Direct-current micro-grid adaptive hierarchical control method based on deep learning

Abstract

The invention discloses a DC micro-grid adaptive hierarchical control method based on deep learning, and the method comprises the steps: enabling a DC micro-grid stability boundary online recognition model based on deep learning to be used for capturing the dynamic characteristics of a system in real time; and a distributed real-time control strategy in a stable boundary and a neural network-based self-adaptive stabilization control strategy outside the stable boundary are provided. By constructing a multi-time-scale hierarchical control framework and combining system dynamic characteristic analysis, fully-distributed real-time power coordination control based on bus voltage signal sharing in the stability boundary of the direct-current micro-grid is realized, and meanwhile, active stabilization management is performed on an abnormal operation state outside the stability boundary by using a deep learning method. According to the invention, the robustness and adaptability of the system under variable operation conditions can be improved, safe and reliable operation of the DC micro-grid is ensured, and the flexibility and expansibility of the DC micro-grid are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of direct current power grids, and in particular to a direct current microgrid adaptive hierarchical control method based on deep learning, which has both system stabilization and power coordination optimization. Background Art

[0002] With the development of renewable energy and energy storage technology, DC microgrids, as a small power system, play an increasingly important role in distributed energy integration and flexible power supply. Compared with traditional AC microgrids, DC microgrids have significant advantages in improving energy efficiency, simplifying structure and reducing conversion losses. However, since DC microgrids are mainly based on power electronic devices for energy conversion, the system exhibits weak inertia. At the same time, the system is also affected by multiple dynamic factors such as load fluctuations, plug-and-play of distributed power sources, and the fast response characteristics of power electronic devices, which further aggravates the challenges faced in system stability and control strategies.

[0003] At present, the control strategy of DC microgrid can be divided into centralized, distributed and decentralized control modes according to different communication methods. In centralized control, the central controller collects and processes the information of power supply and load, and transmits the power command to the local controller through the communication network. Although centralized control is effective, it has inherent limitations, including susceptibility to single point failure, slow dynamic response caused by communication delay, and poor flexibility and scalability of the system. Although the distributed control mode enhances the fault tolerance of the system, it is difficult to ensure the global optimal control effect. In contrast, decentralized control that does not rely on communication networks or public signals is becoming a more suitable solution in DC microgrids due to its plug-and-play, high flexibility and good scalability. However, current related research usually ignores the online identification of the system stability boundary, pays too much attention to the real-time control of the system within the stability boundary, and lacks effective management of the operating state outside the stability boundary. This limitation makes it difficult for existing control strategies to cope with the dynamic changes of the system at multiple time scales, resulting in insufficient robustness of the system under different operating conditions, thereby increasing the risk of system instability.

[0004] With the rapid development of artificial intelligence technology, deep learning, as a powerful data-driven modeling tool, has shown unique advantages in dealing with complex nonlinear systems. Applying deep learning to the control strategy design of DC microgrids can achieve in-depth mining and accurate modeling of the system's dynamic characteristics, providing new ideas for real-time identification of stability boundaries and stabilizing control. However, there is currently a lack of a global adaptive control framework based on deep learning that can identify stability boundaries in complex dynamic environments and adjust control strategies in real time to cope with internal and external disturbances in the system.

[0005] Therefore, it is urgent to develop an adaptive hierarchical control strategy for DC microgrids based on deep learning, and combine deep learning to model system stability and identify boundaries, so as to realize decentralized real-time control of the system within the stability boundary and adaptive stabilizing control outside the stability boundary. Summary of the invention

[0006] The purpose of the present invention is to address the deficiencies of the prior art and propose a deep learning-based adaptive hierarchical control method for a DC microgrid. The method uses an artificial neural network to fit the implicit stability boundary of the system and give stabilizing control actions to achieve adaptive decentralized control of the system inside and outside the stability boundary.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme:

[0008] A DC microgrid adaptive hierarchical control method based on deep learning is characterized by:

[0009] Step 1: Construct a fully decentralized control based on bus voltage sharing, divide and control the working mode of each unit in the system according to the fluctuation range of the DC microgrid bus voltage, wherein each unit in the system includes a photovoltaic unit for generating electric energy under light conditions; an energy storage unit for storing and releasing electric energy to balance the power demand in the power grid; and a load unit for consuming electric energy in the power grid;

[0010] Step 2: Construct a small signal system model A of the implicit stability boundary of the system:

[0011]

[0012] Among them, B 1 , B 2 , …B n are the small signal models of n bus control units, and C is the equivalent small signal model of the constant power load unit;

[0013] Step 3. Collect the droop coefficient and bus voltage value of the DC bus voltage control unit in the underlying fully distributed control model, calculate the static operating point of the system, and insert it into the small signal model A to solve the minimum damping ratio of the system.

