Improved adaptive control method based on radial basis function neural network

Abstract

The invention discloses an improved adaptive control method based on a radial basis function neural network, and the method comprises the following steps: collecting a frequency variable quantity and an active power variable quantity, inputting the frequency variable quantity and the active power variable quantity into an input layer of the RBF neural network, and outputting a rotational inertia J and a damping coefficient D from an output layer of the RBF neural network to a VSG control link; wherein in the RBF neural network, a dynamic Gaussian basis function is adopted as a hidden layer activation function; according to the method, a multi-objective optimization function including four-dimensional dynamic performance indexes including active power overshoot, adjustment time, maximum frequency deviation and recovery time is constructed, an optimization objective is implicitly defined through dynamic gradient calculation, and a momentum item is introduced in combination with a momentum gradient descent method to accelerate convergence and reduce oscillation. And deeply coupling to an RBF (Radial Basis Function) neural network weight updating mechanism, and realizing dynamic collaborative optimization of active power response on the premise of ensuring the frequency stability.

Description

Technical Field

[0001] This invention relates to the field of new energy grid connection technology, and specifically to an improved adaptive control method based on radial basis function neural networks. Background Technology

[0002] With the deepening implementation of the "dual-carbon" strategy, the power system is undergoing a transformation towards a high proportion of new energy access and a high proportion of power electronic equipment connected to the grid ("dual-high" characteristics). The random fluctuations of new energy sources (such as wind power and photovoltaics), as well as the inherent weak damping and lack of rotational inertia support of power electronic equipment such as grid-connected inverters, make the system highly susceptible to high-amplitude oscillations in inverter output power and system frequency when encountering power disturbances or load changes, seriously threatening the safe and stable operation of the power grid.

[0003] To enhance the grid support capability of power electronic equipment, Virtual Synchronous Generator (VSG) technology has emerged. This technology simulates the rotor motion equations of a synchronous generator, endowing the inverter with virtual inertia (J) and virtual damping (D) characteristics, thereby enhancing the dynamic stability of the system. Compared to traditional synchronous generators, VSGs theoretically leverage the programmability of power electronic equipment to flexibly and dynamically configure inertia and damping parameters to optimize system response. However, traditional VSG control typically uses fixed J and D parameters, making it difficult to adapt to strong fluctuations or large disturbances in renewable energy sources. This leads to problems such as increased power oscillations, excessive frequency deviations, and excessively long settling times, limiting the full potential of VSGs.

[0004] To address the limitations of fixed parameters, existing technologies have proposed various improvement methods, such as adaptive parameter adjustment, fuzzy control, and neural network methods (e.g., RBF). However, these methods generally suffer from significant shortcomings: some adaptive methods only adjust J while neglecting the collaborative optimization of D, or the adjustment rules are too simple and lack precision; fuzzy control relies on expert experience to design the rule base, resulting in weak generalization ability and complex rules affecting real-time performance; existing neural network methods (including extended RBF networks) mostly focus on single-objective optimization (e.g., minimizing frequency deviation), failing to effectively consider the dynamic response characteristics of active power (e.g., suppressing oscillations, reducing overshoot and settling time), and their evaluation functions fail to fully coordinate the multi-objective collaborative optimization of power response and frequency stability. Furthermore, some accuracy-enhancing schemes (e.g., variable universe of discourse fuzzy control) increase the complexity of hardware implementation.

[0005] In summary, the main shortcomings of existing VSG control technology are: insufficient real-time collaborative optimization of virtual inertia (J) and damping coefficient (D); a single optimization objective, failing to comprehensively consider the collaborative optimization of key dynamic performance indicators such as power regulation time / overshoot and frequency regulation time / overshoot; and limited adaptability and generalization ability, making it difficult to ensure the overall optimal dynamic performance of the system under complex and ever-changing new energy high penetration conditions.

