Harmonic suppression and unbalanced load compensation method for parallel active power filter

Abstract

The invention discloses a harmonic suppression and unbalanced load compensation method for a parallel active power filter. The method comprises the following steps: firstly, establishing a noise model containing various noise sources and statistical characteristics thereof; then combining the noise model to construct an active power filter containing noise interference; then selecting a self-adaptive neural network algorithm, wherein the self-adaptive neural network algorithm comprises a recursive least square method self-adaptive linear element and a Kalman filtering self-adaptive linear element; an active power filter control strategy adopting an adaptive neural network algorithm is established; and finally, based on the adaptive neural network algorithm, harmonic current detection and load unbalance current compensation are carried out, a recursive least square adaptive linear element is adopted to extract harmonic current, and a Kalman filtering adaptive linear element is adopted to carry out unbalance compensation. According to the method, the harmonic suppression and load imbalance compensation capabilities of the parallel APF are improved by using a noise condition adaptive neural network technology.

Description

Technical Field

[0001] This invention relates to the field of active power filter harmonic suppression technology, specifically a method for harmonic suppression and unbalanced load compensation using a parallel active power filter. Background Technology

[0002] Harmonic pollution and load imbalance in power systems are long-standing key issues affecting power quality. Harmonic currents not only cause overheating, aging, and even damage to power equipment, but also induce voltage distortion in the power grid, affecting the normal operation of other power equipment. Load imbalance leads to three-phase current imbalance in the power system, increasing power equipment losses, reducing system efficiency, and potentially causing a series of safety hazards. Therefore, effectively suppressing harmonics and compensating for load imbalance are crucial for ensuring stable operation of the power system and improving power quality. Parallel active power filters (APFs), as an advanced power quality improvement device, have become an effective way to solve these problems due to their advantages such as fast response and flexible control. However, traditional APF control strategies often struggle to adapt to the complex and ever-changing noise environment of power systems, and their control performance is easily affected by noise interference.

[0003] Traditional APF control strategies, such as proportional-integral (PI) control and hysteresis control based on harmonic current detection, can effectively suppress harmonics and compensate for load imbalances under ideal conditions. However, real power systems are subject to various noise interferences, such as measurement noise and system parameter disturbances. These noises can severely affect the accurate detection of harmonic currents and the stability of the controller, thereby reducing the control accuracy of the APF.

[0004] To overcome this challenge, adaptive neural network control technology has been increasingly applied to the field of APF control in recent years. Adaptive neural networks possess powerful nonlinear approximation and learning capabilities, enabling them to adapt to complex and ever-changing system environments and effectively suppress the effects of noise. However, they have shortcomings in terms of noise filtering accuracy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and propose a method for harmonic suppression and unbalanced load compensation of parallel active power filters. This method utilizes noise condition adaptive neural network technology to improve the harmonic suppression and load imbalance compensation capabilities of parallel APFs.

[0006] To achieve the above objectives, the technical solution specifically adopted by the present invention is as follows:

[0007] A method for harmonic suppression and unbalanced load compensation using a parallel active power filter includes the following steps:

[0008] Step 1: Model the noise in the power system and establish a noise model that includes multiple noise sources and their statistical characteristics;

[0009] Step 2: Establish an LCL-type parallel active power filter system model, and construct an active power filter with noise interference based on the noise model.

[0010] Step 3: Select an adaptive neural network algorithm, which includes a recursive least squares adaptive linear element and a Kalman filter adaptive linear element;

[0011] Step 4: Establish an active power filter control strategy using an adaptive neural network algorithm;

[0012] Step 5: Based on the aforementioned adaptive neural network algorithm, harmonic current detection and load imbalance current compensation are performed. The harmonic current is extracted using a recursive least squares adaptive linear element, and the imbalance compensation is performed using a Kalman filter adaptive linear element.

[0013] Preferably, the noise includes Gaussian white noise and colored noise, and a noise model is established through analysis of actual power system data.

[0014] Preferably, the adaptive neural network algorithm employs an adaptive learning algorithm to adjust the neural network parameters in real time and introduces an exponential weighting factor to adjust the network's weights on historical data.

