Multi-inverter control mode adaptive switching method based on operation short circuit ratio

Abstract

The multi-inverter control mode self-adaptive switching method based on the operation short circuit ratio comprises the steps that an optical storage station grid-connected model is built, and inverter control modes are set to be a GFL mode and a GFM mode; acquiring operation data of a grid-connected point of the optical storage station, and calculating a numerical value of an operation short-circuit ratio (OSCR); when the operation short circuit ratio OSCR is smaller than C1 or larger than C2, the multi-inverter control mode switching optimization requirement is triggered; meanwhile, initializing a to-be-optimized variable; an MATD3 algorithm is adopted to determine an optimal switching combination of a multi-inverter control mode; issuing the output execution scheme of the optimal switching combination in the control period to a station inverter; when the operation short circuit ratio OSCR is between (C1 and C2), triggering the multi-inverter control mode switching optimization requirement again, outputting an optimal switching combination scheme by the MATD3 algorithm, and issuing the optimal switching combination scheme to the station inverter; wherein a part of the inverters are switched to GFL operation according to an optimization result, and the other part of the inverters are kept to control GFM operation in a network construction mode. According to the method, the collaborative minimization of dynamic voltage support and operation loss is realized.

Description

Technical Field

[0001] This invention relates to the field of new energy grid-connected control technology, specifically to an adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio. Background Technology

[0002] As the proportion of new energy sources such as photovoltaics and energy storage continues to increase, the power grid faces increasingly complex operational challenges. Especially in weak grid environments, insufficient short-circuit capacity and high line impedance make the grid connection point voltage more susceptible to fluctuations in photovoltaic output and load changes, leading to overvoltage or undervoltage issues that affect power supply security and power quality.

[0003] Currently, inverters in photovoltaic and energy storage power plants mainly adopt fixed control modes: grid-following (GFL) is suitable for maximum power point tracking but lacks voltage support capabilities; grid-forming (GFM) has the ability to actively regulate voltage, but sacrifices some active power output and increases grid losses. A single fixed mode cannot simultaneously meet the requirements of voltage stability and operational economy, becoming a prominent challenge for grid-connected operation under weak grid conditions.

[0004] The commonly used method is to set thresholds based on indicators such as grid connection point voltage or short-circuit ratio to trigger control mode switching. However, this method mainly relies on static thresholds or empirical rules, which has three shortcomings: First, the thresholds are difficult to dynamically reflect complex grid conditions, and the switching timing is not precise enough; second, it often uses unified switching of all inverters, leading to overcompensation or voltage oscillation; and third, it lacks global optimization and cannot intelligently decide on specific switching combinations and the number of inverters to be switched based on real-time system status.

[0005] To address this issue, researchers introduced the Operating Short-Circuit Ratio (OSCR) as a dynamic quantification of grid strength and designed a control mode switching strategy triggered by OSCR. OSCR, by combining the total active power at the grid connection point with the equivalent grid impedance in real time, can more sensitively reflect voltage exceedance risks. However, relying solely on OSCR triggering cannot solve the combined optimization problem of which inverters should switch and how many should be switched.

[0006] In traditional approaches, some studies have attempted to use heuristic algorithms (such as particle swarm optimization, PSO) to perform global search optimization of multi-inverter switching strategies, which can reduce voltage fluctuations and grid losses. However, standard PSO also suffers from problems such as being prone to getting trapped in local optima, having fixed parameters that are difficult to adapt to dynamic grid disturbances, and having high computational complexity, making it difficult to balance real-time performance and global performance in complex scenarios.

[0007] Therefore, for the optimization problem of multi-inverter control mode combination in weak power grids, it is urgent to introduce more advanced adaptive optimization methods, such as multi-agent deep reinforcement learning (MADRL). Through a structure of centralized training and distributed execution, each inverter can dynamically decide its switching behavior based on local and global information, which can achieve synergistic optimization of voltage stability, transient support capability and network loss economy, thereby improving the operating performance and intelligence level of photovoltaic and energy storage stations under complex power grid conditions. Summary of the Invention

[0008] This invention provides an adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio. This method establishes a grid-connected model of a photovoltaic-storage power station using Python and employs the Multi-Agent Dual-Delay Deep Deterministic Strategy Gradient Algorithm (MATD3) to intelligently determine the optimal switching combination of multiple inverters, thereby achieving the synergistic minimization of dynamic voltage support and operating losses.

[0009] The technical solution adopted in this invention is as follows:

[0010] The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio includes the following steps:

[0011] Step 1: Build a grid-connected model for the photovoltaic and energy storage power station. Set the control mode of each photovoltaic and energy storage inverter to: grid-following control GFL and grid-connecting control GFM mode.

[0012] Step 2: Collect the operation data of the photovoltaic and energy storage power station grid connection point, calculate the value of the operating short-circuit ratio (OSCR), and set the upper and lower thresholds C1 and C2;

[0013] Step 3: When the operating short-circuit ratio OSCR is less than C1 or greater than C2, trigger the multi-inverter control mode switching optimization requirement; at the same time, initialize the variables to be optimized;

[0014] Step 4: Use the Multi-Agent Dual-Delay Deep Deterministic Policy Gradient (MATD3) algorithm to determine the optimal switching combination of multi-inverter control modes;

[0015] Step 5: Send the execution plan of the optimal switching combination for this control cycle output in Step 4 to the inverters in the field station; some inverters switch to the grid-connected control GFM mode according to the optimization results, while the remaining inverters keep running the grid-connected control GFL mode.

[0016] Step 6: When the operating short-circuit ratio OSCR is between (C1, C2), the multi-inverter control mode switching optimization requirement is triggered again. The MATD3 algorithm outputs the optimal switching combination scheme and sends it to the station inverters. Some inverters switch to grid-following control GFL operation according to the optimization results, while the other inverters maintain grid-connected control GFM operation.

[0017] In step 1, a grid-connected model of the photovoltaic and energy storage power station is built according to the typical values ​​of the preset control parameters, and the grid impedance is set so that the short-circuit ratio of the grid connection point of the photovoltaic and energy storage power station is 1.5.

[0018] The control equations for the grid-following control (GFL) and grid-connected control (GFM) of the photovoltaic-storage grid-connected inverter are expressed as follows:

[0019]

[0020] In equation (1): i d i q The inverter output current i are respectively abc dq axis components; e d e q These are the grid voltages e abc dq axis components; u d u q The bridge arm voltages u are respectively abc The dq-axis components; ω is the grid angular frequency; L is the inductance.

