Modular multilevel railway power regulator active thermal control strategy based on neural network

Abstract

An active thermal control strategy of a modular multilevel railway power regulator based on a neural network belongs to the technical field of regulators and control, and a modular multilevel converter has active and reactive flexible regulation and is widely applied to the field of high-voltage direct-current power transmission. In a three-phase system, the optimal capacitor voltage fluctuation and loss can be realized by using the control freedom degree of the circulating current. And aiming at a single-phase system of railway traction, the characteristics of circulating current are greatly different. The invention provides an active thermal control strategy of frequency doubling circulation injection in combination with a neural network aiming at the characteristic that the operation of a modular multi-level railway power regulator has an intermittent working system and the reliability problem caused by junction temperature fluctuation due to frequent start and stop, and analyzes the reliability of devices before and after improvement; the device aging is effectively delayed, so that the service life of the power device is prolonged; the size and weight of the radiator or other cooling devices are reduced, the total cost is reduced, and the design of a radiating system is facilitated.

Description

Technical Field

[0001] The invention belongs to the field of regulator and control technology, and in particular relates to an active thermal control strategy of a modular multi-level railway power regulator based on a neural network. Background Art

[0002] Modular multilevel converters have flexible active and reactive power regulation and are widely used in the field of high-voltage direct current transmission. In three-phase systems, the control freedom of circulating current can achieve optimal capacitor voltage fluctuations and losses. However, for single-phase systems of railway traction, the characteristics of circulating current are very different.

[0003] In view of the intermittent working characteristics of modular multi-level railway power conditioners, frequent start-stopping leads to reliability problems caused by junction temperature fluctuations. The existing technology uses a table lookup polling method. When the number of variables increases, its data table grows exponentially, and the table lookup process becomes more complicated. In addition, although the traditional analytical method can accurately solve the double frequency circulating current, the transcendental equation needs to be solved during the solution process, and the huge amount of calculation leads to a significant increase in the burden on the controller. Therefore, there is an urgent need for an optimization method that can quickly solve the injection of the optimal double frequency component for the load power of the railway power conditioner and improve its loss. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes: a modular multi-level railway power regulator active thermal control strategy based on a neural network, comprising the following steps:

[0005] S1. Analyze and determine the loss of each power semiconductor device in the submodule is affected by the amplitude of the double frequency I 2m and phase angle θ together;

[0006] S2. Formulate the principle for selecting the angle of the double frequency circulation;

[0007] S3. Establish principles for selecting the amplitude of the double frequency circulating current.

[0008] Furthermore, in step S1, the specific analysis process is as follows:

[0009] For the single-phase system, in order to ensure that the circulating current only flows inside the phase unit of the MMC converter, the circulating current angle difference between the two phase units is π, that is, the expression of the circulating current of the two phase units is:

[0010]

[0011] Among them, i 2a and i 2b are the magnitudes of the double frequency circulating current injected into phase units A and B, I 2m is the amplitude of the double frequency circulation, θ is the initial phase of the double frequency circulation;

[0012] The currents of the left and right upper bridge arms are further obtained as:

[0013]

[0014] Assume that the traction network supply voltage is:

[0015] u sa =U m sin(ωt)

[0016] The traction network current is:

[0017]

[0018] The current expressions of the left and right upper bridge arms are obtained:

[0019]

[0020] When the load operating conditions are determined, the DC current component in the above formula and the magnitude of the current on the traction arm side are also determined, and the loss of each power semiconductor device in the submodule is affected by the amplitude of the double frequency I 2m and phase angle θ together.

[0021] Furthermore, in step S2, the amplitude of the double frequency circulating current is fixed, and a curve graph of the bridge arm current changing with the double frequency initial phase angle is drawn. It is found that when the traction network current lags behind the voltage by an angle φ, as long as the double frequency circulating current injection angle θ=-2φ is maintained, the stress of the left and right bridge arm currents is minimized, and the injection angle is θ=-2φ+π / 2, and the stress difference between the left and right bridge arm currents is maximized. Based on the above analysis, the double frequency circulating current injection angle θ=-2φ is selected for subsequent analysis.

