Modular multi-level DC / DC converter prediction control method based on neural network
Through the modular multi-level DC/DC converter prediction control method based on long and short-term memory network, the problems of inaccurate models and slow response in dual active bridge converter control are solved, and more efficient control performance is achieved.
Patent Information
- Application Number
- CN202411137304.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-12-03
AI Technical Summary
The existing modular multi-level dual-active bridge converter control model under multiple degrees of freedom has problems such as inaccurate model, large calculation amount and slow dynamic response of PI controllers.
Using a modular multi-level DC/DC converter prediction control method based on long and short-term memory network, the input and output relationship is established through the nonlinear autoregressive sliding average model, and combined with rolling optimization and feedback correction, the dual active bridge converter is realized. Predictive control.
It significantly improves the accuracy and response speed of traditional control methods, improves control performance, and solves the problems of inaccurate model and complex calculations.
Smart Images

Figure CN119070590A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of dual active bridge converter control, and in particular to a neural network-based modular multi-level DC / DC converter predictive control method. Background Art
[0002] With the rapid development of power electronics technology, the medium-voltage power system of ships has gradually developed from a medium-voltage AC power system to a medium-voltage DC power system. Among them, high-capacity and high-power DC transformers, as an important part of the medium-voltage DC power system of ships, have the advantages of bidirectional power flow and electrical isolation. By performing predictive control based on the long short-term memory network on the output voltage of the dual active bridge converter, the output voltage can be kept high-efficiency while ensuring the stability of the output voltage waveform.
[0003] The long short-term memory network is a variant of the recurrent neural network. It solves the problem of the association between input information and can link previous information to the current task. Based on this, the long short-term memory network also solves the problem that the traditional recurrent neural network loses the ability to link distant information when the time interval between previous information and the current task is large. The long short-term memory network uses the "gate" structure to add, subtract or delete information to the cell state. The structure of the long short-term memory network mainly includes input gate, forget gate, output gate and cell state.
[0004] The establishment of the control model of the modular multi-level dual active bridge converter under multiple degrees of freedom and the design of the controller have the following problems: (1) The power transmission model used in the modeling process is a fundamental equivalent model, ignoring the higher harmonics, so the established model is not accurate enough. (2) Under the control mode of multiple degrees of freedom, the modeling process is relatively complicated and the amount of calculation is large. (3) The PI controller has slow dynamic response and poor control performance. Summary of the invention
[0005] According to the technical problems raised above, a predictive control method for modular multi-level DC / DC converter based on neural network is provided. The present invention collects the input data and output response after the effective value of the converter inductor current is optimized, and then obtains the predictive model of the converter through training of the long short-term memory network. Finally, the predictive control of the converter is realized by combining the rolling optimization and feedback correction methods in the traditional model predictive control, which can significantly improve the above three deficiencies of the traditional control method. The present invention mainly uses the predictive model, rolling optimization and feedback correction to realize the predictive control of the modular multi-level mid-point clamped dual active bridge converter of the long short-term memory network.
[0006] The technical means adopted by the present invention are as follows:
[0007] A modular multi-level DC / DC converter predictive control method based on a neural network, comprising:
[0008] S1. The input-output relationship of the modular multi-level midpoint clamped dual active bridge converter system is established by using a nonlinear autoregressive moving average model, and the input and output data are sampled, and the long short-term memory network is trained offline and online using the sampled data;
[0009] S2. According to the training result of the long short-term memory network in S1, a long short-term memory network model function is established, and the output voltage at the future moment is predicted by rolling optimization and recursive multi-step prediction method;
[0010] S3, the rolling optimization in S2 is performed by calculating the minimum value of the value function to obtain the optimal control amount of the system;
[0011] S4. Optimize and solve the rolling optimization by using the golden section method;
[0012] S5. Use feedback correction and compensation mechanisms to adjust the prediction model and eliminate the prediction errors generated in the dynamic process.
