Fundamental wave extraction method under frequency offset based on linear neural network

Abstract

The application discloses a fundamental wave extraction method under frequency offset based on a linear neural network, and has the advantages of high extraction accuracy, adaptability to frequency disturbance and the like. A traditional digital band-pass filter cannot change its center frequency after design, when the frequency and amplitude of an input signal are offset, the output signal of the traditional digital filter will be phase-shifted and attenuated to different degrees in amplitude, and even waveform distortion will occur, so the traditional digital band-pass filter needs to be redesigned, which is complicated and greatly increases the design cost. The band-pass digital filter against frequency offset based on the adaptive linear neural network can adaptively filter according to the frequency and amplitude changes of the input signal, and thus has strong robustness for frequency-amplitude random fluctuation signal detection, and can be applied to fundamental wave signal extraction of power grid voltage / current and the like.

Description

Technical Field

[0001] This invention relates to the field of digital filters, and more specifically to a frequency-shift resistant digital bandpass filter based on an adaptive linear neural network. Background Technology

[0002] Digital bandpass filters are widely used for extracting the fundamental frequency of grid voltage. Modifying a digital filter only requires program changes, without adding components, thus avoiding the influence of component errors and reducing costs. This solves the component limitations of analog filters. The frequency range of the grid voltage fundamental is 50±1Hz. However, due to limitations in center frequency and bandwidth, traditional digital bandpass filters experience attenuation when the grid voltage signal deviates from its fundamental frequency. In such cases, traditional digital bandpass filters cannot accurately extract the fundamental signal of the grid voltage. Summary of the Invention

[0003] In view of the above problems, this invention proposes a fundamental frequency extraction method based on frequency shift of linear neural network, which solves the problem that traditional digital filters cannot accurately and completely extract the fundamental frequency of voltage when the fundamental frequency of grid voltage shifts.

[0004] To address the aforementioned technical problems, this invention provides a fundamental frequency extraction method based on frequency shift using a linear neural network, comprising the following steps:

[0005] Step 1: Acquire the input signal u in real time in , input signal u in With a pair of sinusoidal reference signals x that are 90° out of phase IA (t) and x IB (t) Input to the first layer of the adaptive linear neural network (ADN-Ⅰ); where the reference signal x IA (t) and x IB The expression for (t) is:

[0006]

[0007] Where C is the amplitude and ω0 is the angular frequency. Its initial phase;

[0008] Step 2: The output of the first layer of the adaptive linear neural network is u f , will u f x serves as the first reference signal for the second layer of neurons. IIA (t), i.e., u f =x IIA (t); x IIA (t) is obtained through the integrator with x IIA(t) Signal S with a phase difference of 90° s ;

[0009] Step 3: Obtain signal x using an amplitude extraction algorithm. IIA (t) and x IIB The amplitude of (t) is calculated by subtracting the above signals and inputting the result into the PI controller, which then outputs the signal S. s The signal x is obtained by multiplying it by the output of the PI controller. IIB (t); x IIA (t) and x IIB (t) have equal amplitudes and a 90° phase difference;

[0010] Step 4: Set x IIA (t) and x IIB (t) serves as a set of reference signals and input signals u in Simultaneously, the signal is input to the second layer of the adaptive linear neural network (ADN-II), and the output is the filtered signal u. o .

[0011] In a preferred embodiment, the training method for the adaptive linear neural network is a least squares algorithm, specifically including:

[0012] Input signal u in With bandpass output signal u f The difference is used to obtain the error signal ε(t). This error signal ε(t) is then compared with two sinusoidal reference signals x. IA (t) and x IB After multiplying by (t), multiply by the gain 2μ to generate the intermediate term 2με(t)x IA (t) and 2με(t)x IB (t);

[0013] Let the intermediate term 2με(t)x IA (t) and 2με(t)x IB (t) and the weight of the previous sampling period A (t) and B Add (t) together to get the new weights. A (t+1) and B (t+1), its expression is

[0014]

[0015] Will A (t+1) and B (t+1) and the corresponding sinusoidal signal x IA (t) and x IB Multiplying (t) and adding the products yields the bandpass output signal u. o Its expression is

[0016] u o =λ A (t+1)x IA (t)+λ B (t+1)x IB (t) (3)

[0017] Bandpass output signal u o Frequency and input signal u in The frequencies are the same.