[0014]

[0015] The stability label is marked according to the system instability standard. When the minimum damping ratio corresponding to the system characteristic value exceeds 3%, the system is considered to be in a stable state; conversely, when the minimum damping ratio is less than 3%, the system is considered to be in an unstable state.

[0016] Step 4. Taking the droop coefficient of each unit under bus voltage control and the bus voltage value as input and the stability label as output, the artificial neural network 1 is trained to fit the implicit stability boundary of the system;

[0017] Step 5. Based on the existing system based on bus voltage signal sharing, build an over / under voltage control strategy for the expansion unit to achieve stabilization control when the system exceeds the stability boundary;

[0018] Step 6. Build a small signal model A' based on the structure and control strategy of the expanded system:

[0019]

[0020] Among them, D 1 , D 2 ,…D k They are small signal models of k expansion control units respectively;

[0021] Step 7: For the case of exceeding the stability boundary, search for an optimal value within the range of the expansion unit droop coefficient so that the system can return to the stability boundary after expansion and the power value borne by the expansion unit is minimized;

[0022] Step 8: When the stability boundary is exceeded, the droop coefficient of each bus voltage unit in the system, the DC bus voltage value, and the minimum damping ratio of the system before and after expansion are taken as input, and the optimal value of the droop coefficient of the expansion unit is taken as output. The artificial neural network 2 is trained to fit the relationship between the expansion control effect and the droop coefficient, so as to facilitate the subsequent guidance of the system to achieve real-time adaptive stabilization control.

[0023] Step 9: In the real-time control of the DC microgrid, the trained neural network 1 is used to determine the stability of the system (whether it exceeds the stability boundary) in real time; if it does not exceed the stability boundary, the system operates in a fully decentralized real-time control mode based on bus voltage sharing before capacity expansion; if it exceeds the stability boundary, the artificial neural network 2 will adaptively guide the expansion unit to participate in the system stabilization control and restore the system stability.

[0024] Further, the working mode in step 1 includes:

[0025] According to the rated value of bus voltage ±5%, set the maximum value of voltage fluctuation U H2 Minimum value U L2 ;

[0026] According to the rated value of the bus voltage ±3.33%, set the maximum voltage U of the energy storage adjustment range H1 and the minimum value U L1 ;

[0027] In mode 1, the voltage fluctuation range is U H1 <Udc ≤U H2 , the photovoltaic unit operates in droop mode to maintain the bus voltage stability, and the output voltage and droop coefficient expression are:

[0028] U dc =U H2 -k pv P pv

[0029]

[0030] Among them, k pv is the droop coefficient of the photovoltaic unit, P pv_max is the maximum photovoltaic output power;

[0031] In mode 2, the voltage fluctuation range is U N dc ≤U H1 , the energy storage unit operates in the droop charging mode to maintain the bus voltage stability, and its output voltage and droop coefficient expression is:

[0032] U dc =U H1 +k b P bat

[0033]

[0034] Among them, k b is the droop coefficient of the energy storage unit, P cha_lim is the maximum charging power of the energy storage unit;

[0035] In mode 3, the voltage fluctuation range is U L1 dc ≤U N , the energy storage unit operates in the droop discharge mode to maintain the bus voltage stability, and the output voltage and droop coefficient expressions are:

[0036] U dc =U N +k b P bat

[0037]

[0038] In mode 4, the bus voltage fluctuation range is U L2 dc ≤U L1 , the photovoltaic unit and the energy storage unit are both in limited power control mode, and the voltage fluctuation is dominated by the load unit.

