Abstract: Based on the instantaneous reactive power harmonic dynamic detection method, the detection accuracy and real-time performance are high. This paper elucidates the basic principle of the power grid harmonic dynamic detection method based on the combination of instantaneous reactive power and neural network control, analyzes the harmonic detection method based on instantaneous reactive power, introduces neural network to improve accuracy and real-time performance during load changes, and conducts simulation experiments in conjunction with active power filters to observe the dynamic detection of harmonics. Keywords: Instantaneous reactive power, neural network, active power filter. Harmonic filtering of the power grid is of great significance for maintaining the power supply quality of the power grid, especially since critical power supply equipment is greatly affected by harmonics and requires high-quality power supply. To ensure the efficiency and accuracy of filtering in the power grid, it is necessary to accurately determine the order and amplitude of harmonics. Therefore, the key issue lies in the detection and judgment of harmonics, especially when the power grid load fluctuates greatly, the rapid and accurate detection of harmonics is of great value. I. Traditional Detection Methods of Power Grid Harmonics There are many traditional methods for harmonic detection, among which the focus is on the detection of current harmonics. In summary, the main harmonic detection methods are as follows: 1) Fourier analysis: This method mainly performs Fourier analysis on the detected current and voltage signals to decompose the spectrum of higher-order harmonics, and finally calculates the parameters of the filter device to be added; 2) Wavelet analysis: Essentially an extension of the Fourier analysis method, but with more distinctive features in spectrum extraction, and the calculation method is similar to the Fourier method; 3) Zero-sequence current analysis: Harmonic analysis is performed by detecting zero-sequence current and instantaneous reactive power; 4) PI-controlled filter method: This method uses PI control technology combined with filtering technology to detect the content of specific harmonics; 5) Equal power method: This method determines grid harmonics by detecting the average voltage of the DC-side capacitor; 6) Other harmonic detection technologies based on advanced control theory, such as neural network theory. II. BP Neural Network Technology The BP (Backpropagation) network is a multi-layer feedforward network with unidirectional propagation. It is a neural network that uses error backpropagation training algorithms. It solves the problem of learning the connection weights of hidden units in multi-layer networks. For the system input M and output L, the BP neural network can be viewed as a mapping from M-dimensional Euclidean space to L-dimensional Euclidean space. A major characteristic of this mapping is its high nonlinearity, making it widely applicable. Its structure is shown in Figure 2.1, including: 1) Data compression technology: compressed storage of image data and extraction of image features; 2) Pattern recognition technology: used for information recognition such as text and language, particularly suitable for feature judgment; 3) Function approximation: used for modeling and control of complex industrial systems, especially suitable for strongly coupled and nonlinear systems. [align=center] Figure 2.1 BP Neural Network Structure[/align] The consideration of using BP network technology to detect harmonics is mainly based on the characteristics of dynamic harmonics in power systems. Power grids have many harmonic orders, and a key piece of equipment in the past was a parallel filtering detection device. The principle block diagram of this device is shown in Figure 2.2. Its real-time performance and accuracy are quite good, but due to the large number of filters, the entire device is very complex, with a relatively low failure rate and reliability. Furthermore, it cannot be functionally modified as needed, resulting in poor adaptability. [align=center] Figure 2.2 Block diagram of parallel filter type harmonic detection device[/align] Comparing Figure 2.1 and Figure 2.2, we can consider using a BP network to replace the hardware circuit in Figure 2.2 and use the learning method of neural network to replace it. The input signal of the actual power grid can be regarded as the signal to be measured, and a series of sampling and self-learning are performed. The output of the BP network is the amplitude of the harmonic current signal to be detected. These current signals are an important part of the command current signal required by the compensation current generation circuit of the active power filter. Based on this idea, we can consider using a BP neural network to detect harmonic current. In this paper, we will conduct a simulation study. III. BP Neural Network and Power Grid Harmonic Detection 3.1 Harmonic Characteristics of Power System In the power system, there is a quantitative index to judge the magnitude of harmonics. The total harmonic distortion (THD) is commonly used to measure the quality of voltage and current. (3.1) In the above formula, Ah represents the amplitude of the h-th harmonic; A1 represents the amplitude of the fundamental wave I1(t). Active power filters essentially reduce the total harmonic distortion (THD) of a power system to achieve the national power quality standards. The characteristics of harmonics in a power system can be summarized as follows: 1) Odd-order harmonics are the main components