[0022] Furthermore, the step S3 includes the following steps:

[0023] S31. Determine the impact of double frequency circulating current injection on loss calculation;

[0024] S32, through loss calculation based on neural network;

[0025] S33. Determine the amplitude of the double frequency circulating current.

[0026] Furthermore, the step S31 includes the following steps: in the absence of double frequency circulating current injection, by calculating the intersection of the bridge arm current and the x-axis, the conduction time of each power device is determined, and the loss of each power device is obtained by integration, and the loss calculation is simple and feasible;

[0027] However, when the double frequency circulating current is injected, the double frequency circulating current changes in real time with the load change. After the double frequency circulating current is superimposed on the bridge arm current, the intersection of the bridge arm current and the x-axis includes three situations, namely: 2 intersections with the x-axis, 3 intersections with the x-axis, and 4 intersections with the x-axis. Through the thermoelectric combined simulation, the loss of each power device is obtained. This avoids the complex process of solving the zero crossing point of the bridge arm current, finding the intersection coordinates, further determining the positive and negative of the integral interval, and then finding the loss of each power device.

[0028] Furthermore, the step S32 includes the following steps: a loss calculation model based on ANN, the model consists of an input layer, one or more hidden layers, and an output layer; the input of each layer is multiplied by a weight value, and then the bias is added to obtain the output of each layer, and the activation function selects a sigmoid type;

[0029] MMC-RPC realizes active power transfer and reactive power compensation of the two power supply arms by receiving current instructions in real time according to different working conditions of the load;

[0030] Injecting double frequency circulating current will increase additional loss, but the increased additional loss can be used to reduce the junction temperature fluctuation of the sub-module device;

[0031] By determining the phase of the double frequency circulating current, the known variable at the input layer of the ANN is the active current command I d , reactive current command I q And the amplitude of the double frequency circulation current I 2m and initial phase angle;

[0032] In order to obtain the relationship between the amplitude and loss of the double frequency circulating current, different active and reactive currents and the amplitude of the double frequency circulating current are input in SIMULINK / PLECS thermoelectric co-simulation, and the conduction loss P of the submodule is sampled in real time. T1 , P D1 , P T2 , P D2 And the switching loss P sw_T1 , P sw_T2 , P sw_D1 , P sw_D2 ;

[0033] The original data under various working conditions are obtained by combined thermal and electrical simulation. The network models of the input and output layers are obtained by training the simulation data. The accuracy of the model is judged by comparing the differences between the predicted data and the actual values.

[0034] Furthermore, the step S33 includes the following steps: after obtaining the network model, in order to reduce the junction temperature fluctuation of the power semiconductor device and keep the switching loss of the device unchanged, calculating the amplitude of the required double frequency circulating current;

[0035] Using the double frequency amplitude I 2m A network ANN model between the loss and the output is proposed, and a control method for obtaining the amplitude of the double frequency circulating current based on the ANN model is proposed:

[0036] First, detect the load current i of the two power supply arms La 、i Lb , thereby obtaining the active current and reactive current instructions of the converters on both sides, sending the active and reactive current instructions to the trained ANN model, and obtaining the loss value P in real time loss , and the loss value is compared with P loss_ref The error is processed by PI to output the current amplitude with double frequency, and then output through the limiting link and sent to the converter as the command value of the circulating current for control.

[0037] The beneficial effects of the present invention are as follows: the present invention aims at the reliability problem caused by junction temperature fluctuation due to the intermittent working characteristics of the modular multi-level railway power conditioner and the frequent start and stop, and invents an active thermal control strategy of double frequency circulating current injection combined with a neural network, and analyzes the reliability of the devices before and after improvement; the present invention can effectively delay device aging, thereby extending the service life of power devices (such as IGBT, MOSFET); reduce thermal shock and local overheating, and improve the overall reliability of the system; reduce the volume and weight of the radiator or other cooling devices, thereby reducing the overall cost and volume of the system, and facilitating the design of the heat dissipation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the topological structure diagram of the MMC-RPC traction power supply system;

