CN-122026584-A - Control method of mining lithium ion battery charger and computer program product

CN122026584ACN 122026584 ACN122026584 ACN 122026584ACN-122026584-A

Abstract

The invention relates to a control method of a mining lithium ion battery charger and a computer program product. The method comprises the steps of constructing a discretization physical model of the relation between output voltage and current of a core circuit of the mining lithium ion battery charger and IGBT duty ratio, constructing a feature vector comprising the historical state and the current error of the battery charger, constructing a CNN-BiLSTM classification model, optimizing a plurality of super parameters of the classification model by adopting a self-adaptive step-size improved dung beetle optimization algorithm, inputting the feature vector into the classification model, obtaining classification prediction results of an adjustment strategy and working conditions, quantifying the classification prediction results into execution instructions according to preset rules, and performing IGBT control on the battery charger. According to the invention, the local feature extraction capacity of CNN and the bidirectional time sequence dependency capture capacity of BiLSTM are fused, the model super-parameters are optimized by adopting IDBO algorithm, and the accurate mapping relation of the historical state, the current error and the future control quantity of the charger is established, so that the model can adapt to the nonlinear characteristic of the charger, and the control precision under the nonlinear working condition is improved.

Inventors

XU ZHIPENG
CHEN HU
ZHAO MING
YIN YUXING
ZHANG YONG
WANG WANFENG
CHENG HONGZHONG
MENG LINGYU

Assignees

中煤科工(上海)新能源有限公司

Dates

Publication Date: 20260512
Application Date: 20251230

Claims (11)