[0018] In a preferred embodiment, step 3 specifically includes:

[0019] x IIA (t) and x IIB (t) is squared and then passed through a low-pass filter LPF. The output of the low-pass filter is x. IIA (t) and x IIB The amplitude M of (t) A and M B The difference is then input into the PI controller, and the transfer function G of the PI controller is... v The expression for (s) is:

[0020]

[0021] The above steps yield another reference signal and x for the second layer of the adaptive linear neural network. IIB (t), whose specific expression is:

[0022]

[0023] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0024] This invention proposes a frequency-offset-resistant digital bandpass filter based on an adaptive linear neural network. When the grid voltage frequency fluctuates, it can accurately and completely extract the fundamental frequency signal of the grid voltage signal, improving the ability of grid-connected converters and other devices that require real-time acquisition of the grid voltage fundamental signal to adapt to the grid's operating conditions. The algorithm is relatively simple and easy to implement, containing only two layers of neural network, an integrator, and a PI closed loop to achieve real-time acquisition of the grid voltage fundamental signal without redundant calculations. The algorithm is robust to grid voltage amplitude fluctuations, accurately extracting the fundamental frequency signal even when both the grid voltage frequency and amplitude fluctuate uncertainly. Attached Figure Description

[0025] Figure 1 This is a block diagram of the construction of a frequency-shift-resistant digital bandpass filter based on an adaptive linear neural network proposed in this invention.

[0026] Figure 2 This is a block diagram of the adaptive linear neural network algorithm used.

[0027] Figure 3 This is a diagram illustrating the amplitude extraction process of the reference signal.

[0028] Figure 4 This is a schematic diagram illustrating the application of this method in controlling an active filter after extracting the fundamental voltage of the power grid.

[0029] Figure 5 These are a pair of sinusoidal signals input to the second layer of the neural network;

[0030] Figure 6 When the simulated power grid frequency is 51Hz, the simulated power grid voltage signal U o And the extracted fundamental frequency. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," "top / bottom," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0033] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installed", "equipped", "sleeved / connected", "connected", etc., should be interpreted broadly. For example, "connection" can be a wall-mounted connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0034] This embodiment provides a method for extracting the fundamental frequency of a power grid voltage signal based on a linear neural network, including a two-layer adaptive linear neural network, an amplitude extraction algorithm, an integrator, and a PI control loop. The training method for the adaptive linear neural network is the least squares algorithm. The integration, amplitude extraction algorithm, and PI control loop are used to generate a pair of sinusoidal signals with the same frequency and amplitude.

[0035] Figure 1 This is a block diagram of the proposed adaptive linear neural network-based anti-frequency offset digital bandpass filter, which is specifically constructed as follows:

[0036] (1) Real-time acquisition of input signal u in , input signal u in With a pair of sinusoidal reference signals x that are 90° out of phase IA (t) and x IB (t) Input to the first layer of the adaptive linear neural network (ADN-I). Where the reference signal x... IA (t) and x IB The expression for (t) is:

[0037]

[0038] Where C is the amplitude and ω0 is the angular frequency. Its initial phase.

[0039] (2) The output of the first layer of the adaptive linear neural network is u f , will u f x serves as the first reference signal for the second layer of neurons. IIA (t), i.e., u f =x IIA (t). x IIA (t) is obtained through the integrator with x IIA (t) Signal S with a phase difference of 90° s .

[0040] (3) Signal x IIA (t) and x IIB (t) is squared and then passed through a low-pass filter (LPF) to obtain the amplitude. The output of the low-pass filter is x. IIA (t) and x IIB The amplitude M of (t) A and M B The amplitude M A and M B The difference is then input to the PI controller, where the PI controller's transfer function G... v The expression for (s) is:

[0041]

[0042] Signal S s The signal x is obtained by multiplying it by the output of the PI controller. IIB (t), whose specific expression is:

[0043]

[0044] At this time x IIA (t) and x IIB (t) have equal amplitudes and a 90° phase difference.

[0045] (4) x IIA (t) and x IIB (t) serves as a set of reference signals and input signals u in Simultaneously, the signal is input to the second layer of the adaptive linear neural network (ADN-II), and the output is the filtered signal u. o .

[0046] Figure 2 for Figure 1 The algorithm flowchart of a linear neural network (ADN) is as follows:

[0047] Input signal u in With bandpass output signal u f The difference is used to obtain the error signal ε(t), and the error signal ε(t) is then compared with the two sinusoidal reference signals x as described in claim 1. IA (t) and x IB After multiplying by (t), multiply by the gain 2μ to generate the intermediate term 2με(t)x IA (t) and 2με(t)x IB (t);

[0048] Let the intermediate term 2με(t)x IA (t) and 2με(t)x IB (t) and the weight λ of the previous sampling period A (t) and λ B Adding (t) together yields a new weight λ. A (t+1) and λ B (t+1), its expression is

[0049]

[0050] λ A (t+1) and λ B (t+1) and the corresponding sinusoidal signal x IA (t) and x IB Multiplying (t) and adding the products yields the bandpass output signal u. o Its expression is

[0051] u o =λ A (t+1)x IA (t)+λ B (t+1)x IB (t)

[0052] At this time, the bandpass output signal u o Frequency and input signal u in The frequencies are the same.