[0013] Furthermore, the step S1 specifically includes:
[0014] S11. The input-output relationship of the modular multi-level midpoint clamped dual active bridge converter system is established using a nonlinear autoregressive moving average model:
[0015] V o (k+1)=f[V o (k),···,V o (k-h+1),d f (kd),···,d f (kd-m+1)]]
[0016] Where f is the nonlinear autoregressive moving average model function; V o is the output voltage of the modular multi-level midpoint clamped dual active bridge converter; d f is the ratio of the primary to secondary side shift of the transformer; k is the current moment; h is the output dimension of the system; m is the input dimension of the system; d is the delay time between the control input and the system response;
[0017] S12, predictive control of dual active bridges is based on the comparison of the primary and secondary side shifts of the transformer. f As input, the secondary output voltage is used as output. After the system open-loop operation is stable, the primary and secondary shift phases of different transformers are compared with the corresponding secondary output voltages, and the sampled data are input into the long short-term memory network for offline training.
[0018] For closed-loop controller output sampling with different inputs and different loads, the transformer primary-secondary shift ratio d is obtained. f , the output voltage under different operating conditions is compared with the transformer primary and secondary side shift f Input the long short-term memory network and perform online training to achieve predictive control under changing working conditions.
[0019] Furthermore, the step S2 specifically includes:
[0020] S21. When establishing the model function of the long short-term memory network, since f is a nonlinear function, the LSTM network is used to fit and approximate f, and the formula in S11 is rewritten as:
[0021] V o (k+1)=f LSTM [V o (k),···,V o (k-h+1),d f (kd),···,d f (kd-m+1)]
[0022] Among them, f LSTM is the long short-term memory network model function; V o (k+1) is the predicted value at time k+1 predicted by the system based on the information at the current time and past time;
[0023] In predictive control, the prediction time domain is the number of steps used to predict the future output or state of the system in each operation cycle, and the predicted output voltage is V H , which is expressed as follows:
[0024]
[0025] Among them, H is the prediction time domain of the system; V o (k+i|k) is the predicted value at time k+i obtained by the system based on the information of the current time and past time;
[0026] S22. In the rolling optimization, the control time domain is the time length of the control quantity to be solved; the control increment ΔD in the system control time domain is f It is expressed as:
[0027]
[0028] Where, Δd f The ratio of the primary and secondary side shifts of the transformer d f The change in , M is the control time domain, M ≥ 1, and the constraint condition for the prediction time domain H is:
[0029] H≥d+M+1
[0030] S23, assuming the output voltage V of the modular multi-level midpoint clamped dual active bridge converter o Depends on the historical input and output information of the converter, the fitting formula of the long short-term memory network for the modular multi-level midpoint clamped dual active bridge converter is:
[0031] V o (k+1)=f LSTM [z(k),…,z(k-T+1)]
[0032] In the formula, T represents the time step of the LSTM input vector, z(k) represents the input and output information of the LSTM at time k, and the expression of z(k) is as follows:
[0033] z(k)=[V o (k),d f (kd)]
[0034] When the system predicts the next H steps, the expression is:
[0035]
[0036]
[0037] Establish a long short-term memory network and use the voltage prediction result V at time k+1 o (k+1) as V o (k+2) part of the input, thereby constructing a recursive multi-step prediction model, which is recursive step by step until the output voltage of the next H moments is predicted.
[0038] Furthermore, the step S3 specifically includes:
[0039] S31, for a given output voltage reference value V ref ,The goal of predictive control is to minimize the error between the output voltage and the reference value in the ,prediction time domain and smooth the regulation process.,In order to optimize the fast response characteristics of the ,modular multi-level midpoint clamped dual active bridge converter, a value function is constructed ,to evaluate the effect of predictive control, and the value function J is defined as:
[0040]
[0041] Among them, V ref (k+i|k) represents the output voltage reference value at time k+i, r j is the suppression weight coefficient of the control quantity;
[0042] S32, construct the value function matrix, smooth the rising process of the output voltage, and convert V ref(k+i|k) forms a reference trajectory sequence, expressed as V r (k)=[V ref (k+1|k),…,V ref (k+H|k)], and let R r =[r0,…r M-1 ], then the matrix form of the value function is:
[0043] J=[V r (k)-V H (k)] T [V r (k)-V H (k)]+[Δd f (k)] T R r [Δd f (k)]
[0044] R r The suppression weight coefficient row representing the control amount;
[0045] S33, the reference trajectory output by the system is expressed as:
[0046]
[0047] Among them, r(k) is the set value of the system at time k, and α is the softening factor of the output voltage rising process.
