Power intelligent energy-saving compensation control method and system for terahertz transmission

Abstract

This application discloses a power intelligent energy-saving compensation control method and system based on terahertz transmission, relating to the field of power intelligent energy-saving compensation control. By collecting and fusing noise-reduced electrical parameters and environmental status data at the end of the distribution network, a multimodal sensing data stream is generated. This stream is then transmitted via high-speed terahertz transmission to achieve low-latency and high-reliability transmission. Based on the sensing data, load trend prediction is completed, and a voltage and power factor collaborative compensation command with voltage deviation priority adjustment as a constraint is generated. Simultaneously, a deep reinforcement learning algorithm is used to iteratively optimize the load prediction model and compensation strategy, and zero-crossing switching actions are executed to achieve precise compensation, ensuring the safe and stable operation of equipment. This achieves precise voltage control and reactive power compensation collaborative optimization at the end of the distribution network, improving power supply reliability and reducing line losses. It is applicable to various energy-saving compensation scenarios at the end of the distribution network.

Description

Technical Field

[0001] This application relates to the field of intelligent power energy-saving compensation control, and in particular to a method and system for intelligent power energy-saving compensation control using terahertz transmission. Background Technology

[0002] As the final link in the power system, the distribution network directly supplies power to users. Its operational stability and power quality directly affect users' electricity experience and the energy-saving effect of the power system. Currently, the distribution network terminal generally suffers from problems such as untimely voltage deviation control, low reactive power compensation efficiency, data transmission lag, and insufficient operational stability of compensation equipment, which seriously restrict the energy-saving level and power supply reliability of the distribution network.

[0003] Existing technologies for data acquisition at the end of traditional distribution networks mostly use single-type sensors, resulting in limited data dimensions. Furthermore, data is often transmitted via traditional wireless frequency bands or wired methods, which suffers from drawbacks such as slow transmission speed, insufficient bandwidth, and high latency. This fails to meet the timeliness requirements of real-time compensation control for data transmission, leading to untimely issuance of compensation commands and difficulty in quickly responding to changes in end-load and voltage fluctuations.

[0004] Existing technologies for load forecasting and compensation control mostly employ traditional load forecasting models, which are difficult to accurately capture load change patterns at different time scales and do not fully consider the impact of environmental factors on the load, resulting in low forecast accuracy. At the same time, compensation control is mostly centered on reactive power compensation and does not prioritize voltage deviation regulation, which makes it easy for the terminal voltage to deviate from the qualified range. This makes it impossible to achieve synergistic optimization of voltage regulation and reactive power compensation, resulting in poor energy-saving effects. Summary of the Invention

[0005] The purpose of this application is to provide a power intelligent energy-saving compensation control method for terahertz transmission to solve the problems mentioned in the background art, including: The electrical parameters and environmental status data of the power supply branch line of the distribution network are collected, and multimodal data fusion and noise reduction processing is used to generate a multimodal sensing data stream that characterizes the operating conditions of the end. The multimodal sensing data stream is transmitted over a preset distance using a wireless transmission link in the terahertz band. Based on multimodal sensing data stream, the load forecasting model is used to predict the trend of the end load and generate a voltage and power factor collaborative compensation command with voltage deviation priority adjustment as the constraint. Based on the trend prediction results and voltage and power factor collaborative compensation instructions, a deep reinforcement learning algorithm is used to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation instructions to obtain the optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy. Based on the voltage and power factor collaborative compensation command, the contactless switching unit based on power electronic devices is controlled to perform zero-crossing switching action in order to prioritize the adjustment of voltage deviation and use the remaining capacity to compensate for reactive power. When a preset fault condition is detected, a protection action is performed, and thermal management is carried out based on the phase change heat dissipation structure and intelligent air cooling linkage.

[0006] In a preferred embodiment of this scheme, the acquisition of electrical parameters and environmental status data at the end of the power distribution network branch line, followed by multimodal data fusion and denoising processing to generate a multimodal sensing data stream characterizing the terminal operating conditions, includes: By deploying voltage transformers, current transformers, temperature and humidity sensors, and image acquisition devices at the end of the power distribution network branch lines, electrical parameters, environmental status data, and equipment appearance images are collected synchronously with a unified time base. The electrical parameters include the effective value of voltage, the effective value of current, active power, reactive power, and power factor. The environmental status data includes ambient temperature and ambient humidity. The collected electrical parameters and environmental status data are time-series aligned according to timestamps to obtain aligned electrical time-series data and environmental time-series data; The adaptive Kalman filter algorithm is used to remove impulse noise and random fluctuation interference from the electrical timing data to obtain clean electrical timing data; The environmental time series data and the device appearance image are denoised using a nonlocal mean denoising algorithm, which combines spatial and temporal domains to obtain clean environmental time series data and clean device image data. The clean electrical time-series data, clean environmental time-series data, and clean equipment image data are spliced ​​together at the feature layer, and an end position identifier is injected to generate the multimodal sensing data stream.

[0007] In a preferred embodiment of this scheme, transmitting the multimodal sensing data stream over a preset distance range via a wireless transmission link in the terahertz band includes: The multimodal sensing data stream is segmented into data frames and prioritized for transmission. Sensing data reflecting voltage drops and fault symptoms at the end are marked as high-priority transmission frames. The photonics-assisted terahertz modulation method is adopted, which uses two laser beams with a frequency interval in the terahertz band to beat in a photoconductive antenna to generate a terahertz carrier. The high-priority transmission frame and other transmission frames are respectively mapped onto a constellation diagram of quadrature amplitude modulation. After digital-to-analog conversion, they are loaded onto the terahertz carrier to generate a terahertz modulation signal. The terahertz modulated signal is directionally transmitted through a high-gain antenna with beamforming function, and after reflection or forwarding by a terahertz relay node, a terahertz wireless transmission link is constructed on a preset transmission path. The multimodal sensing data stream is recovered by downconverting and digitally demodulating the received terahertz modulated signal using a terahertz coherent detection receiver.

[0008] In a preferred embodiment of this scheme, the step of performing trend prediction of end-point load based on the multimodal sensing data stream using a load forecasting model includes: Active power time series, reactive power time series, ambient temperature series and humidity series are extracted from the multimodal sensing data stream to form an input feature vector; The input feature vector is input into a pre-constructed temporal convolutional attention network model, which includes multiple dilated causal convolutional layers and a channel attention mechanism. The dilated causal convolutional layers are used to capture load change patterns at different time scales, and the channel attention mechanism is used to adaptively assess the contribution of weighted electrical parameters and environmental state parameters to load changes. The model outputs a predicted active power sequence and a predicted reactive power sequence for a future preset time period, which serve as the end-load trend prediction result. By combining the predicted end-load trend with the current measured end-voltage value, the expected voltage deviation range and the expected reactive power demand range are calculated.

