Grid-connected converter adaptive control method based on multi-expert fusion and decision optimization

Abstract

The invention discloses a grid-connected converter adaptive control method based on multi-expert fusion and decision optimization, and belongs to the field of electrical engineering. The invention provides a grid-connected converter self-adaptive control method based on multi-expert fusion and decision optimization aiming at the multi-dimensional state adaptation requirement of a complex power grid and the difficulty in adaptation of a fixed control model to complex and variable working conditions. According to the method, the multi-expert model and the decision optimization layer are established, the Transform time domain model is constructed, the adaptability of the converter is improved on the long time scale, and the collaborative optimization capability of the grid-connected converter in the complex heterogeneous power grid environment is achieved. The method has the capabilities of quickly adapting to known working conditions and adaptively optimizing unknown working conditions, so that the grid-connected converter continuously keeps the optimal control performance in a complex disturbance environment.

Description

Technical Field

[0001] This invention relates to the field of electrical engineering, and in particular to an adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization. Background Technology

[0002] The installation ratio of grid-connected converters in power systems is increasing, and their control characteristics have a profound impact on high-penetration renewable energy power generation systems, posing a huge challenge to the stable operation of grid-connected converters.

[0003] In weak grid environments, multi-machine parallel systems face random reconfigurations of the power grid structure, heterogeneous distribution of grid-connected converter characteristics, and transient changes in operating states, making it difficult for fixed control models to adapt to complex and changing operating conditions. Traditional control methods rely on prior knowledge and static models, which are insufficient to cover all potential operating condition combinations and lack the ability to adaptively adjust to unknown states, making it difficult to guarantee the control effect of the system in dynamic and complex environments. With the continuous increase in the penetration rate of new energy sources, this structural contradiction between the static model system and the dynamic and complex environment is becoming increasingly prominent. For example:

[0004] 1) The paper "Seamless Switching Method Between Grid-Following and Grid-Forming Control for Renewable Energy Conversion Systems" published in IEEE Transactions on Industry Applications, Volume 61, Issue 5 in 2025, proposed a dual-mode switching method combining a unified current inner loop design with pre-synchronous control. The unified current inner loop ensures that the current source mode and voltage source mode are consistent in control structure, reducing dynamic shocks during switching. However, while the method has verified the feasibility of efficient dual-mode switching in a single converter experiment, it does not consider the interactive coupling problem when multiple converters are running in parallel. Its applicability in multi-machine systems still needs further research.

[0005] 2) The article "Optimal Configuration Method of Switchable Units in New Energy Power Plants Based on Operational Short-Circuit Ratio (SCR)" published in the journal *Power System Technology*, Volume 48, Issue 3, 2024, proposed a configuration principle based on dual-mode constraints, deriving the minimum voltage source mode proportion, thus avoiding the over-switching problem caused by relying solely on the SCR index. However, the experimental verification of this study was based on a simplified equivalent model and did not fully consider the dispersion of multi-machine control parameters in actual new energy power plants and their interactive effects on the dominant mode. Therefore, its applicability in complex power grid environments still needs further verification.

[0006] Therefore, by leveraging deep reinforcement learning to drive the automatic generation and dynamic updating of a multi-model controller library, a control strategy with self-learning, self-adaptation, and iterative optimization capabilities is constructed. This enables the grid-connected converter to autonomously optimize its operating strategy according to changes in the power grid environment and achieve coordinated optimization of the system under complex grid conditions. For example:

[0007] 1) The paper "Fault Diagnosis of Photovoltaic Grid-Connected Converters Based on Incremental Learning" published in the Journal of Electric Power Systems and Automation, Vol. 37, No. 2 in 2025, proposes an incremental learning method based on multi-scale filtering. After segmenting the processed signal using a sliding window, a fault dataset is obtained. A one-dimensional convolutional neural network is used to learn historical data, and the type of historical fault is identified by the nearest mean classifier, thereby improving the fault identification capability.

[0008] 2) The paper "Data Driven Decentralized Control of Inverter Based Renewable EnergySources Using Safe Guaranteed Multi-Agent Deep Reinforcement Learning", published in IEEE Transactions on Sustainable Energy, Volume 15, Issue 2 in 2024, presents an intelligent converter control method based on multi-agent deep reinforcement learning. By introducing a safe policy projection mechanism, the action space of the DRL agent is effectively limited to the range allowed by physical constraints. This smoothly and effectively avoids any constraint violations in the operation of the distribution network during the training process, ensuring the safety of training and optimizing the operation of the distribution system.