[0021]

[0022] In the formula: f * U * These represent the frequency and voltage values, respectively; f and U represent the frequency and voltage ratings for droop control, respectively; P and Q represent the active and reactive power ratings, respectively; m and n represent the frequency and voltage droop coefficients, respectively; P * Q * These are the real-time detection values ​​of the active and reactive power outputs of the inverter, respectively.

[0023] In step 2, the observed operating data includes active power P, reactive power Q, and grid connection point voltage U. The operating short-circuit ratio (OSCR) is calculated based on the real-time collected active power P, and upper and lower thresholds C1 and C2 are set to form the trigger basis for inverter control mode switching.

[0024] The formulas for calculating the grid connection voltage and short-circuit ratio of photovoltaic-storage grid connection include:

[0025] The formula for calculating the voltage at the photovoltaic-storage grid connection point is:

[0026]

[0027] In the formula: Indicates the grid connection point voltage; R eq Indicates the equivalent resistance on the grid side at the grid connection point; X eq P represents the equivalent reactance on the grid side at the grid connection point. PCC Consider the injected active power for the photovoltaic system; Q PCC Reactive power; This represents the voltage at the end of the line.

[0028] The formula for calculating the Operating Short-Circuit Ratio (OSCR) is as follows:

[0029]

[0030] In the formula: S ac For the short-circuit capacity of the AC system; S N Rated capacity of new energy equipment; P * Z represents the per-unit value of the actual output power of new energy equipment. * This represents the per-unit value of the equivalent impedance of the power grid.

[0031] In step 3, when the operating short-circuit ratio OSCR is less than C1 or greater than C2, the multi-inverter control mode switching optimization requirement will be triggered; at the same time, combined with parameters such as voltage fluctuation at the grid connection point of the photovoltaic-storage power station and grid load, the variables to be optimized are initialized, and the control mode flag bit of each inverter is: GFL / GFM.

[0032] In step 4, the Multi-Agent Dual-Delay Deep Deterministic Policy Gradient (MATD3) algorithm designs the reward function as a weighted sum of grid connection point voltage deviation and line network loss. Through a centralized training-distributed execution (CTDE) architecture, each agent learns the optimal control mode switching strategy based on local observations and partial global information, thereby achieving combined optimization of inverter control modes.

[0033] Step 4 includes:

[0034] Step 4.1: Initialize the MATD3 algorithm network structure, including two Critic networks and one Actor network. The two Critic networks are fitted with the Q-value function respectively. To avoid overestimating the Q value; an Actor network outputs the switching actions of each inverter. i The Critic network updates its parameters by minimizing the loss function, aiming to make the Q-value approach the target value. The Actor network, on the other hand, updates its policy parameters by maximizing the Q-value evaluated by the Critic network, using the policy gradient method, as shown in the following formula:

[0035]

[0036] In the formula: L θ Let θ represent the Q-network loss in terms of parameter θ; θ is the Critic network parameter; E is the expected return; E{i} represents the expectation; r t Let be the reward value at time t; γ be the discount factor, typically 0.95–0.99; Q θ For the value function; s t+1 Indicates the state at the next moment; Indicates by parameters The deterministic strategy is represented in state s. t+1 The next output is an action; Let Q be the target Q value at time t; θ (s t ,a t ) represents the estimated Q value at time t; s t Indicates the state at time t; a t This represents the action actually performed in the environment at time t; Representation Strategy The gradient; Represents the policy objective function; For p π Expected return under the state distribution. Let represent the gradient of action a; π(s) represent the value of policy π in state s; For state s t The actor's actions; These are the parameters of the Actor network;

[0037] Step 4.2: Three complementary technical means were introduced to solve the problems of dimensionality curse and suboptimal solutions, thereby improving the algorithm training effect.

[0038] ①. Cutting double-Q learning, using two sets of Critic networks When calculating the target Q-value, the smaller of the two values ​​is taken. This mechanism can significantly alleviate the problem of Q-value overestimation, improve the accuracy of policy evaluation and the convergence stability of training. The updated formula is as follows:

[0039]

[0040] In the formula: y t The objective function; r(s) t ,a t ) represents state s t and action a t The following rewards; θ n ′ represents the target Critic network parameters; The objective value function.

[0041] ②. Target network and delay strategy update: A delayed update strategy is adopted. The Critic network is updated at each training step, the Actor network is updated every d iterations, and the target network μ is maintained simultaneously. φ″ and And smooth the target value using soft updates:

[0042] θ n ′←τθ n +(1-τ)θ n ′(8);

[0043] In the formula: τ represents the update factor; θ n This represents the critic parameter.

[0044] ③. Target policy smoothing regularization: By adding clipped Gaussian noise to the target action, the policy output action is perturbed within a certain range, thereby smoothing the target Q-value. A Gaussian-distributed noise term is added to the action at the next time step, and the perturbation result is clipped with upper and lower limits to restrict its perturbation amplitude to a reasonable range. The formula is shown below:

[0045]

[0046] In the formula: ε represents randomly added noise;

[0047] ε~clip[N(0,σ)-c,c] (10);

[0048] In the formula: clip[N(0,σ)-c,c] indicates that the noise follows a truncated normal distribution; N(0,σ) indicates sampling from a normal distribution with a mean of 0 and a standard deviation of σ; (-c,c) is the truncation interval.

[0049] Step 4.3: Design a reward function, using the inverse of the absolute value of the voltage deviation and the magnitude of the network loss as the reward function;

[0050] r n =-(α△V) t +βP loss,t (11);

[0051] △V t =|V t -V ref | (12);

[0052] Where: r n The reward function for each agent; ΔV is the voltage deviation at the grid connection point at step t; P loss,t V represents the active power loss of the line at that moment; α and β are the weighting coefficients for voltage deviation and active power loss, respectively; t Indicates the voltage at the grid connection point of the photovoltaic and energy storage power station; V ref The indicated reference value for the grid connection point voltage;

[0053] Step 4.4: Employ a centralized training-distributed execution structure. During the training phase, all agents can share the global observation state S to improve policy learning effectiveness. During the execution phase, each agent independently decides on control mode switching based on its own local state (e.g., local voltage, current, power). Based on the current state S... t The Actor network outputs switching commands for multiple inverters;

[0054] at =μ φ (s t (13);

[0055] Where: a t For time t, the joint action of all agents is the control mode switching decision; μ φ For an Actor network, based on the current state s t Output action; s t The combined state at time t includes grid connection point voltage, active power, OSCR value, etc.