[0009] In a preferred embodiment of this scheme, the step of executing the command to generate voltage and power factor collaborative compensation with voltage deviation priority adjustment as a constraint includes: Based on the line impedance parameters at the end of the power supply branch of the distribution network and the predicted load trend at the end, a voltage and reactive power sensitivity matrix is ​​constructed. Based on the voltage and reactive power sensitivity matrix, a multi-objective optimization model is established with the expected voltage deviation range as the primary constraint and the expected reactive power demand range as the secondary constraint. The first priority objective of the multi-objective optimization is to maintain the end voltage deviation within the preset qualified voltage range during the prediction period, and the second priority objective is to maximize the average power factor after satisfying the first priority objective. Solving the multi-objective optimization model yields the minimum reactive power compensation capacity requirement under the condition of voltage qualification and the additional reactive power compensation allocation value based on the remaining compensation capacity. Based on the minimum reactive power compensation capacity requirement and the additional reactive power compensation allocation value, a voltage and power factor coordinated compensation instruction is generated, which includes a capacitor bank switching scheme or a static var generator reactive power output instruction. In each control cycle, the capacitor bank switching scheme or reactive power output instruction is executed first to meet the minimum reactive power compensation capacity that meets the voltage qualification, and then the remaining capacity is used to improve the power factor.

[0010] In a preferred embodiment of this scheme, based on the trend prediction results and the voltage and power factor collaborative compensation command, a deep reinforcement learning algorithm is used to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation command to obtain optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy, including: A deep deterministic policy gradient proxy model is constructed, which includes an online policy network, a target policy network, an online value network, and a target value network. The state space consists of historical multimodal sensing data, structural hyperparameters of the edge-side load prediction model, and weight coefficients in the voltage and power factor collaborative compensation command generation strategy. The action space is the adjustment amount for the hyperparameters and weight coefficients. The reward function is determined based on the changes in the voltage qualification rate, average power factor, and line loss rate at the end of the distribution network. An experience replay pool is constructed using uploaded long-term historical operational data. The online value network is trained by randomly sampling batches of data from the experience replay pool, and the target value network parameters are synchronized using a soft update method. Based on the trained deep deterministic policy gradient proxy model, the optimized load prediction model hyperparameters and compensation policy weight coefficients are output.

[0011] In a preferred embodiment of this scheme, controlling the contactless switching unit based on power electronic devices to perform zero-crossing switching based on the voltage and power factor collaborative compensation command, so as to prioritize adjusting voltage deviation and utilize remaining capacity to compensate reactive power, includes: The voltage and power factor collaborative compensation command is analyzed to obtain the capacitor branch numbers that need to be put into or cut into the current control cycle and the corresponding switching direction. For each capacitor branch to be operated, the phase of the grid voltage is tracked in real time through a phase-locked loop. When the voltage difference across the bidirectional thyristor to be switched is detected to be within a preset window, a trigger pulse sequence is generated. The trigger pulse sequence is amplified and applied to the gate of the contactless switching unit composed of anti-parallel thyristors, so that the thyristors are turned on or off in a zero voltage or zero current state, and the corresponding capacitor branch is connected to the distribution network or disconnected from the distribution network. After the capacitor branch is put into operation, the voltage amplitude and power factor at the end are monitored in real time. If the voltage at the end is still lower than the preset voltage range lower limit, more capacitor branches will be put into operation until the voltage recovers to the qualified range or all available branches have been put into operation. If the terminal voltage is qualified, the remaining unused capacitor branches will be used to continue to improve the power factor until the power factor reaches the set target or the remaining available capacity is exhausted. After completing all switching actions in this control cycle, the updated effective value of the terminal voltage, power factor, and effective value of the current in each capacitor branch are read from the multimodal sensing data stream. The total reactive power compensation capacity actually switched in this cycle is calculated, and the total reactive power compensation capacity actually switched, the updated effective value of the terminal voltage, the updated power factor, and the timestamp of the action completion time are encapsulated into a compensation execution feedback message.

[0012] In a preferred embodiment of this solution, the step of performing a protection action upon detecting a preset fault condition, and thermal management of the intelligent arc-free compensation execution step based on the phase-change heat dissipation structure and intelligent air-cooling linkage, includes: The surface temperature of the thyristor heat sink and the temperature of the capacitor casing are collected by a temperature sensor array, the instantaneous value of the current in each branch is collected by a Hall current sensor, and the harmonic content of the bus voltage is collected by a voltage transformer. The instantaneous values ​​of temperature, current and voltage harmonic content are compared with preset over-temperature threshold, over-current threshold and harmonic distortion rate threshold. When any parameter exceeds the corresponding threshold and the duration reaches the set delay, a fault is determined and a protection command to trip or disconnect part of the capacitor branch is output. In the absence of faults, the temperature distribution data collected by the temperature sensor array is input to the phase change heat dissipation control logic to maintain the temperature of the thyristor and capacitor within the specified operating range.

[0013] To achieve the above objectives, this application also provides the following technical solution: a terahertz transmission-based intelligent energy-saving compensation control system, comprising the following modules: End-point multi-source sensing module: used to collect electrical parameters and environmental status data at the end of the power supply branch of the distribution network, and generate a multi-modal sensing data stream characterizing the end-point operating conditions through multi-modal data fusion and noise reduction processing; Terahertz high-speed transmission module: used to transmit the multimodal sensing data stream within a preset distance range via a wireless transmission link in the terahertz frequency band; Edge intelligent collaboration module: Based on multimodal sensing data stream, it is used to predict the trend of end load through load prediction model and generate voltage and power factor collaborative compensation instructions with voltage deviation priority adjustment as the constraint. Cloud-based optimization module: Based on the trend prediction results and voltage and power factor collaborative compensation instructions, the module uses a deep reinforcement learning algorithm to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation instructions, thereby obtaining the optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy. Intelligent arc-free compensation execution module: used to control the contactless switching unit based on power electronic devices to perform zero-crossing switching action based on the voltage and power factor collaborative compensation command, so as to prioritize the adjustment of voltage deviation and use the remaining capacity to compensate reactive power. Integrated protection and heat dissipation module: used to perform protection actions when preset fault conditions are detected, and to perform thermal management based on phase change heat dissipation structure and intelligent air cooling linkage.