[0009] In summary, the existing technology has the following problems:

[0010] (1) Existing literature on traditional control methods relies on prior knowledge and static models, which are difficult to cover all potential working condition combinations and lack the ability to adaptively adjust to unknown states, making it difficult to guarantee the control effect of the system in dynamic and complex environments.

[0011] (2) Existing literature on artificial intelligence research still mainly focuses on single-machine or local parameter optimization. Fixed model libraries based on empirical knowledge have fundamental limitations and lack the ability to adaptively construct unknown states. Summary of the Invention

[0012] To overcome the limitations of the above-mentioned technical solutions, this invention proposes an adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization. By establishing a multi-expert model and a decision optimization layer, a Transformer time-domain model is constructed to improve the adaptability of the converter over a long time scale, thus solving the problem of collaborative optimization capability of grid-connected converters in complex heterogeneous power grid environments.

[0013] The technical solution of the present invention is as follows:

[0014] An adaptive control method for a grid-connected converter based on multi-expert fusion and decision optimization is disclosed. The grid-connected converter is an LC-type grid-connected converter, comprising a DC source, an inverter, an LC-type filter, a line equivalent inductance, and a three-phase power grid connected in series. The specific steps are as follows:

[0015] Step 1: Collect real-time operating data of the grid-connected converter and calculate the three-phase grid voltage amplitude V. m Phase angle θ, equivalent short-circuit ratio (SCR), and total harmonic distortion (THD) of three-phase grid current form a state characteristic sequence;

[0016] Step 2: Set up the primary control layer, including establishing a multi-expert fusion model composed of current source power tracking experts and voltage source network experts, inputting the state feature sequence into the softened normalized weight allocation network, dynamically allocating the weights of the two types of experts and weighting and summing their control outputs to obtain the first control quantity u1(t).

[0017] Step 3: Set up the secondary control layer for event triggering, including virtual impedance expert, harmonic damping expert and virtual inertia expert. When the phase margin, specific subharmonic content or frequency deviation exceeds the preset threshold, the expert is activated as needed to generate the compensation control quantity u2(t).

[0018] Step 4: Set up a decision optimization layer, including using a Transformer-based time-domain model, utilizing a multi-head attention mechanism and a softened normalized weight allocation network to jointly encode the state feature sequence and the first control variable u1(t), complete the online identification of unknown working conditions and output targeted strategy weights u3(t).

[0019] Step 5: Introduce a voting confidence weighted strategy to dynamically fuse the first control variable u1(t), the compensation control variable u2(t), and the targeted strategy weight u3(t) to achieve adaptive switching of the dominant expert and optimal impedance control under different operating conditions.

[0020] Preferably, the implementation process of step 2 is as follows:

[0021] Step 2.1: Establish a multi-expert fusion model;

[0022] The current-source power point tracking expert is based on an improved power synchronization control principle, wherein the output reference current i dq Represented as:

[0023] i dq =G PLL (s)(ΔP+H damper (s)Δθ)

[0024] In the formula, G PLL (s) is the PLL transfer function, ΔP is the power during the PLL dynamic process, and Δθ is the phase deviation during the PLL dynamic process; H damper (s) represents the damping compensation element;

[0025] The voltage source type grid construction expert uses a virtual synchronous machine model to enhance the grid support capability by simulating the electromechanical transient characteristics of synchronous generators;

[0026] Step 2.2, obtaining the first control variable u1(t);

[0027] The softened normalized weight allocation network receives the state feature sequence obtained in step 1 as input x, x = [V m [θ, SCR, THD], and activate the current source power tracking expert and the voltage source network construction expert, outputting the multi-expert fusion weight g(x), g(x) = [g c (x), g v (x)],g c (x) represents the expert weights for current-source power point tracking, g v (x) represents the expert weights for voltage source network construction. Each expert calculates the input x and generates a current source output y. c and voltage source type output y v ;

[0028] Let g v (x) satisfies the relation g c (x)+g v (x) = 1, according to the current source type output y c and voltage source type output y v The first control variable u1(t) is calculated, u1(t) = y c ×g c (x)+y v ×g v (x).

[0029] Preferably, the specific method for activating experts on demand in step 3 is as follows:

[0030] When the phase margin is detected to be below the threshold, the virtual impedance expert is activated, and a virtual impedance compensation term Z is added. vir (s), whose expression is:

[0031]

[0032] In the formula, R is the resistance value of the bridge arm side resistor in the LC filter; L is the inductance value of the bridge arm side in the LC filter; and C is the capacitance value of the filter capacitor in the LC filter.