1. The control method of the mining lithium ion battery charger is characterized by comprising the following steps of: Constructing a discretization physical model of the relation between the output voltage and the current of a core circuit of the mining lithium ion battery charger and the duty ratio of the IGBT, and constructing a characteristic vector comprising the historical state and the current error of the mining lithium ion battery charger; Constructing a CNN-BiLSTM classification model containing a convolutional neural network and a two-way long-short-term memory network; Adopting self-adaptive step-size improved dung beetle optimizing algorithm, taking the mean square error of the CNN-BiLSTM classification model prediction result as the fitness value, obtaining the optimal superparameter combination of a plurality of superparameters of the CNN-BiLSTM classification model, automatically optimizing the superparameters, and And inputting the feature vector into the CNN-BiLSTM classification model, obtaining a classification prediction result of an adjustment strategy and working conditions, quantizing the classification prediction result into an execution instruction according to a preset rule, and performing IGBT control on the mining lithium ion battery charger.
2. The method according to claim 1, wherein the step of constructing a discretized physical model of the relation between the output voltage and the current of the core circuit of the lithium ion battery charger for the mine and the duty ratio of the IGBT, and constructing a feature vector including the historical state and the current error of the lithium ion battery charger for the mine specifically comprises: Establishing a continuous time domain state equation of a core circuit of the mining lithium ion battery charger: , Wherein t is a discrete time variable, L is a filter inductance value, C is a filter capacitance value, D (t) is an IGBT duty cycle at time t, I L (t) is a filter inductance current at time t, V dc (t) is a direct current bus voltage at time t, V C (t) is a filter capacitance voltage at time t, and R load (t) is a load equivalent resistance at time t; Discretizing the continuous time domain state equation by adopting a forward Euler method according to the sampling period of execution control to obtain a discretized state equation: , Where k is a discrete time variable, T s is a sampling period, t=kt s , k=0, 1,2,..d (k) is an IGBT duty cycle at time k, I L (k) is a filter inductor current at time k, V dc (k) is a dc bus voltage at time k, V C (k) is a filter capacitor voltage at time k, and R load (k) is a load equivalent resistance at time k; obtaining a current error according to the actual output voltage and current: , , , , Wherein I out (k) is the actual output current of the time k charger, V out (k) is the actual output voltage of the time k charger, I ref (k) is the given output current of the time k charger, V ref (k) is the given output voltage of the time k charger, e V (k) is the voltage error of the time k charger, e I (k) is the current error of the time k charger; constructing a feature matrix according to the historical states and the current errors of the past N moments of the charger N is the length of a historical sampling window, CD is the characteristic dimension, and the characteristic comprises given output voltage and current, actual output voltage and current, voltage error, current error, DC bus voltage and load equivalent resistance.
3. The method according to claim 1, wherein the step of constructing a CNN-BiLSTM classification model including a convolutional neural network and a two-way long-short term memory network, specifically comprises: Constructing a convolutional neural network, wherein the convolutional neural network comprises two parallel convolutional layers, initial parameters of the two parallel convolutional layers are the same, and each convolutional layer is connected with a maximum pooling layer and outputs a characteristic sequence through a full connection layer; Constructing a two-way long-short-term memory network, wherein the two-way long-short-term memory network consists of LSTM in two directions, one LSTM is responsible for processing an input sequence from front to back, the other LSTM is responsible for processing the input sequence from back to front, and hidden states in the two directions output characteristic sequences in a connecting or combining mode; and outputting a classification prediction result from the characteristic sequences output by the convolutional neural network and the two-way long-short-term memory network through a full-connection layer to form the CNN-BiLSTM classification model.
4. The method according to claim 1, wherein the step of using the adaptive step size to improve a dung beetle optimization algorithm, taking a mean square error of a prediction result of the CNN-BiLSTM classification model as an fitness value, obtaining an optimal superparameter combination of a plurality of superparameters of the CNN-BiLSTM classification model, and automatically optimizing the plurality of superparameters specifically includes: taking the mean square error of the CNN-BiLSTM classification model prediction result as the adaptive value of the adaptive step-length improved dung beetle optimization algorithm; Defining a plurality of super parameters to be optimized as dung beetle position parameters, wherein the plurality of super parameters comprise the number H of convolution layer convolution kernels of the convolution neural network, the number J of hidden layer neurons of the two-way long-short-term memory network, the learning rate lr and the model training iteration number M; Determining a parameter optimization range, setting a maximum iteration number tau max and a population scale N p , initializing a dung beetle population through chaotic mapping, wherein each dung beetle position in the population corresponds to a group of super-parameter combinations; performing iterative optimization, calculating the fitness value of each dung beetle position in each iteration, and sequencing according to the fitness value to determine the current global optimal position; In each iteration, updating the position and the speed of the dung beetle based on the self-adaptive step length mechanism, updating the global optimal position, judging whether the termination condition is met, outputting an optimal super-parameter combination corresponding to the global optimal position if the termination condition is met, and otherwise, returning to the next iteration.