[0053] Figure 3 for Figure 1 Reference signal x IIA (t) and x IIB The amplitude extraction process of (t) will extract x IIA (t) and x IIB (t) is squared and then passed through a low-pass filter LPF. The output of the low-pass filter is x. IIA (t) and x IIB The amplitude M of (t) A and M B ,

[0054] like Figure 4 As shown, taking distributed power generation grid connection as an example, the grid voltage signal u is first collected. in Then, using the proposed frequency-offset-resistant digital bandpass filter based on an adaptive linear neural network, the fundamental voltage signal u of the grid in the distributed generation is extracted. o The fundamental frequency signal u o After calculation, a PWM waveform is generated by the controller to control the switching transistors of the active power filter. Once operational, the active power filter suppresses secondary current ripple in the bus, thereby improving the power quality of the power branch. Wherein, U i For input DC power, S a S b S1, S2, S3, and S4 are switching transistors, and L... r For the energy storage inductor of the active filter, C and C r For the active filter capacitor, L f and C f These are the inverter filter inductor and capacitor, respectively.

[0055] Select parameters and perform circuit simulation. The design parameters for this embodiment are shown in the table, including the distributed power supply voltage U. i 250V, capacitor C r With C = 150μF, inductance L r 120μH, switching frequency f s The frequency is 10kHz, and the learning rate μ is 1×10. -4 L f 2.5mH, Cf It is 10μF.

[0056]

[0057] Figure 5 It is the sinusoidal reference signal x input to the second layer of the neural network. IIA (t), x IIB The simulation results (t) show that the two sinusoidal signals have a 90° phase difference and equal amplitude, which meets the ADN input conditions.

[0058] Figure 6 (a) and (b) are the simulated grid voltage signal u when the simulated grid frequency is 51Hz. in and the extracted fundamental frequency u o The simulation results are shown in Figure (a). Figure (b) is the waveform diagram, and Figure (a) is the Fourier analysis diagram. Combining the two figures, it can be seen that when the grid frequency is relatively high, the fundamental frequency extraction effect of the simulated grid voltage signal is good.

[0059] according to Figure 6 When the grid frequency deviates within the allowable range, the fundamental frequency extraction method based on linear neural networks proposed in this invention can accurately extract the fundamental frequency of the grid voltage signal.

[0060] The above description is merely a preferred embodiment of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention by those skilled in the art within the scope of the technology disclosed in the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A fundamental frequency extraction method based on frequency shift using a linear neural network, characterized in that... Includes the following steps: Step 1: Acquire the input signal u in real time in , input signal u in With a pair of sinusoidal reference signals x that are 90° out of phase IA (t) and x IB (t) Input to the first layer of the adaptive linear neural network (ADN-Ⅰ); where the reference signal x IA (t) and x IB The expression for (t) is: Where C is the amplitude and ω0 is the angular frequency. Its initial phase; Step 2: The output of the first layer of the adaptive linear neural network is u f , will u f x serves as the first reference signal for the second layer of neurons. IIA (t), i.e., u f =x IIA (t); x IIA (t) is obtained through the integrator with x IIA (t) Signal S with a phase difference of 90° s ; Step 3: Obtain signal x using an amplitude extraction algorithm. IIA (t) and x IIB The amplitude of (t) is calculated by subtracting the above signals and inputting the result into the PI controller, which then outputs the signal S. s The signal x is obtained by multiplying it by the output of the PI controller. IIB (t); x IIA (t) and x IIB (t) have equal amplitudes and a 90° phase difference; Step 4: Set x IIA (t) and x IIB (t) serves as a set of reference signals and input signals u in Simultaneously, the signal is input to the second layer of the adaptive linear neural network (ADN-II), and the output is the filtered signal u. o .

2. The fundamental frequency extraction method based on frequency shift using a linear neural network according to claim 1, characterized in that, The training method for the adaptive linear neural network is the least squares algorithm, specifically including: Input signal u in With bandpass output signal u f The difference is used to obtain the error signal ε(t). This error signal ε(t) is then compared with two sinusoidal reference signals x. IA (t) and x IB After multiplying by (t), multiply by the gain 2μ to generate the intermediate term 2με(t)x IA (t) and 2με(t)x IB (t); Let the intermediate term 2με(t)x IA (t) and 2με(t)x IB (t) and the weight λ of the previous sampling period A (t) and λ B Adding (t) together yields a new weight λ. A (t+1) and λ B (t+1), its expression is λ A (t+1) and λ B (t+1) and the corresponding sinusoidal signal x IA (t) and x IB Multiplying (t) and adding the products yields the bandpass output signal u. o Its expression is u o =λ A (t+1)x IA (t)+λ B (t+1)x IB (t) (3) Bandpass output signal u o Frequency and input signal u in The frequencies are the same.

3. The fundamental frequency extraction method based on frequency shift using a linear neural network according to claim 1, characterized in that, Step 3 specifically includes: x IIA (t) and x IIB (t) is squared and then passed through a low-pass filter LPF. The output of the low-pass filter is x. IIA (t) and x IIB The amplitude M of (t) A and M B The difference is then input into the PI controller, and the transfer function G of the PI controller is... v The expression for (s) is: The above steps yield another reference signal and x for the second layer of the adaptive linear neural network. IIB (t), whose specific expression is:

Images