[0033] When a specific harmonic exceeds the limit, the harmonic damping expert is activated, and a harmonic compensation term u is added as follows:

[0034]

[0035] In the formula, ω n ζ is the angular frequency of the nth harmonic; n Let i be the damping ratio of the nth harmonic; n The nth harmonic current component; A n is the control gain for the nth harmonic; s is the Laplace operator; n = 2 indicates starting from the 2nd harmonic; N is the Nth harmonic;

[0036] When a rapid frequency disturbance or excessive frequency amplitude deviation is detected, the virtual inertial expert is activated, and an inertial compensation term P is added. inertia Its expression is:

[0037]

[0038] In the formula, M v ω represents the virtual inertial gain, and ω represents the angular velocity.

[0039] The expression for the compensation control quantity u2(t) is: u2(t) = [Z vir (s), u, P inertia ].

[0040] Preferably, the implementation process of step 4 is as follows:

[0041] Step 4.1: Construct a multi-scale prediction model based on time-series Transformer, denoted as prediction model Y; use the state feature sequence obtained in step 1 and the first control quantity u1(t) of the grid-connected converter obtained in step 2 as the input to prediction model Y, denoted as input Q:

[0042] Step 4.2: Utilize a multi-head attention mechanism to extract key dynamic patterns and correlations of the input Q from multiple subspaces and time scales, outputting the multidimensional temporal representation MHA(Q, K, V) of the current state, whose expression is as follows:

[0043] MHA(Q,K,V)=Concat(head1,…head i …,head hW O

[0044] In the formula, i is the index of the attention head, i = 1, 2, ..., h; h is the total number of attention heads. i For the i-th attention head, single-head attention is performed on the subspace; K represents the historical features matching the input Q, K = QW. i K W i K Let K be the linear projection matrix; V be the actual extracted value, V = QW i V W i V W is the linear projection matrix of V; O To output the projection matrix;

[0045] The i-th attention head i The expression is:

[0046]

[0047] In the formula, W i Q Let Q be the linear projection matrix of the input Q; This is a scaling factor, used to prevent gradient explosion caused by large orders of magnitude (KW). i K ) T For (KW) i K The transpose of )

[0048] Step 4.3: Based on the multidimensional time-series expression MHA(Q, K, V) in Step 4.2, call the current source power tracking expert and voltage source network construction expert in Step 2 to construct the current source intelligent agent and the voltage source intelligent agent. Each intelligent agent evaluates the strategy according to its own local objective function to complete the online identification of unknown operating conditions.

[0049] The local objective function J of a current-source type intelligent agent current for:

[0050]

[0051] In the formula, P ref (t) represents the given reference active power; P(t) represents the actual output power; Q ref Q(t) is the given reference reactive power; Q(t) is the actual output reactive power; Δi(t) is the rate of change of current command; t is any moment during operation; T1 is the upper limit of the time range; α1 is the active power error voting weight factor; α2 is the reactive power error voting weight factor; λ1 is the smoothing voting weight factor;

[0052] The local objective function J of a voltage source type intelligent agent voltage for:

[0053]

[0054] In the formula, To provide the desired bus voltage; v bus (t) represents the actual measured bus voltage; f ref f(t) is the given reference frequency; f(t) is the actual frequency measurement value; Δu(t) is the rate of change of the controller output voltage; β1 is the voltage stability voting weight factor; β2 is the frequency stability voting weight factor; λ2 is the smoothing voting weight factor;

[0055] Each voting weight factor dynamically reflects the adaptability of each strategy. Based on the final decision and the fusion of multi-party evaluation results, the time-domain model Y outputs the targeted strategy weight u3(t).

[0056] Preferably, step 5 is implemented as follows:

[0057] Step 5.1: Drive the targeted policy weights u3(t) output by the time-domain model Y to obtain the final weighting coefficients [γ1(t), γ2(t), γ3(t)]. The specific process is as follows:

[0058] The time-domain model Y outputs a condition classification vector p(t), expressed as:

[0059] p(t) = [p1(t), p2(t), ... p a (t)…,p H (t)],[ a (t)∈[0,1]

[0060] In the formula, a is the sequence number of the working condition, H is the total number of working conditions, and a = 1, 2, ..., H; p a (t) represents the confidence weight for the a-th working condition;

[0061] Introduce a predefined control strategy weight Γ, as follows:

[0062]

[0063] In the formula, m is the expert's number, m = 1, 2, 3; where 1 is the virtual impedance expert, 2 is the harmonic damping expert, and 3 is the virtual inertia expert. Corresponding to the a-th working condition and the m-th expert;

[0064] The final weighting coefficients are expressed as follows:

[0065]

[0066] In the formula, γ1(t) is the confidence weight of the primary control layer; γ2(t) is the confidence weight of the secondary control layer; and γ3(t) is the confidence weight of the decision control layer.

