Method and system for cooperatively inhibiting overvoltage of power distribution network through photovoltaic inverter cluster
By dynamically adjusting the voltage threshold and sensitivity prediction of photovoltaic inverters using reinforcement learning and LSTM models, the adaptability problem of traditional photovoltaic inverter control methods under grid changes is solved, and efficient and precise control of photovoltaic inverter clusters to collaboratively suppress overvoltage is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-03-17
AI Technical Summary
Traditional photovoltaic inverter control methods cannot adapt to changes in grid impedance and photovoltaic fluctuations, resulting in over-adjustment or under-adjustment. Furthermore, they lack intelligent and learning capabilities for coordinated control, leading to high malfunction rates and low control accuracy.
The voltage threshold of the photovoltaic inverter is dynamically adjusted using a Q-function model and an LSTM model based on reinforcement learning. Combined with grid impedance and communication quality, the sensitivity change is predicted. The optimal power regulation is solved by optimizing the objective and constraints, thereby realizing the coordinated regulation of the photovoltaic inverter cluster.
It improves the accuracy and stability of the suppression voltage of photovoltaic inverters, reduces the false alarm rate and prediction error, enhances the system's adaptability and robustness, and improves the efficiency and accuracy of photovoltaic inverter cluster collaborative control.
Smart Images
Figure CN121689028A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of distributed photovoltaic grid-connected power generation, and more particularly to a photovoltaic inverter cluster cooperative suppression of power grid overvoltage method and system. BACKGROUND
[0002] The traditional mechanical voltage regulating device has slow response speed and cannot track the minute-level fluctuation of photovoltaic power, while the local inverter control is fast in response but lacks cooperation, which easily leads to oscillation. The traditional control method adopts fixed droop coefficient, sensitivity matrix and starting threshold, and cannot adapt to the time-varying characteristics such as grid impedance change and photovoltaic fluctuation, resulting in over-regulation under weak grid and insufficient regulation under strong grid, and the sensitivity matrix calculated offline is invalid with the change of network topology; the existing cooperative control is mostly based on simple rules or static optimization, and lacks evaluation feedback and learning evolution ability of control effect. SUMMARY
[0003] In order to solve the problems in the prior art, the application provides a photovoltaic inverter cluster cooperative suppression of power grid overvoltage method and system, which dynamically adjusts the voltage threshold of the photovoltaic inverter according to the Q function model of reinforcement learning, first introduces the grid impedance and communication quality into the threshold decision, realizes the multi-dimensional fusion of electrical quantity, communication quantity and time quantity, and reduces the misoperation rate of the photovoltaic inverter; the sensitivity of the photovoltaic inverter at the next moment is predicted according to the LSTM (Long Short Term Memory, Long Short Term Memory Network) model, the evolution law of the sensitivity with time, weather and load is captured, and the precision of the photovoltaic inverter for suppressing voltage is improved.
[0004] The application adopts the following technical solutions.
[0005] The first aspect of the application provides a photovoltaic inverter cluster cooperative suppression of power grid overvoltage method, comprising: A state space input Q function model is constructed according to the real-time measured voltage of the photovoltaic inverter cluster, a dynamic voltage threshold is output, a fusion voltage of each photovoltaic inverter is combined, and a set of out-of-limit photovoltaic inverters is obtained; The sensitivity of the photovoltaic inverter at the next moment is predicted according to the LSTM model, an optimization objective and constraint condition with the purpose of minimizing the cooperative regulation voltage cost of the photovoltaic inverter cluster are constructed in combination with the set of out-of-limit photovoltaic inverters, and the optimal power regulation amount of each photovoltaic inverter is solved; According to the optimal power regulation amount of each photovoltaic inverter, the final active power output instruction and the final active power output instruction of each photovoltaic inverter are determined, and the modulation signal of each photovoltaic inverter is obtained by modulation; The modulation signal of each photovoltaic inverter is executed, and when the grid voltage reaches a new steady state, the actual control effect vector is measured to determine whether the overvoltage of the power distribution network is successfully suppressed. If the suppression fails, the LSTM model and the Q function model are updated online, and the next round of collaborative control cycle is started, and the iteration is repeated until the overvoltage of the power distribution network is successfully suppressed.
[0006] Preferably, obtaining the set of out-of-limit photovoltaic inverters comprises: Fusing the real-time measured voltage of the photovoltaic inverter cluster and the historical voltage value of the photovoltaic inverter cluster in the last period to obtain the fused voltage of each photovoltaic inverter; According to the cumulative duration that the real-time measured voltage of the photovoltaic inverter exceeds the set threshold, the grid impedance of the photovoltaic inverter access point, the communication delay of the photovoltaic inverter, and the signal-to-noise ratio, a state space of the photovoltaic inverter is constructed; the state space is combined with the action space generated by the adjustment amount of the voltage threshold, input into the Q function model, and the optimal threshold adjustment amount selected from the action space under the state space is obtained; and the dynamic voltage threshold is calculated according to the optimal threshold adjustment amount selected from the action space under the state space; According to the fused voltage of each photovoltaic inverter and the dynamic voltage threshold, it is judged whether the photovoltaic inverter is out of limit. If the fused voltage of the photovoltaic inverter is greater than the dynamic voltage threshold, it is determined that the voltage is out of limit, and the set of out-of-limit photovoltaic inverters is obtained.