[0039] ​​​Furthermore, the step 2 constructs a small signal system model A of the implicit stability boundary of the system, specifically including:

[0040] Step 2.1: Equivalent the models of each part in the DC microgrid, and obtain the relationship between its voltage and current according to the bidirectional buck-boost circuit structure of the DC bus voltage control unit in the droop mode;

[0041]

[0042] Among them, C i , L i , R i and d i They represent the output voltage stabilizing capacitor, filter inductor, equivalent resistance and duty cycle of the i-th buck-boost converter respectively; u oi 、i Li 、i oi and u dc It indicates the converter output voltage, input side inductor current, output current and bus voltage value;

[0043] Step 2.2: Based on the voltage and current double closed loop, the DC bus voltage control unit adopts droop control to achieve the coordinated power distribution among multiple units. The corresponding control model can be expressed as:

[0044]

[0045] Where u refi , k i is the output voltage reference value and droop coefficient value of the i-th buck-boost converter; k pi_i , k ii_i 、u ir_i , G i_i (s) = k pi_i +k ii_i / s are the proportional integral parameter integral term output and transfer function of the inner loop PI controller; k pu_i , k iu_i 、u ur_i , G u_i (s) = k pu_i +k iu_i / s are the proportional integral parameter, integral term output and transfer function of the outer loop PI controller respectively;

[0046] Step 2.3: The model of the control unit in power-limited operation and multiple constant-power loads in parallel is simplified as follows:

[0047]

[0048] Step 2.4: Combine the above equations and add small signal perturbations to obtain the small signal model A of the system:

[0049]

[0050] Among them, B 1 , B 2 , …B n are the small signal models of n bus control units, and C is the equivalent small signal model of the constant power load unit.

[0051] When U dc ≥U H2 , within this range, the expansion unit converter is in the droop charging mode, the overvoltage controller reaches the output upper limit, and the expansion unit droop coefficient needs to meet:

[0052]

[0053] When U dc ≤U L2 , in this range, the converter is in droop discharge mode, the undervoltage controller reaches the output upper limit, and the droop coefficient of the expansion unit needs to meet:

[0054]

[0055] When U L2 ≤U dc ≤U H2 At this time, the overvoltage controller and the undervoltage controller both exceed the output limit. In this range, the expansion unit can adaptively adjust the charge and discharge power according to the droop coefficient. The current reference value is simplified after sorting:

[0056]

[0057] Among them I' cha_lim , I' dis_lim They correspond to the maximum charging current for energy storage expansion and the maximum discharging current for energy storage expansion respectively.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] 1) The present invention constructs an artificial neural network model to fit the implicit stability boundary of the system. On the basis of ensuring the calculation accuracy, the system stability margin can be calculated in real time and the system stability can be judged, which is used to guide the subsequent stabilization control of the system.

[0060] 2) The present invention constructs a completely decentralized underlying control model based on bus voltage signal sharing. When the system is within the stability boundary, decentralized real-time control within the stability boundary is realized under the proposed bus voltage signal sharing control. When the system exceeds the stability boundary, the expansion unit is guided by the artificial neural network 2 to participate in the decentralized control of the underlying layer, realizing adaptive stabilizing control outside the stability boundary and helping the system to restore stability.

[0061] 3) When training the artificial neural network 2 to fit the droop coefficient and stabilizing control effect of the expansion control unit, the present invention not only ensures the effectiveness of its stabilizing control, but also takes into account the economic efficiency of the expansion cost of the expansion unit. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic flow chart of a DC microgrid adaptive hierarchical control method based on deep learning according to an embodiment of the present invention;

[0063] Figure 2 It is the structural framework of the 600V low voltage DC microgrid system in this embodiment;

[0064] Figure 3 This is a diagram showing the division of working modes of each unit under shared control based on bus voltage signals before expansion of the DC microgrid in this embodiment;

[0065] Figure 4 Schematic diagram of implicit stable boundary fitting of neural network 1 in this embodiment

[0066] Figure 5 This is a block diagram of the over / under voltage control strategy for the expansion unit in this embodiment.

[0067] Figure 6 The diagram for dividing the working modes of the expansion unit at each voltage level in this embodiment is as follows:

[0068] Figure 7 This is a flowchart for obtaining the neural network 2 data set in this embodiment.

[0069] Figure 8 Schematic diagram of adaptive hierarchical real-time control of a DC microgrid based on deep learning in an embodiment of the present invention; DETAILED DESCRIPTION

[0070] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited thereby.