[0014] Compared with the prior art, the beneficial effects of this application are: This application achieves synchronous acquisition of multi-dimensional data at the distribution network end through a multi-source sensing module at the end of the network. It combines adaptive Kalman filtering and non-local mean denoising algorithms to complete data fusion and denoising, ensuring data authenticity and usability. A low-latency, high-reliability wireless transmission link is constructed using a terahertz high-speed transmission module to address the shortcomings of traditional transmission methods and ensure real-time data and command interaction. Simultaneously, a temporal convolutional attention network model is constructed to improve load forecasting accuracy. A multi-objective optimization model is established with voltage deviation priority adjustment as a constraint to achieve synergy between voltage regulation and reactive power compensation, effectively reducing distribution network line losses and improving energy efficiency. Through a deep deterministic strategy gradient proxy model on a cloud-based optimization platform, dynamic iterative optimization of the load forecasting model and compensation strategy is achieved, adapting to dynamic operating conditions at the end of the network and avoiding the problem of declining adaptability over long-term operation.

[0015] This application significantly improves the safety, stability, and operational flexibility of compensation equipment. The intelligent arc-free compensation execution module employs an anti-parallel thyristor contactless switching unit, using zero-crossing switching to avoid arcing, protecting equipment safety and reducing grid impact. The integrated protection and heat dissipation module enables timely fault handling through real-time multi-parameter monitoring, combined with a phase-change cooling and intelligent air-cooling linkage mechanism to maintain normal operating temperature and extend equipment lifespan. The edge intelligent collaboration module enables rapid response in localized areas, reducing reliance on cloud platforms. The rationally designed structure of each module allows for flexible parameter adjustment based on actual needs, adapting to various distribution network end-user scenarios. Ultimately, it achieves the goals of precise voltage control, reasonable reactive power compensation, and safe and stable equipment operation at the distribution network end, significantly improving power supply reliability and energy efficiency, and possesses significant engineering application value and promising prospects for widespread adoption. Attached Figure Description

[0016] The present application will be further illustrated with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present application. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a power intelligent energy-saving compensation control method for terahertz transmission in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a terahertz transmission power intelligent energy-saving compensation control system in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Example 1, please refer to Figure 1 This application provides a power intelligent energy-saving compensation control method for terahertz transmission, the method comprising: It is used to collect electrical parameters and environmental status data at the end of the power supply branch of the distribution network, and generate a multimodal sensing data stream that characterizes the operating conditions of the end through multimodal data fusion and noise reduction processing; Furthermore, including: First, multi-source data is collected synchronously. Voltage transformers, current transformers, temperature and humidity sensors, and image acquisition devices are deployed at the end of the power distribution network branch lines. All sensors operate synchronously using a unified time base to ensure that the collected data are aligned in the time dimension. Specifically, voltage transformers are used to collect the effective value of the voltage at the end, and current transformers are used to collect the effective value of the current at the end. Combining voltage and current data, active power, reactive power, and power factor can be further calculated. Temperature and humidity sensors are used to collect the ambient temperature and humidity around the end equipment as environmental factor data affecting load changes. Image acquisition devices are used to collect external images of the end compensation equipment and lines to help determine the operating status of the equipment.

[0021] After data acquisition is completed, time-series alignment is performed first. The acquired electrical parameters (RMS voltage, RMS current, active power, reactive power, and power factor) and environmental status data (ambient temperature and humidity) are matched one-to-one according to the acquisition timestamps to form aligned electrical time-series data and environmental time-series data. This ensures the consistency of different types of data in the time dimension and lays the foundation for subsequent fusion processing.

[0022] Subsequently, data denoising processing is performed. Different denoising algorithms are used for different types of data based on their noise characteristics, ensuring effective denoising while preserving the data's valid features. For electrical time-series data, the main noise sources are impulse noise and random fluctuation interference. An adaptive Kalman filter algorithm is used for denoising processing, and the specific process is as follows: Establish the state equation and observation equation for the electrical time series data. The state equation is as follows: The observation equation is .in, Let k be the state vector of the electrical parameters at time k. Here is the state transition matrix. Let be the process noise vector, with a mean of 0 and a covariance of . Gaussian distribution; Let k be the observation vector at time k. For the observation matrix, The observed noise vector follows a pattern with a mean of 0 and a covariance of . The Gaussian distribution.

[0023] The adaptive Kalman filter algorithm estimates the process noise covariance in real time. and observation noise covariance The filter gain is dynamically adjusted, specifically based on the filter residual from the previous time step. Calculate residual covariance ,in Estimate the covariance of the state from the previous time step; adaptively adjust based on the residual covariance. and Then, through Kalman prediction and update steps, the clean electrical timing data at time k is obtained. This eliminates impulse noise and random fluctuation interference.

[0024] For environmental time-series data and equipment appearance images, the noise mainly consists of spatial and temporal noise. A nonlocal mean denoising algorithm is used for joint spatial and temporal denoising. For environmental time-series data, the temporal and spatial neighborhoods of each data point are selected, and the similarity between the data points in the neighborhood and the target data point is calculated. The data points in the neighborhood are then weighted and averaged using similarity as the weight to obtain the denoised clean environmental time-series data. For equipment appearance images, the image is divided into multiple image blocks, and the similarity between each image block and other image blocks is calculated. Image blocks with high similarity are selected and weighted averaged to replace the original image blocks, thereby removing noise from the image and obtaining clean equipment image data.

[0025] Finally, multimodal data fusion processing is performed, concatenating clean electrical time-series data, clean environmental time-series data, and clean equipment image data at a feature layer. Specifically, the electrical and environmental time-series data are extracted to obtain one-dimensional feature vectors, while the equipment appearance images are processed using convolution to extract two-dimensional feature vectors. These two-dimensional feature vectors are then dimensionality-reduced and concatenated with the one-dimensional feature vectors to form a fused feature vector. Simultaneously, end-point location identifiers are injected into the fused feature vector to distinguish sensing data from different power supply branch ends. This process ultimately generates a multimodal sensing data stream, providing a unified data format for subsequent transmission and processing.

[0026] Used to transmit the multimodal sensing data stream within a preset distance range via a wireless transmission link in the terahertz frequency band; Furthermore, including: First, the multimodal sensing data stream is preprocessed, including data frame segmentation and transmission priority labeling. Based on the importance and real-time requirements of the data, the multimodal sensing data stream is segmented into several data frames. Among them, sensing data reflecting voltage drops and fault symptoms at the end point (such as sudden changes in effective voltage values, excessive instantaneous current values, etc.) are marked as high-priority transmission frames. This type of data directly affects the timeliness of compensation control and the effectiveness of fault handling, and must be transmitted first. The remaining sensing data (such as normal ambient temperature, normal appearance images of equipment, etc.) are marked as ordinary-priority transmission frames and transmitted in the normal order.

[0027] Terahertz modulation processing is then performed, using a photonics-assisted terahertz modulation method to generate a terahertz modulation signal. The specific process is as follows: Two laser beams with a frequency interval in the terahertz band are selected and input into a photoconductive antenna. A terahertz carrier is generated through the beat frequency effect. The frequency of the terahertz carrier is equal to the frequency difference between the two laser beams. The frequency interval between the two laser beams can be adjusted according to the transmission requirements to achieve flexible adjustment of the terahertz carrier frequency.