[0048] Furthermore, the step S4 specifically includes:
[0049] S41, predict the output voltage at the next moment:
[0050] By collecting the output voltage at the current moment and the previous moment under different transmission powers and comparing the primary and secondary side shifts of the transformer, the output voltage at the next moment is obtained; the long-term and short-term memory prediction network is optimized in a rolling manner using the golden section method;
[0051] The golden section method sets two test points in the initial search range, and gradually eliminates the area with non-optimal solutions by comparing the function values of the two points, while retaining the part with the optimal solution. This process is repeated until the search interval is reduced to a specific length, and the midpoint of the interval is used as an approximation of the optimal value. When selecting the test points, the golden section point is used.
[0052] S42. At the initial moment, the golden section interval is determined according to the control value of the previous cycle and the maximum change rate of the control value set manually. When the upper and lower limits of the control value are met, the control value is solved by iterative elimination.
[0053] Furthermore, the step S5 specifically includes:
[0054] S51. Reduce the impact of prediction model errors on the system through feedback correction:
[0055] At time k+1, the actual output voltage of the system is V o (k+1), and the output voltage V predicted at time k o (k+1), the error e(k+1) of the long short-term memory network prediction model at time k+1 is:
[0056] e(k+1)=V o (k+1)-V o (k+1)
[0057] S52, using the error in S41 to correct the long short-term memory network prediction model, rewrite the predicted voltage obtained by the prediction model as:
[0058]
[0059] Among them, λ i is the weighted correction coefficient of the predicted voltage obtained by the prediction model, i=1,2,…,H.
[0060] Compared with the prior art, the present invention has the following advantages:
[0061] The neural network-based modular multi-level DC / DC converter predictive control method provided by the present invention introduces more control freedom by introducing multi-levels, and reduces the current stress of power electronic devices in the original traditional dual active bridge topology by introducing multi-level bridge arm submodules.
[0062] The neural network-based modular multi-level DC / DC converter predictive control method provided by the present invention performs predictive control on the dual active bridge converter based on the long short-term memory network, thereby solving the problem of inaccurate model caused by traditional fundamental wave modeling ignoring each harmonic; and also solving the shortcoming of slow response speed of traditional PI control.
[0063] Based on the above reasons, the present invention can be widely promoted in the field of dual active bridge converter control. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0065] Figure 1This is a flow chart of the predictive control method of a modular multi-level DC / DC converter based on a neural network according to the present invention.
[0066] Figure 2 This is a topology diagram of a modular multi-level midpoint clamped dual active bridge in the present invention.
[0067] Figure 3 This is the predictive control block diagram of the present invention.
[0068] Figure 4 It is the multi-step recursive prediction model in the present invention.
[0069] Figure 5 This is the optimization process of the golden section method in the present invention.
[0070] Figure 6 This is the prediction model building process in the present invention.
[0071] Figure 7 is the prediction model training error in the embodiment of the present invention.
[0072] Figure 8 The figure shows the output comparison of different modeling methods in the embodiments of the present invention.
[0073] Fig. 9 1 is a comparison of output voltages under load mutation in an embodiment of the present invention.
[0074] Fig.10 The output voltage comparison when the reference voltage is changed in the embodiment of the present invention.
[0075] Fig.11 The invention provides an optimization of the effective value of the inductor current under a sudden load drop in an embodiment of the present invention.
[0076] Fig.12 This is the current optimized waveform under the load sudden increase condition in the embodiment of the present invention. DETAILED DESCRIPTION
[0077] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0078] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0079] like Figure 1 As shown, the present invention provides a modular multi-level DC / DC converter predictive control method based on a neural network, comprising:
[0080] S1. The input-output relationship of the modular multi-level midpoint clamped dual active bridge converter system is established using a nonlinear autoregressive moving average model, such as Figure 2 As shown, the input and output data are sampled, and the sampled data is used to train the long short-term memory network offline and online;
[0081] In specific implementation, as a preferred embodiment of the present invention, the step S1 specifically includes:
[0082] S11. The input-output relationship of the modular multi-level midpoint clamped dual active bridge converter system is established using a nonlinear autoregressive moving average model:
[0083] V o (k+1)=f[V o (k),···,V o (k-h+1),d f (kd),···,d f (kd-m+1)]
[0084] Where f is the nonlinear autoregressive moving average model function; V o is the output voltage of the modular multi-level midpoint clamped dual active bridge converter; d f is the ratio of the primary to secondary side shift of the transformer; k is the current moment; h is the output dimension of the system; m is the input dimension of the system; d is the delay time between the control input and the system response;
[0085] S12, for the closed-loop operation of the system, since the duty ratio of the primary and secondary side shift angle is not considered to be set but directly output by the controller. The predictive control of the dual active bridge is based on the primary and secondary side shift angle of the transformer. fAs input, the secondary output voltage is used as output. After the system open-loop operation is stable, the primary and secondary shift phases of different transformers are compared with the corresponding secondary output voltages, and the sampled data are input into the long short-term memory network for offline training.
