Energy recovery method and system for high-voltage wire tensioner driven by direct current power supply

Abstract

This application discloses a method and system for energy recovery from a DC-powered high-voltage line tensioning machine. It includes: constructing a hybrid energy storage system comprising a battery pack and a supercapacitor, and multimodal sensing hardware; processing images and vibration information collected from the line-laying trolley using a spatiotemporal fusion deep prediction model to predict mechanical impact characteristics in advance; using a deep reinforcement learning algorithm for active, operation-level energy allocation and scheduling of the hybrid energy storage; and finally, employing dual-timescale control to isolate electrical and mechanical disturbances. This application can actively decouple and eliminate bus voltage oscillations and tension control misalignment caused by sudden load changes such as the traction plate passing over the trolley, reducing the risk of line drop and improving the efficiency of regenerative energy recovery during line-laying construction.

Description

Technical Field

[0001] This application relates to the field of high-voltage line stringing construction technology, specifically to a method and system for energy recovery of a high-voltage line stringing tensioner driven by a DC power supply. Background Technology

[0002] In the conductor laying construction of high-voltage and ultra-high-voltage power transmission projects, electric tensioners are gradually replacing traditional hydraulic tensioners due to their advantages such as environmental friendliness and high control precision. During the tensioning and laying process, the drive motor of a DC-powered electric tensioner operates in a regenerative power generation state, converting the mechanical energy released from the conductor into electrical energy. The efficient and safe handling of this regenerative electrical energy is crucial to the performance of the tensioner.

[0003] However, existing tensioners driven by DC power sources (such as battery packs) face technical challenges in practical applications. The complex environment of high-voltage line construction means that when the conductor passes through the laying trolley, the traction plate, splicing tube, or anti-torsion device connected to it experiences severe mechanical impact. This impact causes a significant sudden change in the tensioner's load within milliseconds, generating a high-power regenerative energy pulse. In the traditional "battery plus energy-dissipating resistor" energy processing architecture, this sudden high-power energy pulse directly impacts the DC bus, causing a momentary voltage surge. To protect power devices such as the motor controller, existing systems must employ passive response measures such as consuming energy with energy-dissipating resistors or forcibly reducing the motor's output torque. These measures directly lead to severe fluctuations or lag in the electromagnetic torque output by the motor, disrupting the constancy of the conductor tension.

[0004] In such systems, there is a close, strong coupling relationship between the stability of the DC bus voltage and the accuracy of the constant tension control of the motor. On the one hand, drastic fluctuations in the DC bus voltage can interfere with the normal operation of the motor vector control algorithm, reducing the stability of torque control; on the other hand, passive protection measures taken to stabilize the bus voltage, in turn, disrupt the smooth output of torque. This negative coupling effect is particularly prominent under low-speed, high-tension continuous winding conditions, easily causing abnormal conductor vibration and even potential safety accidents such as conductor detachment. Existing technologies using conventional proportional-integral-derivative (PID) control or fixed-rule-based energy management strategies cannot proactively sense and decouple millisecond-level impact loads, thus failing to fundamentally solve the aforementioned technical problems. Summary of the Invention

[0005] The technical problem to be solved by this application is to overcome the defect of existing DC power supply driven high-voltage line tensioning machines, which have poor system performance when encountering impact loads due to the strong coupling relationship between DC bus voltage instability and constant tension control accuracy.

[0006] To address the aforementioned technical problems, this application provides a method for energy recovery from a DC-powered high-voltage line tensioning machine, comprising the following steps: Step 1: Construct a hybrid energy storage system and a multimodal sensor array. The hybrid energy storage system includes a DC bus, a battery pack branch, a supercapacitor pack branch, a motor, a motor controller, and a power-consuming resistor. The battery pack branch includes a battery pack and a first bidirectional DC / DC converter, with the battery pack connected to the DC bus via the first bidirectional DC / DC converter. The supercapacitor pack branch includes a supercapacitor pack and a second bidirectional DC / DC converter, with the supercapacitor pack connected to the DC bus via the second bidirectional DC / DC converter. The multimodal sensor array is located at the wire-laying trolley of the tension machine and is used to collect image information of the conductor and splicing fittings, as well as vibration signals from the wire-laying trolley. Step 2: Based on the image information and vibration signal collected by the multimodal sensor group, the impact load within the future time window is predicted by a pre-built prediction model to obtain load prediction information including the time of impact and the magnitude of impact energy. Step 3: Using a deep reinforcement learning agent, based on the real-time voltage of the DC bus, the state of charge of the battery pack, the state of charge of the supercapacitor pack, the feedback power of the motor, and the load prediction information, generate and output optimal control commands for the battery pack branch, the supercapacitor pack branch, and the energy-consuming resistor to perform dynamic energy allocation. Step four: When the magnitude of the impact energy predicted in step two exceeds the maximum energy threshold that the hybrid energy storage system can absorb, a safety protection strategy is executed to dissipate the overflow energy through the energy-consuming resistor.

[0007] Optionally, the multimodal sensor group mentioned in step one includes an industrial camera for acquiring the image information and an accelerometer for acquiring the vibration signal.

[0008] Optionally, the prediction model in step two is a spatiotemporal fusion prediction network. The spatiotemporal fusion prediction network uses a convolutional neural network to extract spatial features from the image information and combines a long short-term memory network to process time-related sequence features. At the same time, the spatiotemporal fusion prediction network introduces an adaptive attention mechanism to dynamically weight and fuse features from different modalities to output quantitative prediction values ​​for the load torque prediction curve, the predicted time of impact, and the magnitude of the pulse energy generated by the impact within the future time window.

[0009] Optionally, the spatiotemporal fusion prediction network has online learning capabilities, which can compare the real load data collected during the actual impact with the predicted data of the impact, calculate the prediction error, and update the weight parameters of the spatiotemporal fusion prediction network through the backpropagation algorithm.

[0010] Optionally, the method further includes implementing dual-time-scale decoupling control: on the fast time scale, a current loop control strategy is adopted for the second bidirectional DC / DC converter, and energy is absorbed or released according to the instructions for the supercapacitor branch in the optimal control instructions output in step three, so as to suppress the voltage fluctuation of the DC bus within a preset range; on the slow time scale, the motor controller executes a field-oriented control algorithm under a stable DC bus voltage within the preset range, thereby outputting a constant electromagnetic torque.

[0011] Optionally, in step three, when the deep reinforcement learning agent predicts that an impact is about to occur based on the load prediction information, and determines that the remaining absorbable energy of the supercapacitor bank is less than the impact energy, the deep reinforcement learning agent outputs a command in advance to control the supercapacitor bank to pre-discharge, thereby freeing up the absorption capacity of the supercapacitor bank.

[0012] Optionally, the safety protection strategy described in step four is a multi-level safety protection, including: a first-level protection, which consumes the overflow energy that the supercapacitor group cannot fully absorb by adjusting the pulse width modulation duty cycle of the energy-consuming resistor; and a second-level protection, which triggers a mechanical braking auxiliary mechanism to apply physical braking damping to the drum of the tension machine when the DC bus voltage continues to rise after the first-level protection is implemented.

[0013] Optionally, the supercapacitor bank in the supercapacitor bank branch described in step one is composed of a graphene-based supercapacitor module.

[0014] Optionally, before inputting the image information and the vibration signal into the spatiotemporal fusion prediction network, a data preprocessing step is further included: performing frame difference processing on the image information to extract the motion trajectory information of the connecting hardware; and performing wavelet packet decomposition on the vibration signal to extract time-frequency domain features.