[0028] High-priority and ordinary-priority transmission frames are mapped onto a quadrature amplitude modulation (QAM) constellation diagram. High-priority transmission frames employ high-order QAM to improve transmission rate and bandwidth utilization, while ordinary-priority transmission frames use low-order QAM to ensure transmission reliability. The mapped digital signal is then converted to an analog signal via digital-to-analog conversion and loaded onto the generated terahertz carrier to complete terahertz modulation, generating a terahertz modulated signal.

[0029] The transmission of terahertz modulated signals utilizes a high-gain antenna with beamforming capabilities. Beamforming technology focuses the terahertz modulated signal into a directional beam, enhancing signal transmission power and anti-interference capabilities while reducing signal attenuation. Based on a pre-defined transmission path, terahertz relay nodes are deployed within the terahertz transmission range. After the terahertz modulated signal is directionally transmitted by the high-gain antenna, it is reflected or forwarded by the terahertz relay nodes, establishing a stable terahertz wireless transmission link and ensuring efficient data transmission within the pre-defined distance range.

[0030] The reception of terahertz modulated signals employs a terahertz coherent detection receiver. The received terahertz modulated signals undergo down-conversion processing, converting the terahertz frequency band signals into intermediate frequency signals. The intermediate frequency signals are then digitally demodulated to restore the digital signals. Subsequently, the digital signals are reassembled, splicing the segmented data frames in their original order to recover the complete multimodal sensing data stream. This ensures the accuracy of data transmission and provides reliable data support for subsequent load forecasting and compensation command generation.

[0031] It is used to predict the trend of end load based on multimodal sensing data stream through load prediction model, and generate voltage and power factor collaborative compensation command with voltage deviation priority adjustment as constraint. Furthermore, including: First, input feature extraction is performed. Core data related to load changes are extracted from the multimodal sensing data stream, including active power time series, reactive power time series, ambient temperature series, and humidity series. The active power and reactive power time series directly reflect the real-time changes in the end-point load, while the ambient temperature and humidity series reflect the impact of environmental factors on the load. These four types of series are aligned by timestamps to form the input feature vector. ,in Let be the active power at time i. Let be the reactive power at time i. Let be the ambient temperature at time i. Let be the ambient humidity at time i, and n be the length of the time series.

[0032] The input feature vector is fed into a pre-built temporal convolutional attention network model, which is one of the core innovations of this method. This model accurately captures load change patterns at different time scales and adaptively adjusts the contribution of each parameter to load changes, thereby improving load prediction accuracy. The structure of the temporal convolutional attention network model is as follows: The model includes multiple dilated causal convolutional layers and a channel attention mechanism. The dilated causal convolutional layers employ different dilation coefficients to capture load variation patterns across different time scales. The calculation formula for dilated causal convolution is as follows: ,in Let be the convolution output at time t. The input features are at time td·k. Here, K represents the kernel weights, K is the kernel size, and d is the dilation coefficient. By setting different dilation coefficients, the convolutional layer can capture short-term, medium-term, and long-term load change patterns, avoiding information loss. The channel attention mechanism is used to adaptively determine the contribution of weighted electrical parameters and environmental state parameters to load changes. The specific implementation process is as follows: First, global average pooling is performed on the output feature map of the dilated causal convolutional layer to obtain the channel feature vectors. Where c is the number of feature channels, , Let H be the i-th feature value of the j-th channel, and H and W be the height and width of the feature map, respectively.

[0033] The channel feature vector U is input into a two-layer fully connected network. The first fully connected layer compresses the channel feature vector to 1 / r of its original dimension (where r is the compression coefficient), and the second fully connected layer restores it to its original dimension, resulting in the channel attention weight vector. ,in , It is the sigmoid activation function. and These are the first and second layer fully connected networks, respectively.

[0034] The channel attention weight vector W is multiplied channel by channel with the output feature map of the dilated causal convolutional layer to obtain a weighted feature map. This achieves adaptive weighting of features from different channels, making the model pay more attention to parameters that have a greater impact on load changes (such as active power and voltage) and reducing interference from secondary parameters.

[0035] The weighted feature map is input to the output layer, and the active power prediction sequence and reactive power prediction sequence for a preset future time period are output through a linear activation function as the end-point load trend prediction result. The preset time period can be adjusted according to the load change characteristics of the distribution network end, and is usually set to 15 minutes to 1 hour to ensure that the prediction result provides sufficient time for the generation of compensation instructions.

[0036] Finally, the predicted load trend at the terminal is combined with the current measured terminal voltage value to calculate the expected voltage deviation range and the expected reactive power demand range. The expected voltage deviation range is calculated as follows: based on the active power prediction sequence and the reactive power prediction sequence, combined with the terminal line impedance, the expected terminal voltage value within a preset time period is calculated and compared with the preset qualified voltage range to obtain the expected voltage deviation range. The expected reactive power demand range is calculated as follows: based on the reactive power prediction sequence, combined with the current measured reactive power value at the terminal, the range of reactive power that needs to be compensated within a preset time period is calculated, providing a basis for the generation of subsequent compensation instructions.

[0037] A voltage and reactive power sensitivity matrix is ​​constructed to characterize the impact of changes in reactive power at the terminal on the terminal voltage, serving as the basis for subsequent multi-objective optimization. Based on the line impedance parameters at the terminal of the distribution network power supply branches and the terminal load trend prediction results, the voltage and reactive power sensitivity matrix is ​​calculated using the nodal voltage method.

[0038] Line impedance parameters, including line resistance R and line reactance X, are obtained in advance using line parameter measurement equipment. The end-load trend prediction results include active power and reactive power prediction sequences for a preset future time period. The calculation formula for the nodal voltage method is as follows: Where U is the terminal voltage vector, Z is the line impedance matrix, and I is the terminal current vector. Differentiating this formula yields the voltage and reactive power sensitivity matrices. The elements in S This indicates the degree of influence of the reactive power change of the j-th reactive power compensation node on the voltage of the i-th terminal node.

[0039] Based on the voltage and reactive power sensitivity matrix, a multi-objective optimization model is established, clarifying the optimization objectives and constraints. Among them, the expected voltage deviation range is the primary constraint, requiring the terminal voltage deviation to be maintained within the preset qualified voltage range (usually ±7% of the rated voltage) within a preset time period in the future; the expected reactive power demand range is the secondary constraint, ensuring that the reactive power compensation capacity is within a reasonable range.