[0006] Therefore, there is an urgent need to develop a new VSG control strategy that can effectively achieve coordinated adaptive regulation of J and D, and simultaneously optimize multiple key performance indicators such as power dynamic response and frequency stability, thereby significantly improving the adaptability and supporting role of VSG in "high-voltage and high-efficiency" power systems. Summary of the Invention

[0007] The purpose of this invention is to provide an improved adaptive control method based on radial basis function (RBF) neural networks. By leveraging the unique nonlinear mapping capability and local approximation characteristics of RBF neural networks, an adaptive adjustment mechanism is constructed between J and D parameters and dynamic performance indicators. This solves the problems of insufficient real-time collaborative optimization of J and D and single optimization objective in existing technologies, and achieves collaborative optimization of frequency dynamic characteristics and active power response. Compared with traditional single-objective optimization methods, it exhibits stronger adaptability to operating conditions.

[0008] To achieve the above objectives, the present invention proposes the following technical solution: An improved adaptive control method based on radial basis function neural networks includes: sampling frequency change and changes in active power ,Will and The input layer of the RBF neural network is fed into the input layer, and the rotational inertia J and damping coefficient D are output from the output layer of the RBF neural network to the VSG control loop. In the RBF neural network, a dynamic Gaussian function is used as the activation function for the hidden layer; Furthermore, the optimization objective was defined through dynamic gradient calculation, and the momentum term was introduced by combining the momentum gradient descent method to accelerate convergence and reduce oscillations.

[0009] Preferably, the RBF neural network consists of an input layer, a hidden layer, and an output layer. The output layer includes two nodes, and the activation function of the hidden layer is a dynamic Gaussian function.

[0010] Preferably, the output function of each node in the input layer of the RBF neural network for,

[0011] in for,

[0012] in, The change in frequency The rate of change of frequency, This represents the change in active power. This represents the rate of change of active power.

[0013] Preferably, the input to the hidden layer of the RBF neural network is,

[0014] The output function of each node in the hidden layer is:

[0015] in, This represents the k-th output function of the i-th node in the hidden layer; here, i represents the number of nodes; k represents the number of outputs of each node, with a total of 2, corresponding to the number of nodes in the output layer; in The dynamic Gaussian function is expressed as follows:

[0016] in, For a predefined center matrix, The standard deviation of the dynamic Gaussian function.

[0017] Furthermore, in the dynamic Gaussian function, , For a predefined center matrix, the standard deviation of the dynamic Gaussian function is,

[0018] according to and Dynamically Adjusted Standard Deviation of Dynamic Gaussian Function and .

[0019] Preferably, the input to the output layer of the RBF neural network is,

[0020] in, This represents the input of the k-th node in the output layer. This represents the k-th output weight of the i-th node in the hidden layer; The output layer outputs J and D expressions respectively,

[0021]

[0022] in, and These represent the output values ​​of the RBF neural network, respectively. and J and D are obtained by scaling and biasing factors, and the output is limited to... , .

[0023] Furthermore, the optimization objective is implicitly defined through dynamic gradient calculation, and the gradient calculation rule is as follows:

[0024] in, and These are the gradient values ​​of J and D, respectively.

[0025] Furthermore, the adaptive momentum gradient descent method is used to update the weights, enabling real-time parameter adjustment. The weight update formula is as follows:

[0026] This is achieved by introducing a momentum term. , To accelerate convergence and reduce oscillations, , For learning rate, As a momentum factor, in the initial stage: β≈0.8, accelerating convergence; in the steady state stage: β→0.92, suppressing oscillations; and It is the output vector of the hidden layer of the RBF neural network.

[0027] Furthermore, for an adaptive learning rate, the learning rate is adjusted under different operating conditions to improve training efficiency and stability. The learning rate Ir is...

[0028] A learning rate enhancement factor of 1.8-2.5 times is triggered when the power change rate exceeds 3000W / s or the frequency deviation is greater than 0.5Hz.