[0015] This application also provides an energy recovery system for a high-voltage line tensioner driven by a DC power supply, comprising: A hybrid energy storage system and a multimodal sensor array are provided. The hybrid energy storage system includes a DC bus, a battery pack branch, a supercapacitor pack branch, a motor, a motor controller, and a power-dissipating resistor. The battery pack branch includes a battery pack and a first bidirectional DC / DC converter, with the battery pack connected to the DC bus via the first bidirectional DC / DC converter. The supercapacitor pack branch includes a supercapacitor pack and a second bidirectional DC / DC converter, with the supercapacitor pack connected to the DC bus via the second bidirectional DC / DC converter. The multimodal sensor array is located at the wire-laying trolley of the tension machine and is used to acquire image information of the conductor and splicing fittings, as well as vibration signals from the wire-laying trolley. The prediction module is used to predict the impact load within a future time window based on the image information and vibration signal collected by the multimodal sensor group, and to obtain load prediction information including the time of impact and the magnitude of impact energy. The energy dynamic allocation module is used to generate and output optimal control commands for the battery pack branch, the supercapacitor branch and the energy-consuming resistor by using a deep reinforcement learning agent based on the real-time voltage of the DC bus, the state of charge of the battery pack, the state of charge of the supercapacitor pack, the feedback power of the motor and the load prediction information, so as to perform dynamic energy allocation. A safety protection module is used to execute a safety protection strategy when the magnitude of the impact energy predicted by the prediction module exceeds the maximum energy threshold that the hybrid energy storage system can absorb, and to consume the overflow energy through the energy-consuming resistor.

[0016] The technical solution provided in this application transforms the traditional passive and responsive energy processing method into an active and predictive energy planning and control approach by introducing multimodal perception and prediction technology based on artificial intelligence and combining it with a hybrid energy storage system and a deep reinforcement learning energy management strategy. This solution decouples the electrical problem of DC bus voltage stability from the mechanical problem of constant tension control, allowing both to simultaneously achieve high performance standards. Therefore, this application effectively solves the control problem of DC power-driven tensioners under complex impact conditions, improves energy recovery efficiency and system stability, and provides high-precision and high-reliability technical support for high-voltage line construction. Attached Figure Description

[0017] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the structure of a DC power supply driven high-voltage line tensioner energy recovery system provided in an embodiment of this application.

[0019] Figure 2This is a flowchart of an energy recovery method for a DC power supply driven high-voltage line tensioner provided in an embodiment of this application.

[0020] Figure 3 This is a block diagram of the impact load prediction and dynamic energy distribution control logic in the embodiments of this application.

[0021] Figure 4 This is a schematic diagram comparing the response of key system parameters of the embodiments of this application and the prior art when encountering impact loads.

[0022] Figure 5 This is a schematic diagram of the deployment of the multi-modal sensor on the tension machine wire feeding trolley in an embodiment of this application.

[0023] Figure 6 This is a cross-sectional view of the electrical cabinet of the hybrid energy storage system in the embodiments of this application.

[0024] Figure 7 This is a comparison diagram of the transient response of DC bus voltage under impact load between the embodiments of this application and the prior art.

[0025] Figure 8 This is a comparison diagram of the electromagnetic torque recovery dynamic characteristics of the embodiments of this application and the prior art when encountering impact loads.

[0026] Figure 9 This is a comparison chart of the energy recovery efficiency and energy flow direction analysis of the embodiments of this application and the prior art in tension impact events. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the specific embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0028] Example 1 This embodiment provides an energy recovery method for a DC-powered high-voltage line tensioning machine. This method actively predicts upcoming impact loads and utilizes a hybrid energy storage system and deep reinforcement learning algorithms for predictive energy scheduling, thus decoupling electrical system stability from mechanical constant tension control. (Refer to...) Figure 2 The method may include the following steps: Step S1: Construct a hybrid energy storage and multimodal sensing system. (Refer to...) Figure 1This embodiment establishes a complete energy recovery and sensing system. The system centers on a DC bus, with multiple electrical units connected in parallel. The power source is a permanent magnet synchronous motor, connected to the DC bus via a motor controller. On the energy storage side, the system constructs a hybrid energy storage unit, specifically including a battery pack and a supercapacitor pack. The battery pack is connected to the DC bus via a first bidirectional DC / DC converter, primarily responsible for energy supply and recovery during long-term system operation, maintaining overall energy balance. Its characteristics include high energy density but slow power response. The supercapacitor pack is connected to the DC bus via a second bidirectional DC / DC converter. This second bidirectional DC / DC converter has extremely high switching frequency and control bandwidth, enabling the supercapacitor pack to perform millisecond-level high-power charging and discharging. Its function is specifically designed to handle high-frequency, high-power pulse energy, characterized by extremely high power density but relatively low energy density. Furthermore, a braking unit consisting of a switching transistor and a power-consuming resistor is connected to the DC bus for energy dissipation under extreme conditions. To achieve advanced sensing of impact loads, a multi-modal sensor array is installed on the wire release trolley at the conductor exit of the tension machine. Specifically, the sensor array includes a high-speed industrial camera and a high-frequency accelerometer. The high-speed industrial camera faces the wire-laying trolley and captures image information of the wire and its accessories, such as splicing tubes, traction plates, or anti-torsion devices, as they pass through the trolley. The high-frequency accelerometer is mounted on the trolley's bearing housing and collects the weak vibration signals generated by the trolley due to mechanical impact.

[0029] As a high-level implementation of the aforementioned hybrid energy storage and multimodal sensing system, the key components in this embodiment have strict specifications in terms of physical properties and structural design. Specifically, the high-speed industrial camera in the multimodal sensor group adopts a charge-coupled device sensor or a complementary metal-oxide-semiconductor sensor with a global shutter mechanism, a physical resolution of not less than 1920×1080 pixels, and a sampling frame rate set between 240 and 500 frames per second to ensure that, under the condition of high-speed deployment of the conductor at 40~50m / min, the displacement blur of accessories such as connectors and traction plates in motion on the image plane is limited to within a single pixel.

[0030] To overcome visual degradation under complex climate and lighting conditions, when the ambient light level is below 500 Lux due to mountain shade, sunset backlight, or heavy rain and fog, the anti-dust-accumulation LED strobe light source array arranged in a ring around the camera lens will automatically trigger supplemental lighting through a photoelectric sensor. The strobe pulse width is synchronized with the camera's global shutter at the microsecond level to ensure the acquisition of high-contrast sharp images without motion blur.

[0031] Furthermore, the high-frequency accelerometer uses a piezoelectric ceramic sensor with an integrated anti-aliasing low-pass filter. Its broadband frequency response curve remains flat in the 0.5Hz to 20kHz range, and its measurement range is set to ±50g. This allows it to not only capture transient low-frequency impacts from large-mass traction skateboards without distortion, but also to sensitively capture high-frequency resonant waves generated by minute deformations of the anti-torsion device. The accelerometer is rigidly coupled to the outer housing of the wire-laying trolley bearing using a flange with an epoxy resin injection layer and multiple high-strength stainless steel bolts, eliminating any low-frequency resonance interference and signal attenuation caused by adhesive aging.

[0032] In terms of hybrid energy storage systems, the main body of the battery pack is constructed from high-capacity lithium iron phosphate square wound cells in a compact topology of 100 series and 10 parallel. To address the heat accumulation issue caused by continuous high-rate venting and regenerative charging during wiring operations, the battery management system integrates a liquid-cooled constant-temperature thermal management strategy, including a microchannel circulating cold plate with perfluoroketone flame-retardant coolant. This strategy can forcibly constrain the temperature difference within the battery pack's core to within a 3°C tolerance range during harsh thermal load cycles of large-scale energy throughput.

[0033] Meanwhile, the battery management system kernel employs a low-level observer structure that deeply integrates the ampere-hour integral method and the adaptive extended Kalman filter algorithm to implement high-precision estimation of the state of charge. Its steady-state estimation error is strictly less than 1%, ensuring the extreme accuracy of the reference energy scheduling information.

[0034] To meet the stringent constraints of ultra-fast, high-power throughput, the supercapacitor bank employs graphene polymer matrix supercapacitor cells, each with a nominal capacitance of 3000F. Dozens of cells are connected in series, along with a corresponding dynamic active voltage equalization circuit, to form an energy storage module. This voltage equalization circuit is built upon the energy transfer mechanism of a multi-winding flyback transformer, ensuring that the terminal voltage of all cells rises synchronously with the voltage equalization state trajectory when subjected to transient, millisecond-level high-intensity energy pulse injections.