[0015] Preferably, the harmonic current detection specifically includes:

[0016] The load current is sampled to obtain a discrete signal, which is then expanded using a Fourier series as follows:

[0017] i L (n) = [a1, b1, ..., a N ,b N ][cos(ωnT s ),sin(ωnT s ),...,cos(NωnT s ),sin(NωnT s )] T

[0018] =W T (n)X(n)

[0019] Among them, a k and b k T is the harmonic coefficient, ω is the angular frequency, and T is the harmonic coefficient. sThe sampling period is k = 2, 3, ..., which represents the harmonic order. X(n) is the input vector of the model, and W(n) is the estimated value of the weight vector at time n.

[0020] The discrete signal is used as the input vector X(n) of the adaptive neural network algorithm, and the weight vector W(n) is updated through the adaptive algorithm.

[0021] The amplitude and phase of each harmonic are calculated based on the updated weight vector W(n), enabling the detection of any harmonic.

[0022] Preferably, the adaptive algorithm is a recursive least squares method, and its cost function is:

[0023]

[0024] Where λ is the exponential weighting factor, d(i) is the desired response, X(i) is the input signal vector, and W(n) is the weight vector.

[0025] Preferably, the method for updating the weight vector W(n) is as follows:

[0026] W(n)=Z(n)R -1 (n)=W(n-1)+k(n)ξ(n)

[0027] Where R(n) represents the time mean of the autocorrelation matrix of the input signal vector X(i), k(n) is the Kalman gain vector, and ξ(n) = d(n) - X(n)W T (n-1) represents the prior error.

[0028] Preferably, the load imbalance current compensation specifically includes:

[0029] The three-phase asymmetrical signal is decomposed into two vertical components by coordinate transformation;

[0030] An adaptive Kalman filter is used to remove noise and interference signals, and to accurately extract the fundamental and arbitrary harmonic currents.

[0031] Preferably, in step 5, the method of using adaptive Kalman filtering to remove noise and interference signals is as follows:

[0032] First, construct the state variance and observation equations for the Kalman filter:

[0033] X k+1|k =Φ k X k|k +B k U k +W k

[0034] Y k+1|k =Hk X k +V k

[0035] In the formula, X k|k X represents the state variable value of the system, which is the optimal value at that moment; k+1|k This indicates the result predicted using the previous state; Y k+1|k U represents the observed variable values ​​of the system; k Φ represents the control quantity of the system. k H represents the state transition matrix; k B represents the observation transition matrix; k W represents the system's control matrix. k V k Let Q and R represent two uncorrelated Gaussian noises with zero mean, respectively;

[0036] The system state values ​​from time k to time k+1 are:

[0037] X k+1|k =Φ k X k|k

[0038] The error covariance matrix from time k to time k+1 is:

[0039]

[0040] In the formula, P k+1|k X represents k+1|k The corresponding error covariance; P k|k X k|k The corresponding error covariance,

[0041] P k|k =E[(X k|k -X k|k-1 (X) k|k -X k|k-1 ) T ]

[0042] The Kalman filter gain at time k+1 is calculated as follows:

[0043] K k+1 =P k+1|k H T HP k+1|k H T +R) -1

[0044] By predicting the states at time k, and then measuring the current state, combining the measured and observed values, we can obtain the optimal state estimate X at time k+1. k+1|k+1 :

[0045] X k+1|k+1 =X k+1|k +K k+1 (Z k+1 -H k+1 X k+1|k )

[0046] To meet the operating conditions of the Kalman filter, X in state k+1 needs to be updated. k+1|k+1 Covariance:

[0047] P k+1|k+1 =(IK k+1 H k )P k+1|k

[0048] Where I is the identity matrix, and the initial filtering value X is given according to the system state. 1|0 and P 1|0 The value of the state variable X can be obtained by repeated filtering and calculation.

[0049] This invention has the following characteristics and beneficial effects:

[0050] This invention proposes a method combining RLS-Adaline (Recursive Least Squares and Adaptive Linear Neural Network) and KF-Adaline (Kalman Filter Adaline) to estimate harmonic components in power systems. The neural network estimator is based on the use of an adaptive perceptron, which includes a linear adaptive neuron called the Adaline. A Kalman filter and recursive least squares algorithm are used for weight updates of the Adaline. The estimator can accurately track DC component signals contaminated by noise and attenuated. Adaptive tracking of power system harmonic components can be easily achieved using these algorithms. The proposed method was tested on static and dynamic signals, and the adaptive neural network in parallel with an active filter performed well in suppressing harmonic content. Simulation results also show that, under conditions of noise, voltage sag, and three-phase imbalance, it outperforms traditional methods based on three-phase instantaneous reactive power theory. p -i q For active power filters, the adaptive neural network method offers higher reliability and accuracy. Attached Figure Description

[0051] Figure 1 This is a framework diagram of the parallel active power filter system under noise interference according to the present invention.