[0071] See also Figure 1, a DC microgrid adaptive hierarchical control method based on deep learning in an embodiment of the present invention comprises the following parts: a system implicit stability boundary neural network fitter based on deep learning, a stabilization control director based on deep learning, a fully decentralized control strategy based on bus voltage sharing within the stability boundary, and a capacity expansion control strategy based on over / under voltage controller outside the stability boundary. Specifically, the embodiment of the present invention is applied to a 600V low-voltage DC microgrid, and its structure is as follows Figure 2 As shown in the figure, the system consists of three energy storage units, one photovoltaic power generation unit, one expansion unit and several DC loads. The photovoltaic unit is connected to the DC bus through a Boost converter, and the energy storage unit and expansion unit are connected to the DC bus through a bidirectional DC-DC converter. The specific control strategy construction steps are as follows:

[0072] Step 1: Construct a fully decentralized control system based on bus voltage sharing. The working modes of each unit are divided as follows: Figure 3 As shown, the specific division is as follows: In mode 1, the bus voltage fluctuation range is U H1 dc ≤U H2 , the photovoltaic unit is in droop working mode, and the other units are in power limit control state. Voltage stability is mainly achieved by the photovoltaic unit, and its output voltage and droop coefficient are set as follows:

[0073] U dc =U H2 -k pv P pv

[0074]

[0075] In mode 2, the bus voltage fluctuation range is U N dc ≤U H1 , the energy storage unit is in droop charging mode, the photovoltaic unit is in power limiting control mode, and voltage stability is mainly achieved by the energy storage unit. Its output voltage and droop coefficient are set as follows:

[0076] U dc =U H1 +k b P bat

[0077]

[0078] In mode 3, the bus voltage fluctuation range is U L1 dc ≤U N ​​​, the energy storage unit is in the droop discharge working mode, the photovoltaic unit is still in the power limit control mode, and the voltage stability is mainly achieved by the energy storage unit. Its output voltage and droop coefficient are set as follows:

[0079] U dc =U N +k b P bat

[0080]

[0081] In mode 4, the bus voltage fluctuation range is U L2 dc ≤U L1 , the photovoltaic unit and the energy storage unit are both in limited power control mode, and the voltage fluctuation is dominated by the load unit.

[0082] Step 2: Establish small signal models of each part according to the above system network structure and control strategy.

[0083] In this embodiment, the system control unit can be divided into a droop working unit and a power limiting working unit. The load is constructed as a constant power load. Since the power limiting working unit is similar to a constant power load, the small signal model construction is mainly divided into two parts:

[0084] First, the bus voltage control unit in droop operation is similar to an ideal voltage source, which is connected to the DC bus through a DC-DC converter to maintain the stability of the bus voltage. Assuming that the switching ripple is small and the base frequency ratio is high, its state space equation is:

[0085]

[0086] Among them, C i , L i , R i and d i They represent the output voltage stabilizing capacitor, filter inductor, equivalent resistance and duty cycle of the i-th buck-boost converter respectively; u oi 、i Li 、i oi and u dc It indicates the converter output voltage, input side inductor current, output current and bus voltage value;

[0087] Then linearize the above equation to obtain the small signal model of the bus voltage control unit:

[0088]

[0089] Among them, I Li , U oi , D i ​They are the steady-state value of inductor current, output voltage and duty cycle respectively; 'Δ' represents the small signal variable.

[0090] The bus voltage control unit in the droop working mode uses droop control to achieve coordinated power distribution based on the voltage and current dual-loop control. The converter control equation can be expressed as:

[0091]

[0092] Where u refi , k i is the output voltage reference value and droop coefficient value of the i-th buck-boost converter; k pi_i , k ii_i 、u ir_i , G i_i (s) = k pi_i +k ii_i / s are the proportional integral parameter integral term output and transfer function of the inner loop PI controller; k pu_i , k iu_i 、u ur_i , G u_i (s) = k pu_i +k iu_i / s are the proportional integral parameter, integral term output and transfer function of the outer loop PI controller respectively;

[0093] Then linearize the above equation to obtain the small signal model of the bus voltage control unit converter control equation:

[0094]

[0095] The equivalent state space model of the constant power load unit is:

[0096]

[0097] Where C eq , L eq , R eq are the equivalent load input side voltage stabilizing capacitor, resistance and inductance of the equivalent circuit respectively.

[0098] Then linearize the above equation to get the small signal model of constant power load:

[0099]

[0100] Finally, the small signal models of each part are combined and the following state variables are defined to obtain the characteristic matrix A of the system:

[0101]

[0102] Among them, B1 , B 2 , B 3 are the state space matrices of the three bus control units, and C is the equivalent state space matrix model of the constant power load unit.