[0086] For closed-loop controller output sampling with different inputs and different loads, the transformer primary-secondary shift ratio d is obtained. f , the output voltage under different operating conditions is compared with the transformer primary and secondary side shift f Input the long short-term memory network and perform online training to achieve predictive control under changing working conditions.
[0087] S2. According to the training result of the long short-term memory network in S1, a long short-term memory network model function is established, and the output voltage at the future moment is predicted by rolling optimization and recursive multi-step prediction method;
[0088] In specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:
[0089] S21. When establishing the model function of the long short-term memory network, since f is a nonlinear function, the LSTM network is used to fit and approximate f, and the formula in S11 is rewritten as:
[0090] V o (k+1)=f LSTM [V o (k),···,V o (k-h+1),d f (kd),···,d f (kd-m+1)]
[0091] Among them, f LSTM is the long short-term memory network model function; V o (k+1) is the predicted value at time k+1 predicted by the system based on the information at the current time and past time;
[0092] In predictive control, the prediction time domain is the number of steps used to predict the future output or state of the system in each operation cycle, and the predicted output voltage is V H , which is expressed as follows:
[0093]
[0094] Among them, H is the prediction time domain of the system; V o (k+i|k) is the predicted value at time k+i obtained by the system based on the information of the current time and past time;
[0095] S22. In the rolling optimization, the control time domain is the time length of the control quantity to be solved; in order to simplify the subsequent rolling optimization calculation, the control quantity in the control time domain is expressed in an incremental manner.
[0096] The control increment ΔD in the system control time domain f It is expressed as:
[0097]
[0098] Where, Δd f The ratio of the primary and secondary side shifts of the transformer d f The change in , M is the control time domain, M ≥ 1, and the constraint condition for the prediction time domain H is:
[0099] H≥d+M+1
[0100] The significance of the system's constraint on the prediction time domain is that the control quantity in the control time domain needs to pass through time d+M to fully affect the system response. Therefore, rolling optimization can only become meaningful when the system's predicted output in the prediction time domain can reflect the effects of M control time domains.
[0101] S23, assuming the output voltage V of the modular multi-level midpoint clamped dual active bridge converter o Depends on the historical input and output information of the converter, the fitting formula of the long short-term memory network for the modular multi-level midpoint clamped dual active bridge converter is:
[0102] V o (k+1)=f LSTM [z(k),…,z(k-T+1)]
[0103] In the formula, T represents the time step of the LSTM input vector, that is, the input and output information of the previous moments contained in the LSTM, z(k) represents the input and output information of the LSTM at moment k, and the expression of z(k) is as follows:
[0104] z(k)=[V o (k),d f (kd)]
[0105] When the system predicts the next H steps, the expression is:
[0106]
[0107]
[0108] Establish a long short-term memory network and use the voltage prediction result V at time k+1 o (k+1) as V o(k+2) part of the input, thereby constructing a recursive multi-step prediction model, such as Figure 4 As shown, the process is recursively performed step by step until the output voltage at the next H moments is predicted.
[0109] In implementation, when the value of the prediction time domain H is greater than 1, the system cannot obtain the prediction value in one step, and multi-step prediction is required. Multi-step prediction can be divided into recursive and non-recursive according to its principle. Since the non-recursive multi-step prediction method needs to establish multiple long-short-term memory network models that match the size of the prediction time domain, the modeling workload is very large, so this paper uses the recursive multi-step prediction scheme.
[0110] S3, the rolling optimization in S2 is performed by calculating the minimum value of the value function to obtain the optimal control amount of the system;
[0111] In specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:
[0112] During implementation, the output voltage control of the modular multi-level mid-point clamped dual active bridge converter mainly involves two aspects: first, the output voltage needs to be adjusted quickly and accurately to reach the reference value and to be as error-free as possible; second, the adjustment process must be smooth to avoid excessive overshoot and oscillation.