[0035] To connect the two types of energy storage systems to the DC bus, both the first and second bidirectional DC / DC converters employ a three-phase interleaved parallel bidirectional controlled buck-boost chopper topology. This interleaved parallel topology effectively increases the equivalent switching frequency by injecting pulse width modulation control signals with a 120° electrical angle difference into each phase arm, thereby reducing the size of the power inductor. Furthermore, it significantly reduces current ripple at the amplitude level, minimizing electromagnetic interference noise induced on the ultra-high voltage bus junction side.

[0036] The switching device is a silicon carbide metal oxide semiconductor field-effect transistor with high frequency and low conduction loss characteristics, ensuring that it can operate in a cold state under hard switching or even soft switching operation at a frequency of not less than 20kHz.

[0037] Furthermore, the traditional low-speed fieldbus was abandoned in favor of a real-time bidirectional distributed communication architecture based on the industrial Ethernet protocol standard, with a 100Mbps operating baseline, connecting the motor controller, the lower-level computers of the two bidirectional DC / DC converters, and the upper-level computer of the multimodal sensing processor. This was achieved by using high-bandwidth optical fiber to cut off electromagnetic coupling paths, supplemented by a microsecond-level distributed network clock synchronization mechanism based on a precision time protocol, thus eliminating timing deviations and command delays caused by multi-node collaborative processing of impact pulses.

[0038] like Figure 5 As shown, the wire-laying trolley 101 is illustrated in a side view, including an inverted U-shaped support, a pulley body, and two bearing seats 107 on both sides. The wire 102 enters from the left, passes over the top of the pulley groove, and extends to the right. The position and direction of movement of the connecting hardware 103 (traction plate / connecting tube) are marked on the wire. A high-speed industrial camera 104 is mounted in front of the trolley, directly facing the pulley. LED strobe light sources 105 are arranged in a ring around the lens, forming a strobe light source array to ensure clear images of irregularly shaped parts even in high-speed motion scenarios. A high-frequency accelerometer 106 is rigidly coupled to the bearing seat 107 via a flange and is used to capture impact vibration signals generated when the connecting hardware passes through the trolley. An optical fiber communication cable 108 leads from and converges between the camera and the accelerometer, enabling high-speed transmission of multimodal sensing data. This deployment scheme ensures that the camera's field of view completely covers the pulley area, and the accelerometer's mounting position effectively collects vibration signals, providing a reliable data foundation for subsequent irregularly shaped part identification and impact detection.

[0039] Step S2, Spatiotemporal Prediction of Impact Load Based on Multimodal Sensing. To achieve accurate prediction of impact load, this step utilizes a deep learning model to fuse the multimodal data collected in Step S1. First, data preprocessing is performed. Specifically, the continuous video stream acquired by the high-speed industrial camera is processed using the frame difference method. By calculating the pixel differences between adjacent frames, moving irregularly shaped parts can be effectively highlighted from a complex background, and their contours, dimensions, and motion trajectory information such as distance and speed relative to the wire-laying trolley can be extracted to form an image feature sequence. Simultaneously, the broadband vibration signal acquired by the high-frequency accelerometer is decomposed into wavelet packet decomposition, breaking it down into different frequency bands. Statistical characteristics such as energy and entropy values ​​of each frequency band are calculated, thereby obtaining a vibration feature vector that can finely characterize the precursors of impact.

[0040] Subsequently, the preprocessed image feature sequence and vibration feature vector are used as dual inputs and fed into a pre-trained spatiotemporal fusion prediction network. (Refer to...) Figure 3The core architecture of this network is a combination of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM). The CNN module is responsible for extracting higher-dimensional spatial representations from the features of each frame of the image, while the LSTM module is responsible for processing the time series composed of the feature vectors output by the CNN, thereby capturing the temporal dependencies of the development of the shock event.

[0041] Furthermore, to enhance the robustness and accuracy of the prediction model in complex on-site environments, this embodiment introduces an adaptive attention mechanism into the spatiotemporal fusion prediction network. Specifically, a modal confidence evaluation network is added to the feature fusion layer of CNN and LSTM. This network, through a channel attention mechanism, can evaluate the quality of image and vibration signals in real time. For example, when a sudden change in on-site lighting causes image overexposure, the confidence of image features decreases. The network automatically learns and assigns smaller weights to the image feature channels while increasing the weights of the vibration feature channels, and vice versa. This dynamic weighted fusion mechanism ensures that the model can intelligently focus on more reliable sources. Simultaneously, the network also includes a temporal attention mechanism, which assigns different levels of attention to different historical moments when analyzing time-series data, enabling the model to focus more on data features within the critical time window before the impact occurs.

[0042] Finally, the spatiotemporal fusion prediction network outputs the load torque prediction curve for a short future time window (e.g., the next 5 seconds), denoted as T_load(t); the precise prediction time of the impact, denoted as t_impact; and the quantitative prediction value of the regenerative pulse energy generated by the impact, denoted as E_pulse.

[0043] In implementing the functional details of the aforementioned spatiotemporal fusion prediction network, the mathematical mechanism of data preprocessing and the basic transformation matrix are core components ensuring the generalization ability of the subsequent feature extractor. For the frame difference processing step in the video stream, the computational flow is not a simple pixel arithmetic subtraction, but rather a deep fusion of the Gaussian mixture background modeling method and a dynamic adaptive threshold segmentation strategy. During the initialization phase, the system collects trolley videos without any metal fittings passing through as a purely passive background dataset. A Gaussian probability density function is constructed based on the temporal brightness distribution of each pixel, and the posterior probability of the current frame's pixel brightness belonging to this Gaussian model is calculated.

[0044] A difference image is extracted from the calculated salient foreground target. Then, a two-dimensional discrete Gaussian kernel is applied for spatial smoothing filtering to remove noise and impurities. This is followed by a morphological closing operation combining dilation and erosion to fill in internal holes and smooth edge burrs, resulting in a binary hardware morphological mask with topological features that perfectly approximate the real contour. Subsequently, a connected component labeling algorithm based on matrix eigenvalues ​​is used to solve for the center coordinates of the mask and the deformation rate of its circumscribed rectangle, generating low-dimensional temporal motion trajectory information.

[0045] In the wavelet packet decomposition algorithm for acceleration signals, a series of wavelets with tightly supported orthogonal properties and approximately smooth phase are selected as the mother wavelet function. The broadband signal undergoes a 4-6 layer complete binary tree architecture multi-scale pyramid decomposition. Based on the Fourier transform time-frequency tradeoff principle (i.e., the Heisenberg uncertainty principle in signal processing, which states that a signal cannot simultaneously achieve arbitrarily high resolution in both the time and frequency domains; the narrower the time window, the lower the frequency resolution), the traditional short-time Fourier transform, with its fixed time-frequency window, cannot accommodate signal components with different time-frequency characteristics. The wavelet packet multi-scale decomposition operation used in this application, through an adaptive variable-scale window, ensures a long time window in the low-frequency band for extremely high frequency resolution and a short time window in the high-frequency band for accurate time resolution. The coefficients of each sub-band at the end are reconstructed, and the sum of squared modes is calculated, transforming them into a mathematical model of node energy entropy that measures the severity of disturbances in each implicit high-frequency resonant mode. This model constitutes a dense vibration feature vector that can be directly mapped nonlinearly by a neural network.

[0046] The convolutional neural network feature extraction layer in the spatiotemporal fusion prediction network mainly consists of three to five consecutive layers of two-dimensional convolutional operations, supplemented by batch normalization layers of the same dimension and an improved nonlinear smooth activation function, to prevent gradient vanishing or gradient explosion during deep backpropagation. Max pooling is used to progressively reduce the dimensionality of the image receptive field without destroying the relative topology of the space.