[0103] Step 3: Collect the droop coefficient and bus voltage value of the bus voltage control unit in the underlying DC microgrid to calculate the static operating point and substitute it into the characteristic matrix A to solve the minimum damping ratio of the system

[0104]

[0105] Where σ and ω are the real and imaginary parts of the eigenvalue of the characteristic matrix A, respectively.

[0106] The stability label is marked according to the system instability standard. When the minimum damping ratio corresponding to the system eigenvalue exceeds 3%, the system is judged to be in a stable state. Conversely, when the minimum damping ratio corresponding to the system eigenvalue is less than 3%, the system is judged to be in an unstable state.

[0107] Step 4: If Figure 4 As shown, the droop coefficient and bus voltage value of each bus voltage control unit are taken as input, and the stability label is taken as output. An artificial neural network based on classification learning is trained to fit the implicit stability boundary of the system for subsequent real-time stability evaluation of the system and to guide the stability control.

[0108] Step 5: Construct the over / under voltage control strategy of the expansion unit under the distributed control strategy based on bus voltage sharing. The control structure block diagram is as follows: Figure 5 shown.

[0109] By setting the droop coefficient within a certain range, the energy storage expansion unit can be ensured to meet the U dc >U H2 When the charging power reaches the upper limit, when U dc L2 The lower limit of discharge power is reached.

[0110] Based on the difference between the bus voltage and the maximum overvoltage and minimum undervoltage values, the expansion unit working mode is set as follows: Figure 6 As shown, it is divided into the following three parts:

[0111] WhenU dc ≥U H2 , within this range, the expansion unit converter is in the maximum charging mode, the overvoltage controller reaches the output upper limit, and the expansion unit droop coefficient must meet:

[0112]

[0113] WhenU​dc ≤U L2 , within this range, the converter is in the maximum discharge working mode, the undervoltage controller reaches the output upper limit, and the droop coefficient of the expansion unit needs to meet:

[0114]

[0115] When U L2 ≤U dc ≤U H2 At this time, the overvoltage controller and the undervoltage controller are both beyond the output limit, the converter is in the droop charge / discharge working state, and the current reference value is simplified after sorting:

[0116]

[0117] I' cha_lim , I' dis_lim They correspond to the maximum charging current for energy storage expansion and the maximum discharging current for energy storage expansion respectively.

[0118] Step 6: Build the small signal model A' of the expanded system.

[0119]

[0120] Among them, D 1 They are small signal models of the expansion control unit respectively.

[0121] Step 7: Figure 7 As shown, the system information (droop coefficient of each bus voltage control unit and bus voltage value, etc.) determined as unstable in step 3 is recorded, and the over / undervoltage expansion unit controller initializes the input expansion unit droop coefficient according to the voltage level within the range of the droop coefficient. Then, the system voltage deviation and stable state are detected to see whether the voltage deviation is less than the maximum allowed value and whether the system has returned to the stable boundary. If all are satisfied, the economy of the expansion unit operation under this droop coefficient input is evaluated, and the droop coefficient value and the system information after expansion control are included in the data set; if not satisfied, a new droop coefficient is re-entered and the above operation is repeated.

[0122] Step 8: Through the data set in step 7, the bus voltage control unit droop coefficient, bus voltage value, minimum damping ratio of the system under unstable conditions before expansion, and the minimum damping ratio of the system after expansion are taken as input, and the expansion unit droop coefficient is taken as output. The mapping between the expansion unit droop coefficient and its stabilizing control effect is fitted through artificial neural network 2.

[0123] Step 9: Figure 8As shown in the figure, in the real-time control of the system, the artificial neural network 1 obtained in Step 4 determines the stable state of the system. If it is determined that the system is within the stable boundary, it operates in a fully decentralized real-time control based on bus voltage sharing. If it is determined that the system exceeds the stable boundary, the droop coefficient k of the expansion unit is adaptively output according to the artificial neural network 2 obtained in Step 8 pi to guide the over / under voltage control strategy of the expansion unit, that is, to guide the system to perform adaptive stability control to make the system return to the stable boundary again.

[0124] As described above, it is only an exemplary example of the present invention. It can be understood that the content described in the examples of this specification is only a list of the implementation forms of the inventive concept, and does not impose any formal restrictions on the invention. Therefore, the protection scope of the present invention should not be limited to the specific forms described in the examples, but should also include the technical means made by those skilled in the art according to the inventive concept.

[0125] S1: Build a fully decentralized coordinated control strategy for the DC microgrid based on DC bus voltage signal sharing.