[0029] Compared with the prior art, the present invention has the following beneficial effects: This method constructs a multi-objective optimization function that includes four-dimensional dynamic performance indicators: active power overshoot, settling time, maximum frequency deviation, and recovery time. It then deeply couples this function to the weight update mechanism of the RBF neural network, thereby achieving dynamic collaborative optimization of active power response while ensuring frequency stability. Attached Figure Description

[0030] Figure 1The complete control block diagram for VSG parameter dual adaptive control based on RBF; Figure 2 This is a diagram of the VSG structure based on RBF; Figure 3 Diagram of the improved RBF neural network structure; Figure 4 The flowchart of the VSG parameter adaptive control algorithm based on the improved RBF is shown below; Figure 5 This is a block diagram of the VSG grid topology and its control system. Figure 6 It is an active power loop; Figure 7 It is a reactive power loop; Figure 8 The active power step response when J changes; Figure 9 The step response of active power when D changes; Figure 10 The step response of angular frequency when J changes; Figure 11 The step response of angular frequency when D changes; Figure 12 The root locus of the system is given by the change of J. Figure 13 The root locus of the system is given by the change in D. Figure 14 The curve showing the change in active power of VSG when active power changes abruptly; Figure 15 The curve showing the change in VSG output frequency when the active power changes abruptly; Figure 16 This is the active power response curve during the grid frequency decline phase; Figure 17 This is the frequency response curve during the power grid frequency decline phase; Figure 18 This is a heatmap of algorithm performance. Detailed Implementation

[0031] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0032] Example 1 RBF neural networks possess a powerful ability to approximate continuous nonlinear functions, handling coupling relationships between system parameters and improving control accuracy. Furthermore, the algorithm is simple, computationally fast, and easy to implement. The complete block diagram and VSG structure diagram of the dual adaptive control of VSG parameters based on the improved RBF neural network for adaptive adjustment of the moment of inertia J and damping coefficient D are attached. Figure 1 and 2As shown, the moment of inertia J and damping coefficient D are fed as outputs into the VSG control loop. The RBF neural network structure is attached. Figure 3 As shown, the neural network consists of an input layer, a hidden layer, and an output layer. The input layer has j nodes, the hidden layer has i nodes, and the output layer has 2 nodes. The activation function of the hidden layer is the dynamic Gaussian function.

[0033] Traditional RBF neural network adaptive control algorithm adopts Using this as a loss function to update the weights, this method suppresses frequency oscillations to some extent, but cannot simultaneously consider active power response. This invention improves the RBF neural network algorithm by using active power deviation and its rate of change, and frequency deviation and its rate of change as inputs. To improve the dynamic response of the VSG, four performance indicators—active power overshoot suppression, active power settling time, maximum frequency deviation, and frequency recovery time—are introduced into the RBF weight update rule. This allows the algorithm to suppress frequency oscillations while also possessing good active power response, enhancing the dynamic stability of the system.

[0034] The algorithm flow of an improved adaptive control method based on radial basis function neural networks is attached. Figure 4 As shown, it includes the following: sampling frequency change and changes in active power ,Will and The input layer of the RBF neural network is fed into the input layer, and the rotational inertia J and damping coefficient D are output from the output layer of the RBF neural network to the VSG control loop. In the RBF neural network, a dynamic Gaussian function is used as the activation function for the hidden layer; Furthermore, the optimization objective is implicitly defined through dynamic gradient calculation, and a momentum term is introduced by combining the momentum gradient descent method to accelerate convergence and reduce oscillations.

[0035] From the appendix Figure 3 It can be seen that the output function of each node in the input layer of the RBF neural network is... for:

[0036] Where x(j) is:

[0037] in, To collect frequency changes, To collect the rate of change of frequency, This represents the change in active power. This represents the rate of change of active power.

[0038] The input to the hidden layer of the RBF neural network is:

[0039] The output function of each node in the hidden layer is:

[0040] in, This represents the k-th output function of the i-th node in the hidden layer. Here, i represents the number of nodes; k represents the number of outputs per node, with a total of 2, corresponding to the number of nodes in the output layer.

[0041]

[0042] in, For a predefined center matrix, The standard deviation of the dynamic Gaussian function.

[0043] in, , For a predefined center matrix, the standard deviation of the dynamic Gaussian function is: .

[0044] according to and Dynamically Adjusted Standard Deviation of Dynamic Gaussian Function and When the power deviation ( As the frequency deviation increases, the σ of the J-network expands, enhancing its robustness to power disturbances. As the σ value increases, the σ value of the D network increases, thus improving its sensitivity to frequency oscillations.