[0040] The first priority objective of the multi-objective optimization is to maintain the terminal voltage deviation within the preset acceptable voltage range during the prediction period. The objective function is: ,in Let be the predicted value of the terminal voltage at time t. The voltage is the rated voltage, and T is the total duration of the prediction period; the second priority objective is to maximize the average power factor after satisfying the first priority objective, and the objective function is: ,in Let be the power factor at time t.

[0041] A hierarchical sequence method is used to solve the multi-objective optimization model. First, the optimal solution for the first priority objective is solved, which is the reactive power compensation capacity range that satisfies the voltage qualification condition. Based on this, the optimal solution for the second priority objective is solved, yielding the minimum reactive power compensation capacity requirement value and the additional reactive power compensation allocation value based on the remaining compensation capacity that satisfies the voltage qualification condition. The minimum reactive power compensation capacity requirement value refers to the minimum reactive power compensation capacity required to maintain the terminal voltage within the qualification range, and the additional reactive power compensation allocation value refers to the remaining compensation capacity allocation value that can be used to improve the power factor after the voltage qualification condition is met.

[0042] Based on the minimum reactive power compensation capacity requirement and the additional reactive power compensation allocation value, voltage and power factor coordinated compensation instructions are generated. These instructions include capacitor bank switching schemes or static var generator (SVR) reactive power output instructions, specifically determined by the type of compensation equipment. The capacitor bank switching scheme specifies the capacitor branch number, switching direction, and switching capacity to be switched in each control cycle; the SVR reactive power output instruction specifies the reactive power output magnitude in each control cycle.

[0043] The principle of executing the compensation command is as follows: In each control cycle, the minimum reactive power compensation capacity that meets the voltage qualification requirement is executed first to ensure that the terminal voltage is quickly restored to the qualified range; after the voltage is qualified, the remaining compensation capacity is used to improve the power factor until the power factor reaches the set target (usually 0.95 and above) or the remaining available capacity is exhausted, so as to achieve the coordinated optimization of voltage regulation and reactive power compensation.

[0044] Based on the trend prediction results and voltage and power factor collaborative compensation instructions, a deep reinforcement learning algorithm is used to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation instructions to obtain the optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy. Furthermore, including: First, a deep deterministic policy gradient surrogate model is constructed. This model is the core of iterative optimization and is used to output the hyperparameters of the optimized load forecasting model and the weight coefficients of the compensation strategy. The deep deterministic policy gradient surrogate model includes an online policy network, a target policy network, an online value network, and a target value network. The structure and function of each network are as follows: The online policy network is used to generate actions in the current state, i.e., adjustments to the hyperparameters of the load prediction model and the weight coefficients of the compensation strategy; the target policy network is used to generate target actions, calculate target value, and reduce variance during training; the online value network is used to evaluate the value of the current state-action pair, and the target value network is used to evaluate the value of the target state-action pair, providing a basis for updating the online policy network.

[0045] The model's state space consists of historical multimodal sensing data, structural hyperparameters of the edge-side load prediction model, and weight coefficients in the voltage and power factor collaborative compensation command generation strategy. Historical multimodal sensing data includes electrical parameters, environmental condition data, and equipment appearance images from the past period. The structural hyperparameters of the load prediction model include the kernel size, dilation coefficient, number of channels, and compression coefficient of the temporal convolutional attention network. The compensation strategy weight coefficients include the weight allocation coefficients for voltage deviation and power factor in the optimization objective.

[0046] The model's action space is the adjustment amount for the aforementioned hyperparameters and weight coefficients. Each action corresponds to a set of adjustment values ​​for hyperparameters and weight coefficients, ensuring that the adjusted parameters can improve the accuracy of load forecasting and the compensation effect.

[0047] The model's reward function is determined based on changes in the distribution network's end-voltage compliance rate, average power factor, and line loss rate. It is used to evaluate the quality of actions and guide the model to iterate towards the optimal direction. The formula for calculating the reward function is as follows: ,in For voltage qualification rate, The average power factor. This represents the change in line loss rate. , , The weighting coefficient is set according to the operational needs of the distribution network to ensure that the reward function can accurately reflect the effect of compensation control.

[0048] Subsequently, an experience replay pool is constructed, utilizing long-term historical operational data uploaded by the edge intelligent collaboration module, including multimodal perception data streams, load forecasting results, compensation instructions, compensation execution feedback data, and operational indicator data such as voltage qualification rate, power factor, and line loss rate. This data is stored in the experience replay pool in chronological order to form a dataset for model training.

[0049] The model training process is as follows: Batch data is randomly sampled from the experience replay pool. Each batch of data contains information such as state, action, reward, and next state. The sampled data is input into the online value network to calculate the value of the current state-action pair, and the target value is calculated based on the target policy network and the target value network. The parameters of the online value network are updated by minimizing the mean square error between the current value and the target value. The parameters of the target value network are synchronized using a soft update method, with the soft update formula being... ,in For the target value network parameters, For online value network parameters, This is a soft update coefficient, with a value range of 0.01 to 0.1, to ensure the stability of parameter updates.

[0050] The online policy network is updated using the policy gradient descent method. The policy gradient is calculated based on the output of the online value network, and the parameters of the online policy network are adjusted to maximize the expected value of the reward function. At the same time, the parameters of the target policy network are synchronized using a soft update method.

[0051] Repeat the above training process until the model converges, that is, the expected value of the reward function tends to stabilize. At this time, based on the trained deep deterministic policy gradient proxy model, output the optimized load prediction model hyperparameters and compensation policy weight coefficients, and send them to the edge intelligent collaboration module to realize the iterative optimization of the load prediction model and compensation policy.

[0052] Based on the voltage and power factor collaborative compensation command, the contactless switching unit based on power electronic devices is controlled to perform zero-crossing switching action, so as to prioritize the adjustment of voltage deviation and use the remaining capacity to compensate reactive power. Furthermore, including: First, the compensation command is analyzed to extract the capacitor branch numbers that need to be put into or cut into the current control cycle, as well as the corresponding switching direction (put into or cut into the current cycle). The action requirements of each capacitor branch are clarified to ensure the accuracy of the switching action.

[0053] For each capacitor branch to be operated, a phase-locked loop (PLL) is used to track the phase of the grid voltage in real time. The output signal of the PLL is kept in phase and frequency with the grid voltage to accurately detect the zero-crossing moment of the grid voltage. When the voltage difference across the bidirectional thyristor to be switched is detected to be close to zero within a preset window (usually 5°~10° before and after the voltage zero crossing), a trigger pulse sequence is generated. The frequency of the trigger pulse sequence is consistent with the grid frequency, and the pulse width ensures that the thyristor can reliably turn on or off.