[0056] Calculate the expected return of the current strategy using the Critic network:

[0057]

[0058] Step 4.5: Iteratively train and update network parameters:

[0059]

[0060] In the formula, L(θ) i ) for Critic network The loss function; θ i is the parameter of the i-th Critic network; B is the size of the mini-batch; D is the experience replay pool, which stores previously collected (s,a,r,s″) transfer samples; y is the current Critic's estimated Q-value for the state-action relationship; γ is the target Q-value; and γ is the discount factor used to balance current and future returns. Use the smaller of the two Critic targets to reduce the overestimation of the Q value; This indicates taking an action in state s′. At time j′, the Q-value given by the Q-network; θ j "" represents the parameters of the target Critic network.

[0061] In step 4.5, the optimal switching combination of the multi-inverter control modes is determined: when the operating short-circuit ratio exceeds the threshold, the MATD3 algorithm simulates multiple switching combinations (such as switching 1 device, 2 devices or even all devices from GFL to GFM), and selects a combination that achieves the goal of minimizing the voltage deviation at the grid connection point of the photovoltaic-storage power station and minimizing network loss for switching.

[0062] Step 7 also includes: Simulation verification of the adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio, and comparison with the traditional full GFL and full GFM modes in summer and winter to determine its effectiveness. Specifically:

[0063] Step 7.1: Based on the photovoltaic and energy storage power station grid connection model built in Step 1, set up 6 inverters (3 photovoltaic + 3 energy storage) to have both GFL and GFM modes in the Python platform; configure the grid equivalent impedance to make the short-circuit ratio 1.5, and import the actual photovoltaic and load data of Tibet in summer and winter.

[0064] Step 7.2: Set the key parameters of the MATD3 algorithm: batch size is set to 64, learning rate of both Actor and Critic networks is 0.0001, experience pool capacity is 10000, and number of interaction steps is 96.

[0065] Step 7.3: Set up three operating scenarios in the simulation environment: 1. Full GFL mode: All inverters always maintain grid-following control to achieve maximum active power point tracking; 2. Full GFM mode: All inverters always maintain grid-following control to provide maximum voltage support; 3. Adaptive switching based on OSCR+MATD3. Collect grid connection point voltage curves, active power curves, and line network loss data respectively, and evaluate the voltage fluctuation amplitude, network loss, and active power output of the three modes under different typical days and different time periods in summer and winter.

[0066] Step 7.4: Set the MATD3 algorithm training rounds to 1000. In each training round, the system uses 24 consecutive hours of training data for complete training, covering the entire daily cycle of photovoltaic power output and load fluctuations to enhance the model's adaptability to varying operating conditions throughout the day. Compare the voltage deviation, network loss, and active power output indices obtained from the MATD3 adaptive switching strategy with the simulation results of full GFL and full GFM modes. Plot the voltage deviation change curves, network loss curves, and active power output curves for the three strategies to verify the superior voltage support effect, smaller voltage fluctuations, and lower system network losses of the proposed method under typical summer and winter conditions.

[0067] A multi-inverter control mode adaptive switching control system based on operating short-circuit ratio includes:

[0068] Data acquisition module: used to collect real-time operating data of the photovoltaic and energy storage power station grid connection point and save it as input dataset. The observed measurements include grid connection point voltage, active power P, reactive power Q and grid equivalent impedance Z.

[0069] OSCR calculation and threshold setting module: used to calculate the running short-circuit ratio based on the collected P and Z values, and set the lower threshold C1 and upper threshold C2 as the basis for mode switching;

[0070] Intelligent agent modeling module: Each photovoltaic and energy storage inverter in the power station is abstracted as an independent intelligent agent, and the state space is constructed including grid connection point voltage U, active power P, OSCR value, current control mode and other information, and the action space is the control mode switching command;

[0071] The strategy training module is used for strategy optimization training based on the MATD3 algorithm. It introduces a dual Critic network to prevent Q-value overestimation and adopts a delayed update and target strategy smoothing regularization mechanism. The training objective is to minimize the weighted sum of grid connection point voltage deviation and line network loss.

[0072] Switching Decision and Execution Module: When the operating short-circuit ratio is less than C1 or greater than C2, the trained strategy network is triggered to output control mode switching instructions for each inverter, enabling some inverters to switch from GFL to GFM to provide voltage support. When the operating short-circuit ratio changes from outside the threshold to between (C1, C2), the trained strategy network is triggered to output control mode switching instructions for each inverter, enabling some inverters to switch from GFM to GFL to ensure active power output.

[0073] This invention discloses an adaptive switching method for control modes of multiple inverters based on the operating short-circuit ratio, with the following technical advantages:

[0074] Real-time operational data, including active power, reactive power, and voltage, is collected from the grid-connected points of the photovoltaic-storage power plant for subsequent control mode optimization. A grid-connected simulation model of the photovoltaic-storage power plant is built, and the OSCR is calculated based on the total active power at the grid connection point and the equivalent impedance of the grid. Upper and lower thresholds are set as switching criteria. Each inverter in the power plant is treated as an independent intelligent agent, and a state space containing voltage, power, and OSCR is constructed. A reward function is designed with the weighted sum of voltage deviation and line network loss as the objective. The MATD3 algorithm is used in combination with a centralized training-distributed execution architecture to intelligently decide the optimal inverter control mode switching combination, which can achieve both voltage support and reduce network losses. This effectively avoids voltage overcompensation and system oscillation caused by fixed thresholds or overall switching, and improves the system's adaptive capability and economy. By introducing multi-agent collaborative control and deep reinforcement learning optimization, the global optimization capability and response speed of the switching strategy are significantly improved, enhancing the voltage stability and operating efficiency of the photovoltaic-storage power plant under weak grid conditions. Attached Figure Description

[0075] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0076] Figure 1 This is a flowchart for adaptive switching of control modes for multiple inverters.

[0077] Figure 2(a) shows the grid connection point voltage response curve under GFL control;

[0078] Figure 2(b) shows the grid connection point voltage response curve under GFM control.

[0079] Figure 3 The flowchart for the MATD3 algorithm is shown below.

[0080] Figure 4 The diagram shows the convergence process of MATD3 training.

[0081] Figure 5(a) shows the photovoltaic output and load curves of a typical summer daytime system.

