TWI658371B - A battery charging algorithm based on model predictive control - Google Patents

Abstract

A battery charging algorithm based on model predictive control is implemented by a control circuit. The battery charging algorithm includes the following steps: setting M charging option combinations, each charging option combination including a preset remaining battery power And N preset current values, where M and N are integers greater than 1. The first to N-1th preset current values of each of the charging option combinations are each subjected to a Coulomb integration operation to obtain N-1 predicted remaining battery power; a type of neural network operation is performed on the preset remaining battery power and the first preset current value of each of the charging option combinations using a type of neural network model to obtain a A first predicted battery temperature rise value, performing the neural network-like operation on the Kth preset current value and the Kth predicted battery remaining power to obtain a Kth predicted battery temperature rise value, K = 2 to N; the preset remaining battery power and N-1 the predicted remaining battery power are used to generate N remaining battery power increments according to each charging option combination; according to each charging option combination N said Performing a weighted scoring operation on the remaining battery power increase and the N predicted battery temperature rise values to obtain a score for each of the M charge charging combinations; and a sum corresponding to the highest of the M scores The combination of charging options determines a charging output current.

Description

A battery charging algorithm based on model predictive control

The invention relates to an algorithm for charging a secondary battery, in particular to an algorithm for charging a secondary battery based on a Model Predictive Control (MPC) combined with a neural network.

With the advancement of secondary battery related technologies, secondary batteries have been widely used in products such as mobile phones, notebook computers and tablet computers. Among secondary batteries, lithium ions can become the mainstream of secondary batteries due to their high energy density, long cycle life, light weight, no memory effect, high average operating voltage, and low self-discharge rate. In addition, many large electric energy storage systems have also begun to invest in research and development of large-capacity lithium-ion batteries.

However, in order to pursue fast charging, the charging current must be increased, which also increases the surface temperature of the lithium-ion battery and shortens the battery life. In other words, if the temperature rises during charging, the higher the loss, the lower the charging efficiency and the safety concerns.

Therefore, charging technology is very important for secondary batteries. Charging technology is related to the charging speed, charging efficiency, battery temperature rise, battery cycle life, and other factors of the secondary battery. There have been many studies exploring how to speed up the charging time, increase the charging efficiency, and reduce the charging temperature rise. The brief introduction is as follows:

At present, the most commonly used charging method is the Constant Current Constant Voltage (CC-CV) method. At the beginning of charging, the battery is charged using a constant current until the rated upper limit voltage of the battery is used to charge the battery. The advantage is that the method is simple, and the disadvantage is that the constant voltage charging time is longer. Many literatures have been improved. Some literatures use dual-loop control to obtain a charging curve similar to the constant-current and constant-voltage charging method without using a current sensor. This method is simple and lower in cost. Some literatures have proposed to use it at the beginning. Step-up constant current-constant voltage charging method for constant voltage charging above the upper limit of the battery, such as 4.3V, to switch to constant current and constant voltage charging method after the constant voltage period. This method can charge the battery to the rated level in a short time. 30% of the capacity, so the charging time is short; some literatures have proposed that the active charging state combined with the fuzzy control charging method and the gray prediction control battery charging system can charge more power in the same time; or some literatures have suggested that The phase-locked loop control-based charging method generates a phase error after comparing the reference phase with the output phase. The phase of the error is transmitted to the current source to generate a suitable current to charge the battery. In order to improve the disadvantages of the phase-locked loop in the constant voltage mode, Some literatures have proposed that the current pump battery charger uses a current pump for charging in the constant current mode, while using a pulsed battery in the constant voltage mode. Charging, the experimental results according to the literature, this method is more efficient than the constant charging current and constant voltage charging method, the charging time is also similar overall.

Because model predictive control has the advantages of being applicable to a variety of models, and taking environmental factors and safety into consideration through a scoring function, it can be used to reduce the temperature rise during battery charging and thus improve charging efficiency. Therefore, there is an urgent need in this field. A novel battery charging algorithm.

One object of the present invention is to disclose a battery charging method based on Model Predictive Control (MPC), which combines a neural network-like operation to establish a battery temperature rise model and model predictive control to select a suitable charging current to reach a region Optimal control to reduce temperature rise during battery charging.

Another object of the present invention is to disclose a battery charging method based on model predictive control, which combines a neural network-like operation to establish a battery temperature rise model and model predictive control to select a suitable charging current to achieve the optimal regional control for Shorten the charging time of the battery and improve the charging efficiency.