[0052] Figure 2 This is a diagram of the active filter control strategy using the adaptive algorithm of the present invention.

[0053] Figure 3This is a block diagram of the harmonic current detection based on the Adaline algorithm of the present invention.

[0054] Figure 4 This is a schematic diagram of the Kalman calculation process of the present invention.

[0055] Figure 5 This is a summary diagram of one phase of the simulation based on the adaptive neural network of the present invention.

[0056] Figure 6 This is the THD diagram of the grid-side current under conventional APF noise conditions according to the present invention.

[0057] Figure 7 This is the THD diagram of the grid-side current under the noise condition of the adaptive neural network APF of the present invention.

[0058] Figure 8 This is a simulation diagram of grid-side voltage drop according to the present invention.

[0059] Figure 9 This is a simulation diagram of the unbalanced current on the load side of the present invention.

[0060] Figure 10 This is the THD diagram of the network-side current of the adaptive neural network APF under interference conditions according to the present invention.

[0061] Figure 11 This is the THD diagram of the traditional APF network-side current under interference conditions according to the present invention. Detailed Implementation

[0062] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0063] A method for harmonic suppression and unbalanced load compensation using a parallel active power filter includes the following steps:

[0064] Step 1: Model the noise in the power system and establish a noise model that includes multiple noise sources and their statistical characteristics.

[0065] In this embodiment, the noise signal used is Gaussian white noise, and the specific method is as follows:

[0066] In practical applications of power systems, Gaussian white noise interference is unavoidable, affecting parameter estimation and leading to a certain degree of randomness in frequency measurement results. Therefore, it is necessary to analyze the accuracy of frequency measurements under the influence of white noise and determine the impact of random white noise on frequency measurement results. In the field of power system signal processing, Gaussian white noise models are commonly used to study the impact of noise on parameter estimation. Gaussian white noise refers to white noise with a normal probability distribution, a mean of 0, and a variance of σ. 2 Its probability density function can be expressed as:

[0067]

[0068] For a real sinusoidal signal with amplitude A, the signal-to-noise ratio (SNR) is expressed as follows:

[0069]

[0070] Step 2: Establish an LCL-type parallel active power filter system model, and construct an active power filter with noise interference based on the noise model.

[0071] The noise model is set to randomly generate Gaussian white noise with a set mean and standard deviation. The following is a set of white noise with a set mean of 0 and a standard deviation of 1. The mean setting affects the base noise value, while the standard deviation affects the peak noise value. In the simulation experiment, noise signals will be added at two points: the load-side current and the harmonic signals extracted by the RLS-adaline module.

[0072] Furthermore, an active power filter (APF) is a highly efficient filter capable of suppressing harmonics and reactive currents in a power system. This embodiment employs a parallel-connected APF device:

[0073] An APF device mainly consists of three modules (highlighted in dashed boxes): a harmonic detection module, a current tracking module, and the APF main circuit module. These modules work together to suppress harmonics. First, the harmonic detection module acquires the load current i through a sensor. L The selected harmonic current detection algorithm is used to separate the harmonic current, which is then modulated to generate a digital signal. Next, based on the system's harmonic current signal, a current with equal amplitude but opposite phase is calculated as a compensation current reference command signal. Output. The current tracking module then selects an appropriate control strategy to output the compensation current signal. This is converted into a control signal for the switching devices in real time. Finally, the APF main circuit receives the control signal and generates the actual compensation current i. c And input it into the system, the actual compensation current i cThe harmonic current components in the load current are canceled out, resulting in a grid current i s It contains only the fundamental component, becoming a standard sinusoidal current, thus achieving the purpose of harmonic suppression.

[0074] Step 3: Select an adaptive neural network algorithm, which includes recursive least squares adaptive linear elements and Kalman filter adaptive linear elements.

[0075] The method used in this embodiment is based on an adaptive linear neural network (Adaline) model for harmonic current detection, which can extract harmonic currents from the power grid more accurately and exhibits strong robustness in the face of system noise and irregular waveforms.