[0045] The input representation of the output layer of the RBF neural network is:

[0046] in, This represents the input of the k-th node in the output layer. This represents the k-th output weight of the i-th node in the hidden layer.

[0047] The output layer outputs J and D expressions respectively,

[0048]

[0049] in, and These represent the output values ​​of the RBF neural network, respectively. and J and D are obtained by scaling and biasing factors, and the output is limited to... , .

[0050] This algorithm does not use an explicit traditional loss function; instead, the optimization objective is implicitly defined through dynamic gradient calculation. Its core idea is to adjust parameters directly based on the real-time state of the system, rather than optimizing a fixed loss function through offline training. The gradient calculation rules are as follows:

[0051] in, and These are the gradient values ​​of J and D, respectively.

[0052] Frequency deviation term ( ): Primarily regulates virtual inertia and suppresses frequency fluctuations. Power deviation term ( ): Enhanced power tracking capability. Differential term ( , ): Predict the dynamic trend of the system and compensate for oscillations in advance.

[0053] Momentum gradient descent is employed, by introducing a momentum term. 、 This accelerates convergence and reduces oscillations. Traditional gradient descent can oscillate during parameter updates due to large changes in gradient direction, especially when the loss function surface is not smooth. The momentum method accumulates previous gradient directions, making the update direction more stable, thereby accelerating the convergence speed. The weight update formula is:

[0054] In the formula: , For learning rate, As the momentum factor, in the initial stage (t→0): β≈0.8, accelerating convergence; in the steady-state stage (t→∞): β→0.92, suppressing oscillations. and It is the output vector of the radial basis function hidden layer.

[0055] To achieve adaptive learning rate, the learning rate is adjusted under different operating conditions to improve training efficiency and stability. Ir is...

[0056] This algorithm constructs a VSG parameter adaptive control system based on an RBF neural network. During system initialization, an RBF network structure containing a four-dimensional center point matrix is ​​established. The weight matrices of the inertia parameter J and damping parameter D are initialized using a normal distribution, and a third-order history array is constructed to achieve parameter trajectory tracking. The input signal is filtered to form a signal containing frequency deviation. Power fluctuation The four-dimensional input vector and its derivative. The Gaussian kernel width is dynamically adjusted during the network inference stage. ,in Adjusted nonlinearly based on the rate of power change The output range is constrained by a hyperbolic tangent function after activation via a radial basis function, dynamically changing with frequency deviation. Parameter updates employ momentum gradient descent, triggering a 1.8-2.5x learning rate enhancement factor when the power change rate exceeds 3000 W / s or the frequency deviation is greater than 0.5 Hz. The dynamic momentum coefficient is used to accelerate decay, and the network stability is ensured by hard limiting [-0.3, 0.3] after weight update. The output stage is double-limited to ensure J∈[0.02, 2.5] and D∈[7, 43], effectively balancing dynamic response speed and system stability, and is suitable for adaptive control requirements in microgrid frequency fluctuation scenarios.

[0057] Example 2 The VSG grid topology is shown in the attached figure. Figure 5 As shown, it mainly consists of four parts: an energy storage system, a three-phase inverter, an LC filter, and a VSG control algorithm. The VSG control algorithm mainly includes two parts: an active power control loop (oscillation equation and virtual speed governor) and a reactive power control loop (virtual exciter). The VSG control algorithm outputs a command voltage, which is controlled by a voltage-current dual closed-loop system to generate a pulse signal. The system pulse trigger signal adopts a Space Vector Pulse Width Modulation (SVPWM) strategy, which has advantages such as fast response, high stability, and low output voltage harmonics.

[0058] The active-frequency regulation of the VSG consists of two parts: the active-frequency droop element and the rotor motion equation. The droop characteristic equation of Pf is: (1) The rotor motion equation (swing equation) of a synchronous generator is: (2) In the formula: To provide active power input to the synchronous motor, To output active power, This is a reference value for active power. This is the droop coefficient of the virtual speed governor. and These are the output angular frequency and the angular frequency reference value, respectively. Let J be the power angle of the synchronous generator, and J and D be the virtual moment of inertia and damping coefficient, respectively. The active power frequency regulation control block diagram is obtained from the droop characteristic equation and the oscillation equation of Pf, as shown in the appendix. Figure 6 As shown.