Another object of the present invention is to disclose a battery charging method based on model predictive control. Compared with the constant current and constant voltage charging method of the conventional technology, the charging time is reduced by 2.8%, and the average temperature rise value is also reduced by 14.45%.

Another object of the present invention is to disclose a battery charging method based on model predictive control, which can extend the cycle life of a battery by reducing the temperature rise during battery charging and shortening the battery charging time.

In order to achieve the aforementioned purpose, a battery charging algorithm based on model predictive control is proposed, which is implemented by a control circuit. The battery charging algorithm includes the following steps: setting M charging option combinations, each charging option combination Each includes a preset remaining battery power and N preset current values, and M and N are integers greater than 1; (step a); the first to the N-1th of each of the charging option combinations are described The preset current value is respectively subjected to a Coulomb integral operation to obtain N-1 predicted remaining battery power; (step b); a type of neural network model is used for the preset remaining battery power of each of the charging option combinations and The first preset current value is subjected to a type of neural network operation to obtain a first predicted battery temperature rise value, and the K-th preset current value and the K-th predicted battery remaining power are performed as described above. Neural network-like operation to obtain a K-th predicted battery temperature rise value, K = 2 to N; (step c); the preset remaining battery power, N-1, according to each of the charging options The predicted remaining battery power generates N batteries Increment of remaining power; (step d); performing a weighted scoring operation based on the N remaining battery power increments and N predicted battery temperature rise values of each of the charging options to make M the charging options Each combination obtains a score; (step e); and determines a charging output current according to a combination of the charging options corresponding to the highest of the M scores; (step f).

In an embodiment, the weighted scoring operation is obtained by multiplying a sum of N remaining battery power increments by a weight value a plus a sum of N predicted battery temperature rise values by (100-a). In the score, a is a positive integer less than 100.

In one embodiment, the weight value a is a variable weight value.

In one embodiment, the neural network model is a reverse-transmission neural network with a four-layer architecture, including one input layer, two hidden layers, and one output layer.

In one embodiment, the input parameters of the input layer are the preset current value and the preset remaining battery power. The two hidden layers each contain 35 neurons. The output parameters of the output layer are The predicted battery temperature rise value.

In one embodiment, it further includes a human-machine interface written by a LabVIEW program to monitor a temperature change.

In order to enable your reviewers to further understand the structure, characteristics, and purpose of the present invention, drawings and detailed descriptions of the preferred embodiments are attached below.

Please refer to FIG. 1, which illustrates a flowchart of steps of a battery charging algorithm based on model predictive control of the present invention.

As shown in the figure, the battery charging algorithm of the present invention based on model predictive control is implemented using a control circuit. The battery charging algorithm includes the following steps:

Set M pen charging option combinations, each of said charging option combinations includes a preset remaining battery power and N preset current values, and M and N are integers greater than 1; (step a);

Performing a Coulomb integration operation on each of the first to N-1th preset current values of each of the charging option combinations to obtain N-1 predicted remaining battery power; (step b);

Use a type of neural network model to perform a type of neural network operation on the preset remaining battery power and the first preset current value of each of the charging option combinations to obtain a first predicted battery temperature rise value. Perform the neural network-like operation on the Kth preset current value and the Kth predicted battery remaining power to obtain a Kth predicted battery temperature rise value, K = 2 to N; (step c );

Generating N battery remaining power increments according to the preset remaining battery power and N-1 the predicted remaining battery power of each of the charging option combinations; (step d);

Perform a weighted scoring operation based on the N remaining battery power increments of the charging option combination and the N predicted battery temperature rise values to obtain a score for each of the M charging option combinations; (step e) ;

And determining a charging output current according to a combination of the charging options corresponding to the highest one of the M scores; (step f).

Wherein, the weighted scoring operation is obtained by multiplying a sum of N remaining battery power increments by a weight value a plus a sum of N predicted battery temperature rise values by (100-a) to obtain the score. , A is a positive integer less than 100.

In another embodiment, the weight value a is a variable weight value, such as, but not limited to, an average current is calculated every three sampling times, and then the weight value a is adjusted according to the interval in which the average current is located. If the current is too large, the weight value will be decreased. If the current is too small, the weight value will be increased.