[0039] Figure 2 It is the voltage and current vector diagram of the traction power supply system;

[0040] Figure 3 It is the task profile diagram of MMC-RPC;

[0041] Figure 4 This is the curve diagram of the influence of the double frequency circulation angle on the bridge arm current;

[0042] Figure 5 The diagram of the bridge arm current variation under different θ;

[0043] Figure 6 It is the intersection diagram of the bridge arm current and the intersection point of the abscissa under different circulating currents;

[0044] Figure 7 This is the ANN loss calculation model diagram;

[0045] Figure 8 It is a line chart of the deviation between the predicted value and the true value;

[0046] Fig. 9 is the regression factor and error simulation diagram;

[0047] Fig.10 This is the block diagram of the double frequency current amplitude control based on ANN;

[0048] Fig.11 This is the system control block diagram. DETAILED DESCRIPTION

[0049] In order to make the technical means and objectives of the present invention easy to understand, the present invention is further described below in conjunction with a specific implementation method. A modular multi-level railway power regulator active thermal control strategy based on a neural network includes the following steps:

[0050] S1. Analyze and determine that the loss of each power semiconductor device of the submodule is jointly affected by the amplitude I2m and the phase angle θ of the double frequency;

[0051] S2. Formulate the principle for selecting the angle of the double frequency circulation;

[0052] S3. Establish principles for selecting the amplitude of the double frequency circulating current.

[0053] 1MMC-RPC topology is as follows:

[0054] The topology of the MMC-RPC traction power supply system is as follows: Figure 1 As shown in the figure, the train traction power supply is obtained by stepping down the three-phase high-voltage power grid through a V / V transformer to obtain the required power supply voltage (27.5kV). The three-phase grid-side currents are i SA 、i SB 、i SC The output of the V / V transformer is the traction power supply arm α and the power supply arm β. The two traction power supply arms supply power to the two section lines respectively. The output voltage of the two traction power supply arms is u sa and u sb , the load power is P La and P Lb , the load current is i La and i Lb The MMC-RPC is located below the traction power supply arm. The MMC-RPC consists of an α converter and a β converter. The α converter is shown on the right side of the figure above. It consists of two phase units connected in parallel to form an intermediate DC side. The DC side voltage is U dc The phase unit consists of two upper and lower bridge arms through the bridge arm inductor L A1 , L A2 (L B1 , L B2 ) in series, and the bridge arm resistance R is negligible. Each bridge arm is composed of N half-bridge sub-modules in cascade, and the current of each bridge arm is i pa 、i na 、ipb 、i nb , the submodule capacitor voltage is u dcxai ,u dcxbi (x=p,n;i=1,…,N). The output of the bridge arm midpoint passes through the grid-side filter inductor L af Connected to the α power supply arm, the β converter is similar to the α converter and will not be described here.

[0055] Taking the α converter as an example, according to the KVL and KCL of the circuit theorem, the following equation can be obtained:

[0056]

[0057] in,

[0058]

[0059] After simplification, we get:

[0060]

[0061] in,

[0062]

[0063] In the above formula, x=a,b.

[0064] It can be seen that by controlling u px and u nx The conversion from the AC side to the DC side of the converter can be realized. The mathematical model of the β converter is similar to it.

[0065] The working principle of MMC-RPC is as follows:

[0066] Considering the real-time changes in the train loads on both sides of the α power supply arm and the β power supply arm, due to the use of V / V transformers, when the loads of the two power supply arms are different, it will cause serious imbalance in the current on the three-phase high-voltage grid side, resulting in power quality problems. MMC-RPC is a power quality management device that spans the two traction power supply arms. By adopting appropriate control methods, it can realize active power transfer between the traction power supply arms and perform comprehensive compensation for reactive power and harmonics on each power supply arm side. Its main working principle is as follows: Figure 2 As shown:

[0067] Figure 2 The voltage and current vector diagram of the secondary and primary sides of the V / V transformer. is the voltage on the three-phase grid side, and are the voltage and current of the traction power supply arm respectively, and Ahead of the times The phase is π / 3. Considering that most existing trains use four-quadrant PWM rectifiers, which basically achieve unity power factor operation, the load current of the traction power supply arm is With traction supply voltage When the loads of the two traction power supply arms are different (assuming I La Lb ), when using MMC-RPC for compensation, first calculate the power difference ΔI on both sides p ,Right now:

[0068]

[0069] And transfer it from the light load to the heavy load side, and then compensate a certain amount of reactive power in the α arm and the β arm respectively. The reactive component to be compensated is:

[0070]

[0071] After compensation, the current of the two power supply arms is equal to the three-phase grid-side current. At this time, the unbalance degree of the primary three-phase load current is 0, and the reactive negative sequence comprehensive compensation of the three-phase current is realized.

[0072] The operating conditions and reliability analysis of MMC-RPC, namely the operating conditions of MMC-RPC and the thermal characteristics of its devices are as follows:

[0073] Considering the randomness of train load, assuming that there is at most one train running in each power supply arm, there are four load combinations of the α and β power supply arms:

[0074] Table 1 Train load combinations

[0075]

[0076] From the above four combinations, it can be seen that: in Case I, both traction power supply arms are unloaded, and the MMC-RPC is in no-load operation; in Case II and Case III, only one side of the traction power supply arm is loaded, and the MMC-RPC needs to transfer active power and compensate for reactive power; in Case IV, the loads on both sides of the power supply arm are the same, and the MMC-RPC only needs to perform reactive power compensation. Therefore, the operating conditions of the MMC are mainly divided into the above four types, and it can be switched arbitrarily under the four operating conditions. Under different operating conditions, the active and reactive power that the MMC-RPC needs to transfer are very different, especially when switching between Case I and Case II and Case III. The switching of the MMC-RPC from no-load to full-load will inevitably bring about changes in the loss of power devices, which will then manifest as large fluctuations in the device temperature. To simplify the analysis, it is advisable to set a switching condition, namely the mission profile, and take the α side as an example for analysis. It is constantly switched among the four operating conditions, such as Figure 3 shown. ​

[0077] When switching between different working conditions, the junction temperature fluctuates greatly. Figure 3 As shown, when Case II is converted to Case I, that is, switching from full load to no load, it will inevitably bring about loss changes, which will lead to increased junction temperature fluctuations. The reliability of MMC-RPC is mainly affected by power semiconductor devices. The junction temperature fluctuations of semiconductor devices will cause thermal fatigue of the materials, thereby accelerating the damage of semiconductor devices. For the topology of MMC, using circulating current injection to improve junction temperature fluctuations has become an important way. The present invention mainly starts from the degree of freedom of the circulating current layer, analyzes the amplitude and phase angle of the injected circulating current, and provides the selection principle.

[0078] Step S1: Thermal stress optimization strategy for annular injection

[0079] For the single-phase system, in order to ensure that the circulating current only flows inside the phase unit of the MMC converter, the circulating current angle difference between the two phase units is π. It is reasonable to assume that the expression of the circulating current of the two phase units is:

[0080]

[0081] Among them, i 2a and i 2b are the magnitudes of the double frequency circulating current injected into phase units A and B, I 2m is the amplitude of the double frequency circulating current, and θ is the initial phase of the double frequency circulating current.

[0082] Further, the current of the left and right upper bridge arms can be obtained as:

[0083]

[0084] Assume that the traction network supply voltage is:

[0085] u sa =U m sin(ωt)

[0086] The traction network current is:

[0087]

[0088] Therefore, the current expressions of the left and right upper bridge arms can be obtained:

[0089]

[0090] When the load operating conditions are determined, the DC current component in the above formula and the magnitude of the current on the traction arm side are also determined, and the loss of each power semiconductor device in the submodule is affected by the amplitude of the double frequency I 2m and phase angle θ together.