[0054] The trigger pulse sequence, after being amplified by the driver, is applied to the gate of the contactless switching unit composed of anti-parallel thyristors, causing the thyristors to turn on or off in a zero-voltage or zero-current state, thus achieving arc-free switching of the capacitor branch. Specifically, when the capacitor branch is switched on, the thyristors are triggered to turn on at the moment the grid voltage crosses zero. At this time, the voltage across the capacitor is close to the grid voltage, avoiding inrush current and arcing. When the capacitor branch is switched off, the thyristors are triggered to turn off at the moment the grid current crosses zero. At this time, the capacitor current is zero, avoiding current-cutting arcing and protecting the thyristors and capacitor equipment.

[0055] During the switching process, the terminal voltage amplitude and power factor are monitored in real time, and the switching strategy is adjusted based on the monitoring results. If the terminal voltage is still lower than the preset lower voltage limit, it indicates that the current reactive power compensation capacity is insufficient, and more capacitor branches need to be added until the voltage recovers to the qualified range or all available branches are in operation. If the terminal voltage is qualified, the capacity of the remaining unoperated capacitor branches is used to continue to improve the power factor until the power factor reaches the set target or the remaining available capacity is exhausted, ensuring that the compensation effect meets expectations.

[0056] After completing all switching operations in this control cycle, the updated RMS values ​​of the terminal voltage, power factor, and RMS current values ​​of each capacitor branch are read from the multimodal sensing data stream. The total reactive power compensation capacity actually switched in this cycle is then calculated. The calculation method for the total reactive power compensation capacity actually switched is as follows: based on the rated capacity of each connected capacitor branch and its actual operating current, the actual reactive power compensation capacity of each branch is calculated, and the summation yields the total reactive power compensation capacity actually switched in this cycle.

[0057] The total reactive power compensation capacity actually switched, the updated effective value of the terminal voltage, the updated power factor, and the timestamp of the action completion time are encapsulated into a compensation execution feedback message and uploaded to the edge intelligent collaboration module and the cloud optimization platform to provide actual operation data support for the optimization of load forecasting models and compensation strategies.

[0058] When a preset fault condition is detected, a protection action is performed, and thermal management is carried out based on the phase change heat dissipation structure and intelligent air cooling linkage.

[0059] Furthermore, including: First, multi-parameter real-time detection is performed. The surface temperature of the thyristor heat sink and the temperature of the capacitor casing are collected by a temperature sensor array. The temperature sensor array is evenly deployed on the key heat dissipation parts of the thyristor and capacitor to ensure the comprehensiveness and accuracy of temperature detection. The instantaneous current value of each capacitor branch is collected by a Hall current sensor to monitor the changes in branch current in real time. The harmonic content of the bus voltage is collected by a voltage transformer to monitor the voltage quality of the power grid.

[0060] The fault detection and protection action implementation process is as follows: The collected instantaneous values ​​of temperature, current, and voltage harmonic content are compared with preset thresholds. The over-temperature threshold is set according to the rated operating temperature of the thyristor and capacitor. Typically, the over-temperature threshold for the surface temperature of the thyristor heat sink is 85℃, and the over-temperature threshold for the capacitor casing temperature is 65℃. The over-current threshold is set according to the rated current of the capacitor branch, which is usually 1.2 times the rated current. The harmonic distortion rate threshold is set according to the power quality standards of the distribution network, which is usually 5%.

[0061] If any parameter exceeds the corresponding threshold and the duration reaches the set delay (usually 1 to 3 seconds), a fault is determined to have occurred, and corresponding protection commands are output according to the fault type: if the temperature exceeds the over-temperature threshold, a cooling protection command is output, and some non-critical capacitor branches are disconnected to reduce the equipment load; if the instantaneous current value exceeds the overcurrent threshold, a trip command is output to disconnect the faulty branch and prevent the fault from spreading; if the voltage harmonic content exceeds the harmonic distortion rate threshold, an alarm command is output, and the compensation strategy is adjusted to reduce harmonic interference.

[0062] Under fault-free conditions, based on temperature distribution data collected by a temperature sensor array, the intelligent arc-free compensation execution module is thermally managed through phase-change heat dissipation control logic to maintain the temperatures of the thyristor and capacitor within the specified operating range (thyristor heatsink surface temperature not exceeding 75℃, and capacitor casing temperature not exceeding 55℃). The specific implementation of the phase-change heat dissipation control logic is as follows: When the temperature collected by the temperature sensor array exceeds the first temperature setpoint (65℃ for the thyristor and 45℃ for the capacitor), the phase change material embedded in the heat dissipation substrate is activated to passively absorb heat. The phase change material is a paraffin-based composite phase change material with high latent heat. This material can absorb a large amount of heat near the phase change temperature and maintain a basically constant temperature. By passively absorbing heat, the device temperature is reduced without consuming additional electrical energy, thus achieving energy-saving heat dissipation.

[0063] When the temperature continues to rise and exceeds the second temperature setpoint (75℃ for the thyristor and 55℃ for the capacitor), the variable-speed DC fan is activated for forced air cooling. The fan speed is continuously adjusted according to the temperature deviation using proportional-integral control. The calculation formula for proportional-integral control is as follows: Where n is the fan speed, This represents the deviation between the actual temperature and the set temperature. This is the proportionality coefficient. This is the integral coefficient. Through proportional and integral control, the fan speed can be adjusted in real time according to temperature changes, ensuring effective heat dissipation while reducing fan energy consumption.

[0064] By linking phase change heat dissipation and intelligent air cooling, a complete thermal management cycle is formed. When the equipment temperature drops below the first temperature setpoint, the variable speed DC fan is turned off, and the heat dissipation relies solely on the phase change material for passive cooling. When the temperature rises again to the second temperature setpoint, the forced air cooling is restarted, realizing intelligent and energy-saving thermal management, maintaining the temperature of the thyristor and capacitor within the specified operating range, and ensuring stable equipment operation.

[0065] To achieve the above objectives, this application also provides the following technical solution: a terahertz transmission-based intelligent energy-saving compensation control system, comprising the following modules: End-point multi-source sensing module: used to collect electrical parameters and environmental status data at the end of the power supply branch of the distribution network, and generate a multi-modal sensing data stream characterizing the end-point operating conditions through multi-modal data fusion and noise reduction processing; Terahertz high-speed transmission module: used to transmit the multimodal sensing data stream within a preset distance range via a wireless transmission link in the terahertz frequency band; Edge intelligent collaboration module: Based on multimodal sensing data stream, it is used to predict the trend of end load through load prediction model and generate voltage and power factor collaborative compensation instructions with voltage deviation priority adjustment as the constraint. Cloud-based optimization module: Based on the trend prediction results and voltage and power factor collaborative compensation instructions, the module uses a deep reinforcement learning algorithm to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation instructions, thereby obtaining the optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy. Intelligent arc-free compensation execution module: used to control the contactless switching unit based on power electronic devices to perform zero-crossing switching action based on the voltage and power factor collaborative compensation command, so as to prioritize the adjustment of voltage deviation and use the remaining capacity to compensate reactive power. Integrated protection and heat dissipation module: used to perform protection actions when preset fault conditions are detected, and to perform thermal management based on phase change heat dissipation structure and intelligent air cooling linkage.