[0076] Specifically, assume the power supply voltage is a standard sinusoidal signal u. S (t)=U s sinωt, where ω is the angular frequency, and the power supply is applied to a nonlinear load, then the load current i L (t) can be represented as:

[0077]

[0078] Where k = 2, 3, ..., represents the harmonic order. 1p (t) is the active component of the fundamental current, i 1q (t) represents the reactive component of the fundamental current, i h (t) represents the harmonic current component. The sampling period T is used as the reference. S The load current is sampled at a sampling frequency of f. s =1 / T S The resulting discrete load current signal, after Fourier series expansion, is shown below:

[0079]

[0080] Equation (4) can be rewritten as:

[0081]

[0082] W(n) = [a1, b1, ..., a N ,b N ] T (7)

[0084] X(n)=[cos(ωnT s ),sin(ωnT s ),...,cos(NωnT s ),sin(NωnT s )] T (8)

[0086] like Figure 3 As shown, X(n) is the input vector of the model, and W(n) is the estimated value of the weight vector at time n. The model also updates the weight coefficients W(n) through an adaptive algorithm to gradually approach the optimal weight coefficients. Then, the amplitude and phase of each harmonic are calculated according to equation (5) to realize the detection of any harmonic.

[0087] Step 4: Establish an active power filter control strategy using an adaptive neural network algorithm.

[0088] Specifically, such as Figure 2 As shown, the active power filter control strategy is as follows: First, for the load current i L A Kalman filter module can be used to accurately extract the harmonic and reactive components in the load current, and obtain the harmonic current reference signal i. r1 Meanwhile, for the DC-side voltage control loop, an adaptive neural network algorithm can be introduced to optimize the PI controller, enabling it to adjust according to the DC-side voltage V. dc With reference voltage V dcref The deviation is adaptively adjusted by the control parameters. After processing by a low-pass filter, and combined with the phase θ of the phase-locked loop output and the dq0 / abc transformation, the fundamental active current reference i is obtained. r2 .

[0089] will i r1 with i r2 The total reference current i is obtained by superposition. ref Then, with the compensation current i c The error current i is obtained by comparison. e Afterwards, the hysteresis current controller, based on i... e A PWM signal is generated to drive the inverter to output compensation current, thereby effectively compensating for grid harmonics and reactive power. This step constitutes the overall process; subsequent steps are simply selections and supplements to the algorithms used in this step.

[0090] Step 5: Based on the active power filter control strategy of the adaptive neural network algorithm, harmonic current detection and load imbalance current compensation are performed. The recursive least squares method adaptive linear element is used to extract the harmonic current, and the Kalman filter adaptive linear element is used for imbalance compensation.

[0091] Specifically,

[0092] The recursive least squares algorithm is an improved version of the steepest descent method. It is an exponentially weighted least squares method and a more efficient data processing algorithm.

[0093] The cost function of the least squares method is defined as follows:

[0094]

[0095] In equation (9), d(i) is the expected response at time i, X(i) = [x(i)x(i-1)...x(i-M+1)] corresponds to the input signal vector at time i, and W(n) = [W0W1...W M-1 [ ] The tap weight vector corresponding to time n. M is the filter length. The parameter λ is called the exponential weighting factor, with a value between 0 and 1, but close to a positive number of 1. The exponential weighting factor is introduced into the algorithm to achieve better signal tracking. For error data closer to time n, the newer the data, the greater the weight is added to the newer data; while for error data farther away from the current time, the added weight is smaller. In short, the algorithm focuses on processing data near the current time, reducing the influence of older data on the present. If the exponential weighting factor λ = 1, the algorithm is no different from the ordinary least squares method.

[0096] According to the steepest descent method theory, in order to obtain the optimal weight coefficient W... opt Setting the derivative of the cost function J(n) with respect to the weight coefficients W(n) to 0, we get:

[0097]

[0098] Rearranging equation (10) yields:

[0099]

[0100] We can solve for W(n):

[0101]

[0102] make:

[0103]

[0104] Substituting equations (13) and (14) into equation (12), equation (12) can be simplified to:

[0105] W(n)=Z(n)R -1 (n) (15)

[0106] R(n) represents the time mean of the autocorrelation matrix of the input vector X(i), and Z(n) represents the time mean of the cross-correlation vector between the desired response and the input vector X(i). By observing R(n) and Z(n), we can respectively use

[0107] Recursive form representation:

[0108] R(n)=λR(n-1)+X T (n)X(n) (16)

[0109] Z(n)=λZ(n-1)+d(n)X(n) (17)