of harmonics in the power grid, and their amplitudes rarely exceed 70% of the fundamental frequency amplitude. Higher-order harmonics have relatively small amplitudes, so filtering of odd-order harmonics is a primary focus; 2) The distorted current in a power system contains many odd-order harmonic components, but several, including the 5th, 7th, 11th, and 13th harmonics, severely impact power quality; 3) In practical operation, the detection stage of an active power filter first detects these harmonics, then controls the compensation and filtering circuits to filter out these more influential harmonics, thereby significantly reducing the harmonic content and thus substantially lowering the THD of the power system. Based on the above analysis, this paper proposes the following detection scheme, as shown in Figure 3.1. This paper describes a BP network with 128 input nodes and 3 output nodes to implement an active power filter for harmonic current detection. The expected outputs of the output layer are the amplitudes of the 5th, 7th, 11th, and 13th harmonics. The input of the input layer consists of 128 sampled values ​​of the distorted wave over one fundamental frequency period (the sampling period is 0.02/127s when the fundamental frequency is 50Hz). [align=center] Figure 3.1 Harmonic Current Detection Scheme Based on BP Network[/align] 3.2 Implementation of BP Harmonic Detection Network A key step in BP network design is the hidden layer design, including the number and correspondence of hidden layers. If each output harmonic in the BP network is connected to the same hidden layer, the connection weights between the output and hidden layers can provide the optimal values ​​for harmonic filtering. However, this places a heavy memory burden on the entire system, reducing system efficiency and potentially causing mutual interference. However, if each harmonic corresponds to a hidden layer, i.e., each has its own hidden layer, and each hidden layer is only responsible for memorizing the implicit mapping relationship of its corresponding harmonic, the problem of mutual influence between harmonics carried by a hidden layer will be better overcome. In this paper, we use a heuristic learning algorithm, namely the Momentum Backpropagation (MOBP) algorithm, to train the multilayer feedforward neural network shown in Figure 3.1. This algorithm uses a momentum adjustment strategy, which can significantly reduce the network's sensitivity to local details of the error surface and effectively suppress the network from getting trapped in local minima. In network training, MOBP uses equations (3.1) and (3.2) to modify the weights and thresholds. In equations (3.1) and (3.2): is the learning rate, γ is the momentum coefficient. m refers to the m-th layer of the network, is the sensitivity of the approximate mean square error to the input of the m-th layer, b is the network weight, w is the network threshold, Y is the output of the network output layer, and T is the matrix transpose. IV. Simulation Study Simulation was conducted using the combined control system described above. Current detection was applied to an active power filter to observe the waveform. The neural network output layer used both the nonlinear activation function logsig and the linear activation function pureline to compare the detection results. The average THD calculated based on the simulation values ​​is shown in Table 1: Table 1 Average THD values ​​before and after compensation for test samples. The current waveform of the simulated filter circuit was detected using a trained BP network. The power supply current waveform and the filter injection current waveform are shown in Figures 4.1 to 4.5, respectively. [align=center] Figure 4.1 Power supply current waveform Figure 4.2 Filter injection current waveform Figure 4.3 System current without filter device Figure 4.4 System current after the hybrid active filter is applied Figure 4.5 Comparison of system current spectrum under different conditions[/align] Table 1 and Figures 4.1 and 4.2 show that the total harmonic distortion (THD) is significantly reduced after harmonic compensation, indicating that this harmonic current detection method can effectively detect and compensate for harmonic currents. Conclusion This paper starts with instantaneous reactive power and combines improvements to the BP network model and detection method for control. By introducing a neural network to improve accuracy and real-time performance during load abrupt changes, a harmonic detection method with good accuracy and real-time performance is obtained. The detection method is then applied to the detection stage using an active filter. Simulation results show that this method provides an effective means and tool for analyzing and designing dynamic harmonic detection. References: [1] Yuan Zengren. Artificial Neural Networks and Their Applications. Beijing: Tsinghua University Press, 1999. [2] Wang Qun, Wu Ning. A Method for Measuring Power Harmonics Based on Artificial Neural Networks. Automation of Electric Power Systems, 1998, 22(11): 35-39. [3] Liu Jianbao, Chen Wei, Zhao Luhuai. A Novel Neural Network Control Method for Active Power Filters [J]. Modern Electronics Technology, 2004, (3): 6-11. [4] He Yihong, Zhuo Fang, Zhou Xin, et al. A Discrete Filtering Method Based on Instantaneous Reactive Power Theory Harmonic Detection [J]. New Technologies for Electrical Engineering and Energy, 2002, 21(4): 13-16. [5] Jiang Junmin, Liu Huikang. Application of Parallel Active Filter Based on PEBB in Vertical Roll Mill [J]. Electric Power Automation Equipment, 2006, 11: 13-16.