[0091] In step S2: Principles for selecting the angle of the double frequency circulation

[0092] Taking the rated working condition as an example for analysis (i.e., the β power supply arm is fully loaded and the α power supply arm is unloaded), the amplitude and initial phase angle of the double frequency circulating current will affect the bridge arm current. In order to study the influence of the initial phase angle on the bridge arm current, the present invention first fixes the amplitude of the double frequency circulating current and makes a curve of the bridge arm current changing with the double frequency initial phase angle. Figure 4 As shown:

[0093] Figure 4 The figure shows the relationship between the maximum and minimum values ​​of the bridge arm current and the initial phase angle of the double frequency circulating current injection. It can be seen that the difference in the initial phase angle of the double frequency circulating current will affect the maximum and minimum values ​​of the left and right bridge arm currents. When θ = 5π / 6, the maximum difference between the left and right bridge arm currents is the largest. The maximum difference in current reflects the current stress of the left and right bridge arms, that is, the difference in the current stress of the left and right bridge arms is the largest at θ = 5π / 6. If it runs in this state for a long time, it will cause the reliability of the left and right bridge arms to be inconsistent. When θ = π / 3, the difference in the current stress of the left and right bridge arms is 0, that is, at this angle, the current stress of the left and right bridge arms can be effectively guaranteed to be the same. Figure 5 Figure 2 is the instantaneous current waveform of the bridge arm current when θ=5π / 6 and θ=π / 3 respectively. The maximum value of the left and right bridge arm current in (a) is ΔI higher than that in (b), and the minimum value is ΔI less than that in (b). The current fluctuations of the left and right bridge arms in (a) are inconsistent, while the current stresses of the left and right bridge arms in (b) are the same, and the consistency is good.

[0094] More generally, when the traction network current lags behind the voltage by an angle of φ, as long as the double frequency circulating current injection angle θ=-2φ is maintained, the stress of the left and right bridge arm currents is minimized, while the injection angle is θ=-2φ+π / 2, and the stress difference of the left and right bridge arm currents is maximized. Based on the above analysis, the present invention selects the double frequency circulating current injection angle θ=-2φ for subsequent analysis.

[0095] In step S3: Principles for selecting the amplitude of the double frequency circulating current

[0096] The previous step analyzed the relationship between the initial phase angle of the double frequency circulating current and the bridge arm current. To calculate the loss of the submodule, it is necessary to further study the relationship between the amplitude of the double frequency circulating current and the loss.

[0097] 1) The impact of double frequency circulating current injection on loss calculation

[0098] Without double frequency circulating current injection, the intersection of the bridge arm current and the x-axis is calculated to determine the conduction time of each power device, and the loss of each power device is obtained by integration. The loss calculation is simple and feasible. However, when double frequency circulating current injection is considered, the situation becomes much more complicated. In order to improve the reliability of the device under various operating conditions, the double frequency circulating current is required to change in real time with the load change. After the double frequency circulating current is superimposed on the bridge arm current, the situation of the intersection of the bridge arm current and the x-axis becomes more complicated, mainly including the following: Figure 6 In the three cases, it becomes particularly difficult to solve the zero-crossing point of the bridge arm current.

[0099] Figure 6 It can be seen that after injecting the double frequency circulating current, there are three situations for the intersection of the bridge arm current and the horizontal coordinate. If you want to obtain the loss of each power device, you need to find out the coordinates of the intersection, further determine the positive and negative of the integral interval, and then analyze and calculate the loss. Since it is necessary to solve the bridge arm current equation with different frequencies superimposed and containing a DC component, the analytical solution is difficult to obtain. Solving it through online iteration will occupy a large amount of memory of the controller, resulting in a slow solution speed. Therefore, in order to quickly obtain the relationship between the loss of each device of the converter and the bridge arm current, the present invention determines the loss of the submodule by combining thermoelectric simulation with a neural network model.

[0100] 2) Loss calculation based on neural network

[0101] like Figure 7 As shown, in order to quickly and flexibly obtain real-time loss, the present invention proposes a loss calculation model based on ANN, which consists of an input layer, one or more hidden layers, and an output layer; the input of each layer is multiplied by a weight value, and then the bias is added to obtain the output of each layer, and the activation function selects the sigmoid type.