[0113] S31, for a given output voltage reference value V ref ,The goal of predictive control is to minimize the error between the output voltage and the reference value in the ,prediction time domain and smooth the regulation process.,In order to optimize the fast response characteristics of the ,modular multi-level midpoint clamped dual active bridge converter, a value function is constructed ,to evaluate the effect of predictive control, and the value function J is defined as:
[0114]
[0115] Among them, V ref (k+i|k) represents the output voltage reference value at time k+i, j is the jth time after time K, r j is the suppression weight coefficient of the control quantity; when r j When r is zero, the value function is only committed to reducing the error of the output voltage, and the size of the control quantity is not considered in the optimization process. j When it is large, it will limit the range of control amount.
[0116] S32, construct the value function matrix, smooth the rising process of the output voltage, and convert V ref (k+i|k) forms a reference trajectory sequence, expressed as V r (k)=[V ref (k+1k),…,V ref(k+Hk)], and let R r =[r0,…r M-1 ], then the matrix form of the value function is:
[0117] J=[V r (k)-V H (k)] T [V r (k)-V H (k)]+[Δd f (k)] T R r [Δd f (k)]
[0118] R r The suppression weight coefficient row representing the control amount;
[0119] S33, the reference trajectory output by the system is expressed as:
[0120]
[0121] Among them, r(k) is the set value of the system at time k, and α is the softening factor of the output voltage rising process. The range of α is 0 to 1. The larger the value of α, the slower the output voltage follows the reference value, and the stability is improved; the smaller the value of α, the faster the output voltage follows the reference value, but it is easy to cause system instability.
[0122] S4. Optimize and solve the rolling optimization by using the golden section method; Figure 5 This is the optimization process of the golden section method.
[0123] In specific implementation, as a preferred embodiment of the present invention, step S4 specifically includes:
[0124] S41, predict the output voltage at the next moment:
[0125] After defining the value function, it is necessary to find the control quantity corresponding to the minimum value function, that is, df. Different from the rolling optimization in traditional model predictive control, the commonly used rolling optimization method in LSTM predictive control is the golden section method.
[0126] By collecting the output voltage at the current moment and the previous moment under different transmission powers and comparing the primary and secondary side shifts of the transformer, the output voltage at the next moment is obtained; the long-term and short-term memory prediction network is optimized in a rolling manner using the golden section method;
[0127] The golden section method sets two test points within the initial search range, and gradually eliminates the area of non-optimal solutions by comparing the function values of the two points, while retaining the part containing the optimal solution. This process is repeated until the search interval is narrowed to a specific length, and the midpoint of the interval is used as an approximation of the optimal value. When selecting the test points, the golden section points are used. This ensures that the proportion of intervals eliminated in each iteration remains consistent, thereby ensuring that the number of iterations for finding the optimal solution remains unchanged.
[0128] S42. At the initial moment, the golden section interval is determined according to the control value of the previous cycle and the maximum change rate of the control value set manually. When the upper and lower limits of the control value are met, the control value is solved by iterative elimination.
[0129] The golden section method has the advantages of fast iteration speed, less calculation and simple calculation, avoiding the complex derivative calculation when solving analytical solutions. However, as a single variable optimization method, the golden section method can only optimize and solve a control time domain, which limits the scale of the control time domain in predictive control and has a certain impact on control performance.
[0130] S5. Use feedback correction and compensation mechanisms to adjust the prediction model and eliminate the prediction errors generated in the dynamic process.
[0131] In specific implementation, as a preferred embodiment of the present invention, step S5 specifically includes:
[0132] S51. Reduce the impact of prediction model errors on the system through feedback correction:
[0133] At time k+1, the actual output voltage of the system is V o (k+1), and the output voltage V predicted at time k o (k+1), the error e(k+1) of the long short-term memory network prediction model at time k+1 is:
[0134] e(k+1)=V o (k+1)-V o (k+1)
[0135] S52, using the error in S41 to correct the long short-term memory network prediction model, rewrite the predicted voltage obtained by the prediction model as:
[0136]
[0137] Among them, λ i is the weighted correction coefficient of the predicted voltage obtained by the prediction model, i=1,2,…,H.
[0138] λ iThe smaller the error, the slower the system reduces the error. On the contrary, the faster the system reduces the error, but λ i The larger it is, the more likely it is to cause system shock.