[0066] Example 2 uses the end of a 10kV distribution network power supply branch as an application scenario. Based on a monitoring and control and data acquisition system, it aims to achieve coordinated compensation of voltage and power factor, as well as equipment safety management, as detailed below: During the data acquisition phase, voltage / current transformers, temperature and humidity sensors, image acquisition devices, and temperature sensor arrays are deployed at the end of the branch line. A unified time base is used for synchronous acquisition to achieve full-dimensional data collection. Real-time acquisition includes electrical parameters such as voltage and current, environmental temperature and humidity data, equipment appearance images, and thyristor and capacitor temperature data. After adaptive Kalman filtering, non-local mean denoising, and multi-modal fusion processing, a standardized sensing data stream is generated, completing data acquisition and preprocessing.

[0067] In the monitoring, control, and transmission links, a terahertz wireless transmission link is used to prioritize data streams, with fault-related data set as high priority. After photonic-assisted modulation, the data is transmitted in a directional manner, and relay nodes forward the data to ensure transmission stability. The receiving end demodulates and reassembles the data to provide support for monitoring and control.

[0068] Compensation control and supervision optimization are based on fused data streams. A temporal convolutional attention network is used to predict load trends, and voltage deviation constraints are combined to generate collaborative compensation commands. The prediction model and compensation strategy are iteratively optimized through deep reinforcement learning. The contactless switching unit responds to commands in real time. The phase-locked loop tracks the zero-crossing moment to achieve arc-free switching, while voltage and power factor are monitored in real time to dynamically adjust the switching strategy.

[0069] Fault monitoring and thermal management involve real-time monitoring of equipment temperature, branch current, and voltage harmonics, triggering protection actions when thresholds are exceeded. Phase-change cooling and intelligent air cooling are linked to maintain constant equipment temperature. All collected data, compensation execution results, and fault information are fed back to the cloud in real time, forming a closed loop of data acquisition, transmission control, monitoring, and optimization, achieving energy conservation, consumption reduction, and stable operation of the power distribution network.

[0070] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A power intelligent energy-saving compensation control method for terahertz transmission, characterized in that, include: The electrical parameters and environmental status data of the power supply branch line of the distribution network are collected, and multimodal data fusion and noise reduction processing is used to generate a multimodal sensing data stream that characterizes the operating conditions of the end. The multimodal sensing data stream is transmitted over a preset distance using a wireless transmission link in the terahertz band. Based on multimodal sensing data stream, the load forecasting model is used to predict the trend of the end load and generate a voltage and power factor collaborative compensation command with voltage deviation priority adjustment as the constraint. Based on the trend prediction results and voltage and power factor collaborative compensation instructions, a deep reinforcement learning algorithm is used to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation instructions to obtain the optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy. Based on the voltage and power factor collaborative compensation command, the contactless switching unit based on power electronic devices is controlled to perform zero-crossing switching action in order to prioritize the adjustment of voltage deviation and use the remaining capacity to compensate for reactive power. When a preset fault condition is detected, a protection action is performed, and thermal management is carried out based on the phase change heat dissipation structure and intelligent air cooling linkage.

2. The terahertz transmission intelligent energy-saving compensation control method as described in claim 1, characterized in that, The collected electrical parameters and environmental status data of the power supply branch lines of the distribution network are processed through multimodal data fusion and denoising to generate a multimodal sensing data stream characterizing the operating conditions of the terminal, including: By deploying voltage transformers, current transformers, temperature and humidity sensors, and image acquisition devices at the end of the power distribution network branch lines, electrical parameters, environmental status data, and equipment appearance images are collected synchronously with a unified time base. The electrical parameters include the effective value of voltage, the effective value of current, active power, reactive power, and power factor. The environmental status data includes ambient temperature and ambient humidity. The collected electrical parameters and environmental status data are time-series aligned according to timestamps to obtain aligned electrical time-series data and environmental time-series data; The adaptive Kalman filter algorithm is used to remove impulse noise and random fluctuation interference from the electrical timing data to obtain clean electrical timing data; The environmental time series data and the device appearance image are denoised using a nonlocal mean denoising algorithm, which combines spatial and temporal domains to obtain clean environmental time series data and clean device image data. The clean electrical time-series data, clean environmental time-series data, and clean equipment image data are spliced ​​together at the feature layer, and an end position identifier is injected to generate the multimodal sensing data stream.

3. The terahertz transmission intelligent energy-saving compensation control method as described in claim 1, characterized in that, The transmission of the multimodal sensing data stream over a preset distance range via a wireless transmission link in the terahertz band includes: The multimodal sensing data stream is segmented into data frames and prioritized for transmission. Sensing data reflecting voltage drops and fault symptoms at the end are marked as high-priority transmission frames. The photonics-assisted terahertz modulation method is adopted, which uses two laser beams with a frequency interval in the terahertz band to beat in a photoconductive antenna to generate a terahertz carrier. The high-priority transmission frame and other transmission frames are respectively mapped onto a constellation diagram of quadrature amplitude modulation. After digital-to-analog conversion, they are loaded onto the terahertz carrier to generate a terahertz modulation signal. The terahertz modulated signal is directionally transmitted through a high-gain antenna with beamforming function, and after reflection or forwarding by a terahertz relay node, a terahertz wireless transmission link is constructed on a preset transmission path. The multimodal sensing data stream is recovered by downconverting and digitally demodulating the received terahertz modulated signal using a terahertz coherent detection receiver.

4. The terahertz transmission intelligent energy-saving compensation control method as described in claim 1, characterized in that, The step of performing trend prediction of end-point load based on the multimodal sensing data stream using a load forecasting model includes: Active power time series, reactive power time series, ambient temperature series and humidity series are extracted from the multimodal sensing data stream to form an input feature vector; The input feature vector is input into a pre-constructed temporal convolutional attention network model, which includes multiple dilated causal convolutional layers and a channel attention mechanism. The dilated causal convolutional layers are used to capture load change patterns at different time scales, and the channel attention mechanism is used to adaptively assess the contribution of weighted electrical parameters and environmental state parameters to load changes. The model outputs a predicted active power sequence and a predicted reactive power sequence for a future preset time period, which serve as the end-load trend prediction result. By combining the predicted end-load trend with the current measured end-voltage value, the expected voltage deviation range and the expected reactive power demand range are calculated.