[0110] Using the theorem of matrix inversion, the inverse matrix R can be calculated from R(n). -1 (n), and denoted as P(n):

[0111] P(n)=λ -1 [P(n-1)-X T [(n)k(n)P(n-1)] (18)

[0112] In equation (18), k(n) is the Kalman gain vector, which can be expressed as:

[0113]

[0114] Substituting equations (13) and (14) into (11), we can derive the weight update formula for the algorithm:

[0115] W(n)=Z(n)R -1 (n)=W(n-1)+k(n)ξ(n) (20)

[0116] Where ζ(n) is the prior error:

[0117] ξ(n)=d(n)-X(n)W T (n-1) (21)

[0118] Based on the above derivation formula, the implementation process of the RLS algorithm can be obtained. The specific algorithm calculation and update process is shown in Table 1. In the RLS algorithm, the exponential weighting factor λ determines the performance of the algorithm. The larger λ is, the greater the weighting of the data in the past time step, and the relatively smaller the steady-state error of the detection result, but the tracking ability of the load is poor. The smaller λ is, the tracking ability is significantly enhanced when the load changes abruptly, but it cannot solve the disadvantage of amplification of steady-state error.

[0119] Table 1: Overview of the RLS Algorithm

[0120]

[0121] Kalman filtering is a real-time recursive filtering method based on linear minimum variance estimation. It utilizes a state-space model to design filters in the time domain, making it widely applicable to the estimation of multidimensional stochastic processes and easy to implement in computers and digital signal processors. Using a known system model, this method can continuously estimate and correct the signal to obtain the optimal filtering result, and has been widely used in power quality detection. Kalman filtering is particularly adept at handling fundamental and harmonic currents (or voltages) in three-phase symmetrical signals, but its accuracy is poor when handling three-phase asymmetrical signals, especially in detecting fundamental positive-sequence currents and harmonic currents. Furthermore, when the system is in a steady state for a long period, the gain coefficient of the Kalman filter gradually decreases, which may lead to data saturation or even data divergence.

[0122] To address this issue, this embodiment proposes an arbitrary harmonic current detection method combining Adaline and adaptive Kalman filtering. This method decomposes the three-phase asymmetrical signal into two vertical components through coordinate transformation, and then uses adaptive Kalman filtering to remove noise and interference signals, thereby accurately extracting the fundamental and arbitrary harmonic currents. This method provides effective support for parallel-type APF to suppress harmonic currents.

[0123] Next, we will briefly introduce the specific implementation process of this algorithm. Assume the stochastic linear difference equation is as follows:

[0124] X k+1|k =Φ k X k|k +B k U k +W k (twenty two)

[0125] Y k+1|k =H k X k +V k (twenty three)

[0126] In the formula, X k|k X represents the state variable value of the system, which is the optimal value at that moment; k+1|k This indicates the result predicted using the previous state; Y k+1|k U represents the observed variable values ​​of the system; k Φ represents the control quantity of the system. k H represents the state transition matrix; k B represents the observation transition matrix; k W represents the system's control matrix. k V k Let U and R represent two uncorrelated Gaussian noises with zero mean, and their covariances be Q and R, respectively. In this embodiment, since the detected grid voltage signal has no control quantity, U... k=0. The system state value from time k to time k+1 is:

[0127] X k+1|k =Φ k X k|k (twenty four)

[0128] The error covariance matrix from time k to time k+1 is:

[0129]

[0130] In the formula, P k+1|k ——X k+1|k The corresponding error covariance; P k|k ——X k|k The corresponding error covariance,

[0131] P k|k =E[(X k|k -X k|k-1 (X) k|k -X k|k-1 ) T (26)

[0132] The Kalman filter gain at time k+1 is calculated as follows:

[0133] K k+1 =P k+1|k H T HP k+1|k H T +R) -1 (27)

[0134] By predicting the states at time k, and then measuring the current state, combining the measured and observed values, we can obtain the optimal state estimate X at time k+1. k+1|k+1

[0135] X k+1|k+1 =X k+1|k +K k+1 (Z k+1 -H k+1 X k+1|k (28)

[0136] To meet the operating conditions of the Kalman filter, X in state k+1 needs to be updated. k+1|k+1 Covariance:

[0137] P k+1|k+1 =(IK k+1 H k )P k+1|k (29)

[0138] Where I is the identity matrix. The initial filter value X is given based on the system state. 1|0 and P 1|0 After multiple filtering iterations and calculations, the value of the state variable X can be obtained. The calculation process is as follows: Figure 4 As shown.