[0007] Preferably, solving the optimal power adjustment amount of each photovoltaic inverter comprises: The historical sequence of the voltage-to-reactive power sensitivity of the photovoltaic inverter in the past 24 time steps is input into the pre-trained LSTM model, and the predicted sensitivity of the photovoltaic inverter at the next time is output. According to the predicted sensitivity of the photovoltaic inverter at the next time and the set of out-of-limit photovoltaic inverters, the photovoltaic inverter weight is solved, and the optimization objective and the constraint condition for the purpose of minimizing the voltage adjustment cost of the photovoltaic inverter cluster are constructed according to the photovoltaic inverter weight. The optimization objective is solved by ADMM (Alternating Direction Method of Multipliers), and the optimal power adjustment amount of each photovoltaic inverter is obtained.
[0008] Preferably, the optimization objective is represented by the following formula:
[0009] In the formula, represents the power adjustment amount of the photovoltaic inverter i, , represents the active power change amount of the photovoltaic inverter i, represents the reactive power change amount of the photovoltaic inverter i, denotes the stage objective function of the photovoltaic inverter i, and N denotes the number of photovoltaic inverters.
[0010] Preferably, denotes the stage objective function of the photovoltaic inverter i, and is expressed by the following formula:
[0011] In the formula, denotes the reactive power of the photovoltaic inverter i, denotes the active power of the photovoltaic inverter i, denotes the coordination parameter, and denotes the upper and lower limits of the voltage in the regulation strategy of reactive power priority in stage 1, and denotes the upper and lower limits of the voltage in the regulation strategy of active power and reactive power coordination in stage 2, and denotes the upper and lower limits of the voltage in the regulation strategy of power factor constraint in stage 3, denotes the upper limit of the voltage in the regulation strategy of output recovery in stage 4, denotes the weight of the photovoltaic inverter i.
[0012] Preferably, denotes the weight of the photovoltaic inverter i, and is expressed by the following formula:
[0013] In the formula, denotes the predicted sensitivity of the photovoltaic inverter i at t+1, denotes the out-of-limit inverter set, denotes the zero-prevention constant.
[0014] Preferably, the method for obtaining the modulation signal of each photovoltaic inverter comprises: predicting the future voltage value of the photovoltaic inverter according to the sensitivity of the voltage of the photovoltaic inverter to the reactive power of other photovoltaic inverters; correcting the future voltage value of the photovoltaic inverter based on the sensitivity prediction according to the model predictive control, to obtain the reactive power instruction of the photovoltaic inverter; performing adaptive droop compensation according to the reactive power instruction of the photovoltaic inverter, to obtain the final reactive power output instruction of the photovoltaic inverter; summing the current value of the active power of the photovoltaic inverter and the optimal active regulation amount, to obtain the final active power output instruction of the photovoltaic inverter; performing PWM (Pulse Width Modulation) modulation on the final active power output instruction and the final active power output instruction, to obtain the PWM modulation signal of each photovoltaic inverter.
[0015] Preferably, the measuring the actual control effect vector to determine whether the power distribution network overvoltage is successfully suppressed comprises: measuring the new voltage of each photovoltaic inverter, obtaining the actual reactive power change amount by subtracting the reactive power reference value at the control cycle start time from the final reactive power output instruction of the photovoltaic inverter, obtaining the response time by subtracting the control start time from the control end time, and constructing the actual control effect vector according to the new voltage of the photovoltaic inverter, the actual reactive power change amount and the response time; determining whether the photovoltaic inverter cluster successfully suppresses the power distribution network overvoltage according to the new voltage of the photovoltaic inverter, and if the maximum value of the new voltage of the photovoltaic inverter in the photovoltaic inverter cluster is less than the set maximum threshold value and the minimum value is greater than the set minimum threshold value, determining that the suppression is successful, otherwise, determining that the suppression fails.
[0016] Preferably, the online learning updates the LSTM model and the Q function model comprises: solving the loss function of the LSTM model according to the predicted value of the sensitivity of the photovoltaic inverter voltage to the reactive power and the actual sensitivity of the photovoltaic inverter, and updating the LSTM model weight online according to the gradient vector of the loss function of the LSTM model; constructing the immediate reward according to the new voltage of the photovoltaic inverter, the active power reduction amount and the response time, and updating the Q function online according to the immediate reward.