[0047] The Long Short-Term Memory (LSTM) network structure utilizes a tightly constructed hidden state transition and gating system to process the temporal sequence of spatial feature vectors from previous outputs. This gating system includes logistic functions as input gates, forget gates, and output gates, which respectively control the proportion of newly introduced modal information, the degree of forgetting of historical long-term memory cells, and the filtering intensity for outputting to the current hidden state and passing it layer by layer. This enables the entire network to possess temporal-domain awareness of millisecond-level temporal correlations and impact evolution.

[0048] As for the adaptive attention mechanism on which multimodal dynamic feature weighted fusion relies, its core is a global pooling self-similarity measurement based on a single-layer perceptron. The signal-to-noise ratio quality evaluation scores of different modal feature blocks are mapped to probability weight constants between 0 and 1 through a normalized exponential function, thereby dynamically shielding distorted signals and amplifying reliable signal features.

[0049] The large-scale offline training and fine-tuning process of this prediction network relies on a massive, multi-dimensional, heterogeneous database with labeled ground truth values. The database exhaustively collects data samples from more than six extreme weather conditions, ranging from extreme heat to extreme cold, as well as twenty occasional eccentric jamming attitudes, including traction plate rotation collisions and splicing steel pipe bounces, and different conductor deployment speed domains.

[0050] In the training phase, the mean squared error function is used as the loss calculation criterion for torque approximation and energy fitting, and a smoothed absolute value error function is introduced to specifically penalize the time-series tolerance at the impact prediction time, forming a joint optimization loss function. Based on this, an adaptive moment optimizer with momentum correction and support for weight decoupling decay is used to perform gradient iterative search in the network parameter space. The initial learning rate is set to 0.001, and during hundreds of iterations, a cosine annealing algorithm with a hot restart mechanism guides the nonlinear stable decay of the learning rate.

[0051] To further prevent the model from mechanically memorizing the training data, a random deactivation technique is used to nest 30% of neurons after the nodes in the fully connected layer. Combined with an early stopping algorithm that automatically terminates training if the validation error on the validation set does not decrease for ten consecutive rounds, the average root mean square error of the network under the new conditions is successfully kept between 2% and 4%, ensuring that the model has both extreme fitting accuracy and general generalization robustness.

[0052] Step S3: Dynamic energy allocation based on deep reinforcement learning. This step employs a deep reinforcement learning algorithm to perform forward-looking optimal energy allocation for the hybrid energy storage system based on the prediction results from step S2. This embodiment preferably uses the Deep Deterministic Policy Gradient (DDPG) algorithm to construct the energy management agent. First, the three elements of reinforcement learning are defined. The state space S is the basis for the agent's decision-making, including the real-time voltage V_dc of the DC bus, the state of charge (SOC) of the battery pack (SOC_bat), the state of charge (SOC_sc) of the supercapacitor pack, the current actual feedback power P_motor of the permanent magnet synchronous motor, and most importantly, the future key load prediction feature T_load_pred output from step S2. This feature is a quantitative indicator extracted from the prediction curve T_load(t) that is crucial for decision-making, such as peak power or average power. Action space A comprises the operations that the agent can execute, including the target power command P_bat_ref issued to the battery branch of the first bidirectional DC / DC converter, the target power command P_sc_ref issued to the supercapacitor branch of the second bidirectional DC / DC converter, and the pulse width modulation (PWM) duty cycle R_brake controlling the energy dissipation resistor chopper circuit. The reward function R is designed as a weighted sum of multiple objectives, in the form R = w1×R_volt + w2×R_tension + w3×R_eff. Here, w1, w2, and w3 are preset weighting coefficients used to adjust the relative importance of each reward item in the total reward function; R_volt is used to penalize deviations of the bus voltage from the target value, R_volt = -|V_dc - V_target|, where V_target is the target bus voltage; R_tension is used to penalize fluctuations in motor output tension, R_tension = -|T_out - T_target|, where T_out is the motor output tension and T_target is the target tension; R_eff is used to encourage improved energy recovery efficiency, i.e., to penalize the use of energy-consuming resistors, R_eff = -k · P_brake, where P_brake is the power consumed by the energy-consuming resistor and k is a proportionality coefficient.

[0053] In practical operation, this deep reinforcement learning agent can achieve proactive energy planning. For example, when the prediction network outputs a prediction at t=0: an impact with energy E_pulse will occur in t_impact=1.5 seconds, the agent receives this information and first checks the state of charge (SOC_sc) of the supercapacitor. If it finds that the remaining absorbable energy is less than E_pulse, the agent will output an action command 1.5 seconds before the impact occurs. For example, it sets P_sc_ref to a negative value to control the supercapacitor to pre-discharge, and simultaneously sets P_bat_ref to a positive value to slowly and efficiently transfer this energy to the battery pack, thus providing sufficient absorption margin for the supercapacitor. When the impact actually occurs at t=1.5 seconds, the motor controller feeds regenerative energy back to the DC bus, and the agent precisely instructs the supercapacitor branch to absorb the pulse energy at its maximum power, while the target power P_bat_ref of the battery branch remains stable or slightly adjusted, thus completely avoiding the impact of large currents on the battery and extending battery life.

[0054] In the training and inference process of deep reinforcement learning agents, refined mathematical modeling of the state space, action space, and multi-objective reward function is the foundation for achieving millisecond-level policy delivery. For the state space, to eliminate numerical calculation biases caused by different physical dimensions, the real-time voltage of the DC bus, the state of charge of the battery pack, the state of charge of the supercapacitor pack, and the motor feedback power are all subjected to linear normalization based on maximum and minimum values, mapped to a dimensionless closed interval from -1 to +1. For the predicted feature vector, an independent principal component analysis dimensionality reduction layer is introduced to remove redundant time series information and extract six orthogonal components representing the impact transient climb rate and steady-state extrema, which are then incorporated into the state matrix.

[0055] Regarding the action space, the system not only defines the static physical upper and lower bounds of each target power command to prevent hardware overheating and damage due to exceeding the converter's rated current capacity, but also introduces a first-order low-pass filter algorithm to smooth and limit the command variables of adjacent control cycles, forcibly constraining their step change rate within a safe boundary, thus eliminating the possibility of pole oscillations at the software source. In the reward function set under this framework, the weighting coefficients of each term are not constant, but are dynamic coefficients that are adaptively adjusted based on the operating stage. For example, in the initial acceleration and expansion stage of the tension machine, the bus voltage is prone to alternating voltage drops and pump rises. At this time, the system will automatically double the weighting coefficient associated with the voltage deviation penalty term. In the uniform and stable pulley passage stage, constant tension assessment becomes the primary requirement, and the weighting center will seamlessly and smoothly shift to the tension fluctuation penalty term.

[0056] The implementation of the deep deterministic policy gradient algorithm relies on an independent yet closely interacting actor network and critic network. Both networks are structured with a dual-hidden-layer cascade architecture containing 300 and 200 neurons respectively. The actor network uses a Tamrheic curve activation function to smoothly compress the final action command into a space with a reasonable duty cycle and per-unit power. The critic network's first hidden layer only accepts state space input, while its second hidden layer nonlinearly weights and concatenates the abstract situational information generated by the first layer with the external physical action space to predict the global long-term action value of taking a specific action under a given situation.

[0057] To break the strong correlation of time-series data and enhance the independent and identically distributed characteristics of network learning experience under different operating conditions, a priority replay buffer with a capacity of up to 2,000,000 experience dictionaries was constructed. This buffer abandoned a purely random sampling strategy, instead assigning extremely high sampling weights to rare and extreme subsets of safe-margin experiences with large temporal difference errors, such as "traction plate jamming causing a huge energy pump." The initial exploration mechanism did not use uniformly distributed Gaussian white noise, but instead specifically introduced Ornstein-Uhlenbeck decayed related noise. This ensured that the generated motion exploration trajectory had a certain degree of temporal coherence and inertia, better simulating the physical exploration characteristics of large-inertia mechanical systems, without generating high-frequency vibration commands that could damage the tension wheel gearbox.

