweight machine learning module configuration step f is also constructed through the lightweight machine learning. A lightweight machine learning module trained by the training framework is installed in each edge estimation device A.

To sum up, the present invention is about an "intelligent power management edge estimation system and construction method", and its constituent devices, systems and construction methods have not been seen in books or publicly used. It sincerely meets the requirements for patent application. We sincerely invite you to apply. We sincerely hope that the patent will be approved as soon as possible. If it is necessary to clarify, the above are the specific embodiments of the present invention application and the technical principles used. If changes are made according to the concept of the present invention application, the As long as the resulting functional effects do not exceed the spirit covered by the description and drawings, they should be stated within the scope of the present invention application.

A: Edge estimation device

1: Memory unit

2: Operational logic unit

3: Input and output unit

4: Communication unit

5: Sensor

Claims (9)

An intelligent power management edge estimation system provides a BMS system embedded or externally connected to a battery module. It includes: a plurality of edge estimation devices, and the edge estimation device further includes a memory unit that pre-stores a charge and discharge algorithm instructions, a manager node demand algorithm instruction, a manager node configuration algorithm instruction, a path algorithm instruction and a lightweight machine learning module instruction; an arithmetic logic unit is electrically connected to the Memory unit; an input-output unit electrically connected to the arithmetic logic unit, and the input-output unit provides several external sensors; and a communication unit electrically connected to the input-output unit or/and the arithmetic logic unit; and The plurality of edge estimation devices can be interconnected, wherein a first path is formed between each two edge estimation devices; and a cloud server is provided with a database containing training data and a lightweight server connected to the database. A training framework for machine learning, and the cloud server is interconnected with the plurality of edge estimation devices, and the training data includes the voltage and current data of each battery module captured by the plurality of edge estimation devices, and A second path is formed between the cloud server and each edge evaluation device; the arithmetic logic unit receives sensor signals including the voltage and current of the battery module according to the instructions of the memory unit and the input and output unit. To perform execution, calculation, transmission path selection or/and data upload, and then output a charge and discharge control signal of the BMS system through the input and output unit. The intelligent power management edge estimation system as described in claim 1, wherein the charge and discharge algorithm instructions are set to first set the initial parameters of SOCH, SOCL, IDH and ICH, etc., and then perform real-time measurement and calculation of SOC, ID and IC data. , and then judge the SOC value. When the SOC value is less than the SOCL value, the charging mechanism is started. Otherwise, the SOC value and the SOCH value are compared. When the SOC value is higher than the high value of the remaining power, Just turn off the charging mechanism; otherwise, compare the instant value of the discharge amount with the high value of the discharge current limit. If the conditions are met, start the charging mechanism; otherwise, compare the IC, that is, the instant value of the charging current with the high value of the IDH charging current limit. If When the comparison result is established, the charging mechanism is turned off; otherwise, the supervision of judging whether the battery management system continues to be executed. The intelligent power management edge estimation system as described in claim 1, wherein the manager node configuration algorithm instructions are set to first check the working node, and then check whether the previous working node has an active manager node. If there is an active manager node, the new manager node changes its status to inactive mode; if the active manager node does not, update the status of the node to active node; after that, the allocated differences are classified and check whether the manager node is in Active state, if it is active, the allocation is sent. The manager node is responsible for checking the work of the load node. When the load is no longer working, the allocation will be performed by the worker node. The intelligent power management edge estimation system of claim 1, wherein the path algorithm instruction is set to determine whether the minimum excess bandwidth of path P is greater than the selected path. If not, then perform the routing process along the selected path. If yes, The candidate P path is the selected path. After entering the (P+1) nested candidate path, it is judged whether the P path is greater than the number of the candidate path. If not, the routing process will be carried out along the selected path. If so, the aforementioned "Path P" will be repeated. Judgment of whether the minimum excess bandwidth is greater than the selected path. The intelligent power management edge estimation system of claim 1, wherein the input and output unit of the edge estimation device also receives sensor signals including the position distance and operating temperature of the battery module. The intelligent power management edge estimation system of claim 1, wherein the communication unit of the edge estimation device has wired communication and wireless communication. The intelligent power management edge estimation system of claim 1, wherein the lightweight machine learning module instructions use a neural algorithm that supports vector regression. A method of building an intelligent power management edge estimation system includes: a step of building several edge estimation devices, and each edge estimation device is embedded or externally connected to a corresponding battery The module's BMS system and several sensors; a cloud server construction step is to build a cloud server and interconnect it with the plurality of edge estimation devices, wherein a first path is formed between each two edge estimation devices , and a second path is formed between the cloud server and each edge estimation device; a training data collection step is to capture the voltage and current data of the battery module from each edge estimation device, and then upload it Go to a database set up in the cloud server; an algorithm building step is to build a charging and discharging algorithm, a manager node demand algorithm, a manager node configuration algorithm and a Path algorithm; a lightweight machine learning training framework construction step is to build a lightweight machine learning training framework in the cloud server, and the lightweight machine learning training framework interfaces with the data The training data of the library; and a lightweight machine learning module configuration step, the system also builds a lightweight machine learning module trained through the lightweight machine learning training framework on each edge estimation device within. The method of building an intelligent power management edge estimation system as described in claim 8, wherein the lightweight machine learning training framework in the lightweight machine learning training framework building step uses a neural algorithm that supports vector regression. .