[0017] The second aspect of the present application provides a photovoltaic inverter cluster collaborative suppression of power distribution network overvoltage system, which runs the photovoltaic inverter cluster collaborative suppression of power distribution network overvoltage method described in the first aspect, comprising: an out-of-limit photovoltaic inverter solving module for constructing a state space input Q function model according to the real-time measured voltage of the photovoltaic inverter cluster, outputting a dynamic voltage threshold, and obtaining an out-of-limit photovoltaic inverter set in combination with the fusion voltage of each photovoltaic inverter; an optimal power adjustment amount solving module for predicting the sensitivity of the photovoltaic inverter at the next time according to the LSTM model, constructing an optimization objective and constraint condition for the purpose of minimizing the voltage adjustment cost of the photovoltaic inverter cluster in combination with the out-of-limit photovoltaic inverter set to solve the optimal power adjustment amount of each photovoltaic inverter; a modulation signal solving module for determining the final active power output instruction and the final active power output instruction of each photovoltaic inverter according to the optimal power adjustment amount of each photovoltaic inverter, and modulating to obtain the modulation signal of each photovoltaic inverter; an output module for executing the modulation signal of each photovoltaic inverter, measuring the actual control effect vector to determine whether the power distribution network overvoltage is successfully suppressed when the grid voltage reaches a new stable state, updating the LSTM model and the Q function model online when the suppression fails, starting the next round of collaborative control cycle, and repeating iteration until the power distribution network overvoltage is successfully suppressed.
[0018] Compared with the prior art, the beneficial effects of the present application at least include: The Q function model of reinforcement learning dynamically adjusts the voltage threshold of the photovoltaic inverter, solves the problems of poor adaptability and high malfunction rate of the traditional fixed threshold in complex scenarios such as sudden change of light and change of grid impedance, improves the intelligent level of voltage control and the robustness to different operating conditions, reduces the malfunction rate of the photovoltaic inverter, and reduces the adjustment oscillation or response delay of the photovoltaic inverter caused by unreasonable threshold setting; The space-time prediction of voltage-reactive power sensitivity is performed through LSTM, the problems of the offline sensitivity calculation model being unable to adapt to the change of grid operation mode and large prediction error are solved, the accuracy of photovoltaic inverter group collaborative optimization is improved, and the sensitivity prediction error is reduced; The parameters are online self-adaptive set through fuzzy control, the problems of the PID and droop coefficient control parameters being dependent on experience setting and being unable to be optimized online are solved, the adaptive ability and dynamic adjustment precision of the system are improved, the influence of system instability caused by parameter mismatch is reduced, and the precision and stability of voltage suppression of the photovoltaic inverter are improved. BRIEF DESCRIPTION OF DRAWINGS
[0019] Figure 1 is a schematic diagram of a photovoltaic inverter cluster collaborative overvoltage suppression process for a power distribution network provided according to an embodiment of the present application. DETAILED DESCRIPTION
[0020] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application, not all the embodiments. All other embodiments obtained by those skilled in the art without creative labor based on the spirit of the present application are within the protection scope of the present application.
[0021] As Figure 1 shown, embodiment 1 of the present application provides a photovoltaic inverter cluster collaborative overvoltage suppression method, including the following steps: Step 1, constructing a state space input Q function model according to real-time measured voltages of the photovoltaic inverter cluster, outputting a dynamic voltage threshold, and combining the fusion voltages of each photovoltaic inverter to obtain an out-of-limit photovoltaic inverter set.
[0022] In a preferred but non-limiting embodiment of the present application, step 1 includes: Step 1.1, fusing the real-time measured voltages of the photovoltaic inverter cluster and the historical voltage values of the photovoltaic inverter cluster from the previous period to obtain the fusion voltage of each photovoltaic inverter, which is expressed by the following formula:
[0023] wherein, represents the fusion voltage of photovoltaic inverter i, represents the real-time measured grid-connected point voltage value of photovoltaic inverter i, represents the historical voltage value of photovoltaic inverter cluster from the last cycle, if there is no historical voltage value of photovoltaic inverter cluster from the last cycle, and is a weight for adjusting whether to rely on real-time measurement or historical data.
[0024] Step 1.2, constructing a state space of the photovoltaic inverter according to the cumulative duration that the real-time measured voltage of the photovoltaic inverter exceeds a set threshold value, the grid impedance of the photovoltaic inverter access point, the communication delay of the photovoltaic inverter, and the signal-to-noise ratio; combining the state space with the action space generated by the adjustment amount of the voltage threshold value, inputting the Q function model, and obtaining the optimal threshold adjustment amount selected from the action space A under the state space, which is expressed as follows:
[0025] wherein, represents the state space under which the optimal threshold adjustment amount is selected from the action space A, represents the state space of photovoltaic inverter i at the current time t, represents the state vector under which the expected cumulative reward of performing action a is executed, represents finding the action a in the set A that makes maximize, and A represents the action space composed of all possible threshold adjustment amounts.
[0026] calculating the dynamic voltage threshold according to the optimal threshold adjustment amount selected from the action space A under the state space, which is expressed as follows:
[0027] wherein, represents the dynamic voltage threshold, represents the upper limit of the distribution network voltage, represents the state vector under which the expected cumulative reward of performing action is executed, represents the maximum threshold adjustment range.
[0028] Step 1.3, judging whether the photovoltaic inverter is out of limit according to the fusion voltage of the photovoltaic inverter obtained in step 1.1 and the dynamic voltage threshold obtained in step 1.2, if the fusion voltage of the photovoltaic inverter is greater than the dynamic voltage threshold, it is determined that the voltage is out of limit, and the out-of-limit photovoltaic inverter set is obtained.