[0067] Step 5.2: Dynamically fuse the first control variable u1(t), the compensation control variable u2(t), and the targeted strategy weight u3(t) to obtain the dynamic fusion output u. final Its expression is:

[0068] u final (t)=γ1(t)×u1(t)+γ2(t)×u2(t)+γ3(t)×u3(t)

[0069] Step 5.3, dynamically fused output u final As the final control output, it is connected to the controller to issue real-time control commands to the PWM module, thereby completing adaptive switching and optimal impedance control.

[0070] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0071] 1. This invention not only constructs a multi-expert fusion control architecture to ensure stability on a short time scale, but also utilizes deep reinforcement learning to optimize control strategies to improve adaptability on a long time scale;

[0072] 2. The control strategy of the present invention has the ability to quickly adapt to known operating conditions and adaptively optimize unknown operating conditions, enabling the grid-connected converter to maintain optimal control performance in complex disturbance environments.

[0073] 3. This invention can be updated and upgraded by adding expert models and improving the decision optimization layer algorithm to cope with the increasingly complex multi-disturbance coupling problem of power grids; Attached Figure Description

[0074] Figure 1 This is a topology diagram of the grid-connected converter described in this invention.

[0075] Figure 2 This is a control architecture diagram of the method of the present invention.

[0076] Figure 3 This is a flowchart illustrating the implementation steps of the present invention.

[0077] Figure 4 The Bode plot of the inverter output impedance after the adaptive control of this invention is completed with the addition of compensation. Detailed Implementation

[0078] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0079] Figure 1This is a topology diagram of the grid-connected converter described in this invention. Figure 1 As can be seen, the grid-connected converter is an LC-type grid-connected converter, which includes a DC source, an inverter, an LC-type filter, a line equivalent inductance, and a three-phase power grid connected in series. The LC-type filter includes a bridge arm-side inductor, a bridge arm-side inductor, and a filter capacitor.

[0080] Figure 1 In the middle, L g For the equivalent inductance of the line, C dc It is a DC source.

[0081] In this embodiment, L = 2mH, R = 0.1Ω, and C = 30μF.

[0082] Figure 2 This is a simplified control diagram of the optimization method of the present invention, by Figure 2 As can be seen, this method architecture incorporates various classic expert models that can meet different power grid conditions and performance requirements. Different experts are suitable for different control tasks, each with its own advantages and disadvantages. First, a primary control layer is constructed, consisting of current-source tracking experts and voltage-source network construction experts. Then, the current-source and voltage-source expert branches can be further divided into two main categories: structural optimization and parameter optimization. To address specific harmonics and frequency fluctuations in the power grid, a secondary control layer is constructed based on "event triggering," consisting of experts in virtual impedance, harmonic damping, and virtual inertia.

[0083] Figure 3 This is a flowchart illustrating the implementation steps of the present invention. Figure 3 As can be seen, this invention provides an adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization. This adaptive control method constructs a state-aware system, establishes a multi-expert fusion model composed of current-source power tracking experts and voltage-source grid-building experts, sets up an event-triggered secondary control layer, employs a Transformer-based time-domain model for decision optimization, and dynamically fuses the outputs of each layer according to a voting confidence weighting strategy to achieve adaptive switching of the dominant expert and optimal impedance control under different operating conditions. The specific steps are as follows:

[0084] Step 1: Collect real-time operating data of the grid-connected converter and calculate the three-phase grid voltage amplitude V. m Phase angle θ, equivalent short-circuit ratio (SCR), and total harmonic distortion (THD) of three-phase grid current form a state characteristic sequence.

[0085] In this embodiment, a state-sensing system is used to collect the above data in real time. The state-sensing system includes obtaining the three-phase power grid voltage amplitude of 311V, phase angle of 18°, equivalent short-circuit ratio of 2.1, and total harmonic distortion rate of three-phase power grid current of 5%.

[0086] Step 2: Set up the primary control layer, including establishing a multi-expert fusion model composed of current source power tracking experts and voltage source network experts. Input the state feature sequence into the softened normalized weight allocation network, dynamically allocate the weights of the two types of experts, and perform weighted summation of their control outputs to obtain the first control quantity u1(t).