[0059] The reactive power loop mainly simulates the excitation system of a synchronous generator, and its expression is: (3) In the formula, E, , These represent the single-phase effective value of the inverter's electromotive force, the effective phase voltage of the grid, and the reference phase voltage, respectively. To output the reactive power measurement value, For reactive power setpoint, Let K be the reactive power droop coefficient, and K be the integral coefficient. The reactive power-voltage control block diagram can be obtained from equation (3) as shown in the appendix. Figure 7 As shown The output active and reactive power of VSG can be calculated based on the main circuit system topology, as shown in equation (4).

[0060] (4) In the formula: This is the effective value of the grid-side voltage; The impedance of the filter circuit, and ,in For the impedance of the filter circuit Resistance value For the impedance of the filter circuit The reactance value; α represents the phase of the inverter output voltage; α is the impedance angle of the filter circuit. The reactance value in the impedance of a typical filter circuit is... Much greater than the resistance value Therefore, it is believed ,Right now With line reactance E The relationship between δ and δ can be expressed as: (5) Based on the analysis method of the traditional small-signal model of synchronous generator, and combined with equations (2) and (5), the active power reference value can be obtained. arrive The second-order transfer function is: (6) from The second-order transfer function to ω is: (7) In the formula

[0061] Therefore, in order to study J and The mechanism by which VSG output active power and frequency are affected will be explored in this embodiment by observing the effects of different values ​​of J and on the VSG output active power and frequency. The dynamic response under the given conditions is analyzed. Equation (6) is rearranged into the standard form of a second-order system as follows: (8) Among them, natural frequency Damping ratio .

[0062] The time-domain response expression, obtained through the inverse Laplace transform, is as follows: (9) Among them, the damped oscillation frequency Natural frequency Damping ratio .

[0063] The active power step response curve for J variation is plotted using equation (9), as shown in the appendix. Figure 8 As shown.

[0064] From the appendix Figure 8 It can be seen that as the moment of inertia J increases, the system's natural frequency... Because it decreases inversely proportional to J, it directly leads to a decrease in response speed and a lengthening of settling time; simultaneously, the damping ratio... As the moment of inertia J decreases, the system tends towards an underdamped state, characterized by a significant increase in overshoot, aggravated oscillation amplitude, and a decreased convergence rate. Conversely, as the moment of inertia J decreases, the natural frequency... Increasing the damping ratio ζ accelerates system response and shortens settling time, while increasing the damping ratio ζ causes the system to transition to an overdamped state, at which point overshoot is completely eliminated and oscillation characteristics are effectively suppressed. Therefore, while small moment of inertia systems achieve rapid response, they must be wary of phase lag and resonance risks; while large moment of inertia systems can achieve smooth motion through critical damping, dynamic hysteresis reduces control efficiency.

[0065] Similarly, the active power step response curve for D changes can be plotted using equation (9) as shown in the attached figure. Figure 9 As shown.

[0066] When the damping coefficient D increases, the damping ratio As D increases, the system tends towards an overdamped state, at which point the overshoot and oscillation of active power decrease; when D decreases, the damping ratio... As the damping coefficient decreases, the system tends to become underdamped, with increased overshoot and oscillations. For large damping coefficients, the response is smooth and oscillatory, but may be too slow; while small moments of inertia D result in a fast response but may oscillate violently, leading to instability. The response characteristics of active power with changes in J and D are summarized in Table 1 below.

[0067] Table 1. Response characteristics of active power when J and D change.

[0068] Similarly, by applying the inverse Laplace transform to equation (6), the time-domain response is: (10) Among them, attenuation coefficient , actual frequency

[0069] Natural frequency

[0070] Equivalent damping ratio

[0071] Adjusting time

[0072] Ascent Time

[0073] when At that time, an engineering approximation can be made.

[0074] With D fixed, plot the step response curve of angular frequency as J changes.