5. The method of claim 1, wherein the step of inputting the feature vector into the CNN-BiLSTM classification model to obtain a classification prediction result of an adjustment strategy and a working condition, quantizing the classification prediction result into an execution instruction according to a preset rule, and performing IGBT control on a mining lithium ion battery charger specifically comprises: inputting the feature vector into the CNN-BiLSTM classification model, and outputting an adjustment strategy classification result and a working condition classification result by the CNN-BiLSTM classification model, wherein the adjustment strategy classification result is a discrete label; obtaining continuous control quantity according to a preset mapping rule of a discrete adjustment strategy classification result and the continuous control quantity, adjusting the control quantity by combining the working condition classification result, and outputting an IGBT duty ratio adjustment instruction; The control circuit converts the IGBT duty ratio adjusting instruction into the on-off time of the IGBT, and adjusts the output voltage and the current to given values.
6. A method according to claim 3, wherein the convolutional neural network is trained using a hybrid loss of a multi-class cross entropy loss function and a focus loss function, the hybrid loss calculated by: , wherein Loss is a mixing Loss, CE_loss is a multi-class cross entropy Loss, focal_loss is a focusing Loss function, and lambda is a mixing proportion parameter; The convolution calculation of the convolution layer in the CNN-BiLSTM classification model meets the following conditions: , Wherein i is an input sample index, j is a width dimension index of an output feature map, k is a convolution kernel index, H i,j,k is an output feature value of an i sample, a k convolution kernel and a j position, L is a channel dimension index of the input feature map, M and N are space dimension indexes of the convolution kernel, L is the channel number of the input feature map, M and N are the height and width of the convolution kernel, W is a convolution kernel weight matrix, X is the input feature map of the convolution neural network, and b k is a bias value of the k convolution kernel.
7. A method according to claim 3, wherein the CNN-BiLSTM classification model is subjected to a parameter dynamic adjustment: , Wherein, the , , Wherein θ is a model parameter to be dynamically adjusted, comprising convolution kernel weight and bias of the convolution neural network, gate weight and bias of the two-way long-short-term memory network, θ t 、θ t+1 is a θ parameter value at time t and time t+1, η is a basic learning rate, For the loss function gradient of the time t parameter theta, alpha (S t ) is the working condition dynamic adjustment coefficient, S t is the working condition characteristic vector of the time t, The variable amplitude of the working condition is shown, and lambda is the working condition sensitivity coefficient.
8. The method according to claim 4, wherein the termination condition is that a maximum number of iterations is reached and/or an fitness value is not higher than a fitness threshold; The adaptive step size mechanism comprises: The individual position update of the dung beetles meets the following conditions: , In the formula, And Is the position of individual i at the t and t +1 generations, Is the speed of the individual at the t+1 generation; The individual speed update of the dung beetles meets the following conditions: , In the formula, And Is the speed of the individual i in the t and t+1 generation, alpha is the step attenuation factor, the alpha value range is [0.5, 0.9], beta is the learning factor, the beta value range is [0.5, 0.9], Is the global optimal position of the generation t; The individual self-adaptive step length of the dung beetles meets the following conditions: , In the formula, And Is the step size of individual i at t-1 and t generation.
9. The method according to claim 4, wherein the step of performing iterative optimization comprises: Before each iteration starts, classifying the dung beetle population according to the fitness value sequencing result, wherein the categories comprise propagation dung beetle X rep , foraging dung beetle X for and thief dung beetle X thie ; In each iteration, updating the positions of various dung beetles; after each round of updating is finished, calculating the fitness values of all the dung beetles again, sequencing, updating the global optimal position, and entering the next round of iteration; The breeding dung beetles are the first K individuals with the optimal applicability, K=round (N p multiplied by 25%), and round () is a rounding function, the thief dung beetles are Q non-breeding dung beetle individuals selected randomly, and Q=round (N p multiplied by 8%); the position update of the propagation dung beetles meets the following conditions: , Wherein X rep (t+1)、X rep (t) is a position vector of the propagation dung beetles at the time t+1 and t, the dimension of the position vector is d, X best (t) is a global optimal position at the time t, s i (t) is a self-adaptive step length, S 0 is the initial step size, rand (1, d) is the uniform random number matrix that generates 1 row and d column, The absolute value of Euclidean distance between the dung beetles and the global optimal position is propagated; the position update of foraging dung beetles meets the following conditions: , Wherein X for (t+1)、X for (t) is the position vector of foraging dung beetles at the time t+1 and t, X avg (t) is the average position of the population at the time t, randn (1, d) is a standard normal distribution random number matrix for generating 1 row and d columns, Deviation of foraging dung beetles and average positions of groups; The position update of the dung beetles meets the following conditions: , Wherein X thie (t+1)、X thie (t) is the position vector of the dung beetles at the time t+1 and t, X rep (t) is the position of one reproduction dung beetle selected randomly at the time t, each of the dung beetles corresponds to different reproduction dung beetles, The method is the deviation of the dung beetles and the global optimal position.
10. The method of claim 5, wherein the regulatory policy classification result is selected from a plurality of discrete regulatory policy categories pre-divided, the regulatory policy categories comprising "fine-tune+", "fine-tune-", "medium-tune+", "medium-tune-", "large-tune+", "large-tune-"; the mapping rule of the discrete adjustment strategy classification result and the continuous control quantity comprises the following steps: the "fine tuning+" category corresponds to a duty cycle increase of 0.01 to 0.03, The "fine tuning-" category corresponds to a duty cycle reduction of 0.01 to 0.03, The "medium plus" category corresponds to a duty cycle increase of 0.04 to 0.08, The "medium-" category corresponds to a duty cycle reduction of 0.04 to 0.08, The "major plus" category corresponds to a duty cycle increase of 0.09 to 0.15, The "major-" category corresponds to a duty cycle reduction of 0.09 to 0.15.
11. A computer program product comprising a computer program, characterized in that the computer program when executed realizes the steps of the method according to any one of claims 1-10.