[0087] In this embodiment, the implementation process of step 2 is as follows:

[0088] Step 2.1: Establish a multi-expert fusion model;

[0089] The current-source power point tracking expert is based on an improved power synchronization control principle, wherein the output reference current i dq Represented as:

[0090] i dq =G PLL (s)(ΔP+H damper (s)Δθ)

[0091] In the formula, G PLL (s) is the PLL transfer function, ΔP is the power during the PLL dynamic process, and Δθ is the phase deviation during the PLL dynamic process; H damper (s) represents the damping compensation element;

[0092] The voltage source type grid construction expert uses a virtual synchronous machine model to enhance the grid support capability by simulating the electromechanical transient characteristics of synchronous generators;

[0093] Step 2.2, obtaining the first control variable u1(t);

[0094] The softened normalized weight allocation network receives the state feature sequence obtained in step 1 as input x, x = [V m [θ, SCR, THD], and activate the current source power tracking expert and the voltage source network construction expert, outputting the multi-expert fusion weight g(x), g(x) = [g c (x), g v (x)],g c (x) represents the expert weights for current-source power point tracking, and gv(x) represents the expert weights for voltage-source network construction. Each expert calculates the input x and generates the current-source output y. c and voltage source type output y v ;

[0095] Let g v (x) satisfies the relation g c (x)+g v (x) = 1, according to the current source type output y c and voltage source type output y vThe first control variable u1(t) is calculated, u1(t) = y c ×g c (x)+y v ×g v (x).

[0096] Step 3: Set up the secondary control layer for event triggering, including virtual impedance expert, harmonic damping expert and virtual inertia expert. When the phase margin, specific subharmonic content or frequency deviation exceeds the preset threshold, the expert is activated as needed to generate the compensation control quantity u2(t).

[0097] In this embodiment, the specific method for activating experts on demand in step 3 is as follows:

[0098] When the phase margin is detected to be below the threshold, the virtual impedance expert is activated, and a virtual impedance compensation term Z is added. vir (s), whose expression is:

[0099]

[0100] In the formula, R is the resistance value of the bridge arm side resistor in the LC filter; L is the inductance value of the bridge arm side in the LC filter; and C is the capacitance value of the filter capacitor in the LC filter.

[0101] When a specific harmonic exceeds the limit, the harmonic damping expert is activated, and a harmonic compensation term u is added as follows:

[0102]

[0103] In the formula, ω n ζ is the angular frequency of the nth harmonic; n Let i be the damping ratio of the nth harmonic; n The nth harmonic current component; A n is the control gain for the nth harmonic; s is the Laplace operator; n = 2 indicates starting from the 2nd harmonic; N is the Nth harmonic;

[0104] When a rapid frequency disturbance or excessive frequency amplitude deviation is detected, the virtual inertial expert is activated, and an inertial compensation term P is added. inertia Its expression is:

[0105]

[0106] In the formula, M v ω represents the virtual inertial gain, and ω represents the angular velocity.

[0107] The expression for the compensation control quantity u2(t) is: u2(t) = [Z vir (s), u, P inertia ].

[0108] Step 4: Set up a decision optimization layer, adopt a Transformer-based time-domain model, and use a multi-head attention mechanism and a softened normalized weight allocation network to jointly encode the state feature sequence and the first control quantity u1(t) to complete the online identification of unknown working conditions and output targeted strategy weights u3(t).

[0109] In this embodiment, the implementation process of step 4 is as follows:

[0110] Step 4.1: Construct a multi-scale prediction model based on time-series Transformer, denoted as prediction model Y; use the state feature sequence obtained in step 1 and the first control quantity u1(t) of the grid-connected converter obtained in step 2 as the input to prediction model Y, denoted as input Q:

[0111] Step 4.2: Utilize a multi-head attention mechanism to extract key dynamic patterns and correlations of the input Q from multiple subspaces and time scales, outputting the multidimensional temporal representation MHA(Q, K, V) of the current state, whose expression is as follows:

[0112] MHA(Q,K,V)=Concat(head1,…head i …,head h W O

[0113] In the formula, i is the index of the attention head, i = 1, 2, ..., h; h is the total number of attention heads. i For the i-th attention head, single-head attention is performed on the subspace; K represents the historical features matching the input Q, K = QW. i K W i K Let K be the linear projection matrix; V be the actual extracted value, V = QW i V W i V W is the linear projection matrix of V; O To output the projection matrix;

[0114] The i-th attention head i The expression is:

[0115]

[0116] In the formula, W i Q Let Q be the linear projection matrix of the input Q; The scaling factor is used to prevent gradient bursts caused by large magnitudes (KW). i K )T For (KW) i K The transpose of )

[0117] Step 4.3: Based on the multidimensional time-series expression MHA(Q, K, V) in Step 4.2, call the current source power tracking expert and voltage source network construction expert in Step 2 to construct the current source intelligent agent and the voltage source intelligent agent. Each intelligent agent evaluates the strategy according to its own local objective function to complete the online identification of unknown operating conditions.