[0075] From the appendix Figure 10 It can be seen that increasing the moment of inertia J leads to increased inertia, resulting in a slower system response and a longer rise time. Equivalent damping ratio. Decreasing J causes the system to become underdamped, increasing oscillations. This underdamping can lead to sustained oscillations. Decreasing J reduces inertia, accelerating the system response and shortening the rise time. Equivalent damping ratio As the damping increases, the number of overdamped oscillations in the system decreases.

[0076] Similarly, with J fixed, the step response curve of the angular frequency as D changes is shown in the attached figure. Figure 11 As shown.

[0077] Increasing D increases the oscillation frequency. Decreasing D increases the oscillation period; reduces overshoot (increased damping suppresses oscillation); and shortens the settling time. Decreasing D increases the oscillation frequency. An increase in overshoot shortens the oscillation period. An increase in overshoot (weakened damping leads to more intense oscillation) increases the settling time.

[0078] Table 2 Frequency response characteristics when J and D change

[0079] The response characteristics of active power and frequency when J and D change are summarized in Table 3.

[0080] Table 3. Response characteristics of active power and frequency when J and D change.

[0081] Table 3 shows that decreasing the moment of inertia J speeds up the response speed of both active power and frequency, reduces the overshoot of active power, but increases the overshoot of frequency. Increasing the moment of inertia J slows down both frequency and power responses, reduces the overshoot of frequency, but increases the overshoot of active power. Therefore, J needs to be chosen with an appropriate value to balance response speed and overshoot. For D, increasing the moment of inertia D reduces the overshoot of both active power and frequency, speeds up the frequency response, and slows down the response speed of active power.

[0082] To ensure the rationality of the selected control parameters, a stability analysis was performed according to equation (6), and the parameters were taken as follows: E = U g = 311 V, active power droop factor = 1000, equivalent reactance X f = 0.942 Ω. (Plotting) G ( s The pole distribution diagram under varying parameters, and the root locus of the system when the virtual inertia J and damping D change are respectively as shown in the figure. Figure 12 and Figure 13 As shown.

[0083] Figure 12 The diagram shows a fixed virtual damping value D = 25 (N·m·rad) / s, with the virtual inertia parameter J varying from 0.2. Increase to 2 The pole distribution diagram of G(s) is shown. It can be seen that as J increases, a pair of conjugate complex roots gradually approach the imaginary axis, and the system stability deteriorates. The system can converge as J increases, but the convergence speed slows down. The larger J is, the slower the poles move, indicating that when the virtual inertia is large, the change in the J parameter has a smaller impact on the system's dynamic characteristics. Therefore, to ensure a certain stability margin for the system and to ensure good control effect of parameter adjustment, the magnitude of the J parameter should be limited in VSG adaptive control, and the virtual inertia adjustment process should satisfy... (11) Figure 13 The figure shows the virtual inertia J = 0.25. With the virtual damping D constant, the pole distribution of G(s) is shown when it increases from 5 to 100 (N·m·rad) / s. The graph shows that as D gradually increases, a pair of conjugate complex roots move away from the imaginary axis, enhancing system stability and accelerating convergence. When D further increases, the poles change from conjugate complex roots to two real roots, indicating a shift from underdamped to overdamped state. One of the real roots gradually moves closer to the imaginary axis, weakening system stability. The overdamped state also prolongs the system's transient settling time. Therefore, the virtual damping parameter should be set relatively large, but not excessively so. To ensure good control performance, a damping parameter D < 40 (N·m·rad) / s is chosen.

[0084] Example 3 To verify the superiority of the proposed adaptive control method for VSG virtual inertia and dynamic compensation damping based on RBF neural network, a single VSG system was built in Matlab / Simulink software. The simulation results were compared with the RBF neural network-based adaptive control method for virtual inertia and damping coefficient under power fluctuation and low-pressure ride-through conditions. The simulation parameters are shown in Table 4.