[0139] To further verify the effectiveness of this embodiment, a simulation model based on an adaptive neural network was constructed, which includes a noise generation module, an RLS-adaline harmonic extraction module, a KF-adaline imbalance compensation module, a voltage regulation module, a gate signal generation module, and a DSOGI phase-locked loop module. Figure 5 The image shown is a simulation result diagram of one phase of the three-phase system.

[0140] To compare the adaptive neural network method with the i-based method based on the three-phase instantaneous reactive power theory... p -i q The harmonic detection method APF has different effects on harmonic suppression current compensation under noise conditions. A Gaussian white noise signal with a mean of 3 and a standard deviation of 2 is added to the load side current signal of the traditional APF. Figure 6 The THD diagrams of the current before and after adding noise to a traditional APF are shown. Similarly, using an adaptive neural network method, a noise signal with the same parameter settings is added at the same location, and the simulation results are as follows. Figure 7 As shown.

[0141] Through comparison, it is not difficult to see the traditional i p -i q In normal mode, the APF does not effectively eliminate noise interference, and its harmonic suppression and current compensation effects are affected by noise interference, potentially exceeding the standard limit of 5% for total harmonic distortion (THD), resulting in unsatisfactory performance. In contrast, the adaptive neural network method provides dual filtering effects of RLS-adaline and KF-adaline for noise signals, allowing the APF to function normally even under noisy conditions, and its current compensation function remains unaffected by fluctuations in noise signals.

[0142] Two APF control strategies were tested under voltage dip and three-phase imbalance conditions. A 20% voltage dip was set within a time period of 0.15s-0.35s. Figure 8 This is a simulation diagram of the grid-side voltage. The load side is configured with three parallel resistors with resistances of 10Ω, 100Ω, and 200Ω. Due to load imbalance, the three-phase currents will be unbalanced, as shown below. This imbalance will cause significant changes in the amplitude and phase of each phase current, resulting in them no longer being consistent. Figure 9 As shown.

[0143] The simulation results of the two APF methods are shown below. Figure 10 and Figure 11 The comparison shows that the current compensation function of traditional APF is significantly affected by voltage dips and three-phase imbalance, resulting in the harmonic suppression effect not meeting expectations. In such cases, traditional APF cannot function properly. However, the combination of adaptive neural network and APF can still maintain normal operation in harmonic suppression even under voltage dips and three-phase imbalance, and it is conducive to further reducing harmonic content, thus having a more positive effect on current compensation.

[0144] Simulation analysis demonstrates that the adaptive neural network-based PAPF active power filter exhibits better harmonic suppression under noisy conditions. The analysis focuses on comparing the APF using a strategy combining adaptive neural networks, the RLS algorithm, and Kalman filtering principles with the traditional i-type filter based on three-phase instantaneous reactive power theory. p -i q The operating simulation of the APF under noise conditions, voltage dips, and three-phase imbalance was conducted. Analysis shows that the adaptive neural network method exhibits higher reliability and accuracy under conditions of high noise interference and constantly changing voltage and load.

[0145] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for harmonic suppression and unbalanced load compensation using a parallel active power filter, characterized in that, Includes the following steps: Step 1: Model the noise in the power system and establish a noise model that includes multiple noise sources and their statistical characteristics; Step 2: Establish an LCL-type parallel active power filter system model, and construct an active power filter with noise interference based on the noise model. Step 3: Select an adaptive neural network algorithm, which includes a recursive least squares adaptive linear element and a Kalman filter adaptive linear element; Step 4: Based on the active power filter with noise interference, establish an active power filter control strategy using an adaptive neural network algorithm; Step 5: Based on the active power filter control strategy using the adaptive neural network algorithm, harmonic current detection and load imbalance current compensation are performed. The recursive least squares method is used to extract the harmonic current, and the Kalman filter adaptive linear element is used for imbalance compensation.

2. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 1, characterized in that, The noise includes Gaussian white noise and colored noise, and a noise model is established through analysis of actual power system data.

3. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 1, characterized in that, The adaptive neural network algorithm employs an adaptive learning algorithm to adjust the neural network parameters in real time and introduces an exponential weighting factor to adjust the network's weights on historical data.

4. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 3, characterized in that, In step 3, the active power filter control strategy is as follows: First, for the load current i L A Kalman filter module is used to extract the harmonic and reactive components in the load current to obtain the harmonic current reference signal i. r1 Meanwhile, for the DC-side voltage control loop, an adaptive neural network algorithm can be introduced to optimize the PI controller, enabling it to adjust according to the DC-side voltage V. dc With reference voltage V dcref The deviation is adaptively adjusted by the control parameters. After processing by a low-pass filter, and combined with the phase θ of the phase-locked loop output and the dq0 / abc transformation, the fundamental active current reference i is obtained. r2 ; will i r1 with i r2 The total reference current i is obtained by superposition. ref Then, with the compensation current i c The error current i is obtained by comparison. e After that, the hysteresis current controller based on i e A PWM signal is generated to drive the inverter to output compensation current, thereby achieving effective compensation for grid harmonics and reactive power.

5. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 4, characterized in that, The harmonic current detection specifically includes: The load current is sampled to obtain a discrete signal, which is then expanded using a Fourier series as follows: i L (n)=[a1,b1,...,a N ,b N ][cos(ωnT s ),sin(ωnT s ),...,cos(NωnT s ),sin(NωnT s )] T =W T (n)X(n) Among them, a k and b k T is the harmonic coefficient, ω is the angular frequency, and T is the harmonic coefficient. s The sampling period is k = 2, 3, ..., which represents the harmonic order. X(n) is the input vector of the model, and W(n) is the estimated value of the weight vector at time n. The discrete signal is used as the input vector X(n) of the adaptive neural network algorithm, and the weight vector W(n) is updated through the adaptive algorithm. The amplitude and phase of each harmonic are calculated based on the updated weight vector W(n), enabling the detection of any harmonic.

6. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 5, characterized in that, The adaptive algorithm is a recursive least squares method, and its cost function is: Where λ is the exponential weighting factor, d(i) is the desired response, X(i) is the input signal vector, and W(n) is the weight vector.

7. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 5, characterized in that, The update method for the weight vector W(n) is as follows: W(n)=Z(n)R -1 (n)=W(n-1)+k(n)ξ(n) Where R(n) represents the time mean of the autocorrelation matrix of the input signal vector X(i), k(n) is the Kalman gain vector, and ξ(n) = d(n) - X(n)W T (n-1) represents the prior error.

8. The method for harmonic suppression and unbalanced load compensation of a parallel active power filter according to claim 1, characterized in that, The load imbalance current compensation specifically includes: The three-phase asymmetrical signal is decomposed into two vertical components by coordinate transformation; An adaptive Kalman filter is used to remove noise and interference signals, and to accurately extract the fundamental and arbitrary harmonic currents.

9. A method for harmonic suppression and unbalanced load compensation in a parallel active power filter according to claim 8, characterized in that, In step 5, the method for removing noise and interference signals using adaptive Kalman filtering is as follows: First, construct the state variance and observation equations for the Kalman filter: X k+1|k =Φ k X k|k +B k U k +W k Y k+1|k =H k X k +V k In the formula, X k|k X represents the state variable value of the system, which is the optimal value at that moment; k+1|k This indicates the result predicted using the previous state; Y k+1|k U represents the observed variable values ​​of the system; k Φ represents the control quantity of the system. k H represents the state transition matrix; k B represents the observation transition matrix; k W represents the system's control matrix. k V k Let Q and R represent two uncorrelated Gaussian noises with zero mean, respectively; The system state values ​​from time k to time k+1 are: X k+1|k =Φ k X k|k The error covariance matrix from time k to time k+1 is: In the formula, P k+1|k X represents k+1|k The corresponding error covariance; P k|k X k|k The corresponding error covariance, P k|k =E[(X k|k -X k|k-1 )(X k|k -X k|k-1 ) T ] The Kalman filter gain at time k+1 is calculated as follows: K k+1 =P k+1|k H T (HP k+1|k H T +R) -1 By predicting the states at time k, and then measuring the current state, combining the measured and observed values, we can obtain the optimal state estimate X at time k+1. k+1|k+1 : X k+1|k+1 =X k+1|k +K k+1 (Z k+1 -H k+1 X k+1|k ) To meet the operating conditions of the Kalman filter, X in state k+1 needs to be updated. k+1|k+1 Covariance: P k+1|k+1 =(I-K k+1 H k )P k+1|k Where I is the identity matrix, and the initial filtering value X is given according to the system state. 1|0 and P 1|0 The value of the state variable X can be obtained by repeated filtering and calculation.

Images