[0118] The local objective function J of a current-source type intelligent agent current for:

[0119]

[0120] In the formula, P ref (t) represents the given reference active power; P(t) represents the actual output power; Q ref (t) represents the given reference reactive power; Q(t) represents the actual output reactive power; Δi(t) represents the rate of change of the current command; t represents any moment during operation; T1 represents the upper limit of the time range; α1 represents the active power error voting weight factor; α2 represents the reactive power error voting weight factor; λ1 represents the smoothing voting weight factor.

[0121] The local objective function J of a voltage source type intelligent agent voltage for:

[0122]

[0123] In the formula, To provide the desired bus voltage; v bus (t) represents the actual measured bus voltage; f ref f(t) is the given reference frequency; f(t) is the actual frequency measurement value; Δu(t) is the rate of change of the controller output voltage; β1 is the voltage stability voting weight factor; β2 is the frequency stability voting weight factor; λ2 is the smoothing voting weight factor;

[0124] Each voting weight factor dynamically reflects the adaptability of each strategy. Based on the final decision and the fusion of multi-party evaluation results, the time-domain model Y outputs the targeted strategy weight u3(t).

[0125] Step 5: Introduce a voting confidence weighted strategy to dynamically fuse the first control variable u1(t), the compensation control variable u2(t), and the targeted strategy weight u3(t) to achieve adaptive switching of the dominant expert and optimal impedance control under different operating conditions.

[0126] In this embodiment, step 5 is implemented as follows:

[0127] Step 5.1: Drive the targeted policy weights u3(t) output by the time-domain model Y to obtain the final weighting coefficients [γ1(t), γ2(t), γ3(t)]. The specific process is as follows:

[0128] The time-domain model Y outputs a condition classification vector p(t), expressed as:

[0129] p(t) = [p1(t), p2(t), ... p a (t)…,p H (t)],p a (t)∈[0,1]

[0130] In the formula, a is the sequence number of the working condition, H is the total number of working conditions, and a = 1, 2, ..., H; p a (t) represents the confidence weight for the a-th working condition;

[0131] Introduce a predefined control strategy weight Γ, as follows:

[0132]

[0133] In the formula, m is the expert's number, m = 1, 2, 3; where 1 is the virtual impedance expert, 2 is the harmonic damping expert, and 3 is the virtual inertia expert. This corresponds to the a-th working condition and the m-th expert.

[0134] In implementation, if the virtual impedance expert is not enabled, take If virtual impedance expert is enabled, take

[0135] The final weighting coefficients are expressed as follows:

[0136]

[0137] In the formula, γ1(t) is the confidence weight of the primary control layer; γ2(t) is the confidence weight of the secondary control layer; and γ3(t) is the confidence weight of the decision control layer.

[0138] Step 5.2: Dynamically fuse the first control variable u1(t), the compensation control variable u2(t), and the targeted strategy weight u3(t) to obtain the dynamic fusion output u. final Its expression is:

[0139] u final (t)=γ1(t)×u1(t)+γ2(t)×u2(t)+γ3(t)×u3(t)

[0140] Step 5.3, dynamically fused output u finalAs the final control output, it is connected to the controller to issue real-time control commands to the PWM module, thereby completing adaptive switching and optimal impedance control.

[0141] This invention focuses on an instance where, as grid impedance increases, the SCR decreases, leading to a negative impedance region in the inverter output. A state characteristic sequence is acquired through a state-aware system. This sequence is then used to call upon the current-source power point tracking expert and voltage-source network configuration expert at the main controller layer, as well as the virtual impedance expert at the secondary controller layer, to obtain the first control quantity of the grid-connected converter and the compensation quantity Z obtained from the virtual impedance expert at this point. vir (s);

[0142] The first control variable and the compensation variable are input into the time-domain model Y and jointly encoded. After testing, a current-source local objective function is selected for evaluation, and the calculated result is 0.05, close to 0, indicating that the current-source control method is better. The targeted strategy weights output by the time-domain model Y are used to update the weights of the gating network, resulting in new weights. Finally, the targeted policy weights output by the time-domain model Y are used to drive the confidence weights, which are: γ1 = 0.3, γ2 = 0.2, γ3 = 0.5.