[0029] Notably, the application dynamically adjusts the threshold value through reinforcement learning, avoids the premature action or delayed action of the photovoltaic inverter caused by the fixed threshold value in scenarios such as rapid changes in illumination, and reduces the misoperation rate of the photovoltaic inverter cluster in the process of collaborative suppression of the power grid overvoltage; by fusing multi-dimensional state information such as communication quality and grid impedance, the system can perceive the current operating environment, thereby making more reasonable control start decisions and improving the adaptive ability of the photovoltaic inverter cluster in different operating conditions in the process of collaborative suppression of the power grid overvoltage, such as strong / weak grid and good / poor communication, but not limited to; the application ensures that the photovoltaic inverter cluster collaborative optimization calculation is only for the nodes that really need to be adjusted through the out-of-limit judgment and node set screening, avoids unnecessary waste of computing resources and control interference, and improves the efficiency of the photovoltaic inverter cluster in the process of collaborative suppression of the power grid overvoltage.
[0030] Step 2, according to the predicted sensitivity of the photovoltaic inverter at the next time by the LSTM model, the out-of-limit photovoltaic inverter set is combined to build an optimization objective and constraint condition for the purpose of minimizing the voltage adjustment cost of the photovoltaic inverter cluster, so as to solve the optimal power adjustment amount of each photovoltaic inverter.
[0031] In the preferred but non-limiting embodiments of the application, step 2 includes: Step 2.1, input the historical sequence of voltage versus reactive power sensitivity of each photovoltaic inverter in the past 24 time steps into the pre-trained LSTM model, and output the predicted sensitivity of the photovoltaic inverter at the next time, which is expressed as follows:
[0032] In the formula, represents the predicted sensitivity of the photovoltaic inverter i at t+1, represents the pre-trained LSTM model, represents the historical sequence of voltage versus reactive power sensitivity of the photovoltaic inverter i in the past 24 time steps, represents the historical data from the current time t to the past 24 time steps, represents the voltage U versus reactive power Q sensitivity of the photovoltaic inverter i.
[0033] Step 2.2, according to the predicted sensitivity of the photovoltaic inverter at the next time in step 2.1 and the out-of-limit photovoltaic inverter set obtained in step 1, solve the photovoltaic inverter weight, and build an optimization objective for the purpose of minimizing the voltage adjustment cost of the photovoltaic inverter cluster according to the photovoltaic inverter weight, which is expressed as follows:
[0034] In the formula, represents the power adjustment amount of the photovoltaic inverter i, , represents the active power variation of the photovoltaic inverter i, represents the reactive power variation of the photovoltaic inverter i, represents the stage target function of the photovoltaic inverter i, and is expressed as follows:
[0035] In the formula, represents the reactive power of the photovoltaic inverter i, represents the active power of the photovoltaic inverter i, represents the coordination parameter, and represents the upper and lower limits of the voltage in the first stage with the regulation strategy of reactive power priority, and represents the upper and lower limits of the voltage in the second stage with the regulation strategy of active and reactive power coordination, and represents the upper and lower limits of the voltage in the third stage with the regulation strategy of power factor constraint, represents the upper limit of the voltage in the fourth stage with the regulation strategy of output recovery, represents the weight of the photovoltaic inverter i, and is expressed as follows:
[0036] In the formula, represents the out-of-limit inverter set, represents the zero-prevention constant.
[0037] The constraint condition of the optimization target is expressed as follows:
[0038] In the formula, represents the voltage of the photovoltaic inverter i, and respectively represent the lower and upper limits of the voltage safety constraint, represents the power factor of the photovoltaic inverter i.
[0039] The optimization target is solved by ADMM, and is expressed as follows:
[0040] In the formula, represents the optimization result of the photovoltaic inverter i in the k+1th iteration, represents the penalty coefficient of the consistency constraint, represents the global consensus value of the kth iteration, represents the dual variable of the photovoltaic inverter i in the kth iteration, represents the square of the L2 norm.
[0041] When the ADMM converges, i.e. , the optimal power regulation amount of the photovoltaic inverter i is obtained , represents the optimal active regulation amount, represents the optimal reactive regulation amount.