5. The terahertz transmission intelligent energy-saving compensation control method as described in claim 4, characterized in that, The process of executing the command to generate a voltage and power factor collaborative compensation system with voltage deviation as the primary constraint includes: Based on the line impedance parameters at the end of the power supply branch of the distribution network and the predicted load trend at the end, a voltage and reactive power sensitivity matrix is ​​constructed. Based on the voltage and reactive power sensitivity matrix, a multi-objective optimization model is established with the expected voltage deviation range as the primary constraint and the expected reactive power demand range as the secondary constraint. The first priority objective of the multi-objective optimization is to maintain the end voltage deviation within the preset qualified voltage range during the prediction period, and the second priority objective is to maximize the average power factor after satisfying the first priority objective. Solving the multi-objective optimization model yields the minimum reactive power compensation capacity requirement under the condition of voltage qualification and the additional reactive power compensation allocation value based on the remaining compensation capacity. Based on the minimum reactive power compensation capacity requirement and the additional reactive power compensation allocation value, a voltage and power factor coordinated compensation instruction is generated, which includes a capacitor bank switching scheme or a static var generator reactive power output instruction. In each control cycle, the capacitor bank switching scheme or reactive power output instruction is executed first to meet the minimum reactive power compensation capacity that meets the voltage qualification, and then the remaining capacity is used to improve the power factor.

6. The terahertz transmission intelligent energy-saving compensation control method as described in claim 1, characterized in that, Based on the trend prediction results and the voltage and power factor collaborative compensation command, a deep reinforcement learning algorithm is used to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation command, resulting in optimized load prediction model hyperparameters and compensation strategy weight coefficients. These are used for iterative optimization of the load prediction model and compensation strategy, including: A deep deterministic policy gradient proxy model is constructed, which includes an online policy network, a target policy network, an online value network, and a target value network. The state space consists of historical multimodal sensing data, structural hyperparameters of the edge-side load prediction model, and weight coefficients in the voltage and power factor collaborative compensation command generation strategy. The action space is the adjustment amount for the hyperparameters and weight coefficients. The reward function is determined based on the changes in the voltage qualification rate, average power factor, and line loss rate at the end of the distribution network. An experience replay pool is constructed using uploaded long-term historical operational data. The online value network is trained by randomly sampling batches of data from the experience replay pool, and the target value network parameters are synchronized using a soft update method. Based on the trained deep deterministic policy gradient proxy model, the optimized load prediction model hyperparameters and compensation policy weight coefficients are output.

7. The terahertz transmission intelligent energy-saving compensation control method as described in claim 1, characterized in that, The method of controlling the contactless switching unit based on power electronic devices to perform zero-crossing switching based on the voltage and power factor coordinated compensation command, in order to prioritize the adjustment of voltage deviation and utilize remaining capacity to compensate for reactive power, includes: The voltage and power factor collaborative compensation command is analyzed to obtain the capacitor branch numbers that need to be put into or cut into the current control cycle and the corresponding switching direction. For each capacitor branch to be operated, the phase of the grid voltage is tracked in real time through a phase-locked loop. When the voltage difference across the bidirectional thyristor to be switched is detected to be within a preset window, a trigger pulse sequence is generated. The trigger pulse sequence is amplified and applied to the gate of the contactless switching unit composed of anti-parallel thyristors, so that the thyristors are turned on or off in a zero voltage or zero current state, and the corresponding capacitor branch is connected to the distribution network or disconnected from the distribution network. After the capacitor branch is put into operation, the voltage amplitude and power factor at the end are monitored in real time. If the voltage at the end is still lower than the preset voltage range lower limit, more capacitor branches will be put into operation until the voltage recovers to the qualified range or all available branches have been put into operation. If the terminal voltage is qualified, the remaining unused capacitor branches will be used to continue to improve the power factor until the power factor reaches the set target or the remaining available capacity is exhausted. After completing all switching actions in this control cycle, the updated effective value of the terminal voltage, power factor, and effective value of the current in each capacitor branch are read from the multimodal sensing data stream. The total reactive power compensation capacity actually switched in this cycle is calculated, and the total reactive power compensation capacity actually switched, the updated effective value of the terminal voltage, the updated power factor, and the timestamp of the action completion time are encapsulated into a compensation execution feedback message.

8. The terahertz transmission intelligent energy-saving compensation control method as described in claim 1, characterized in that, The provision for executing protective actions upon detecting a preset fault condition, and for thermal management based on the phase-change heat dissipation structure and intelligent air cooling linkage, includes: The surface temperature of the thyristor heat sink and the temperature of the capacitor casing are collected by a temperature sensor array, the instantaneous value of the current in each branch is collected by a Hall current sensor, and the harmonic content of the bus voltage is collected by a voltage transformer. The instantaneous values ​​of temperature, current and voltage harmonic content are compared with preset over-temperature threshold, over-current threshold and harmonic distortion rate threshold. When any parameter exceeds the corresponding threshold and the duration reaches the set delay, a fault is determined and a protection command to trip or disconnect part of the capacitor branch is output. In the absence of faults, the temperature distribution data collected by the temperature sensor array is input to the phase change heat dissipation control logic to maintain the temperature of the thyristor and capacitor within the specified operating range.

9. A terahertz transmission intelligent energy-saving compensation control system for power, applied to the terahertz transmission intelligent energy-saving compensation control method for power as described in any one of claims 1-8, characterized in that: include: End-point multi-source sensing module: used to collect electrical parameters and environmental status data at the end of the power supply branch of the distribution network, and generate a multi-modal sensing data stream characterizing the end-point operating conditions through multi-modal data fusion and noise reduction processing; Terahertz high-speed transmission module: used to transmit the multimodal sensing data stream within a preset distance range via a wireless transmission link in the terahertz frequency band; Edge intelligent collaboration module: Based on multimodal sensing data stream, it is used to predict the trend of end load through load prediction model and generate voltage and power factor collaborative compensation instructions with voltage deviation priority adjustment as the constraint. Cloud-based optimization module: Based on the trend prediction results and voltage and power factor collaborative compensation instructions, the module uses a deep reinforcement learning algorithm to iteratively optimize the load prediction model and the generation strategy of the voltage and power factor collaborative compensation instructions, thereby obtaining the optimized load prediction model hyperparameters and compensation strategy weight coefficients, which are used to iteratively optimize the load prediction model and compensation strategy. Intelligent arc-free compensation execution module: used to control the contactless switching unit based on power electronic devices to perform zero-crossing switching action based on the voltage and power factor collaborative compensation command, so as to prioritize the adjustment of voltage deviation and use the remaining capacity to compensate reactive power. Integrated protection and heat dissipation module: used to perform protection actions when preset fault conditions are detected, and to perform thermal management based on phase change heat dissipation structure and intelligent air cooling linkage.

Images