This type of neural network model is a backward-transmission neural network with a four-layer architecture, including one input layer, two hidden layers, and one output layer. The input parameters of the input layer are the preset current. Value and the preset remaining battery power, the two hidden layers each include 35 neurons, and the output parameter of the output layer is the predicted battery temperature rise value.

The following will explain the principle of the present invention:

Model predictive control Significance:

Model Predictive Control (MPC) was developed in the 20th century. Because many application scenarios in the industry are nonlinear and time-varying, it is relatively difficult to build accurate models. Known control technologies such as Proportional-Integral and Derivative (PID) control and modern control theory are difficult to obtain ideal control results.

Model predictive control is a multi-step prediction, scoring function, and feedback correction strategy. It has the advantages of good control effect and robustness. Therefore, it is widely used in industrial control. Success stories.

Please refer to FIG. 2a to FIG. 2c together, where FIG. 2a is a schematic diagram of the model predictive control, FIG. 2b is a schematic diagram of the control signal of the model predictive control, and FIG. 2c is a schematic diagram of the output signal of the model predictive control.

As shown in the figure, the model predictive control is to compare the predicted output with the expected value to obtain the prediction error and optimize the output to adjust the output. Therefore, the predictive control can predict the process output based on the system's current control input and historical information of the process. Future value. Among them, k is the sampling time point, u is the input, w is the curve of the expected output, y is the predicted output, and H _p is the length of the predicted steps. The error is obtained by subtracting the y and w sequences. The optimizer can decide the next input.

The algorithms for predictive control include model algorithmic control (MAC), dynamic matrix control (DMC) based on non-parametric models, and general predictive control (GPC) based on parametric models. .

The present invention adopts a General Predictive Control (GPC) based on a parameter model:

Generalized predictive control was invented by French scholar J. Richalet. It optimizes the performance of predictive control through the scoring function J, and combines the mechanism of self-correction with low robustness and low model requirements.

In prediction theory, a model describing dynamic behavior is needed, which can predict the future output through the system's historical output and future input. The generalized predictive control model can obtain the prediction model through the controlled regression integral moving average model, as shown in equation (1) .

(1)

Among them, y is an output, u is an input, and ε is a white noise sequence with a mean value of zero, and A, B, C, and Δ are as shown in equation (2).

(2)

For convenience, let c (z ^-1 ) = 1. In order to use equation (1) to obtain the output after step j, the Diophantine equation needs to be introduced, as shown in equation (3).

(3) where

Multiply both sides of the equation (1) by E _j (z ^-1 ) Δ, and then arrange the equation (3). Substituting it into j steps, the prediction equation (4) is shown.

(4)

The future output forecast can ignore the influence of noise, and let E _j (z ^-1 ) B _j (z ^-1 ) = G _j (z ^-1 ), then equation (4) can be rewritten as shown in equation (5) .

(5)

G _j (z ^-1 ) can be rewritten as shown in equation (6).

(6)

Among them, the equation (6) is composed of a known quantity and an unknown quantity, and f (k + n) is used to represent the known quantity, and the term is rewritten as shown in equation (7).

(7)

Rearranged into a matrix form, it can be written as shown in equation (8).

(8)

among them,

Sort equation (5) into equation (9)

(9) of which

In order to enhance the robustness of the system, a scoring function needs to be established, in which the influence of the current control u (k) on the future output of the system needs to be considered. The scoring function is shown in equation (10).

(10)

The latter term is mainly used to suppress excessively drastic control increments to prevent the system from exceeding the limit range, where n is the predicted length and is generally greater than the order of B (Z ^-1 ); m is the control length and m n; l is a non-zero weighting coefficient, usually let l (j) = l.

In order to make the predicted output approach the expected output, the expected curve is shown by equation (11).

(11)

Among them, y _r is a set value, y (k) is an output function, and w (k) is an expected curve.

If equation (10) is expressed as a matrix type, equation (12) is obtained.

(12)

Among them, [] ^T table Transposed Matrix, and then substituting equation (9) into equation (13).

(13)

The smaller the scoring function J, the better. Equation (14) is obtained.

(14)

However, only the first input is used in actual control, such as equation (15).

(15)

Where g ^T is In the first line.

Because the optimization goal of predictive control will go over time, there are local optimization goals at every moment. The optimization process is not performed offline, so it can repair and reduce deviations according to the time-varying, non-linear or interference problems of the model.