[0042] Notably, based on the dynamic weight allocation mechanism of the predicted sensitivity, the fair responsibility allocation in the photovoltaic inverter cluster collaboration is realized, the technical problem that the traditional average weight or fixed weight algorithm cannot distinguish the node voltage regulation capacity difference and leads to low collaboration efficiency is solved, the accuracy of reactive resource utilization and the overall regulation efficiency of the cluster are improved, and the total reactive regulation cost and equipment action loss are reduced; By constructing a multi-stage adaptive objective function, the smooth switching of the incremental optimization strategy is realized, the technical problem that a single optimization target cannot adapt to different overvoltage severity and cannot balance between control safety and economy is solved, the adaptive ability of the photovoltaic inverter cluster collaboration control in different scenes such as mild overrunning and severe overrunning is improved, the active power generation loss caused by improper control strategy is reduced, the total regulation cost of the cluster is minimized to achieve the voltage control target, the optimization efficiency of the photovoltaic inverter cluster collaboration to suppress overvoltage is improved, and the total cost of the photovoltaic inverter cluster collaboration to suppress overvoltage is reduced; The sensitivity dynamic allocation optimization weight based on LSTM prediction makes the inverters with greater influence on voltage bear more voltage regulation responsibility, solves the technical problem that the traditional optimization is based on current time information and causes decision lag and control overshoot due to system inertia, improves the forwardness and stability of the photovoltaic inverter cluster collaboration to suppress overvoltage, and reduces the oscillation risk of the photovoltaic inverter cluster collaboration to suppress overvoltage; By constructing an optimization framework in which local constraints and coupling constraints are coordinated by a distributed algorithm, effective decomposition of complex cluster optimization problems is realized, the technical problem that centralized optimization has high computational complexity and is difficult to meet the real-time requirements of large-scale clusters is solved, the scalability and calculation speed of the photovoltaic inverter cluster collaboration to suppress overvoltage are improved, and the dependence on the calculation ability of the central processor is reduced.
[0043] Step 3, according to the optimal power regulation amount of each photovoltaic inverter, the final active power output instruction and the final active power output instruction of each photovoltaic inverter are determined, and the modulation signal of each photovoltaic inverter is obtained.
[0044] In the preferred but non-limiting embodiments of the present application, step 3 comprises: Step 3.1, predict the future voltage value of the photovoltaic inverter according to the sensitivity of the voltage of the photovoltaic inverter to the reactive power of other photovoltaic inverters, which is expressed by the following formula:
[0045] wherein, denotes the future voltage value of photovoltaic inverter i based on sensitivity prediction, denotes the actual value of grid-connected point voltage measured by photovoltaic inverter i at current time t, denotes the sensitivity of photovoltaic inverter i voltage to photovoltaic inverter j reactive power, denotes the optimal reactive regulation amount of photovoltaic inverter j, and N denotes the number of photovoltaic inverters.
[0046] The future voltage value of photovoltaic inverter based on sensitivity prediction is corrected according to model predictive control to obtain the reactive instruction of photovoltaic inverter, which is expressed by the following formula:
[0047] wherein, denotes the reactive instruction of photovoltaic inverter i, denotes the proportional correction coefficient, denotes the integral correction coefficient, denotes the voltage measurement value of photovoltaic inverter i at historical time τ, Step 3.2, adaptive droop compensation is performed according to the reactive instruction of photovoltaic inverter to obtain the final reactive power output instruction of photovoltaic inverter, which is expressed by the following formula:
[0048] wherein, denotes the final reactive power output instruction of photovoltaic inverter i, denotes the voltage reference value, denotes the adaptive droop coefficient of photovoltaic inverter i, which is expressed by the following formula:
[0049] wherein, denotes the basic droop coefficient, denotes the sensitivity adaptive gain coefficient.
[0050] The final active power output instruction of photovoltaic inverter i is obtained by summing the active power current value of photovoltaic inverter i at time t and the optimal active regulation amount, which is expressed by the following formula:
[0051] wherein, denotes the active power current value of photovoltaic inverter i at time t.
[0052] The final power instruction is obtained:
[0053] Step 3.3, after the final power instruction is PWM modulated, the PWM modulation signal of each photovoltaic inverter is output, which is expressed as follows:
[0054] In the formula, represents the d-axis current reference value of the photovoltaic inverter i, represents the DC side voltage of the photovoltaic inverter i, represents the q-axis current reference value of the photovoltaic inverter i, represents the PWM modulation signal of the photovoltaic inverter i, which is used to control the on-off timing of the photovoltaic inverter switch tube, represents a space vector pulse width modulation algorithm, which converts the current reference into a switching signal.
[0055] It is worth noting that the present application compensates for the control deviation caused by the sensitivity prediction error and the communication delay by feeding forward correction of the optimization instruction through the robust MPC (Model Predictive Control), improves the voltage overshoot of the photovoltaic inverter cluster in cooperatively suppressing overvoltage, and improves the control accuracy and robustness of the photovoltaic inverter cluster in cooperatively suppressing overvoltage; while executing the optimization instruction, the present application superimposes the local adaptive droop control to provide millisecond-level compensation for fast voltage fluctuations, forming a perfect complement of the optimization instruction solving steady-state deviation and the droop control suppressing dynamic fluctuations, thereby improving the control accuracy of the photovoltaic inverter cluster in cooperatively suppressing overvoltage; through power factor constraints and other restrictions, the present application eliminates secondary problems such as power factor exceeding the standard or power shock caused by voltage regulation, and ensures the safety of the photovoltaic inverter itself and the power quality of the power grid while suppressing overvoltage.
[0056] Step 4, execute the modulation signal of each photovoltaic inverter, when the grid voltage reaches a new stable state, measure the actual control effect vector to determine whether the distribution network overvoltage is successfully suppressed, if not, update the LSTM model and the Q function model online, start the next round of cooperative control cycle, and repeat iteration until the distribution network overvoltage is successfully suppressed.