[0082] Figure 5(b) shows the photovoltaic output and load curves of a typical daytime solar system in winter.

[0083] Figure 6(a) shows the voltage distribution throughout the day on a typical summer day under different scenarios.

[0084] Figure 6(b) shows the voltage distribution throughout the day on a typical winter day under different scenarios.

[0085] Figure 7(a) is a schematic diagram of the typical daytime OSCR and control mode switching in summer.

[0086] Figure 7(b) is a schematic diagram of the typical daytime OSCR and control mode switching in winter.

[0087] Figure 8 This is a diagram showing the total network loss of a system under different control strategies during a typical summer day. Detailed Implementation

[0088] This embodiment proposes a method for adaptive switching of control modes for multiple inverters based on a grid-connected photovoltaic and energy storage power station model built on the Python platform and combined with the short-circuit ratio index. This method intelligently and in real time optimizes the combination of control modes for photovoltaic and energy storage inverters, thereby effectively balancing voltage support capability and system operating economy, and improving voltage stability under weak grid conditions.

[0089] like Figure 1 As shown, the adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio includes the following steps:

[0090] S1: Build a photovoltaic-storage grid-connected simulation model according to the typical values ​​of the preset control parameters, and set the grid impedance so that the short-circuit ratio at the grid connection point of the photovoltaic-storage power station is 1.5. Each photovoltaic and energy storage inverter control mode is set with two control modes: grid-following control (GFL) and grid-connected control (GFM) mode.

[0091] S2: Collects operational data from the grid-connected points of the photovoltaic and energy storage power plants. Measured parameters include active power P, reactive power Q, and grid-connected point voltage U. Based on the real-time collected P, the operating short-circuit ratio (OSCR) is calculated, and upper and lower thresholds C1 and C2 are set to form the triggering basis for control mode switching.

[0092] S3: When OSCR is less than C1 or greater than C2, it will trigger the multi-inverter control mode switching optimization requirement; at the same time, combined with parameters such as grid connection point voltage fluctuation and grid load, it will initialize the variables to be optimized, and the control mode flag bit of each inverter: GFL / GFM).

[0093] S4: Multi-agent dual-delay deep deterministic policy gradient algorithm, with a reward function designed as the weighted sum of grid connection point voltage deviation and line network loss; through a centralized training-distributed execution (CTDE) architecture, each agent learns the optimal control mode switching strategy based on local observations and partial global information, thereby achieving combined optimization of inverter control modes.

[0094] S5: The optimal switching combination output by the MATD3 algorithm is used as the execution plan for this control cycle and sent to the inverters in the field. Some inverters switch to grid-connected (GFM) mode according to the optimization results to improve voltage support and transient stability; the remaining inverters maintain grid-following (GFL) operation to achieve maximum active power output and grid loss control.

[0095] S6: When the OSCR is between (C1, C2), the multi-inverter control mode switching optimization requirement will be triggered again. The MATD3 algorithm will output the optimal switching combination scheme and send it to the inverters in the site. Some inverters will switch to GFL operation according to the optimization results, while the remaining inverters will remain in GFM operation.

[0096] S7: Simulation verification of the adaptive switching method of multi-inverter control mode based on the operating short-circuit ratio is carried out and compared with the traditional full GFL and full GFM modes in summer and winter to determine its effectiveness.

[0097] This embodiment collects real-time system operation data from the grid-connected points of the photovoltaic and energy storage power plant as input datasets for subsequent multi-inverter control mode optimization. A grid-connected simulation model of the photovoltaic and energy storage power plant is built in Python, and the short-circuit ratio is calculated based on the total active power at the grid connection point and the equivalent impedance of the grid. Upper and lower thresholds are set as trigger criteria. Each photovoltaic and energy storage inverter in the power plant is treated as an independent agent, and a state space containing state information such as grid connection point voltage, active power, and OSCR is constructed. The control mode (GFL or GFM) is used as the action space. Subsequently, a reward function with a weighted sum of voltage deviation and grid loss as the objective is designed, and the Multi-Agent Dual-Delay Deep Deterministic Policy Gradient (MATD3) algorithm is used for centralized training and distributed execution to achieve intelligent optimization of the control mode switching strategy. This dynamically determines which inverters switch to GFM mode to improve voltage support capability, while the remaining inverters maintain GFL mode. The FL mode ensures both active power output and economic efficiency. Simulation verification and comparison are conducted under typical summer and winter operating conditions. Voltage curves, network losses, and active power output data at the grid connection point are collected. The adaptive switching strategy based on OSCR+MATD3 is compared with the traditional full GFL and full GFM modes. Voltage deviation, network losses, and system economic indicators are analyzed to further verify the effectiveness and superiority of the proposed algorithm. This embodiment achieves global optimization of multi-inverter control mode combinations by introducing the MATD3 algorithm, effectively avoiding overcompensation and system oscillations caused by traditional empirical rules or fixed threshold switching, thus improving voltage stability and operating efficiency. Simultaneously, a centralized training-distributed execution structure is adopted to balance global information and real-time response, enhancing adaptability and engineering practical value in weak grid scenarios. Simulation comparisons demonstrate that the proposed method outperforms traditional methods in voltage support and network loss optimization, showing strong potential for engineering application.

[0098] In the preferred embodiment, in step S1, the GFL and GFM control equations of the photovoltaic-storage grid-connected inverter are expressed as follows:

[0099]

[0100] Among them, i d i q The inverter output current i abc dq axis components; e d e q For grid voltage e abc dq axis components; u d u q For the bridge arm voltage u abc The dq-axis component; ω is the grid angular frequency; L is the inductance. Where: f * U *Here, f and U are the frequency and voltage values, respectively; f and U are the frequency and voltage ratings for droop control; P and Q are the active and reactive power ratings, respectively; m and n are the frequency and voltage droop coefficients, respectively; P * Q * This provides real-time detection values ​​for the active and reactive power outputs of the inverter.

[0101] Figure 2(a) shows the grid connection point voltage response curve under GFL control mode. It can be observed that with the sharp drop in photovoltaic power, the voltage drops under all SCR conditions, and the lower the SCR (i.e., the weaker the grid), the greater the voltage drop. Especially under the extremely weak grid condition of SCR=1.5, the voltage drop is close to 30%, posing a significant risk of voltage drop. This is because in GFL mode, the inverter only operates as a current source and cannot actively regulate the voltage. When the injected power decreases, the system lacks voltage support capability, and the voltage is prone to large fluctuations or even collapse.