In summary, model predictive control has the following advantages: it can be flexibly applied to various models, such as linear or non-linear systems; it can take into account the situation of multiple variables; various restrictions can be added to the scoring function and it is easy to modify, such as the environment Factors, safety issues.

The invention combines the operation of a neural network : the operation of a neural network is to continuously adjust the weights and partial weights between nodes so that the calculated output value is the target output value, and the purpose is to enable the neural network to map out Correct input-output relationship mode.

Please refer to FIG. 3, which illustrates a schematic diagram of a reverse transfer neural network architecture used in the present invention.

As shown in the figure, the present invention uses a four-layer inverted transfer neural network architecture, including one input layer, two hidden layers, and one output layer.

The number of nodes in the input layer and the output layer in a neural network is determined according to the dimensions of the input parameters and output parameters. The number of neurons in the hidden layer must be determined by trial and error. The parameters are the preset current value and the preset battery remaining power. The two hidden layers each include 35 neurons, and the output parameter of the output layer is the predicted battery temperature rise value.

Neural network-like training estimates are made by changing input parameters and output parameters. If the training samples of the neural network are complete enough, it can be done when any reasonable range of data is input to the neural network such as learning completion. Make proper judgments and produce near correct output results.

When performing a neural network-like operation, the input value and output value processed must be scaled to the range of -1 to 1, and the role of the transfer function is to limit, compress, or process its nonlinear relationship to achieve non-linearity. The linear multiplication operation is input to the next neuron layer. In the present invention, the tangent double bending (Tansig) function is used as the transfer function, which is a conventional technique and will not be described again.

Please refer to FIG. 4, which illustrates a schematic diagram of an experimental platform constructed using a neural network-like operation according to the present invention.

As shown in the figure, the experimental platform is to place the battery in a constant temperature box and control it at 25˚C, then measure the temperature of the battery through a USB NI9211 temperature picker and record the temperature curve through the LabVIEW human-machine interface. After the battery is discharged to the cut-off current, it is used for charging test.

When using a neural network to estimate the temperature rise of the side battery charge, in order to make the estimation more accurate, the experimental platform of the present invention measures it every 1 minute, and the charging range is from 0.5C to 1.0C, with every 0.1C as an interval A total of 6 kinds of charging currents and the corresponding remaining battery power are used, and the battery temperature is individually recorded until the end of the constant current charging phase.

The distribution of neural network-like data is shown in Table 1. There are 331 data in total, of which 307 are training data, and 4 charges are obtained by each charging current in an even distribution manner, and a total of 24 are test data and input it to After training, the neural network compares the output with the actual value to verify the correctness of the neural network.

Table 1 Training materials (307 results) Test data (24 records) 0.5C 0.6C 0.7C 0.8C 0.9C 1.0C 0.5C 0.6C 0.7C 0.8C 0.9C 1.0C 84 64 52 43 35 29 4 4 4 4 4 4

The implementation steps of neural network are as follows:

1. Establishing training database Please refer to FIG. 5, which shows a training database of 0.5C to 1.0C charging current of a neural network such as the present invention.

As shown in the figure, the present invention collects 307 parameter data collected from a charging current of 0.5C to 1.0C, which is represented by a matrix of [2´307]. Each parameter data has a corresponding output battery temperature rise value, that is, output The target matrix is [1´307].

After training the neural network, it will form a [2´24] test data in an evenly distributed manner, and then input it to the trained neural network to test. The temperature rise of the battery output of the neural network is also one. The output matrix of group [1´24] is compared with the temperature rise value obtained from the real test to verify the correctness of the parameters of the neural network.

Second, build a backward transitive neural network model

In the present invention, a neural network graphic interface (Graphic User Interface, GUI) provided by MATLAB is used to establish a reverse transfer neural network, and training and simulation verification are performed based on this. The training function uses the Trainlm (Levenberg-Marquardt) function, and The Mean Squared Error (MSE) function is used as a comparison index. Among them, the Trainlm (Levenberg-Marquardt) function is an algorithm that combines the steepest descent method and Newton method to overcome the slow convergence and easy to fall into. The shortcomings of local minimums can greatly reduce the number of iterations of network training. This is a known technique and will not be described again.

After the matrix classification and network parameter settings are made, a backward transfer neural network model can be established, and then the neural network-like training is performed.