Description

Control method of mining lithium ion battery charger and computer program product Technical Field The invention relates to the technical field of control of mining charging equipment, in particular to a control method of a mining lithium ion charger and a computer program product. Background The traditional linear control is a basic scheme of IGBT control of a charger, a controller is designed based on a linearization model of a power conversion circuit of the charger, typical technologies comprise proportional-integral-derivative (PID) control, proportional-resonance (PR) control and the like, and the principle is that closed-loop adjustment is realized by establishing a linear mapping relation between an output error (a difference value between a given value and an actual value) and an IGBT control quantity (such as a duty ratio). The power conversion process of the charger has obvious nonlinear characteristics, such as nonlinearity of IGBT conduction voltage drop along with junction temperature change, nonlinearity of inductance value caused by saturation of a filter inductance magnetic core, working point jump caused by load jump (such as load resistance jump caused by SOC change in the charging process of a power battery) and the like. The traditional linear control is designed based on a linearization model, has a good control effect only near a specific working point (such as rated load and normal temperature environment), and when the working condition deviates from the design point, the control precision is rapidly reduced, and output overshoot (such as 5% -10% of voltage overshoot) or steady-state error (such as current steady-state error exceeding 2%) is easy to occur. And, the parameters of traditional linear control (such as proportional coefficient and integral time constant of PID) are usually fixed values, so that the quick dynamic change of the charger is difficult to match. Disclosure of Invention In view of the above, the present invention provides a mining lithium ion battery charger control method and computer program product that solves or at least alleviates one or more of the above-identified problems and other problems of the prior art. In order to achieve the foregoing object, a first aspect of the present invention provides a control method for a lithium ion battery charger for a mine, including: Constructing a discretization physical model of the relation between the output voltage and the current of a core circuit of the mining lithium ion battery charger and the duty ratio of the IGBT, and constructing a characteristic vector comprising the historical state and the current error of the mining lithium ion battery charger; Constructing a CNN-BiLSTM classification model containing a convolutional neural network and a two-way long-short-term memory network; Adopting self-adaptive step-size improved dung beetle optimizing algorithm, taking the mean square error of the CNN-BiLSTM classification model prediction result as the fitness value, obtaining the optimal superparameter combination of a plurality of superparameters of the CNN-BiLSTM classification model, automatically optimizing the superparameters, and And inputting the feature vector into the CNN-BiLSTM classification model, obtaining a classification prediction result of an adjustment strategy and working conditions, quantizing the classification prediction result into an execution instruction according to a preset rule, and performing IGBT control on the mining lithium ion battery charger. In the foregoing method, optionally, the step of constructing a discretized physical model of a relation between the output voltage and the current of the core circuit of the mining lithium ion battery charger and the duty ratio of the IGBT, and constructing a feature vector including the historical state and the current error of the mining lithium ion battery charger specifically includes: Establishing a continuous time domain state equation of a core circuit of the mining lithium ion battery charger: , Wherein t is a discrete time variable, L is a filter inductance value, C is a filter capacitance value, D (t) is an IGBT duty cycle at time t, I L (t) is a filter inductance current at time t, V dc (t) is a direct current bus voltage at time t, V C (t) is a filter capacitance voltage at time t, and R load (t) is a load equivalent resistance at time t; Discretizing the continuous time domain state equation by adopting a forward Euler method according to the sampling period of execution control to obtain a discretized state equation: , Where k is a discrete time variable, T s is a sampling period, t=kt s, k=0, 1,2,..d (k) is an IGBT duty cycle at time k, I L (k) is a filter inductor current at time k, V dc (k) is a dc bus voltage at time k, V C (k) is a filter capacitor voltage at time k, and R load (k) is a load equivalent resistance at time k; obtaining a current error according to the actual output voltage and current: ,, ,, Wherein I ou