[0102] MMC-RPC realizes active power transfer and reactive power compensation of the two power supply arms by receiving current commands in real time according to different working conditions of the load. According to the above analysis, injecting double frequency circulating current will increase additional losses, but the increased additional losses can be used to reduce the junction temperature fluctuation of the submodule device. The phase of the double frequency circulating current has been determined in the previous step. The known variable in the input layer of the ANN is the active current command I d , reactive current command I q And the amplitude of the double frequency circulation current I 2m In order to obtain the relationship between the amplitude and loss of the double frequency circulating current, different active and reactive currents and the amplitude of the double frequency circulating current are input in SIMULINK / PLECS thermoelectric co-simulation, and the conduction loss P of the submodule is sampled in real time. T1 , P D1 , P T2 , P D2 And the switching loss Psw_T1 , P sw_T2 , P sw_D1 , P sw_D2 .

[0103] The original data under various working conditions are obtained by combined thermal and electrical simulation. The network models of the input and output layers are obtained by training the simulation data. The accuracy of the model is judged by comparing the differences between the predicted data and the actual values. Figure 8 The figure shows the difference between the device loss value predicted by the ANN network and the actual value. Fig. 9 The regression factor is close to 1, which shows that the error between the predicted value and the true value of the loss is small, and a good prediction can be achieved.

[0104] 3) Determination of the amplitude of the double frequency circulating current

[0105] After obtaining the network model, in order to reduce the junction temperature fluctuation of the power semiconductor device and keep the switching loss of the device unchanged, the required amplitude of the double frequency circulating current is calculated. The double frequency amplitude I has been obtained in 2). 2m The control block diagram of the present invention for obtaining the amplitude of the double frequency circulating current based on the ANN model is as follows: Fig.10 As shown:

[0106] Fig.10 The figure shows the control block diagram for obtaining the amplitude of the double frequency circulating current. First, the load current i of the two power supply arms is detected. La 、i Lb , thereby obtaining the active current and reactive current instructions of the converters on both sides, sending the active and reactive current instructions to the trained ANN model, and obtaining the loss value P in real time loss , and the loss value is compared with P loss_ref The error is processed by PI to output the current amplitude with double frequency, and then output through the limiting link and sent to the converter as the command value of the circulating current for control.

[0107] In summary, the initial phase angle and amplitude of the double frequency circulating current are determined from the two aspects of current stress and ANN network loss model; in practice, the bridge arm current is also affected by the maximum current value allowed by the device and the system modulation, thus determining the final double frequency circulating current injection amount.

[0108] System control block diagram

[0109] Through Section 2 and the above determination principle of injection circulation, the overall control block diagram of MMC-RPC can be obtained as follows: Fig.11 As shown: the α converter side adopts a DC voltage outer loop and a current inner loop decoupling control; the β converter side adopts a single current loop decoupling control; the double frequency circulating current adopts a PR control controller, and the carrier phase shift modulation is used to obtain the output pulse.

[0110] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A modular multi-level railway power regulator active thermal control strategy based on neural network, characterized in that: The steps include: S1. Analyze and determine the loss of each power semiconductor device in the submodule is affected by the amplitude of the double frequency I 2m and phase angle θ together; S2. Formulate the principle for selecting the angle of the double frequency circulation; S3. Establish principles for selecting the amplitude of the double frequency circulating current.

2. A modular multi-level railway power regulator active thermal control strategy based on neural network, characterized in that: In step S1, the specific analysis process is as follows: For the single-phase system, in order to ensure that the circulating current only flows inside the phase unit of the MMC converter, the circulating current angle difference between the two phase units is π, that is, the expression of the circulating current of the two phase units is: Among them, i 2a and i 2b are the magnitudes of the double frequency circulating current injected into phase units A and B, I 2m is the amplitude of the double frequency circulation, θ is the initial phase of the double frequency circulation; The currents of the left and right upper bridge arms are further obtained as: Assume that the traction network supply voltage is: u sa =U m sin(ωt) The traction network current is: The current expressions of the left and right upper bridge arms are obtained: When the load operating conditions are determined, the DC current component in the above formula and the magnitude of the current on the traction arm side are also determined, and the loss of each power semiconductor device in the submodule is affected by the amplitude of the double frequency I 2m and phase angle θ together.