Please refer to FIG. 6, which illustrates a training convergence diagram of a neural network such as the present invention.

As shown in the figure, after the 4th iteration training, the mean square error value is 0.00037911. Please refer to FIG. 7, which illustrates the weights and partial weights of neural networks such as the present invention.

After training the neural network, the weight and partial weight of the neural network can be obtained. As shown in the figure, the weights are iw (1,1), lw (2,1), lw (3,2) ; Partial weights are b (1), b (2), b (3).

Third, verify the completion of the inverted transfer neural network

In order to verify the correctness of the reverse transfer neural network, the temperature rise input parameters in the test matrix are not included in the training parameters. The test matrix is input into the neural network, and then the output of the neural network and the real battery are used. The temperature rise values are compared, and the results are summarized in Table 2 to verify the accuracy of the neural network.

 SOC (%) C rate  5  20  35  50  1C  0.505449  0.117151  0.065983  0.061198  ANN-output  0.55261  0.13239  0.054729  0.056403

Please refer to FIG. 8, which illustrates a bar graph of test data errors of the present invention.

Figure 8 shows the actual temperature rise of Table 2 and the temperature rise estimated by the neural network-like value. The absolute value is taken, and the error is arranged into a bar graph. As shown in the figure, the test data and the neural network-like value are obtained The maximum error is close to 0.05 ° C, so it can be verified that the temperature rise of the battery can be obtained through a neural network-like operation.

Please refer to FIG. 9, which illustrates a man-machine interface for monitoring temperature changes according to the present invention.

As shown in the figure, the present invention further includes a human-machine interface written by a LabVIEW program to monitor a temperature change, and import the weights and partial weights obtained after the training of the neural network, and then the calculation results of each layer need to pass through the tangent The double bending (Tansig) transfer function is used to normalize the normal output value after the transfer, and the battery temperature rise value can be obtained.

The invention performs a Coulomb Counting Method operation on the preset current value during the charging process to obtain the predicted remaining battery power, and records the temperature through a shift register (not shown in the figure). Appreciation information. Among them, the Coulomb integration method is to accumulate the current and time to obtain the amount of power that the battery has consumed or replenished in the state of use. It is a conventional technology and will not be repeated here.

The increment of the remaining battery power and the temperature rise value are used as scoring conditions. The scoring program performs a weighted scoring operation on the increment of the remaining battery power and the temperature rise value, and is selected as the highest score in the score. Charging output current. The weighted scoring operation is obtained by multiplying a sum of N remaining battery power incremental SOCs by a weight value a plus a sum of N predicted battery temperature rise values DT by (100-a) to obtain the score. , A is a positive integer less than 100, as shown in equation (16).

(16)

The present invention proposes an MPC-ANN charging method, wherein the weight value a is, for example, but not limited to, a constant 65.

The present invention further proposes an MPC-ANN-CS charging method, wherein the weight value a is a variable weight value, and its variation criterion is, for example, but not limited to, the average current is calculated every three sampling times, and the sampling time is for example but not The limit is 120 seconds, and the weight is adjusted according to the interval where the average current is located. If the current is too large, the weight is reduced, and if the current is too small, the weight is increased.

The parameter settings of the MPC-ANN charging method and the MPC-ANN-CS charging method proposed in the present invention are shown in Table 3.

table 3 MPC-ANN MPC-ANN-CS Sampling time Weight value a Sampling time Weight value a Variation Δα1 Variation Δα2 Variation Δα3 Variation Δα4 120 (seconds) 65 120 (seconds) 65 5 1.25 -5 -10

Experimental results and comparison of the present invention and the conventional technology:

In the following, experimental tests will be carried out for the MPC-ANN charging method, MPC-ANN-CS charging method and the constant current and constant voltage charging method and the five-phase constant current method proposed in the present invention, and the voltage and current of each charging method will be recorded. And the temperature rise value, and compare the charging speed and temperature rise changes to verify the feasibility and performance improvement of the present invention.

The lithium battery used in the experimental test is the NCR18650B lithium-ion battery produced by PANASONIC. Its specifications are shown in Table 4.