[0139] The output voltage at the next moment is obtained by collecting the output voltage and shift phase ratio at the current moment and the output voltage and shift phase ratio at the previous moment. In order to make the long short-term memory network prediction model more accurate, input and output data can be collected at different transmission powers. The prediction model establishment process is as follows: Figure 6 In addition, the collected data are all data after the effective value of the inductor current is optimized.
[0140] Example
[0141] like Figure 1 As shown, the present invention provides a modular multi-level DC / DC converter predictive control method based on neural network. In this embodiment, since the error of the recursive multi-step prediction model will gradually accumulate, its prediction time domain should not be too large. This paper selects the prediction time domain H as 2 and the control time domain M as 1. The delay time d between the system output and input is 0, and the input and output dimensions of the system are both 2, that is, h=2, m=2. At this time, the formula in S1 can be rewritten as:
[0142] V o (k+1)=f LSTM [V o (k),V o (k-1),d f (k),d f (k-1)]
[0143] Figure 7 is the training error of the prediction model in this embodiment, Figure 8 Output comparison of three different modeling approaches: converter, predictive model, and transfer function.
[0144] In addition, for the three methods of predictive control, load feedforward and PI control, Fig. 9 The output voltage comparison of the three methods when the load changes suddenly is shown. Fig.10 The output voltage comparison of three methods when the reference voltage is changed is shown below: Fig.11 The optimization of the effective value of the inductor current under load sudden reduction in three ways; Fig.12 The current waveform is optimized for three methods under load surge conditions.
[0145] It can be seen from the simulation results that the predictive control based on the long short-term memory network is more accurate than the traditional fundamental wave modeling control method. It also solves the calculation complexity problem under multiple degrees of freedom. Finally, compared with the traditional PI control, the control performance is significantly improved and the response speed is greatly increased.
[0146] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.
[0147] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0148] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.
[0149] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0150] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0151] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.
[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A modular multi-level DC / DC converter predictive control method based on neural network, characterized in that: include: S1. The input-output relationship of the modular multi-level midpoint clamped dual active bridge converter system is established by using a nonlinear autoregressive moving average model, and the input and output data are sampled, and the long short-term memory network is trained offline and online using the sampled data; S2. According to the training result of the long short-term memory network in S1, a long short-term memory network model function is established, and the output voltage at the future moment is predicted by rolling optimization and recursive multi-step prediction method; S3, the rolling optimization in S2 is performed by calculating the minimum value of the value function to obtain the optimal control amount of the system; S4. Optimize and solve the rolling optimization by using the golden section method; S5. Use feedback correction and compensation mechanisms to adjust the prediction model and eliminate the prediction errors generated in the dynamic process.
2. The method for predictive control of a modular multi-level DC / DC converter based on a neural network according to claim 1, characterized in that: The step S1 specifically includes: S11. The input-output relationship of the modular multi-level midpoint clamped dual active bridge converter system is established using a nonlinear autoregressive moving average model: V o (k+1)=f[V o (k),···,V o (k-h+1),d f (k-d),···,d f (k-d-m+1)] Where f is the nonlinear autoregressive moving average model function; V o is the output voltage of the modular multi-level midpoint clamped dual active bridge converter; d f is the ratio of the primary to secondary side shift of the transformer; k is the current moment; h is the output dimension of the system; m is the input dimension of the system; d is the delay time between the control input and the system response; S12, predictive control of dual active bridges is based on the comparison of the primary and secondary side shifts of the transformer. f As input, the secondary output voltage is used as output. After the system open-loop operation is stable, the primary and secondary shift phases of different transformers are compared with the corresponding secondary output voltages, and the sampled data are input into the long short-term memory network for offline training. For closed-loop controller output sampling with different inputs and different loads, the transformer primary-secondary shift ratio d is obtained. f , the output voltage under different operating conditions is compared with the transformer primary and secondary side shift f Input the long short-term memory network and perform online training to achieve predictive control under changing working conditions.