[0126] S2: Build a small-signal stability analysis model for the DC microgrid according to the microgrid structure and control strategy constructed in S1.

[0127] S3: Obtain the droop coefficient of each bus voltage control unit in the DC microgrid and the DC bus voltage value, calculate the static operating point of the DC microgrid system in the corresponding state and substitute it into the small-signal analysis model obtained in S2 to calculate the system eigenvalue.

[0128] S4: Use the stability criterion based on the system eigenvalue calculated in Step S3 to label the stability margin and stability label (stable or unstable) for the system.

[0129] S5: Take the droop coefficient of each bus voltage control unit and the bus voltage value in Step S3 as inputs, and take the stability label obtained in Step S4 as the output, and fit the implicit stability boundary of the system through the artificial neural network 1.

[0130] S6: Build a fully decentralized coordinated control strategy for the expanded system based on DC bus voltage signal sharing.

[0131] S7: Build a small-signal stability analysis model for the expanded DC microgrid according to the structure and control strategy of the expanded system constructed in S6.

[0132] S8: Substitute the droop coefficient of each bus voltage control unit and the DC bus voltage value in the unstable situation judged in Step S44 into the small-signal model of the expanded system, and calculate the droop coefficient that makes the system return to the stable state with the minimum power borne by the expansion unit and the system stability margin in this case.

[0133] S9: The bus voltage control unit droop coefficient, bus voltage value, stability margin of the system under unstable conditions before expansion, and the stability margin of the system after expansion are taken as input, and the expansion unit droop coefficient is taken as output. The mapping between the expansion unit droop coefficient and the stabilizing control effect is fitted through artificial neural network 2.

[0134] S10: The artificial neural network 1 obtained in step S5 determines the system stability in real time. If the system is within the stability boundary, it runs in fully distributed real-time control; if the system exceeds the stability boundary, the artificial neural network 2 obtained in step S9 outputs the expansion unit droop coefficient k 4 To guide the microgrid system to perform adaptive stabilization control and bring the system back to the stability boundary.

Claims

1. A DC microgrid adaptive hierarchical control method based on deep learning, characterized in that: include: Step 1: Construct a fully decentralized control based on bus voltage sharing, divide and control the working mode of each unit in the system according to the fluctuation range of the DC microgrid bus voltage, wherein each unit in the system includes a photovoltaic unit for generating electric energy under light conditions; an energy storage unit for storing and releasing electric energy to balance the power demand in the power grid; and a load unit for consuming electric energy in the power grid; Step 2: Construct a small signal system model A of the implicit stability boundary of the system: Among them, B1, B2, ...B n are the small signal models of n bus control units, and C is the equivalent small signal model of the constant power load unit; Step 3. Collect the droop coefficient and bus voltage value of the DC bus voltage control unit in the underlying fully distributed control model, calculate the static operating point of the system, and insert it into the small signal model A to solve the minimum damping ratio of the system. The stability label is marked according to the system instability standard. When the minimum damping ratio corresponding to the system eigenvalue exceeds 3%, the system is judged to be in a stable state. Conversely, when the minimum damping ratio corresponding to the system eigenvalue is less than 3%, the system is judged to be in an unstable state. Step 4. Taking the droop coefficient and bus voltage value of each unit under bus voltage control as input and the stability label as output, the artificial neural network 1 is trained to fit the implicit stability boundary of the system; Step 5. Based on the existing system based on bus voltage signal sharing, build an over / under voltage control strategy for the expansion unit to achieve stabilization control when the system exceeds the stability boundary; Step 6. Build a small signal model A' based on the structure and control strategy of the expanded system: Among them, D1, D2, ...D k They are small signal models of k expansion control units respectively; Step 7: For the case of exceeding the stability boundary, search for an optimal value within the range of the expansion unit droop coefficient so that the system can return to the stability boundary after expansion and the power value borne by the expansion unit is minimized; Step 8: When the stability boundary is exceeded, the droop coefficient of each bus voltage unit in the system, the DC bus voltage value, and the minimum damping ratio of the system before and after expansion are taken as input, and the optimal value of the droop coefficient of the expansion unit is taken as output. The artificial neural network 2 is trained to fit the relationship between the expansion control effect and the droop coefficient, so as to facilitate the subsequent guidance of the system to achieve real-time adaptive stabilization control. Step 9: In the real-time control of the DC microgrid, the trained neural network 1 is used to determine the stability of the system in real time, that is, to determine whether it exceeds the stability boundary; if it does not exceed the stability boundary, the system operates in a fully decentralized real-time control mode based on bus voltage sharing before expansion; if it exceeds the stability boundary, the artificial neural network 2 will adaptively guide the expansion unit to participate in the system stabilization control and restore the system stability.