[0102] Figure 2(b) shows the voltage response under GFM control mode. Under the same disturbance conditions, GFM control significantly suppresses the voltage drop at the grid connection point, resulting in a smoother overall voltage curve and a significantly reduced fluctuation amplitude. Especially at SCR=1.5, GFM control can still maintain the voltage within a relatively stable range, indicating that this control strategy has strong adaptability to weak grid conditions. This is because the GFM inverter possesses voltage source characteristics and can actively support the system voltage through mechanisms such as virtual synchronous machines or droop control, providing additional regulation margin when output changes.

[0103] Although GFM control has a certain voltage regulation capability, its control structure is complex, its dynamic response is slow, and it poses stability risks in strong power grids. Therefore, both have their applicable scenarios in engineering applications. In environments with a high proportion of renewable energy grid integration, a more meaningful approach is to flexibly switch between GFL and GFM based on the system operating status, so as to ensure voltage stability while also considering control costs.

[0104] In this embodiment, data acquisition and simulation modeling, namely steps S1 and S2, provide crucial foundational data support for the subsequent multi-inverter control mode optimization process. The collected grid-connected operation data of the photovoltaic-storage power station (including voltage, active power, reactive power, etc.) and the grid-connected simulation model built on the Python platform are prerequisites for optimizing and verifying the intelligent switching strategy of control modes.

[0105] In the preferred embodiment, the calculation formulas for the photovoltaic-storage grid connection point voltage and the operating short-circuit ratio in step S2 include:

[0106] S201: The formula for calculating the voltage at the photovoltaic-storage grid connection point is:

[0107]

[0108] In the formula: Indicates the grid connection point voltage; R eq Indicates the equivalent resistance on the grid side at the grid connection point; X eq P represents the equivalent reactance on the grid side at the grid connection point. PCC Consider the injected active power for the photovoltaic system; Q PCC Reactive power; This is the voltage at the end of the line.

[0109] S202: The formula for calculating OSCR is:

[0110]

[0111] In the formula: S ac For the short-circuit capacity of the AC system; S N Rated capacity of new energy equipment; P * Z represents the per-unit value of the actual output power of new energy equipment. * This represents the per-unit value of the equivalent impedance of the power grid.

[0112] In the preferred scheme, in step S3, the two thresholds C1 and C2 of OSCR are obtained through trial and error. By trying different thresholds, it is observed when the voltage deviation at the grid connection point reaches the minimum value.

[0113] In the preferred embodiment, step S4 specifically includes:

[0114] S401: Initialize the MATD3 algorithm network structure, including two Critic networks and one Actor network. The two Critic networks are fitted with the Q-value function respectively. To avoid overestimating the Q value; an Actor network outputs the switching actions of each inverter. i The Critic network updates its parameters by minimizing the loss function, aiming to make the Q-value approach the target value. The Actor network, on the other hand, updates its policy parameters by maximizing the Q-value evaluated by the Critic network, using the policy gradient method, as shown in the following formula:

[0115]

[0116] In the formula: L θ Let θ represent the Q-network loss in terms of parameter θ; θ is the Critic network parameter; E is the expected return; E{i} represents the expectation; r t Let be the reward value at time t; γ be the discount factor, typically 0.95–0.99; Q θ For the value function; s t+1 Indicates the state at the next moment; Indicates by parameters The deterministic strategy is represented in state s. t+1The next output is an action; Let Q be the target Q value at time t; θ (s t ,a t ) represents the estimated Q value at time t; s t Indicates the state at time t; a t This represents the action actually performed in the environment at time t; Representation Strategy The gradient; Represents the policy objective function; For p π Expected return under the state distribution. Let represent the gradient of action a; π(s) represent the value of policy π in state s; For state s t The actor's actions; These are the parameters of the Actor network;

[0117] S402: Introduces three complementary techniques to solve the problems of dimensionality curse and suboptimal solutions, improving the algorithm's training performance. 1. Pruning double Q-learning, employing two sets of Critic networks. When calculating the target Q-value, the smaller of the two values ​​is taken. This mechanism can significantly alleviate the problem of Q-value overestimation, improve the accuracy of policy evaluation and the convergence stability of training. The updated formula is as follows:

[0118]

[0119] In the formula: y t The objective function; r(s) t ,a t ) represents state s t and action a t The following rewards; θ n ′ represents the target Critic network parameters; The objective value function.

[0120] 2. Target network and delayed update strategy: A delayed update strategy is adopted. The Critic network is updated at every training step, the Actor network is updated every d iterations, and the target network μ is maintained simultaneously. φ″ and And smooth the target value using soft updates:

[0121] θ n ′←τθ n +(1-τ)θ n ′;

[0122] In the formula: τ represents the update factor; θ n This represents the critic parameter.

[0123] 3. Target policy smoothing regularization: By adding clipped Gaussian noise to the target action, the policy output action is perturbed within a certain range, thereby smoothing the target Q-value. A Gaussian-distributed noise term is added to the action at the next time step, and the perturbation result is pruned with upper and lower limits to restrict its perturbation amplitude to a reasonable range. The formula is shown below:

[0124]

[0125] ε~clip[N(0,σ)-c,c];

[0126] In the formula: clip[N(0,σ)-c,c] indicates that the noise follows a truncated normal distribution; N(0,σ) indicates sampling from a normal distribution with a mean of 0 and a standard deviation of σ; (-c,c) is the truncation interval.

[0127] S403: Design a reward function that uses the inverse of the absolute value of the voltage deviation and the magnitude of the network loss as the reward function:

[0128] r n =-(αΔV) t +βP loss,t );

[0129] ΔV t =|V t -V ref |;

[0130] Where: r n The reward function for each agent; ΔV is the voltage deviation at the grid connection point at step t; P loss,t V represents the active power loss of the line at that moment; α and β are the weighting coefficients for voltage deviation and active power loss, respectively; t Indicates the voltage at the grid connection point of the photovoltaic and energy storage power station; V ref The indicated reference value for the grid connection point voltage;

[0131] S404: Employs a centralized training-distributed execution structure. During the training phase, all agents can share the global observation state S to improve policy learning. During the execution phase, each agent independently decides on control mode switching based on its local state (e.g., local voltage, current, power). Based on the current state S... t The Actor network outputs switching commands for multiple inverters:

[0132] a t =μ φ (s t );

[0133] Where a t μ represents the joint action of all agents at time t, i.e., the control mode switching decision.φ For an Actor network, based on the current state s t Output action. s t The combined state at time t includes grid connection point voltage, active power, OSCR value, etc.