[0058] The target network and the online master network have abandoned the periodic hard copying of weights and adopted a soft update mechanism with a moving average that is strictly set to 0.005. This measure will reduce the oscillations in the learning curve caused by overvaluation to almost invisible.

[0059] like Figure 6As shown, the electrical cabinet 209 displays its internal layout in a front sectional view. Ventilation holes are located at the top and bottom of the cabinet, a hinge is on the left, and a handle is on the right. The cabinet's interior is divided into four layers from top to bottom: the upper layer is the battery module 201, which uses a lithium iron phosphate square cell array to achieve a 100-in-10-in-parallel topology. A liquid-cooled plate 202 is configured on the right, with a microchannel serpentine pipeline structure inside, allowing the coolant to circulate in the direction of the arrows; the middle layer is the supercapacitor module 203, composed of graphene-based supercapacitor cells arranged in series, with a voltage equalization circuit 204 on the right, achieving voltage balance among the cells through a resistor network; the lower layer has a first bidirectional DC / DC converter 205 on the left and a second bidirectional DC / DC converter 206 on the right, both using a three-phase interleaved parallel topology and silicon carbide MOSFETs as switching transistors, with inductor symbols and simplified switching transistor symbols drawn internally; the bottom layer is the energy-consuming resistor module 207, including a heat sink and a sawtooth waveform resistor, with terminals at both ends. A DC busbar 208, made of copper, runs longitudinally through the right side of the cabinet. It connects to each module via connecting cables, enabling the collection and distribution of electrical energy. This structural design allows for the coordinated operation of the battery pack and supercapacitor, achieving bidirectional energy flow through a bidirectional DC / DC converter. A power-dissipating resistor provides overvoltage protection. The overall layout is compact and rational, facilitating heat dissipation and maintenance.

[0060] Step S4: Implement dual-time-scale decoupled control. This step hierarchically controls the system along the time dimension, achieving decoupling between electrical and mechanical stability. (Refer to...) Figure 4 On a fast timescale, electrical loop control is executed. Its core objective is to maintain the stability of the DC bus voltage V_dc. Specifically, the second bidirectional DC / DC converter controlling the supercapacitor bank employs high-bandwidth closed-loop current control. The response frequency of this current loop is set very high, for example, above 2kHz, which is 5 to 10 times the response frequency of the motor current loop. This ensures that it can precisely track and execute the dynamically changing supercapacitor power command P_sc_ref output by the agent in step S3 with extremely low latency. Therefore, when millisecond-level surge energy is injected into the DC bus, the supercapacitor can absorb it instantaneously, suppressing the bus voltage fluctuation within a preset range, for example, within ±2% of the rated voltage.

[0061] On a slow timescale, mechanical loop control is executed. Its core objective is to achieve a constant output of conductor tension. Since the fast timescale electrical loop has clamped the DC bus voltage V_dc to a stable level, this creates ideal operating conditions for the motor controller to execute a high-precision field-oriented control (FOC) algorithm. In the FOC algorithm, there is a crucial voltage feedforward compensation stage, which adjusts the inverter's modulation ratio based on changes in the bus voltage. In traditional systems, drastic fluctuations in the bus voltage cause frequent and significant adjustments to this compensation term, thus interfering with precise torque control. In this application, because V_dc is constant, this feedforward compensation term is almost constant, resulting in an excellent linear relationship between the motor's electromagnetic torque output and the command value. The end result is that even if the external mechanical load changes drastically in an instant, the motor can smoothly and stably output a constant electromagnetic torque, thereby maintaining constant conductor tension, as shown in the reference... Figure 4 As shown.

[0062] Step S5: Implement multi-level safety protection under extreme operating conditions. To cope with extreme impact events that exceed the system's normal processing capacity, this step designs a multi-level safety protection mechanism. When the impact energy E_pulse predicted in step S2 exceeds the maximum energy threshold that the hybrid energy storage system (battery pack and supercapacitor pack) can safely absorb under the current state, the system automatically triggers the first-level protection. In the first-level protection, the deep reinforcement learning agent, while instructing the supercapacitor and battery to absorb energy at maximum power, calculates the portion of overflow energy that cannot be absorbed and generates a precise PWM duty cycle R_brake instruction for the chopper circuit of the energy-consuming resistor. This fine adjustment method based on the PWM duty cycle makes the energy consumption process smooth and linear, avoiding the secondary voltage surges introduced by traditional relay on / off control. If the system detects that the DC bus voltage continues to rise uncontrollably after the first-level protection is activated due to excessive impact or system failure, the second-level protection is immediately triggered. Secondary protection is a mechanical braking auxiliary mechanism. The control system sends a command to the hydraulic brake caliper of the tension machine drum to apply a certain physical braking resistance. This works in conjunction with the electromagnetic braking torque generated by the motor to counteract external loads and help ensure construction safety.

[0063] Meanwhile, to address the occasional failures of multimodal sensor components that may be caused by extremely harsh on-site environments, this system is designed with a robust dimensionality reduction and fault tolerance mechanism. If the lead wire of the high-frequency accelerometer at the front end is broken by a flying stone, or if the inorganic silicate protective lens of the industrial camera is obscured by dense fog or a large area of ​​mud adhering to the wall, the multimodal comprehensive evaluator will lock the abnormal modal branch within two milliseconds using a time-frequency characteristic mutation algorithm and immediately hard reset the confidence weight of the individual failure source to zero.

[0064] At this point, the prediction network does not completely collapse and exit even when the perceived input is missing. Instead, it relies on the surviving single image or single-dimensional vibration source to retrieve a large number of modally incomplete pre-simulation model parameters injected during the offline training phase. It continues to output the estimation results of the rough boundary with a degraded and limited time prediction window (e.g., reduced from five seconds in advance to 0.5 seconds in advance).

[0065] If both sensor groups experience a common-mode failure such as simultaneous power loss or bus communication timeout, the system's internal watchdog timer will trigger the highest-level hardware reset mechanism. This completely relinquishes the interconnection permissions between the AI ​​control layer and the hybrid energy storage scheduling layer. The primary and secondary control loops simultaneously and seamlessly degrade to the classic multi-loop proportional-integral-derivative reactive power control architecture, relying entirely on the energy-consuming resistor group to perform passive overvoltage chopping and dissipation tasks. This multi-level hot standby fault degradation exercise can withstand hardware failures in any stage, ensuring the requirement of constant tension for conductor deployment under extreme conditions and avoiding the risk of serious construction accidents such as wire breakage or runaway.

[0066] Optionally, the spatiotemporal fusion prediction network in this embodiment possesses online learning capabilities. After each impact event, the system uses the actual collected load torque, impact energy, and other real data as labels, compares them with the predicted data for that event, and calculates the prediction error. This error is used to fine-tune the weight parameters within the network through a backpropagation algorithm. This continuous online self-optimization enables the prediction model to continuously adapt to changes in impact characteristics under different construction stages, different conductor specifications, and different weather conditions, achieving self-evolution of the model.

[0067] Optionally, to meet the high pulse power requirements, the supercapacitor bank in this embodiment is preferably composed of a graphene-based supercapacitor module. This type of supercapacitor has a low equivalent series resistance and a high charge / discharge rate exceeding 100C, and its physical characteristics are highly compatible with the application scenario of absorbing millisecond-level high-power impulse energy in this application.

[0068] Example 2 This embodiment provides an energy recovery system for a high-voltage line tensioner driven by a DC power supply, which is a hardware implementation of the method described in Embodiment 1. (Refer to...) Figure 1 The system includes: The hybrid energy storage and multimodal sensing device forms the physical basis of the system. This device includes a permanent magnet synchronous motor and its controller as the power actuator, a DC bus as the energy exchange bus, a battery pack connected to the DC bus via a first bidirectional DC / DC converter, a supercapacitor group connected to the DC bus via a second bidirectional DC / DC converter with high-frequency response capability, an energy-dissipating resistor and its control circuit for dissipating excess energy, and a multimodal sensor group installed at the wire-laying trolley location. This sensor group includes a high-speed industrial camera and a high-frequency accelerometer.