2. The DC microgrid adaptive hierarchical control method based on deep learning according to claim 1 is characterized by: The working modes in step 1 include: According to the rated value of bus voltage ±5%, set the maximum value of voltage fluctuation U H2 Minimum value U L2 ; According to the rated value of the bus voltage ±3.33%, set the maximum voltage U of the energy storage adjustment range H1 and the minimum value U L1 ; In mode 1, the voltage fluctuation range is U H1 dc ≤U H2 , the photovoltaic unit operates in droop mode to maintain the bus voltage stability, and the output voltage and droop coefficient expression are:​ IN dc =U H2 -k pv P pv Among them, k pv is the droop coefficient of the photovoltaic unit, P pv_max is the maximum photovoltaic output power; In mode 2, the voltage fluctuation range is U N dc ≤U H1 , the energy storage unit operates in the droop charging mode to maintain the bus voltage stability, and its output voltage and droop coefficient expression is:​ IN dc =U H1 +k b P bat Among them, k b is the droop coefficient of the energy storage unit, P cha_lim is the maximum charging power of the energy storage unit; In mode 3, the voltage fluctuation range is U L1 dc ≤U N , the energy storage unit operates in the droop discharge mode to maintain the bus voltage stability, and the output voltage and droop coefficient expressions are:​ IN dc =U N +k b P bat In mode 4, the bus voltage fluctuation range is U L2 dc ≤U L1 , the photovoltaic unit and the energy storage unit are both in limited power control mode, and the voltage fluctuation is dominated by the load unit.​ 3. The DC microgrid adaptive hierarchical control method based on deep learning according to claim 1, characterized in that: The step 2 constructs a small signal system model A of the implicit stability boundary of the system, specifically including: Step 2.1: Equivalent the models of each part in the DC microgrid, and obtain the relationship between its voltage and current according to the bidirectional buck-boost circuit structure of the DC bus voltage control unit in the droop mode; Among them, C i , L i , R i and d i They represent the output voltage stabilizing capacitor, filter inductor, equivalent resistance and duty cycle of the i-th buck-boost converter respectively; u oi 、i Li 、i oi and u dc It indicates the converter output voltage, input side inductor current, output current and bus voltage value; Step 2.2: Based on the voltage and current double closed loop, the DC bus voltage control unit adopts droop control to achieve the coordinated power distribution among multiple units. The corresponding control model can be expressed as: Where u refi , k i is the output voltage reference value and droop coefficient value of the i-th buck-boost converter; k pi_i , k ii_i 、u ir_i , G i_i (s) = k pi_i +k ii_i / s are the proportional integral parameter integral term output and transfer function of the inner loop PI controller; k pu_i , k iu_i 、u ur_i , G u_i (s) = k pu_i +k iu_i / s are the proportional integral parameter, integral term output and transfer function of the outer loop PI controller respectively; Step 2.3: The model of the control unit in power-limited operation and multiple constant-power loads in parallel is simplified as follows: Step 2.4: Combine the above equations and add small signal perturbations to obtain the small signal model A of the system: Among them, B1, B2, ...B n are the small signal models of n bus control units, and C is the equivalent small signal model of the constant power load unit.

4. The DC microgrid adaptive hierarchical control method based on deep learning according to claim 1, characterized in that: When U dc ≥U H2 , within this range, the expansion unit converter is in the droop charging mode, the overvoltage controller reaches the output upper limit, and the expansion unit droop coefficient needs to meet: When U dc ≤U L2 , in this range, the converter is in droop discharge mode, the undervoltage controller reaches the output upper limit, and the droop coefficient of the expansion unit needs to meet: When U L2 ≤U dc ≤U H2 At this time, the overvoltage controller and the undervoltage controller both exceed the output limit. In this range, the expansion unit can adaptively adjust the charge and discharge power according to the droop coefficient. The current reference value is simplified after sorting: Among them I' cha_lim , I' dis_lim They correspond to the maximum charging current for energy storage expansion and the maximum discharging current for energy storage expansion respectively.

Images