[0085] Table 4 Main Control Parameters of VSG and Grid Side

[0086] To verify the impact of active power command switching on the VSG, the active power command change of the VSG was set as follows: it abruptly changed from 10 kW to 14 kW at 0.8 s, and then switched back to 12 kW at 1.4 s. The active power response, frequency response, and other indicators under this condition are shown in the attached figure. Figure 14 Appendix Figure 15 As shown in Table 5.

[0087] Table 5 Analysis Indicators for Operating Condition 1

[0088] The curve of VSG active power change when active power changes abruptly is shown in the attached figure. Figure 14As shown in the figure, at 0.8s, the active power command switches from 10kW to 14kW. Due to the influence of inertia, the overshoot of active power under traditional non-adaptive VSG control is relatively large, with a peak value of approximately 14.860kW and an overshoot of about 6.1%. After adopting adaptive control strategy 1, the active power overshoot is suppressed, with a peak value of approximately 14.440kW and an overshoot of about 3.1%. After adopting the traditional RBF control strategy, the suppression effect of active power overshoot is significant, with a peak value of approximately 14.359kW and an overshoot of about 2.5%. After adopting the control strategy of this invention, the maximum active power overshoot is approximately 14.104kW, and the overshoot is about 0.7%. Compared with traditional fixed damping, this method improves the suppression effect of active power by approximately 87.9%.

[0089] The curve showing the change in VSG output frequency during a sudden change in active power is attached. Figure 15 As shown in the figure, at 0.8s, under the traditional non-adaptive VSG, the sudden power change causes the VSG output frequency to deviate rapidly from the rated value, with a maximum amplitude of 50.067Hz, and the time to recover to the rated value is approximately 0.42s. After adopting adaptive control strategy 1, the maximum frequency deviation is 50.048Hz, and the time to recover to the rated value is approximately 0.37s. When using traditional RBF control, the maximum frequency deviation is 50.077Hz, indicating that this strategy has a good effect on suppressing active power oscillations, but the frequency deviation is the largest, and the time to recover to the rated value is approximately 0.30s. After adopting the control strategy of this method, the maximum frequency deviation is 50.052Hz, and the time to recover to the rated value is approximately 0.25s.

[0090] In summary: Adaptive control 1 has the best frequency suppression effect, but active power oscillations are still significant; traditional RBF adaptive control will… As a loss function, it can suppress active power oscillations, but the frequency deviation is the largest. After adopting this method, the suppression effect of active power oscillations is the best, the maximum frequency deviation is also greatly reduced, and the recovery time is the fastest.

[0091] The impact of grid-side frequency disturbances on the VSG was verified under the parameters specified in Table 4. Initially, the grid-connected power of the VSG was 15 kW, and the grid-side frequency was 50 Hz. At 0.5 s, the grid-side frequency decreased by 0.2 Hz, and at 1.5 s, the grid-side frequency recovered to 50 Hz. The active power response, frequency response, and other parameters under this condition are shown in the attached table. Figure 16 Appendix Figure 17 As shown in Table 6.

[0092] Table 6 Analysis Indicators for Operating Condition 2

[0093] During the grid frequency decrease phase, the active power response curve is as follows: Figure 16 As shown, the active power output of the VSG increased to 21.120kW, verifying its primary frequency regulation function. At this time, the power overshoot controlled by fixed parameter control, adaptive control 1, traditional RBF control, and the control method of this invention were 4.4%, 3.75%, 2.75%, and 1.34%, respectively.

[0094] During the phase of grid frequency decline, the frequency response curve is as follows: Figure 17 As shown, under fixed parameter control, the maximum frequency deviation is 0.12Hz and the frequency settling time is 0.38s. Under adaptive control 1, the maximum frequency deviation is 0.03Hz and the frequency settling time is 0.48s; under conventional RBF control, the maximum frequency deviation is 0.02Hz and the frequency settling time is 0.38s. Under the adaptive control of this invention, the maximum frequency deviation is 0 and the frequency settling time is 0.24s.

[0095] Simulation results show that the improved control strategy proposed in this invention is significantly better than other strategies in suppressing active power oscillations and frequency oscillations, with a shorter oscillation stabilization time, effectively suppressing grid frequency interference, and optimizing the dynamic response process of VSG active power and frequency.