[0057] In the preferred but non-limiting embodiments of the present application, step 4 includes: Step 4.1, the PWM signal is converted into the conduction timing of the photovoltaic inverter switch tube, the DC power is converted into a high-frequency pulse voltage through the switching action of the IGBT bridge arm according to the conduction timing of the photovoltaic inverter switch tube, the high-frequency component is filtered out through the LCL filter, and the 50Hz sine wave voltage is output. The alternating voltage generated by the transformer is injected into the power distribution network, the injected reactive power changes the line voltage drop, thereby adjusting the PCC point voltage, waiting for 20-30ms for the grid voltage to reach a new stable state, performing ADC sampling, measuring the new voltage of each photovoltaic inverter, calculating the actual reactive power change amount according to the difference between the final reactive power output instruction of the photovoltaic inverter and the reactive power reference value at the beginning of the control period, calculating the response time according to the control end time minus the control start time, and constructing the actual control effect vector according to the new voltage of the photovoltaic inverter, the actual reactive power change amount and the response time.
[0058] According to the new voltage of the photovoltaic inverter, it is determined whether the photovoltaic inverter cluster cooperatively suppresses the overvoltage of the power distribution network successfully. If the maximum value of the new voltage of the photovoltaic inverter in the photovoltaic inverter cluster is less than the set maximum threshold value, and the minimum value is greater than the set minimum threshold value, it is determined that the suppression is successful, otherwise the suppression fails.
[0059] Step 4.2, according to the actual control effect vector, the LSTM model is updated online to obtain the LSTM model weight updated online.
[0060] Further preferably, step 4.2 comprises: Step 4.2.1, solving the loss function of the LSTM model according to the predicted value of the sensitivity of the photovoltaic inverter voltage to the reactive power and the actual sensitivity of the photovoltaic inverter, and updating the LSTM model weight online according to the gradient vector of the loss function of the LSTM model, which is expressed as follows:
[0061] In the formula, LSTM model loss function, training sample number, the predicted value of the sensitivity of the photovoltaic inverter i voltage to the reactive power of the mth sample, the new voltage of the photovoltaic inverter i, actual reactive power change amount, the actual sensitivity of the photovoltaic inverter i; LSTM model reward and punishment coefficient, voltage suppression success, LSTM model weight updated online, initial LSTM model weight, LSTM model learning rate, denotes a gradient vector of
[0062] Step 4.2.2, constructing an immediate reward according to the new voltage of the photovoltaic inverter, the active power reduction amount and the response time, updating the Q function online according to the immediate reward.
[0063] More preferably, step 4.2.2 comprises: updating the Q function online, which is expressed as follows:
[0064] wherein, denotes a Q function learning rate, denotes a discount factor, denotes, denotes an immediate reward, which is expressed as follows:
[0065] wherein, denotes a voltage deviation weight, denotes an active power reduction weight, denotes an active power reduction amount of the photovoltaic inverter i, denotes a response time.
[0066] Step 4.3, according to the new voltage of the photovoltaic inverter and the state vector of the next time, the system restarts the next round of collaborative control cycle, and repeats iteration until the photovoltaic inverter cluster successfully collaboratively suppresses the overvoltage of the power distribution network. If successful, it can enter a low-power monitoring mode until the voltage is out of limit again.
[0067] It is worth noting that the actual control effect (V , ) obtained by step 4.1 is used as a real label, and the pre-trained LSTM model is updated by online gradient descent with a prediction error as a loss function, which solves the technical problem of time-varying voltage-reactive sensitivity caused by changes in illumination conditions, network topology and load, thereby reducing the control instruction error caused by model mismatch and improving the accuracy of collaborative optimization; the present application solves the technical problem that traditional open-loop or semi-closed-loop control cannot learn from experience and cannot cope with long-term evolution of the system, improves the lifelong learning ability and long-term performance stability of the entire control system, realizes the self-evolution characteristics of getting smarter, reduces the operation and maintenance cost and the dependence on artificial intervention during the life cycle, and realizes the autonomous and intelligent operation of the system.
[0068] Embodiment 2 of the present application provides a photovoltaic inverter cluster cooperative suppression power grid overvoltage system, which runs the photovoltaic inverter cluster cooperative suppression power grid overvoltage method of embodiment 1, comprising: An out-of-limit photovoltaic inverter solving module is configured to construct a state space input Q function model according to real-time measurement voltage of the photovoltaic inverter cluster, output a dynamic voltage threshold, and obtain an out-of-limit photovoltaic inverter set in combination with fusion voltages of the photovoltaic inverters; An optimal power adjustment amount solving module is configured to predict the sensitivity of the photovoltaic inverter at the next moment according to the LSTM model, construct an optimization target and constraint condition for the purpose of minimizing the voltage regulation cost of the photovoltaic inverter cluster in combination with the out-of-limit photovoltaic inverter set, and solve the optimal power adjustment amount of each photovoltaic inverter; A modulation signal solving module is configured to determine the final active power output instruction and the final active power output instruction of each photovoltaic inverter according to the optimal power adjustment amount of each photovoltaic inverter, and modulate the modulation signal of each photovoltaic inverter; An output module is configured to execute the modulation signal of each photovoltaic inverter, measure the actual control effect vector to determine whether the power grid overvoltage is successfully suppressed when the grid voltage reaches a new stable state, update the LSTM model and the Q function model online when the suppression fails, start the next round of cooperative control cycle, and repeat iteration until the power grid overvoltage is successfully suppressed.