[0143] A simulation model built using Simulink based on MATLAB was used to simulate u. final The adaptive switching and impedance-optimal control implemented by inputting into the controller were verified; the results are as follows. Figure 4 As shown, the negative resistance region of the inverter output impedance disappears, and it exhibits positive resistance characteristics in the low-frequency range. Figure 4 The results show that the adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization proposed in this invention is feasible and effective.

Claims

1. An adaptive control method for a grid-connected converter based on multi-expert fusion and decision optimization, wherein the grid-connected converter is an LC-type grid-connected converter, comprising a DC source, an inverter, an LC-type filter, a line equivalent inductance, and a three-phase power grid connected in series, characterized in that, The specific steps are as follows: Step 1: Collect real-time operating data of the grid-connected converter and calculate the three-phase grid voltage amplitude V. m Phase angle θ, equivalent short-circuit ratio (SCR), and total harmonic distortion (THD) of three-phase grid current form a state characteristic sequence; Step 2: Set up the primary control layer, including establishing a multi-expert fusion model composed of current source power tracking experts and voltage source network experts, inputting the state feature sequence into the softened normalized weight allocation network, dynamically allocating the weights of the two types of experts and weighting and summing their control outputs to obtain the first control quantity u1(t). Step 3: Set up the secondary control layer for event triggering, including virtual impedance expert, harmonic damping expert and virtual inertia expert. When the phase margin, specific subharmonic content or frequency deviation exceeds the preset threshold, the expert is activated as needed to generate the compensation control quantity u2(t). Step 4: Set up a decision optimization layer, including using a Transformer-based time-domain model, utilizing a multi-head attention mechanism and a softened normalized weight allocation network to jointly encode the state feature sequence and the first control variable u1(t), complete the online identification of unknown working conditions and output targeted strategy weights u3(t). Step 5: Introduce a voting confidence weighted strategy to dynamically fuse the first control variable u1(t), the compensation control variable u2(t), and the targeted strategy weight u3(t) to achieve adaptive switching of the dominant expert and optimal impedance control under different operating conditions.

2. The adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization according to claim 1, characterized in that, The implementation process of step 2 is as follows: Step 2.1: Establish a multi-expert fusion model; The current-source power point tracking expert is based on an improved power synchronization control principle, wherein the output reference current i dq Represented as: i dq =G PLL (s)(ΔP+H damper (s)Δθ) In the formula, G PLL (s) is the PLL transfer function, ΔP is the power during the PLL dynamic process, and Δθ is the phase deviation during the PLL dynamic process; H damper (s) represents the damping compensation element; The voltage source type grid construction expert uses a virtual synchronous machine model to enhance the grid support capability by simulating the electromechanical transient characteristics of synchronous generators; Step 2.2, obtaining the first control variable u1(t); The softened normalized weight allocation network receives the state feature sequence obtained in step 1 as input x, x = [V m [θ, SCR, THD], and activate the current source power tracking expert and the voltage source network construction expert, outputting the multi-expert fusion weight g(x), g(x) = [g c (x), g v (x)],g c (x) represents the expert weights for current-source power point tracking, g v (x) represents the expert weights for voltage source network construction. Each expert calculates the input x and generates a current source output y. c and voltage source type output y v ; Let g v (x) satisfies the relation g c (x)+g v (x) = 1, according to the current source type output y c and voltage source type output y v The first control variable u1(t) is calculated, u1(t) = y c ×g c (x)+y v ×g v (x).

3. The adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization according to claim 1, characterized in that, The specific method for activating experts on demand as described in step 3 is as follows: When the phase margin is detected to be below the threshold, the virtual impedance expert is activated, and a virtual impedance compensation term Z is added. vir (s), whose expression is: In the formula, R is the resistance value of the bridge arm side resistor in the LC filter; L is the inductance value of the bridge arm side in the LC filter; and C is the capacitance value of the filter capacitor in the LC filter. When a specific harmonic exceeds the limit, the harmonic damping expert is activated, and a harmonic compensation term u is added as follows: In the formula, ω n ζ is the angular frequency of the nth harmonic; n Let i be the damping ratio of the nth harmonic; n The nth harmonic current component; A n The control gain for the nth harmonic; s is the Laplace operator; n = 2 indicates starting from the 2nd harmonic; N is the Nth harmonic; When a rapid frequency disturbance or excessive frequency amplitude deviation is detected, the virtual inertial expert is activated, and an inertial compensation term P is added. inertia Its expression is: In the formula, M v ω represents the virtual inertial gain, and ω represents the angular velocity. The expression for the compensation control quantity u2(t) is: u2(t) = [Z vir (s), u, P inertia ].