[0069] The impact load spatiotemporal prediction module executes the prediction algorithm in step S2 of Embodiment 1. This module is typically implemented using a high-performance embedded controller or industrial computer. It communicates at high speed with the cameras and sensors in the multimodal sensing device, receiving real-time image streams and vibration signals. Internally, the module contains a pre-trained spatiotemporal fusion prediction network model that integrates convolutional neural networks, long short-term memory networks, and adaptive attention mechanisms. This model processes the input data and outputs quantitative prediction information for future impact loads, including the predicted torque curve, impact time, and impact energy.

[0070] The energy dynamic allocation module executes the energy management strategy in step S3 of Embodiment 1. This module is connected to the impact load spatiotemporal prediction module and the various controllers in the hybrid energy storage device. It receives future load information output by the prediction module and real-time status information (such as bus voltage, SOC of each energy storage unit, etc.) collected from the system, and runs a deep reinforcement learning algorithm (such as the DDPG algorithm). Based on the current state and future predictions, this module generates optimal control commands for the battery branch, supercapacitor branch, and energy-consuming resistor, such as power reference values ​​or PWM duty cycles.

[0071] The dual-timescale decoupling control module executes the hierarchical control logic in step S4 of Example 1. This module receives instructions from the energy dynamic allocation module. It comprises two sub-parts: a fast-timescale controller for high-bandwidth current loop control of the second bidirectional DC / DC converter to strictly execute the supercapacitor's power command, thereby quickly stabilizing the DC bus voltage; and a slow-timescale controller, i.e., the motor controller itself, which executes a field-oriented control algorithm under the stable bus voltage environment created by the fast-timescale controller to achieve precise and smooth control of the motor's output torque, thereby ensuring constant conductor tension.

[0072] The multi-level safety protection module performs the safety measures in step S5 of Embodiment 1. This module continuously monitors whether the predicted impact energy exceeds the system's processing capacity. In the event of an over-limit event, the module first activates the precise PWM control function of the energy-consuming resistor to dissipate the overflow energy. If the primary protection measures fail to effectively suppress the voltage rise, the module immediately issues a command to activate the mechanical braking device on the tensioner, providing additional physical braking to ensure the ultimate safety of the system.

[0073] Example 3 This embodiment illustrates the necessity of the method and system provided in this application through a specific, high-difficulty application scenario. This scenario involves the overhead line construction of a large-scale cross-river project, requiring the deployment of eight-split ultra-high voltage conductors over a span exceeding 2000 meters. Such applications have high technical requirements. First, to control the sag of long-distance conductors, a high and constant tension value must be applied and maintained at extremely low speeds. Second, the size and weight of accessories such as traction plates and splicing pipes used with eight-split conductors far exceed those of conventional conductors, generating high-energy, short-duration impact loads when passing through the laying pulley. Conventional energy handling solutions, namely the "battery plus energy-dissipating resistor" architecture, cannot meet this challenge. Because its control logic is passively responsive, the system only begins to operate after the impact occurs, by which time the high-power regenerative energy has already caused a sudden increase in the DC bus voltage. To protect power devices, the system can only take measures such as torque derating or activating energy-dissipating resistors. Both of these measures significantly change the motor's output torque within milliseconds, leading to the failure of constant tension control. Under long span and high tension conditions, such instantaneous tension fluctuations can cause the conductor to generate uncontrollable low-frequency vibrations, posing a direct threat to construction safety. Therefore, conventional methods are not applicable in this scenario.

[0074] The above-mentioned problems can be solved by adopting the technical solution of this application. After construction begins, the system operates stably according to steps S1 to S5. When the heavy traction plate approaches the wire-laying trolley with the conductor, the spatiotemporal fusion prediction network in step S2 accurately predicts 2.8 seconds in advance that the traction plate will touch the trolley at time t_impact by fusing visual data from a high-speed industrial camera and vibration precursor signals from an accelerometer, and quantifies and predicts that a high-amplitude pulse energy E_pulse will be generated. Upon receiving this prediction information, the deep reinforcement learning energy management agent in step S3 immediately enters the active planning stage. The agent first checks that the state of charge (SOC_sc) of the supercapacitor bank is 80%, and its remaining absorbable capacity is insufficient to fully accommodate the predicted pulse energy E_pulse. Within the 2.8-second time window before the impact, the agent outputs a command to control the second bidirectional DC / DC converter to recharge part of the electrical energy stored in the supercapacitor back to the battery bank at a stable power, reducing SOC_sc to 35% before the impact, thereby reserving sufficient absorption margin for the upcoming energy pulse. When the traction plate actually impacts the trolley, a regenerative energy pulse fed back by the motor is injected into the DC bus. At this time, the dual-timescale decoupling control mechanism in step S4 comes into play. The fast-timescale controller responds with extremely high bandwidth, instructing the second bidirectional DC / DC converter to guide all of this energy pulse rapidly to the ready supercapacitor bank. During this process, fluctuations in the DC bus voltage are suppressed within ±1.5% of the target voltage, remaining stable throughout. Due to the stable bus voltage, the slow-timescale motor controller can execute its constant tension control algorithm accurately and undisturbed, maintaining a smooth electromagnetic torque output by the motor, ultimately ensuring that the conductor tension remains constant throughout the impact process. This implementation process demonstrates that this application, by transforming passive response into active prediction and planning, fundamentally decouples the contradiction between electrical stability and mechanical control, meeting the high-precision control requirements under extreme operating conditions.

[0075] To verify the practical technical effectiveness of the proposed DC-powered high-voltage line tensioning machine energy recovery method and system in controlling constant tension and stabilizing the DC bus voltage boundary layer, the research team constructed a proportional energized test platform simulating a three-span continuous heavy-duty line laying condition at the UHVDC transmission test base. This physical testing system includes a 1000m full-scale large-scale laying test line, each equipped with a stainless steel main laying trolley assembly with four pulley blocks.

[0076] All cables used in the installation process have a cross-sectional area of ​​630mm². 2The system uses steel-cored aluminum stranded wire with extremely high tensile strength, densely connected in series at 30m intervals with anti-torsion flywheels and extension splicing hardware traction assemblies weighing over 70kg. The basic cable laying operation was strictly anchored as follows: an initial constant tension of 40kN and a constant cable speed of 40m / min. Under these stringent experimental conditions, two comparative examples were designed. Comparative example one represents the limitations of traditional, outdated technology, equipped only with a battery pack and employing a pure feedback compensation proportional-integral-derivative control architecture lacking predictive capabilities. Comparative example two represents a cutting-edge solution; although it adds a supercapacitor bank to form a hybrid energy storage system, its energy dispatch system relies entirely on fixed binary fuzzy logic rules with manually set thresholds, lacking artificial intelligence predictive intervention.

[0077] A high-speed multi-channel waveform recorder captured two seconds of data waveform at the moment the traction plate struck the pulley system. Detailed data comparison revealed significant differences between the three technical approaches. When a sudden tensile force of up to 45 kN·m was generated and instantly converted back into a massive regenerative fall-off electromagnetic braking pulse injection system, the comparative example, lacking an absorption buffer, experienced a sudden surge in the nominal voltage of the DC powertrain bus from a smooth 600V to approximately 780V, triggering an alarm, within 10 ms. This voltage fluctuation amplitude exceeded the ±80V limit threshold. This forced its reactive safety system to maintain full load on the energy-consuming resistor for an extended period, causing the motor torque to drop sharply from 40,000 N·m to below 20,000 N·m, resulting in a severe infrasonic whipping effect with an amplitude exceeding 4 m on the overhead conductors. Comparative Example 2 passively identified the rise in bus voltage 10ms after being impacted before starting to drive the supercapacitor to smooth out peaks and valleys. The bus fluctuation range still reached ±50V, and the recovery time of the static error after the sudden change in rotor electromagnetic torque was a slow 1.2s.