[0069] Compared with the prior art, the present application has at least the following beneficial effects: The Q function model of reinforcement learning is used to dynamically adjust the voltage threshold of the photovoltaic inverter, solving the problems of poor adaptability and high misoperation rate of the traditional fixed threshold in complex scenarios such as sudden change of light and change of grid impedance, improving the intelligent level of voltage control and the robustness to different operating conditions, reducing the misoperation rate of the photovoltaic inverter, and reducing the regulation oscillation or response delay of the photovoltaic inverter caused by unreasonable threshold setting; The LSTM is used for spatio-temporal prediction of voltage-reactive power sensitivity, solving the problems of large prediction error and inability of offline sensitivity calculation model to adapt to changes in grid operation mode, improving the accuracy of photovoltaic inverter cluster cooperative optimization, and reducing the sensitivity prediction error; The fuzzy control is used for online adaptive setting of parameters, solving the problems of fixed parameters such as PID and droop coefficient that depend on experience setting and cannot be optimized online, improving the adaptive ability and dynamic regulation accuracy of the system, reducing the influence of system instability caused by parameter mismatch, and improving the accuracy and stability of photovoltaic inverter voltage suppression. The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for causing a processor to implement various aspects of the present disclosure.
[0070] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application but not to limit it. Although the present application has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the specific embodiments of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and any modification or equivalent replacement should be covered within the protection scope of the claims of the present application.
Claims
1. A method for photovoltaic inverter cluster collaborative suppression of overvoltage in power distribution network, characterized in that: a state space input Q function model is constructed according to real-time measured voltage of the photovoltaic inverter cluster, a dynamic voltage threshold value is output, a set of out-of-limit photovoltaic inverters is obtained in combination with the fusion voltage of each photovoltaic inverter; the sensitivity of the photovoltaic inverter at the next time is predicted according to the LSTM model, the optimization objective and the constraint condition for the purpose of minimizing the voltage adjustment cost of the photovoltaic inverter cluster are constructed in combination with the set of out-of-limit photovoltaic inverters to solve the optimal power adjustment amount of each photovoltaic inverter; the final active power output instruction and the final active power output instruction of each photovoltaic inverter are determined according to the optimal power adjustment amount of each photovoltaic inverter, and the modulation signal of each photovoltaic inverter is obtained by modulation; the modulation signal of each photovoltaic inverter is executed, when the grid voltage reaches a new stable state, the actual control effect vector is measured to determine whether the overvoltage in the power distribution network is successfully suppressed, if the suppression fails, the LSTM model and the Q function model are updated online, the next round of collaborative control cycle is started, and the iteration is repeated until the overvoltage in the power distribution network is successfully suppressed. 2.The method according to claim 1, characterized in that: obtaining the set of out-of-limit photovoltaic inverters comprises: fusing the real-time measured voltage of the photovoltaic inverter cluster and the historical voltage value of the photovoltaic inverter cluster from the last cycle to obtain the fusion voltage of each photovoltaic inverter; constructing the state space of the photovoltaic inverter according to the cumulative duration of the real-time measured voltage of the photovoltaic inverter exceeding the set threshold value, the grid impedance of the photovoltaic inverter access point, the communication delay of the photovoltaic inverter, and the signal-to-noise ratio; inputting the state space into the Q function model in combination with the action space generated by the adjustment amount of the voltage threshold value to obtain the optimal threshold adjustment amount selected from the action space under the state space; calculating the dynamic voltage threshold value according to the optimal threshold adjustment amount selected from the action space under the state space; determining whether the photovoltaic inverter is out of limit according to the fusion voltage of each photovoltaic inverter and the dynamic voltage threshold value, if the fusion voltage of the photovoltaic inverter is greater than the dynamic voltage threshold value, it is determined that the voltage is out of limit, and the set of out-of-limit photovoltaic inverters is obtained. 3.The method according to claim 1, characterized in that: solving the optimal power adjustment amount of each photovoltaic inverter comprises: inputting the historical sequence of the voltage versus reactive power sensitivity of the photovoltaic inverter at the past 24 time steps into the pre-trained LSTM model to output the predicted sensitivity of the photovoltaic inverter at the next time; solving the photovoltaic inverter weight according to the predicted sensitivity of the photovoltaic inverter at the next time and the set of out-of-limit photovoltaic inverters, constructing the optimization objective and the constraint condition for the purpose of minimizing the voltage adjustment cost of the photovoltaic inverter cluster according to the photovoltaic inverter weight; solving the optimization objective by ADMM to obtain the optimal power adjustment amount of each photovoltaic inverter. 4.The method according to claim 3, characterized in that: the optimization objective is represented by the following formula: wherein represents the power regulation amount of the photovoltaic inverter i, , represents the active power variation amount of the photovoltaic inverter i, represents the reactive power variation amount of the photovoltaic inverter i, represents the stage objective function of the photovoltaic inverter i, and N represents the number of photovoltaic inverters.