[0134] Calculate the expected return of the current strategy using the Critic network:

[0135]

[0136] S405: Iteratively train and update network parameters:

[0137]

[0138] In the formula, L(θ) i ) for Critic network The loss function; θ i is the parameter of the i-th Critic network; B is the size of the mini-batch; D is the experience replay pool, which stores previously collected (s,a,r,s″) transfer samples; y is the current Critic's estimated Q-value for the state-action relationship; γ is the target Q-value; and γ is the discount factor used to balance current and future returns. Use the smaller of the two Critic targets to reduce the overestimation of the Q value; This indicates taking an action in state s′. At time j′, the Q-value given by the Q-network; θ j "" represents the parameters of the target Critic network.

[0139] In the preferred scheme, step S7 specifically includes: performing simulation verification on the adaptive switching method of multi-inverter control mode based on the operating short-circuit ratio and comparing it with the traditional full GFL and full GFM modes in summer and winter to determine its effectiveness, specifically:

[0140] S701: A grid-connected model of a photovoltaic and energy storage power station built on the S1 platform in Python. It is set that all 6 inverters (3 photovoltaic + 3 energy storage) have both GFL and GFM modes. The grid equivalent impedance is configured to make the short-circuit ratio 1.5, and the actual photovoltaic and load data of Tibet in summer and winter are imported.

[0141] S702: Set the key parameters of the MATD3 algorithm: batch size is set to 64, learning rate of both Actor and Critic networks is 0.0001, experience pool capacity is 10000, and number of interaction steps is 96.

[0142] S703: Three operating scenarios are set in the simulation environment: 1. Full GFL mode: All inverters always maintain grid-following control to achieve maximum active power point tracking; 2. Full GFM mode: All inverters always maintain grid-following control to provide maximum voltage support; 3. Adaptive switching based on OSCR+MATD3. Voltage curves, active power curves, and line loss data at the grid connection point are collected respectively to evaluate the voltage fluctuation amplitude, network loss, and active power output of the three modes under different typical days and time periods in summer and winter.

[0143] S704: The MATD3 algorithm training rounds were set to 1000. In each training round, the system used 24 hours of continuous training data for complete training, covering the entire daily cycle of photovoltaic power output and load fluctuations, to enhance the model's adaptability to varied operating conditions throughout the day. The voltage deviation, network loss, and active power output indices obtained from the MATD3 adaptive switching strategy were compared with the simulation results of the full GFL and full GFM modes. The voltage deviation change curves, network loss curves, and active power output curves of the three strategies were plotted to verify that the proposed method has better voltage support, smaller voltage fluctuations, and lower system network losses under typical summer and winter conditions.

[0144] In this embodiment, the MATD3 training convergence graph is as follows: Figure 4 The solid line represents the reward value, and the shaded area indicates the error range. In the early stages of training, the agent explores the environment and accumulates experience through random actions. As iterations progress, its policy is gradually optimized, the accumulated reward continues to rise, and eventually converges to a certain level.

[0145] In this embodiment, the typical daily voltage extreme value curves under different scenarios are shown in Figure 5, and the average voltage deviation index is defined as follows:

[0146]

[0147] In the formula: ρ is the average voltage deviation index; N T This indicates the number of time steps. Typical daytime voltage regulation effects are shown in Table 1 below:

[0148] Table 1 Comparison of typical summer day control results under different scenarios

[0149]

[0150] Comparative scenarios reveal that in Scenario 1 (full grid connection mode), from 10:00 to 15:00, the grid connection voltage shows a significant upward trend with changes in photovoltaic output, peaking around 12:00. As photovoltaic output gradually decreases and user load increases, the voltage level begins to decline. From 17:00 to 24:00, due to the gradual disappearance of sunlight, photovoltaic output drops to zero, while user load peaks at night, resulting in undervoltage at the grid connection point from 18:00, reaching a voltage trough around 20:00. Subsequently, as user electricity consumption gradually decreases, the undervoltage problem is alleviated around 23:00, and the system voltage gradually returns to normal. In Scenario 2 (full grid connection mode), although voltage fluctuations approach the lower limit at certain times, the overall grid connection voltage remains within a safe operating range. In Scenario 3, after adopting the MATD3 adaptive control strategy, the voltage at each time period is effectively suppressed within a safe range, demonstrating superior voltage stability and control performance.

[0151] In this embodiment, Figures 7(a) and 7(b) show two time-series curves under scenario 3 on different typical days: The upper graph in Figures 7(a) and 7(b) shows the real-time change of the OSCR index at the grid connection point. The horizontal dashed lines mark the thresholds of 2.5 and 11, respectively. Considering that the photovoltaic output may approach zero in the evening and at night, which may easily cause the OSCR value to increase abnormally, this paper introduces an upper limit limit on the OSCR, limiting its maximum value to 20, in order to enhance the numerical stability and control feasibility of the index. When the OSCR falls within the range, the system maintains the maximum power point tracking mode; when the OSCR is lower or higher, the grid connection mode switching requirement is triggered. The lower graph corresponds to the control mode switching of each inverter. The green step line records the mode decision made by the MATD3 agent for the 6 inverters at each time step. That is, when the OSCR exceeds the limit, the agent automatically selects the optimal equipment switching combination (such as full switching, partial switching, etc.) according to the real-time strategy output, and determines the final switching action by maximizing the cumulative reward, so that the grid connection point voltage is always effectively constrained within the safe range.

[0152] Figures 7(a) and 7(b) show that, regardless of whether it is summer or winter, when the OSCR exceeds the threshold, the MATD3 algorithm can select the optimal device switching combination for switching. Further analysis on a typical summer day, as shown in Table 1, reveals that compared to the methods proposed in Scenario 1 and Scenario 2, the average voltage deviation of the MATD3-based dual-mode switching method in Scenario 3 is reduced by 56.34% and 14.94%, respectively, significantly improving voltage quality. The minimum voltages in Scenario 1 and Scenario 2 are 0.9349 and 0.9400, respectively, both exceeding the voltage threshold. Although the voltage in Scenario 2 is within the safe operating constraint range for most of the time, from the perspective of control effect, the MATD3-based dual-mode switching method proposed in this invention can more effectively suppress voltage fluctuations and better improve power quality compared to the methods used in Scenario 1 and Scenario 2.