3. The neural network-based modular multi-level DC / DC converter predictive control method according to claim 1, characterized in that: The step S2 specifically includes: S21. When establishing the model function of the long short-term memory network, since f is a nonlinear function, the LSTM network is used to fit and approximate f, and the formula in S11 is rewritten as: V o (k+1)=f LSTM [V o (k),···,V o (k-h+1),d f (k-d),···,d f (k-d-m+1)] Among them, f LSTM is the long short-term memory network model function; V o (k+1) is the predicted value at time k+1 predicted by the system based on the information at the current time and past time; In predictive control, the prediction time domain is the number of steps used to predict the future output or state of the system in each operation cycle, and the predicted output voltage is V H , which is expressed as follows: Among them, H is the prediction time domain of the system; V o (k+i|k) is the predicted value at time k+i obtained by the system based on the information of the current time and past time; S22. In the rolling optimization, the control time domain is the time length of the control quantity to be solved; the control increment ΔD in the system control time domain is f It is expressed as: Where, Δd f The ratio of the primary to secondary side shift of the transformer d f The change in , M is the control time domain, M ≥ 1, and the constraint condition for the prediction time domain H is: H≥d+M+1 S23, assuming the output voltage V of the modular multi-level midpoint clamped dual active bridge converter o Depends on the historical input and output information of the converter, the fitting formula of the long short-term memory network for the modular multi-level midpoint clamped dual active bridge converter is: V o (k+1)=f LSTM [z(k),…,z(k-T+1)] In the formula, T represents the time step of the LSTM input vector, z(k) represents the input and output information of the LSTM at time k, and the expression of z(k) is as follows: z(k)=[V o (k),d f (k-d)] When the system predicts the next H steps, the expression is: Establish a long short-term memory network and use the voltage prediction result V at time k+1 o (k+1) as V o (k+2) part of the input, thereby constructing a recursive multi-step prediction model, which is recursive step by step until the output voltage of the next H moments is predicted.
4. The method for predictive control of modular multi-level DC / DC converters based on neural networks according to claim 1, characterized in that: The step S3 specifically includes: S31, for a given output voltage reference value V ref ,The goal of predictive control is to minimize the error between the output voltage and the reference value in the ,prediction time domain and smooth the regulation process. ,In order to optimize the fast response characteristics of the ,modular multi-level midpoint clamped dual active bridge converter, a value function is constructed to ,evaluate Among them, V ref (k+i|k) represents the output voltage reference value at time k+i, r j is the suppression weight coefficient of the control quantity; S32, construct the value function matrix, smooth the rising process of the output voltage, and convert V ref (k+i|k) forms a reference trajectory sequence, expressed as V r (k)=[V ref (k+1|k),…,V ref (k+H|k)], and let R r =[r0,…r M-1 ], then the matrix form of the value function is: J=[V r (k)-V H (k)] T [V r (k)-V H (k)]+[Δd f (k)] T R r [Δd f (k)[ R r The suppression weight coefficient row representing the control amount; S33, the reference trajectory output by the system is expressed as: Among them, r(k) is the set value of the system at time k, and α is the softening factor of the output voltage rising process.
5. The method for predictive control of modular multi-level DC / DC converters based on neural network according to claim 1, characterized in that: The step S4 specifically includes: S41, predict the output voltage at the next moment: By collecting the output voltage at the current moment and the previous moment under different transmission powers and comparing the primary and secondary side shifts of the transformer, the output voltage at the next moment is obtained; the long-term and short-term memory prediction network is optimized in a rolling manner using the golden section method; The golden section method sets two test points in the initial search range, and gradually eliminates the area with non-optimal solutions by comparing the function values of the two points, while retaining the part with the optimal solution. This process is repeated until the search interval is reduced to a specific length, and the midpoint of the interval is used as an approximation of the optimal value. When selecting the test points, the golden section point is used. S42. At the initial moment, the golden section interval is determined according to the control value of the previous cycle and the maximum change rate of the control value set manually. When the upper and lower limits of the control value are met, the control value is solved by iterative elimination.
6. The method for predictive control of modular multi-level DC / DC converters based on neural network according to claim 1, characterized in that: The step S5 specifically includes: S51. Reduce the impact of prediction model errors on the system through feedback correction: At time k+1, the actual output voltage of the system is V o (k+1), and the output voltage V predicted at time k o (k+1), the error e(k+1) of the long short-term memory network prediction model at time k+1 is: e(k+1)=V o (k+1)-V o (k+1) S52, using the error in S41 to correct the long short-term memory network prediction model, rewrite the predicted voltage obtained by the prediction model as: Among them, λ i is the weighted correction coefficient of the predicted voltage obtained by the prediction model, i=1,2,…,H.