5. A method for collaboratively suppressing overvoltage in a distribution network using a cluster of photovoltaic inverters according to claim 4, characterized in that: The stage objective function representing the photovoltaic inverter i is expressed as follows: wherein, represents the reactive power of the photovoltaic inverter i, represents the active power of the photovoltaic inverter i, represents the coordination parameter, and represents the voltage upper and lower limits for phase 1 with reactive power priority as the regulation strategy, and represents the voltage upper and lower limits for phase 2 with active and reactive power coordination as the regulation strategy, and represents the voltage upper and lower limits for phase 3 with power factor constraint as the regulation strategy, represents the voltage upper limit for phase 4 with power output recovery as the regulation strategy, represents the weight of the photovoltaic inverter i.
6. A method for collaboratively suppressing overvoltage in a distribution network using a cluster of photovoltaic inverters according to claim 5, characterized in that: The weight of the photovoltaic inverter i is represented by the following equation: wherein, denotes the predicted sensitivity of the photovoltaic inverter i at time t+1, denotes the set of out-of-limit inverters, denotes the anti-zero constant.
7. The method for collaborative suppression of overvoltage in a distribution network by a cluster of photovoltaic inverters according to claim 1, characterized in that: Obtaining the modulation signals of each photovoltaic inverter includes: The future voltage value of the photovoltaic inverter is predicted based on the sensitivity of the photovoltaic inverter's voltage to the reactive power of other photovoltaic inverters; the future voltage value of the photovoltaic inverter based on the sensitivity prediction is corrected by model predictive control to obtain the reactive power command of the photovoltaic inverter; Adaptive droop compensation is performed based on the reactive power command of the photovoltaic inverter to obtain the final reactive power output command of the photovoltaic inverter; the final active power output command of the photovoltaic inverter is obtained by summing the current active power value and the optimal active power regulation. The final active power output command and the final active power output command are modulated by PWM and then output to obtain the PWM modulation signal of each photovoltaic inverter.
8. The method for collaborative suppression of overvoltage in a distribution network by a cluster of photovoltaic inverters according to claim 1, characterized in that: Measuring the actual control effect vector to determine whether overvoltage in the distribution network has been successfully suppressed includes: The new voltage of each photovoltaic inverter is measured, the actual reactive power change is obtained by subtracting the final reactive power output command of the photovoltaic inverter from the reactive power reference value at the beginning of the control cycle, the response time is obtained by subtracting the control start time from the control end time, and the actual control effect vector is constructed based on the new voltage of the photovoltaic inverter, the actual reactive power change and the response time. The success of the photovoltaic inverter cluster's collaborative suppression of overvoltage in the distribution network is determined by the new voltage of the photovoltaic inverters. If the maximum new voltage of the photovoltaic inverters in the photovoltaic inverter cluster is less than the set maximum value threshold and the minimum new voltage is greater than the set minimum value threshold, the suppression is considered successful; otherwise, the suppression fails.
9. A method for collaboratively suppressing overvoltage in a distribution network using a cluster of photovoltaic inverters according to claim 1, characterized in that: Online learning updates for LSTM and Q-function models include: The loss function of the LSTM model is solved based on the predicted value of the photovoltaic inverter voltage sensitivity to reactive power and the actual sensitivity of the photovoltaic inverter. The weights of the LSTM model are then updated online based on the gradient vector of the loss function of the LSTM model. An instant reward is constructed based on the new voltage, active power reduction, and response time of the photovoltaic inverter, and the Q function is updated online based on the instant reward.
10. A photovoltaic inverter cluster collaborative overvoltage suppression system for distribution networks, comprising the photovoltaic inverter cluster collaborative overvoltage suppression method for distribution networks as described in any one of claims 1-9, characterized in that: The over-limit photovoltaic inverter solution module is used to construct a state-space input Q-function model based on the real-time measured voltage of the photovoltaic inverter cluster, output a dynamic voltage threshold, and obtain the over-limit photovoltaic inverter set by combining the fused voltage of each photovoltaic inverter. The optimal power adjustment amount solving module is configured to predict the sensitivity of the photovoltaic inverter at the next moment according to the LSTM model, combine the out-of-limit photovoltaic inverter set, construct an optimization target and constraint condition with the minimum voltage regulation cost of the photovoltaic inverter cluster for the purpose of solving the optimal power adjustment amount of each photovoltaic inverter; The modulation signal solving module is configured to determine the final active power output instruction and the final active power output instruction of each photovoltaic inverter according to the optimal power adjustment amount of each photovoltaic inverter, and modulate the modulation signal of each photovoltaic inverter; The output module is configured to execute the modulation signal of each photovoltaic inverter, measure the actual control effect vector when the grid voltage reaches a new stable state to determine whether the overvoltage of the distribution network is successfully suppressed, update the LSTM model and the Q function model online when the suppression fails, start the next round of collaborative control cycle, and repeat the iteration until the overvoltage of the distribution network is successfully suppressed.