[0153] In this embodiment, the total network loss of the system under different control strategies on a typical summer day is as follows: Figure 8 As shown. The results show that the proposed MATD3 adaptive control reduces network losses by 42.34% and 7.72% compared to full grid-following control and full grid-connection control, respectively. The proposed method has the lowest network loss, mainly due to the fact that the MATD3 algorithm can intelligently select the inverter switching control mode according to the system state, flexibly coordinate voltage regulation, significantly reduce the ineffective flow of reactive power in the system, and thus minimize energy loss.

Claims

1. A multi-inverter control mode adaptive switching method based on operating short-circuit ratio, characterized in that... Includes the following steps: Step 1: Build a grid-connected model for the photovoltaic and energy storage power station. Set the control mode of each photovoltaic and energy storage inverter to: grid-following control GFL and grid-connecting control GFM mode. Step 2: Collect the operation data of the photovoltaic and energy storage power station grid connection point, calculate the value of the operating short-circuit ratio (OSCR), and set the upper and lower thresholds C1 and C2; Step 3: When the operating short-circuit ratio OSCR is less than C1 or greater than C2, trigger the multi-inverter control mode switching optimization requirement; at the same time, initialize the variables to be optimized; Step 4: Use the multi-agent dual-delay deep deterministic strategy gradient MATD3 algorithm to determine the optimal switching combination of multi-inverter control modes; Step 5: Send the execution plan of the optimal switching combination for this control cycle output in Step 4 to the inverters in the field station; some inverters switch to the grid-connected control GFM mode according to the optimization results, while the remaining inverters keep running the grid-connected control GFL mode. Step 6: When the operating short-circuit ratio OSCR is between (C1, C2), the multi-inverter control mode switching optimization requirement is triggered again. The optimal switching combination scheme is output by the multi-agent dual-delay deep deterministic strategy gradient MATD3 algorithm and sent to the station inverters. Some inverters switch to grid-following control GFL operation according to the optimization results, while the other part of the inverters maintains grid-connected control GFM operation.

2. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 1, characterized in that: In step 1, a grid-connected model of the photovoltaic and energy storage power station is built according to the typical values ​​of the preset control parameters, and the grid impedance is set so that the short-circuit ratio of the grid connection point of the photovoltaic and energy storage power station is 1.

5. The control equations for the grid-following control (GFL) and grid-connected control (GFM) of the photovoltaic-storage grid-connected inverter are expressed as follows: In equation (1): i d i q The inverter output current i are respectively abc dq axis components; e d e q These are the grid voltages e abc dq axis components; u d u q The bridge arm voltages u are respectively abc dq-axis components; ω is the grid angular frequency; L is the inductance; In the formula: f * U * These represent the frequency and voltage values, respectively; f and U represent the frequency and voltage ratings for droop control, respectively; P and Q represent the active and reactive power ratings, respectively; m and n represent the frequency and voltage droop coefficients, respectively; P * Q * These are the real-time detection values ​​of the active and reactive power outputs of the inverter, respectively.

3. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 2, characterized in that: In step 2, the observed operating data includes active power P, reactive power Q, and grid connection point voltage U. The operating short-circuit ratio (OSCR) is calculated based on the real-time collected active power P, and upper and lower thresholds C1 and C2 are set to form the trigger basis for inverter control mode switching.

4. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 3, characterized in that: The formula for calculating the voltage at the photovoltaic-storage grid connection point is: In the formula: Indicates the grid connection point voltage; R eq Indicates the equivalent resistance on the grid side at the grid connection point; X eq Indicates the grid-side equivalent at the grid connection point Reactance; P PCC Consider the injected active power for the photovoltaic system; Q PCC Reactive power; This refers to the voltage at the end of the line. The formula for calculating the Operating Short-Circuit Ratio (OSCR) is as follows: In the formula: S ac For the short-circuit capacity of the AC system; S N Rated capacity of new energy equipment; P * Z represents the per-unit value of the actual output power of new energy equipment. * This represents the per-unit value of the equivalent impedance of the power grid.

5. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 4, characterized in that: In step 3, when the operating short-circuit ratio OSCR is less than C1 or greater than C2, the multi-inverter control mode switching optimization requirement will be triggered; at the same time, combined with parameters such as voltage fluctuation at the grid connection point of the photovoltaic-storage power station and grid load, the variables to be optimized are initialized, and the control mode flag bit of each inverter is: GFL / GFM.

6. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 5, characterized in that: In step 4, the multi-agent dual-delay deep deterministic policy gradient (MATD3) algorithm is designed with a reward function that is a weighted sum of grid connection point voltage deviation and line network loss. Through a centralized training-distributed execution (CTDE) architecture, each agent learns the optimal control mode switching strategy based on local observations and some global information, thereby achieving combined optimization of inverter control modes.

7. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 6, characterized in that: Step 4 includes: Step 4.1: Initialize the MATD3 algorithm network structure, including two Critic networks and one Actor network. The two Critic networks are fitted with the Q-value function respectively. To avoid overestimating the Q value; an Actor network outputs the switching actions of each inverter. i The Critic network updates its parameters by minimizing the loss function, aiming to make the Q-value approach the target value. The Actor network, on the other hand, updates its policy parameters by maximizing the Q-value evaluated by the Critic network, using the policy gradient method, as shown in the following formula: In the formula: L θ Let θ represent the Q-network loss in terms of parameter θ; θ is the Critic network parameter; E{i} represents the expectation; r t Let be the reward value at time t; γ be the discount factor; Q be the reward value at time t; γ be θ For the value function; s t+1 Indicates the state at the next moment; Indicates by parameters The deterministic strategy is represented in state s. t+1 The next output is an action; Let Q be the target Q value at time t; θ (s t ,a t ) represents the estimated Q value at time t; s t Indicates the state at time t; a t This represents the action actually performed in the environment at time t; Representation Strategy The gradient; Represents the policy objective function; For p π Expected return under state distribution; Let represent the gradient of action a; π(s) represent the value of policy π in state s; For state s t The actor's actions; Step 4.2: ①. Cutting double-Q learning, using two sets of Critic networks When calculating the target Q value, the smaller of the two values ​​is taken, and the updated formula is as follows: In the formula: y t The objective function; r(s) t ,a t ) represents state s t and action a t The following rewards; θ n ′ represents the target Critic network parameters; The objective value function; ②. Target network and delay strategy update: A delayed update strategy is adopted. The Critic network is updated at each training step, the Actor network is updated every d iterations, and the target network μ is maintained simultaneously. φ "and And smooth the target value using soft updates: θ′ n ←tth n +(1-τ)θ′ n (8); In the formula: τ represents the update factor; θ n Indicates the critical parameter; ③. Target policy smoothing regularization: By adding clipped Gaussian noise to the target action, the policy output action is perturbed within a certain range, thereby smoothing the target Q value; a Gaussian-distributed noise term is added to the action at the next time step, and the perturbation result is clipped with upper and lower limits, as shown in the following formula: In the formula: ε represents randomly added noise; ε~clip[N(0,σ)-c,c] (10); In the formula: clip[N(0,σ)-c,c] indicates that the noise follows a truncated normal distribution; N(0,σ) indicates sampling from a normal distribution with a mean of 0 and a standard deviation of σ; (-c,c) is the truncation interval; Step 4.3: Design a reward function, using the inverse of the absolute value of the voltage deviation and the magnitude of the network loss as the reward function; r n =-(α△V t +βP loss,t ) (11); △V t =|V t -V ref | (12); Where: r n The reward function for each agent; ΔV is the voltage deviation at the grid connection point at step t; P loss,t V represents the active power loss of the line at that moment; α and β are the weighting coefficients for voltage deviation and active power loss, respectively; t Indicates the voltage at the grid connection point of the photovoltaic and energy storage power station; V ref The indicated reference value for the grid connection point voltage; Step 4.4: Employ a centralized training-distributed execution structure. During the training phase, all agents can share the global observation state S to improve policy learning performance. During the execution phase, each agent independently decides on control mode switching based on its local state; according to the current state S... t The Actor network outputs switching commands for multiple inverters; a t =μ φ (s t ) (13); Where: a t For time t, the joint action of all agents is the control mode switching decision; μ φ For an Actor network, based on the current state s t Output action; s t The joint state at time t includes grid connection point voltage, active power, OSCR value, etc. Calculate the expected return of the current strategy using the Critic network: Step 4.5: Iteratively train and update network parameters: In the formula, L(θ) i ) for Critic network Loss function; θ i Here are the parameters of the i-th Critic network; B is the size of the mini-batch. This serves as an experience replay pool, storing previously collected (s,a,r,s″) transfer samples. y is the current Critic's estimated Q-value for the state-action relationship; γ is the target Q-value; and γ is the discount factor used to balance current and future returns. Use the smaller of the two Critic targets to reduce the overestimation of the Q value; This indicates taking an action in state s′. At time j′, the Q-value given by the Q-network; θ″ j The parameters are those of the target Critic network.

8. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 7, characterized in that: In step 4.5, the optimal switching combination of the multi-inverter control modes is determined: when the operating short-circuit ratio exceeds the threshold, the MATD3 algorithm simulates multiple switching combinations and selects a combination that achieves the goal of minimizing the voltage deviation at the grid connection point of the photovoltaic-storage power station and minimizing network loss for switching.

9. The adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio according to claim 8, characterized in that: Step 7 also includes: Simulation verification of the adaptive switching method for multi-inverter control modes based on the operating short-circuit ratio, and comparison with the traditional full GFL and full GFM modes in summer and winter to determine its effectiveness. Specifically: Step 7.1: Based on the photovoltaic and energy storage power station grid connection model built in Step 1 in the Python platform, set all 6 inverters (3 photovoltaic + 3 energy storage) to have both GFL and GFM modes; configure the grid equivalent impedance to make the short-circuit ratio 1.5, and import the actual photovoltaic and load data of Tibet in summer and winter. Step 7.2: Set the key parameters of the MATD3 algorithm: batch size is set to 64, learning rate of both Actor and Critic networks is 0.0001, experience pool capacity is 10000, and number of interaction steps is 96; Step 7.3: Set up three operating scenarios in the simulation environment:

1. Full GFL mode: All inverters always maintain grid-following control to achieve maximum active power tracking; 2. Full GFM mode: All inverters always maintain grid-type control, providing maximum voltage support; 3. Adaptive switching based on OSCR+MATD3; Collect grid connection point voltage curves, active power curves, and line network loss data respectively, and evaluate the voltage fluctuation amplitude, network loss, and active power output of the three modes under different typical days and different time periods in summer and winter. Step 7.4: Set the training rounds of the MATD3 algorithm to 1000. In each training round, the system uses 24 hours of continuous training data for complete training, covering the entire daily cycle of photovoltaic power output and load fluctuations, to enhance the model's adaptability to various operating conditions throughout the day. Compare the voltage deviation, network loss, and active power output indices obtained by the MATD3 adaptive switching strategy with the simulation results of the full GFL and full GFM modes. Plot the voltage deviation change curves, network loss curves, and active power output curves of the three strategies to verify that the proposed method has better voltage support effect, smaller voltage fluctuations, and lower system network losses under typical summer and winter operating conditions.

10. A multi-inverter control mode adaptive switching control system based on operating short-circuit ratio, characterized in that... The system includes: Data acquisition module: used to collect real-time operating data of the photovoltaic and energy storage power station grid connection point and save it as input dataset. The observed measurements include grid connection point voltage, active power P, reactive power Q and grid equivalent impedance Z. OSCR calculation and threshold setting module: used to calculate the running short-circuit ratio based on the collected P and Z values, and set the lower threshold C1 and upper threshold C2 as the basis for mode switching; Intelligent agent modeling module: Each photovoltaic and energy storage inverter in the power station is abstracted as an independent intelligent agent, and the state space is constructed including grid connection point voltage U, active power P, OSCR value, current control mode and other information, and the action space is the control mode switching command; The strategy training module is used for strategy optimization training based on the MATD3 algorithm. It introduces a dual Critic network to prevent Q-value overestimation and adopts a delayed update and target strategy smoothing regularization mechanism. The training objective is to minimize the weighted sum of grid connection point voltage deviation and line network loss. Switching Decision and Execution Module: When the operating short-circuit ratio is less than C1 or greater than C2, the trained strategy network is triggered to output the control mode switching command for each inverter, enabling some inverters to switch from GFL to GFM to provide voltage support; when the operating short-circuit ratio changes from outside the threshold to between (C1, C2), the trained strategy network is triggered to output the control mode switching command for each inverter, enabling some inverters to switch from GFM to GFL to ensure active power output.

Images