4. The adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization according to claim 1, characterized in that, The implementation process of step 4 is as follows: Step 4.1: Construct a multi-scale prediction model based on time-series Transformer, denoted as prediction model Y; use the state feature sequence obtained in step 1 and the first control quantity u1(t) of the grid-connected converter obtained in step 2 as the input to prediction model Y, denoted as input Q: Step 4.2: Utilize a multi-head attention mechanism to extract key dynamic patterns and correlations of the input Q from multiple subspaces and time scales, outputting the multidimensional temporal representation MHA(Q, K, V) of the current state, whose expression is as follows: MHA(Q,K,V)=Concat(head1,…head i …,head h )W O In the formula, i is the index of the attention head, i = 1, 2, ..., h; h is the total number of attention heads. i For the i-th attention head, single-head attention is performed on the subspace; K represents the historical features matching the input Q, K = QW. i K W i x Let K be the linear projection matrix; V represents the actual extracted value, V = QW i V W i V W is the linear projection matrix of V; O To output the projection matrix; The i-th attention head i The expression is: In the formula, W i Q Let Q be the linear projection matrix of the input Q; This is a scaling factor, used to prevent gradient explosion caused by large orders of magnitude (KW). i K ) T For (KW) i K The transpose of ) Step 4.3: Based on the multidimensional time-series expression MHA(Q, K, V) in Step 4.2, call the current source power tracking expert and voltage source network construction expert in Step 2 to construct the current source intelligent agent and the voltage source intelligent agent. Each intelligent agent evaluates the strategy according to its own local objective function to complete the online identification of unknown operating conditions. The local objective function J of a current-source type intelligent agent current for: In the formula, P ref (t) represents the given reference active power; P(t) represents the actual output power; Q ref (t) represents the given reference reactive power; Q(t) represents the actual output reactive power; Δi(t) represents the rate of change of current command; t represents any moment during operation; T1 represents the upper limit of the time range; α1 represents the active power error voting weight factor; α2 represents the reactive power error voting weight factor; λ1 is the smoothing voting weight factor; The local objective function J of a voltage source type intelligent agent voltage for: In the formula, To provide the desired bus voltage; v bus (t) represents the actual measured bus voltage; f ref f(t) is the given reference frequency; f(t) is the actual frequency measurement value; Δu(t) is the rate of change of the controller output voltage; β1 is the voltage stability voting weight factor; β2 is the frequency stability voting weight factor; λ2 is the smoothing voting weight factor; Each voting weight factor dynamically reflects the adaptability of each strategy. Based on the final decision and the fusion of multi-party evaluation results, the time-domain model Y outputs the targeted strategy weight u3(t).

5. The adaptive control method for grid-connected converters based on multi-expert fusion and decision optimization according to claim 4, characterized in that, The implementation process for step 5 is as follows: Step 5.1: Drive the targeted policy weights u3(t) output by the time-domain model Y to obtain the final weighting coefficients [γ1(t), γ2(t), γ3(t)]. The specific process is as follows: The time-domain model Y outputs a condition classification vector p(t), expressed as: p(t)=[p1(t),p2(t),…p a (t)…,p H (t)],p a (t)∈[0,1] In the formula, a is the sequence number of the working condition, H is the total number of working conditions, and a = 1, 2, ..., H; p a (t) represents the confidence weight for the a-th working condition; Introduce a predefined control strategy weight Γ, as follows: In the formula, m is the expert's number, m = 1, 2, 3; where 1 is the virtual impedance expert, 2 is the harmonic damping expert, and 3 is the virtual inertia expert. Corresponding to the a-th working condition and the m-th expert; The final weighting coefficients are expressed as follows: In the formula, γ1(t) is the confidence weight of the primary control layer; γ2(t) is the confidence weight of the secondary control layer; and γ3(t) is the confidence weight of the decision control layer. Step 5.2: Dynamically fuse the first control variable u1(t), the compensation control variable u2(t), and the targeted strategy weight u3(t) to obtain the dynamic fusion output u. final Its expression is: u final (t)=γ1(t)×u1(t)+γ2(t)×u2(t)+γ3(t)×u3(t) Step 5.3, dynamically fused output u final As the final control output, it is connected to the controller to issue real-time control commands to the PWM module, thereby completing adaptive switching and optimal impedance control.

Images