Table 4 Rated Capacity 3350 mAh Minimum rated capacity 3200 mAh Rated voltage 3.6 V Cut-off voltage 2.5 V Standard charging conditions CC-CV, 1625 mA, 4.2 V, 4.0 hr Energy Density 676 Wh / l size 18.5 mm (diameter), 65.3 mm (height) weight 48.5 g Charging temperature 0 ˚C to +45 ˚C Discharge temperature -20 ˚C to +60 ˚C Static temperature -20 ˚C to +50 ˚C

Please refer to FIG. 10a to FIG. 10e together, wherein FIG. 10a shows the charging voltage, current, and temperature rise curve of charging at 0.8C during the constant current and constant voltage charging legal current phase of the conventional technology, and FIG. The fifth stage of technology is the charging voltage, current and temperature rising curve of constant current charging. Figure 10c shows the charging voltage, current and temperature rising curve of the MPC-ANN charging method of the present invention, and Figure 10d shows the MPC-ANN-CS charging method's charging voltage, current and temperature rising curve. Figure 10e shows the comparison of the temperature change of the battery caused by the above charging methods.

As shown in the figure, the temperature rise value of the MPC-ANN charging method of the present invention is similar to the constant current charging method 0.7C charging method; the temperature rising value of the MPC-ANN-CS charging method is about 0.7C and 0.8C charging method.

Therefore, the charging speed and temperature rise value will be compared with the charging time, and the charging time can be easily compared. However, the battery temperature rise value can only be compared with the highest value. In this way, the overall temperature rise change during the charging process cannot be seen, so the present invention points The temperature rise curve area is convenient for comparison, and because the time for each charging current to reach the CV section is different, the shorter integration time is selected to ensure fairness.

The charging time and the maximum temperature rise required for various charging methods to fully charge the battery are shown in Table 5.

table 5 Charging method Charging method Charging method of the present invention Constant current and constant voltage charging method (CC-CV) Five-phase constant current MPC- ANN MPC- ANN- CS 0.6C 0.7C 0.75C 0.8C 0.9C 1C Charging time (seconds) 7580 7022 6860 6616 6318 6103 6138 6909 6663 Temperature rise (° C) 2.45 3.21 3.66 4.07 4.7 5.7 5.66 2.993 3.4318

The comparison of the average temperature rise values of the full-charge battery of the constant current and constant voltage charging method of the present invention and the conventional technology is shown in Table 6.

Table 6 Charging method time Charging method of the present invention Constant current and constant voltage charging method (CC-CV) MPC-ANN MPC-ANN-CS 0.7C 0.75C 0.8C 57 (minutes) 2.255 ° C 2.5072 ° C 48 (minutes) 2.3261 ° C 2.3873 ° C 2.7190 ° C 3.0643 ° C

As can be seen from Tables 5 and 6, the charging speed of the MPC-ANN charging method of the present invention is 1.6% faster than the constant current and constant voltage charging method of the conventional technology, which is 1.6%, and the average temperature rise value is 0.252 ° C lower. The MPC-ANN of the present invention -The charging speed of the CS charging method is 2.9% faster than the constant current constant voltage charging method of conventional technology 0.75C, and the average temperature rise value is 0.393 ° C lower. Compared with the constant current constant voltage charging method of conventional technology 0.7C, the charging speed is faster. 5.4%, the average temperature rise is less than 0.0612 ° C.

With the design disclosed above, the present invention has the following advantages:

1. The battery charging method based on model predictive control disclosed in the present invention combines a neural network-like operation to establish a battery temperature rise model and model predictive control to select a suitable charging current to achieve the optimal area control to reduce the battery charging time. The temperature rises.

2. The battery charging method based on model predictive control disclosed in the present invention combines a neural network-like operation to establish a battery temperature rise model and model predictive control to select a suitable charging current to achieve the optimal area control to shorten the battery charge. Time to improve charging efficiency.

The disclosure of the present invention is a preferred embodiment, and any change or modification that is partly derived from the technical idea of the present invention and easily inferred by those skilled in the art will not depart from the scope of patent rights of the present invention.

To sum up, the present invention, regardless of the purpose, means and effect, is showing its technical characteristics that are quite different from the conventional ones, and its first invention is practical, and it also meets the patent requirements of the invention. Pray for granting patents at an early date.

Step a‧‧‧Set M charging option combinations, each charging option combination includes a preset remaining battery power and N preset current values, and M and N are integers greater than 1.