3. The neural network-based modular multi-level railway power conditioner active thermal control strategy according to claim 1, characterized in that: In the step S2, the amplitude of the double frequency circulating current is fixed, and a curve graph of the bridge arm current changing with the double frequency initial phase angle is drawn. It is found that when the traction network current lags behind the voltage by an angle φ, as long as the double frequency circulating current injection angle θ=-2φ is maintained, the stress of the left and right bridge arm currents is minimized, and when the injection angle is θ=-2φ+π / 2, the stress difference between the left and right bridge arm currents is maximized. Based on the above analysis, the double frequency circulating current injection angle θ=-2φ is selected for subsequent analysis.

4. The neural network-based modular multi-level railway power conditioner active thermal control strategy according to claim 1, characterized in that: The step S3 includes the following steps: S31. Determine the impact of double frequency circulating current injection on loss calculation; S32, through loss calculation based on neural network; S33. Determine the amplitude of the double frequency circulating current.

5. The neural network-based modular multi-level railway power conditioner active thermal control strategy according to claim 4, characterized in that: The step S31 includes the following steps: in the absence of double frequency circulating current injection, by calculating the intersection of the bridge arm current and the x-axis, the conduction time of each power device is determined, and the loss of each power device is obtained by integration, and the loss calculation is simple and feasible; However, when considering the injection of double frequency circulating current, the double frequency circulating current changes in real time with the load change. After the double frequency circulating current is superimposed on the bridge arm current, the intersection of the bridge arm current and the x-axis includes three situations, namely: 2 intersections with the x-axis, 3 intersections with the x-axis, and 4 intersections with the x-axis. Through thermal and electric combined simulation, the loss of each power device is obtained.

6. The neural network-based modular multi-level railway power conditioner active thermal control strategy according to claim 4, characterized in that: The step S32 includes the following steps: a loss calculation model based on ANN, the model consists of an input layer, one or more hidden layers, and an output layer; the input of each layer is multiplied by a weight value, and then the bias is added to obtain the output of each layer, and the activation function is selected as sigmoid type; MMC-RPC realizes active power transfer and reactive power compensation of the two power supply arms by receiving current instructions in real time according to different working conditions of the load; Injecting double frequency circulating current will increase additional loss, but the increased additional loss can be used to reduce the junction temperature fluctuation of the sub-module device; By determining the phase of the double frequency circulating current, the known variable at the input layer of the ANN is the active current command I d , reactive current command I q And the amplitude of the double frequency circulation current I 2m and initial phase angle; In order to obtain the relationship between the amplitude and loss of the double frequency circulating current, different active and reactive currents and the amplitude of the double frequency circulating current are input in SIMULINK / PLECS thermoelectric co-simulation, and the conduction loss P of the submodule is sampled in real time. T1 , P D1 , P T2 , P D2 And the switching loss P sw_T1 , P sw_T2 , P sw_D1 , P sw_D2 ; The original data under various working conditions are obtained by combined thermal and electrical simulation. The network models of the input and output layers are obtained by training the simulation data. The accuracy of the model is judged by comparing the differences between the predicted data and the actual values.

7. The neural network-based modular multi-level railway power conditioner active thermal control strategy according to claim 6, characterized in that: The step S33 includes the following steps: after obtaining the network model, in order to reduce the junction temperature fluctuation of the power semiconductor device and keep the switching loss of the device unchanged, calculating the amplitude of the required double frequency circulating current; Using the double frequency amplitude I 2m A network ANN model between the loss and the output is proposed, and a control method for obtaining the amplitude of the double frequency circulating current based on the ANN model is proposed: First, detect the load current i of the two power supply arms La 、i Lb , thereby obtaining the active current and reactive current instructions of the converters on both sides, sending the active and reactive current instructions to the trained ANN model, and obtaining the loss value P in real time loss , and the loss value is compared with P loss_ref The error is processed by PI to output the current amplitude with double frequency, and then output through the limiting link and sent to the converter as the command value of the circulating current for control.

Images