[0078] In stark contrast, the multi-source data deep mining and active scheduling technology system described in this application accurately predicted the impact energy peak with a lead time of up to 2.8 seconds based on a spatiotemporal fusion network. By issuing energy transfer and reversal commands, the supercapacitor absorbed the total energy of the pulse pump within the nanosecond instant of the impact. Waveform recorders precisely confirmed that, under the control of the system in this application, the DC bus voltage was maintained near the 600V baseline. Due to the effective operation of the decoupling mechanism, the transient peak-valley voltage was locked within a small ±15V absolute control band. The high-level magnetic field orientation algorithm built upon this micro-fluctuation achieved a high degree of cleanliness in its stator current tracking command, and the recovery time of the motor electromagnetic braking torque was controlled within 0.15 seconds with minimal fluctuations. In this complete 1000m discharge cycle, statistical calculations using a high-precision bidirectional DC energy meter showed that the traditional comparative example, due to excessive braking and heating, only maintained a bidirectional conversion and recovery efficiency of 65% for the overall feeder of its discharge system. In contrast, this application, relying on the efficient absorption and conversion of potential energy, achieved a comprehensive green energy recovery efficiency of over 88%. Multiple objective experimental evidence chains demonstrate that this scheme, to a certain extent, reduces the strong coupling threat posed by the feedback of drastic fluctuations in external loads to the stability of the core busbar, ensuring mechanical balance within the extremely confined space of long-span UHV transmission lines.

[0079] like Figure 7 As shown, when a tension impact event occurs at t=500ms, the transient response characteristics of the DC bus voltage of the three control schemes show significant differences. Comparative Example 1 uses traditional PID control, relying solely on the battery pack and energy-consuming resistor for energy management. Within 10ms after the impact, the bus voltage surges to 780V, exceeding the target voltage of 600V by 180V, with a voltage fluctuation range of ±80V. The oscillation decays slowly, returning to around 650V after approximately 1500ms, still exhibiting a 50V deviation. Comparative Example 2 uses a hybrid energy storage system with fuzzy logic control. It begins responding 10ms after the impact, with a peak voltage of approximately 650V and a fluctuation range of ±50V. It returns to 610V after approximately 800ms, showing improved response speed and stability, but still not ideal. The scheme in this application uses an AI prediction module to detect the abnormal tension trend 2.8s before the impact, triggering pre-discharge of the supercapacitor in advance, reducing its SOC from 80% to 35%, reserving sufficient absorption space for the upcoming energy impact. Therefore, at the moment of impact, the bus voltage only experiences a small fluctuation, with a peak value of only 615V and a fluctuation range controlled within ±15V. It also quickly returns to a steady-state range of 600V±5V within 50ms. This result demonstrates that the deep reinforcement learning scheduling strategy combined with predictive control proposed in this application can suppress bus voltage fluctuations to within 18.75% of that of traditional schemes, significantly improving the voltage stability of the DC bus and effectively protecting power electronic devices from overvoltage impacts.

[0080] like Figure 8 As shown, when the tension impact occurs at t=500ms, the electromagnetic torque recovery dynamic characteristics of the three control schemes exhibit significant differences. With the target torque set at 40000 N·m, Comparative Example 1, using traditional PID control, shows a sharp drop in electromagnetic torque to 18000 N·m after the impact, a decrease of 55%. Due to severe bus voltage fluctuations and insufficient energy recovery capacity, torque recovery is extremely slow, only recovering to 35000 N·m by the end of the 2000ms observation window, failing to reach the target torque and severely impacting the continuity and safety of overhead line operations. Comparative Example 2, using a hybrid energy storage system with fuzzy logic control, shows a torque drop to 28000 N·m after the impact, a decrease of 30%. Through fixed-rule fuzzy logic scheduling of the battery pack and supercapacitor working together, the torque recovers to 39000 N·m after 1200ms, reaching 98% of the target torque, with a recovery time of 1200ms. This application employs a deep reinforcement learning scheduling strategy to pre-discharge the supercapacitor before the impact. Upon impact, the supercapacitor immediately absorbs regenerative energy at maximum power, strictly controlling bus voltage fluctuations. The electromagnetic torque drops only to 37,000 N·m, a decrease of only 7.5%, and rapidly recovers to the target value of 40,000 N·m within 150 ms, a recovery time only 12.5% ​​of that of Comparative Example 2. These results demonstrate that the dual-timescale decoupling control strategy of this application can respond rapidly to impact events within milliseconds, significantly shortening the torque recovery time, ensuring continuous and stable control of conductor tension during overhead line operations, and avoiding safety hazards such as conductor slack or over-tension caused by torque fluctuations.

[0081] like Figure 9As shown, the three control schemes exhibit fundamental differences in energy recovery efficiency and energy flow distribution during tension impact events. Comparative Example 1 uses traditional PID control, with only a battery pack and energy-consuming resistor. Only 45% of the regenerated energy is recovered by the battery pack, 35% is passively consumed as heat through the energy-consuming resistor, and 20% is lost in the lines and converters, resulting in a comprehensive energy recovery efficiency of only 65%, with a significant amount of valuable regenerated energy wasted. Comparative Example 2 uses a hybrid energy storage system with fuzzy logic control, equipped with a battery pack and supercapacitor. However, due to the use of a fixed-rule scheduling strategy, the supercapacitor absorbs only 30% of the regenerated energy, the battery pack recovers 25%, 25% is still consumed through the energy-consuming resistor, and 20% is lost in the system, increasing the comprehensive recovery efficiency to 75%, but there is still considerable room for improvement. This application's solution achieves intelligent optimization of energy flow allocation through a deep reinforcement learning scheduling strategy: The supercapacitor, leveraging its high power density, rapidly absorbs 52% of the pulsed regeneration energy, effectively suppressing bus voltage spikes; the battery pack recovers 28% of energy through a stable charging method, avoiding large current surges; after the surge, the supercapacitor transfers 8% of the energy to the battery pack via a bidirectional DC / DC converter, achieving energy rebalancing within the energy storage system; the energy-consuming resistor is used only in small quantities under extreme overflow conditions, reducing its consumption ratio to 4%; and system losses are reduced to 8% through optimized control strategies. The final overall energy recovery efficiency reaches 88%, a 35.4% improvement over Comparative Example 1 and a 17.3% improvement over Comparative Example 2. These results demonstrate that the AI ​​prediction and deep reinforcement learning scheduling strategy of this application can dynamically optimize energy flow based on real-time operating conditions, maximizing the complementary characteristics of the hybrid energy storage system, significantly improving energy recovery efficiency, reducing the energy consumption cost of overhead line operations, and aligning with the development direction of green energy conservation.

[0082] Beyond the aforementioned main embodiment and its direct evolution system, the core defense and general algorithmic paradigm of the "cross-modal prior perception and multi-scale temporal decoupling constant servo control model under the background of multi-terminal multi-state asymmetric energy exchange" proposed in this application, after being extended through module-to-module substitution and superset generalization mapping of parameter arrays, also covers and extends to a series of more extensive alternative and equivalent operational architectures. Analyzing from the alternative dimension of upstream beyond-line-of-sight sensor architecture, when tension machines are deployed for extended periods in coastal high-salt fog, perennial dense fog in high mountains, or in a very few special application scenarios requiring uninterrupted cable laying across midnight in all weather conditions, the confidence level of a single high-speed industrial camera's visual input will drop exponentially or even face complete blockage due to the irreversible physical attenuation of visible spectrum photon penetration and the blinding effect of total internal reflection of white light from a powerful searchlight compensation source in water vapor. Therefore, this optical information capture submodule can be replaced and upgraded in situ at cost to a three-dimensional lidar imager or an alternating system of an uncooled focal plane array long-wave infrared thermal imager with multi-view wavelengths at 905nm and 1550nm.