[0096] After normalization, a comparative analysis of the four control algorithms was conducted from the perspective of dynamic response characteristics. The main evaluation indicators included active power overshoot, maximum frequency deviation, dynamic recovery time, and response rise time. (See attached...) Figure 18 The algorithm performance heatmap shown indicates that darker colors represent better performance. Visual comparison results demonstrate that the proposed method exhibits significant advantages in suppressing active power overshoot and shortening oscillation recovery time, outperforming other algorithms in both aspects. In terms of maximum frequency deviation control, it is second only to adaptive method 1. While its rise time is relatively slower, it is still better than adaptive method 1. Through comprehensive evaluation of multi-dimensional performance indicators, this method demonstrates significant superiority in dynamic response quality and system stability assurance.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An improved adaptive control method based on radial basis function neural networks, characterized in that, include: sampling frequency change and changes in active power ,Will and The input layer of the RBF neural network is fed into the input layer, and the rotational inertia J and damping coefficient D are output from the output layer of the RBF neural network to the VSG control loop. In the RBF neural network, a dynamic Gaussian function is used as the activation function for the hidden layer; Furthermore, the optimization objective was defined through dynamic gradient calculation, and the momentum term was introduced by combining the momentum gradient descent method to accelerate convergence and reduce oscillations.

2. The improved adaptive control method based on radial basis function neural network according to claim 1, characterized in that, The RBF neural network consists of an input layer, a hidden layer, and an output layer. The output layer includes two nodes, and the activation function of the hidden layer is the dynamic Gaussian function.

3. The improved adaptive control method based on radial basis function neural network according to claim 1, characterized in that, The output function of each node in the input layer of the RBF neural network for, in for, in, The change in frequency The rate of change of frequency, This represents the change in active power. This represents the rate of change of active power.

4. An improved adaptive control method based on radial basis function neural networks according to claim 1, characterized in that, The input to the hidden layer of the RBF neural network is, The output function of each node in the hidden layer is: in, This represents the k-th output function of the i-th node in the hidden layer; here, i represents the number of nodes; k represents the number of outputs of each node, with a total of 2, corresponding to the number of nodes in the output layer; in The dynamic Gaussian function is expressed as follows: in, For a predefined center matrix, The standard deviation of the dynamic Gaussian function.

5. An improved adaptive control method based on a radial basis function neural network according to claim 4, characterized in that, In the dynamic Gaussian function , For a predefined center matrix, the standard deviation of the dynamic Gaussian function is, according to and Dynamically Adjusted Standard Deviation of Dynamic Gaussian Function and .

6. An improved adaptive control method based on radial basis function neural networks according to claim 1, characterized in that, The input to the output layer of the RBF neural network is, in, This represents the input of the k-th node in the output layer. This represents the k-th output weight of the i-th node in the hidden layer; The output layer outputs J and D expressions respectively, in, and These represent the output values ​​of the RBF neural network, respectively. and J and D are obtained by scaling and biasing factors, and the output is limited to... , .

7. An improved adaptive control method based on a radial basis function neural network according to claim 6, characterized in that, The optimization objective is implicitly defined through dynamic gradient calculation, and the gradient calculation rule is as follows: in, and These are the gradient values ​​of J and D, respectively.

8. An improved adaptive control method based on a radial basis function neural network according to claim 6, characterized in that, By combining the adaptive momentum gradient descent method to update the weights, real-time parameter adjustment is achieved. The weight update formula is as follows: This is achieved by introducing a momentum term. , To accelerate convergence and reduce oscillations, , For learning rate, As a momentum factor, in the initial stage: β≈0.8, accelerating convergence; in the steady state stage: β→0.92, suppressing oscillations; and It is the output vector of the hidden layer of the RBF neural network.

9. An improved adaptive control method based on a radial basis function neural network according to claim 6, characterized in that, To achieve adaptive learning rate, the learning rate is adjusted under different operating conditions to improve training efficiency and stability. The learning rate Ir is... A learning rate enhancement factor of 1.8-2.5 times is triggered when the power change rate exceeds 3000W / s or the frequency deviation is greater than 0.5Hz.

Images