Step b‧‧‧ performs a Coulomb integration operation on each of the first to N-1 preset current values of each of the charging option combinations to obtain N-1 predicted battery remaining power

Step c‧‧‧ uses a type of neural network model to perform a type of neural network operation on the preset remaining battery power and the first preset current value of each of the charging option combinations to obtain a first prediction Battery temperature rise value, performing the neural network-like operation on the Kth preset current value and the Kth predicted battery remaining power to obtain a Kth predicted battery temperature rise value, K = 2 to N

Step d‧‧‧ generates N remaining battery power increments according to the preset remaining battery power and N-1 the predicted remaining battery power of each of the charging option combinations

Step e‧‧‧ perform a weighted scoring operation based on the N remaining battery power increments and N predicted battery temperature rise values of each of the charging option combinations to obtain a score for each of the M charging options combinations

Step f‧‧‧ determines a charging output current according to the charge option combination corresponding to the highest of the M scores

FIG. 1 is a flowchart of steps of a battery charging algorithm based on model predictive control according to the present invention. FIG. 2a is a schematic structural diagram of model predictive control. FIG. 2b is a schematic diagram of control signals for model predictive control. FIG. 2c is a schematic diagram showing output signals of the model predictive control. FIG. 3 is a schematic diagram of a reverse transfer neural network architecture used in the present invention. FIG. 4 is a schematic diagram of an experimental platform constructed using a neural network-like operation according to the present invention. FIG. 5 shows a training database of a charging current of 0.5C to 1.0C for a neural network such as the present invention. FIG. 6 illustrates a training convergence diagram of a neural network such as the present invention. FIG. 7 illustrates the weights and partial weights of neural networks such as the present invention. FIG. 8 shows a bar graph of the test data error of the present invention. FIG. 9 illustrates a human-machine interface for monitoring temperature changes according to the present invention. FIG. 10a shows the charging voltage, current, and temperature rise curves of the conventional technique of Constant Current Constant Voltage (CC-CV) charging at 0.8C. FIG. 10b shows the charging voltage, current, and temperature rise curves of the five-phase constant current charging of the conventional technology. FIG. 10c shows the charging voltage, current, and temperature rise curves of the MPC-ANN charging method of the present invention. FIG. 10d shows the charging voltage, current, and temperature rise curves of the MPC-ANN-CS charging method of the present invention. FIG. 10e shows a comparison of temperature changes of the batteries caused by the above charging methods.

Claims (6)

A battery charging algorithm based on model predictive control is implemented using a control circuit. The battery charging algorithm includes the following steps: setting M charging option combinations, each charging option combination including a preset remaining battery power And N preset current values, M and N are both integers greater than 1. The first to N-1th preset current values of each of the charging option combinations are respectively subjected to a Coulomb integration operation to obtain N-1 predictions of remaining battery power; using a type of neural network model to perform a type of neural network operation on the preset remaining battery power and the first preset current value for each of the charging option combinations A first predicted battery temperature rise value, performing the neural network-like operation on the Kth preset current value and the Kth predicted battery remaining power to obtain a Kth predicted battery temperature rise value, K = 2 to N; N preset remaining battery power and N-1 predicted battery remaining power are used to generate N remaining battery power increments according to each charging option combination; Performing a weighted scoring operation on the N remaining battery power increase and N predicted battery temperature rise values of the item combination to obtain a score for each of the M charging combination options; and the highest of the M pieces of the ratings Each of the charging option combinations corresponding to a branch determines a charging output current. The battery charging algorithm based on model predictive control as described in item 1 of the scope of the patent application, wherein the weighted scoring operation is calculated by multiplying the sum of the remaining remaining battery power increments by a weight value a plus N said The sum of the predicted battery temperature rise values is multiplied by (100-a) to obtain the score, a is a positive integer less than 100. The battery charging algorithm based on model predictive control as described in item 2 of the scope of patent application, wherein the weight value a is a variable weight value. According to the battery charging algorithm based on model predictive control described in item 2 or item 3 of the scope of the patent application, this type of neural network model is a four-layer reverse transfer neural network, including a 1-layer input layer, 2 hidden layers and 1 output layer. According to the battery charging algorithm based on model predictive control described in item 4 of the scope of patent application, the input parameters of the input layer are the preset current value and the preset remaining battery power, and each of the two hidden layers It includes 35 neurons, and the output parameter of the output layer is the predicted battery temperature rise value. The battery charging algorithm based on model predictive control described in item 1 of the patent application scope further includes a human-machine interface written by a LabVIEW program to monitor a temperature change.