[0083] Based on point cloud 3D matrix projection or thermal imaging that relies on temperature gradients to delineate contour features, the interface of its multimodal attention fusion channel only needs to accept the underlying feature space registration mapping pre-transformation to be seamlessly mounted on the existing multi-feature nonlinear time series mining network. In this way, the entire decoupled prediction assembly can continue to operate without absolute dependence on light and dust, and its anti-slip warning perception boundary is extended from the traditional two-dimensional planar model to a more three-dimensional spatial depth detection.

[0084] The focus then shifts to the variant construction of the deep temporal inference framework of the model. Considering the evolving hardware acceleration preferences of embedded industrial microcomputers and automotive-grade coprocessors, the deep long short-term memory gated recurrent structure of the spatiotemporal fusion prediction network kernel can not only be seamlessly replaced by the gated recurrent unit (GRU), which has lower computational cost and further reduced latency, but can also be completely reconstructed into a visual-temporal transformation network model architecture with a large-scale global parallel self-attention mechanism. By connecting the image spatial domain segmentation module with a time-encoded multi-head attention mechanism, the transformation mechanism model often exhibits a sharper approximation mapping fit than traditional recurrent gating in capturing the macroscopic cognitive scale of extremely long-period, slowly varying physical impact evolution characteristics exceeding 10 seconds, making it the best alternative algorithm engine for solving the prediction of oscillations caused by multi-crossing construction.

[0085] In the replacement strategy of energy storage buffer hardware carriers after decision execution, for some large-tonnage constant tension platform systems that require maintenance-free cycles spanning several decades or even semi-permanent, stationary operation, the organic mixed electrolyte of the supercapacitor cluster still suffers from the drawbacks of evaporation and slow-changing internal resistance. Therefore, the system allows for the physical hard replacement of the second set of bidirectional high-speed energy chopper buffer branches with a vacuum-suspended flywheel energy storage motor array based on carbon fiber polymer as the main axis, or a high-silicon negative electrode lithium titanate battery array with a wider and more comfortable range and extremely high rate of operation at destructive safety limits such as needle penetration and deep charge-discharge. Although these energy transmission relay entities differ significantly in their electromechanical and physical implementation, their microsecond-level charge-discharge time delays, as well as the response characteristics and feedback logic of their electrical interfaces to the dynamic commands of the deep-enhanced algorithm network, are completely equivalent to the aforementioned graphene-structured capacitor clusters, and have not deviated from the general extensional boundary of the preceding control topology.

[0086] Against the backdrop of the macro-application of general cross-domain empowerment in industrial automation, this application addresses the core concept and system of solving complex negative interactive interference: "to resolve and smooth transient pulse energy at the predictive front line to isolate the deadlock loop of mutual deterioration between high-voltage power supply and underlying mechanical resistance." Through specialized adaptation and optimization of the sensing and prediction module and the feedback braking mechanical module, this system can be seamlessly transplanted and applied to other heavy-duty high-end flexible cable equipment fields with large inertial loads and similar gap impact damage capabilities through simple structural modifications. For example, in the heave and lateral wave compensation active decoupling constant tension winch operation system on a deep-sea mining retrieval and lifting mother ship thousands of meters deep, the system only needs to replace the target of feature extraction with ocean current turbulent waves and the ship's six-degree-of-freedom sway data; this framework is also effective in the wire rope weightlessness multi-stage anti-collision constant tension suspension cable laying and submersion device of a large-scale modern deep vertical shaft mine hoisting system. This signifies that the theoretical completeness and generalized adaptability of this technical proposal in overcoming all rope-based precision mechanical maintenance systems with sudden large-level disturbances and nonlinear cross-coupling of multiple variables possess a thorough and extremely high potential for universal coverage. This further demonstrates that the core technical logic, topological architecture, and specific combinations of polymorphic modules described above, without fundamentally altering the original technical intent and underlying principles, constitute a derivative boundary of equal technical inspiration and equivalent protection across many similar or related technical directions.

[0087] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for energy recovery from a DC power supply-driven high-voltage line tensioning machine, characterized in that, Includes the following steps: Step 1: Construct a hybrid energy storage system and a multimodal sensor array. The hybrid energy storage system includes a DC bus, a battery pack branch, a supercapacitor pack branch, a motor, a motor controller, and a power-consuming resistor. The battery pack branch includes a battery pack and a first bidirectional DC / DC converter, with the battery pack connected to the DC bus via the first bidirectional DC / DC converter. The supercapacitor pack branch includes a supercapacitor pack and a second bidirectional DC / DC converter, with the supercapacitor pack connected to the DC bus via the second bidirectional DC / DC converter. The multimodal sensor array is located at the wire-laying trolley of the tension machine and is used to collect image information of the conductor and splicing fittings, as well as vibration signals from the wire-laying trolley. Step 2: Based on the image information and vibration signal collected by the multimodal sensor group, the impact load within the future time window is predicted by a pre-built prediction model to obtain load prediction information including the time of impact and the magnitude of impact energy. Step 3: Using a deep reinforcement learning agent, based on the real-time voltage of the DC bus, the state of charge of the battery pack, the state of charge of the supercapacitor pack, the feedback power of the motor, and the load prediction information, generate and output optimal control commands for the battery pack branch, the supercapacitor pack branch, and the energy-consuming resistor to perform dynamic energy allocation. Step four: When the magnitude of the impact energy predicted in step two exceeds the maximum energy threshold that the hybrid energy storage system can absorb, a safety protection strategy is executed to dissipate the overflow energy through the energy-consuming resistor.

2. The method according to claim 1, characterized in that, The multimodal sensor group mentioned in step one includes an industrial camera for acquiring the image information and an accelerometer for acquiring the vibration signal.

3. The method according to claim 1, characterized in that, The prediction model described in step two is a spatiotemporal fusion prediction network. The spatiotemporal fusion prediction network uses a convolutional neural network to extract spatial features from the image information and combines a long short-term memory network to process time-related sequence features. At the same time, the spatiotemporal fusion prediction network introduces an adaptive attention mechanism to dynamically weight and fuse features from different modalities to output quantitative prediction values ​​for the load torque prediction curve, the predicted time of impact, and the magnitude of the pulse energy generated by the impact within the future time window.

4. The method according to claim 3, characterized in that, The spatiotemporal fusion prediction network has online learning capabilities, which can compare the real load data collected during the actual impact with the predicted data of the impact, calculate the prediction error, and update the weight parameters of the spatiotemporal fusion prediction network through the backpropagation algorithm.

5. An energy recovery system for a high-voltage line tensioning machine driven by a DC power supply, characterized in that, include: A hybrid energy storage system and a multimodal sensor array are provided. The hybrid energy storage system includes a DC bus, a battery pack branch, a supercapacitor pack branch, a motor, a motor controller, and a power-dissipating resistor. The battery pack branch includes a battery pack and a first bidirectional DC / DC converter, with the battery pack connected to the DC bus via the first bidirectional DC / DC converter. The supercapacitor pack branch includes a supercapacitor pack and a second bidirectional DC / DC converter, with the supercapacitor pack connected to the DC bus via the second bidirectional DC / DC converter. The multimodal sensor array is located at the wire-laying trolley of the tension machine and is used to acquire image information of the conductor and splicing fittings, as well as vibration signals from the wire-laying trolley. The prediction module is used to predict the impact load within a future time window based on the image information and vibration signal collected by the multimodal sensor group, and to obtain load prediction information including the time of impact and the magnitude of impact energy. The energy dynamic allocation module is used to generate and output optimal control commands for the battery pack branch, the supercapacitor branch and the energy-consuming resistor by using a deep reinforcement learning agent based on the real-time voltage of the DC bus, the state of charge of the battery pack, the state of charge of the supercapacitor pack, the feedback power of the motor and the load prediction information, so as to perform dynamic energy allocation. The safety protection module is used to execute a safety protection strategy when the magnitude of the impact energy predicted by the prediction module exceeds the maximum energy threshold that the hybrid energy storage system can absorb, and to consume the overflow energy through the energy-consuming resistor.

Images