An Adaptive UHPC-Reinforced Concrete Tower Structure Based on Digital Twin and Intelligent Decision-Making and Its Seismic Control Method
Patent Information
- Application Number
- CN202610439633.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-03
- Publication Date
- 2026-05-26
Smart Images
Figure CN122082937A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wind power structure seismic resistance technology, specifically relating to an adaptive UHPC-steel-concrete tower structure based on digital twin and intelligent decision-making and its seismic control method, which is applicable to tall wind power tower structures in high-intensity earthquake zones. Background Technology
[0002] With the transformation of the global energy structure and the deepening of the "dual carbon" goal, wind power technology is rapidly developing towards larger single-unit capacity and taller hub towers, with tower heights reaching 140 to 200 meters. This development trend places higher demands on the load-bearing capacity and safety of the tower structure, especially in high-seismic-intensity areas, where the seismic challenges faced by tall wind turbine towers are becoming increasingly severe. Ultra-high performance concrete (UHPC), due to its superior compressive strength, durability, and toughness, has been explored for application in steel-concrete composite tower structures in recent years. The aim is to reduce the structural weight while improving the overall stiffness and fatigue resistance of the tower, thereby meeting the support requirements of large wind turbine units.
[0003] However, after in-depth analysis, existing steel-concrete composite towers based on UHPC still suffer from several insurmountable technical defects in seismic design. First, the seismic resistance mechanism of existing towers is essentially a passive response mechanism. Their seismic performance relies entirely on the material's inherent strength, cross-sectional geometry, and passive damping energy dissipation, failing to actively adjust based on the real-time seismic spectral characteristics, duration, and energy distribution. This rigid design is highly susceptible to coupling with the tower's natural frequency when encountering near-fault pulse-type ground motions containing long-period velocity pulses or long-period ground motions with energy concentrated in the low-frequency range, leading to strong resonance damage and premature structural failure. Second, at the structural level, the interface between traditional steel and concrete sections often exhibits significant stiffness abrupt changes. Under seismic cyclic loading, this area experiences severe stress concentration, directly causing premature tensile cracking, crushing, or even peeling of the UHPC layer from the steel cylinder surface, significantly weakening the collaborative performance and durability of the composite structure. Furthermore, existing wind turbine tower health monitoring systems typically only have data recording and threshold alarm functions. The collected strain, acceleration, and other data cannot form a closed-loop control at the structural level. In other words, they lack a comprehensive collaborative control capability that integrates "real-time structural state perception—rapid inversion of seismic characteristics—intelligent decision-making on optimal control parameters—online adjustment of structural physical parameters," resulting in a lack of effective proactive intervention in the structure's response and damage evolution throughout the earthquake process. Finally, traditional towers have serious shortcomings in post-earthquake recovery. Damage assessment heavily relies on manual high-altitude climbing inspections, which is not only time-consuming, inefficient, and risky, but also yields subjective and inaccurate results. Coupled with the lack of a design that facilitates rapid repair, this leads to prolonged turbine shutdowns after earthquakes, resulting in significant power generation losses and increased maintenance costs.
[0004] The inventors of this application discovered in their research that the root cause of the deficiencies in existing technologies lies in the lack of an adaptive system capable of deeply integrating intelligent material regulation, active structural control, and real-time state inversion. Therefore, there is an urgent need to provide a novel UHPC-reinforced concrete tower structure and its seismic control method to solve the aforementioned problems. Summary of the Invention
[0005] This application provides an adaptive UHPC-steel-concrete tower structure and its seismic control method based on digital twins and intelligent decision-making. The purpose is to improve the problems mentioned in the background section. The core is to introduce intelligent variable stiffness materials, active control devices and cloud-based collaborative intelligent decision-making systems to construct a closed-loop intelligent seismic resistance system of "perception-inversion-decision-execution", thereby significantly improving structural safety and post-earthquake recoverability.
[0006] To solve the above-mentioned technical problems, the technical solution proposed in this application is as follows:
[0007] This invention provides an adaptive UHPC-steel-concrete tower structure based on digital twins and intelligent decision-making, comprising the following components connected sequentially from bottom to top: Concrete foundation; The intelligent variable stiffness UHPC functional gradient tower section has a wall consisting of an outer UHPC main load-bearing layer, a variable stiffness intelligent interlayer, and an inner UHPC counterweight energy dissipation layer from the outside to the inside. The variable stiffness intelligent interlayer is made of a controllable stiffness material and has an excitation device embedded inside to achieve continuous adjustment of the overall stiffness of the wall. Top steel section; as well as The top active control device, installed between the top steel section and the nacelle, includes a movable mass block capable of bidirectional movement in the horizontal plane and its driving mechanism. The multi-source sensing and edge computing unit includes a distributed optical fiber sensor array, an acceleration sensor array, and an environmental monitoring sensor located at the base of the tower, all arranged along the tower height, as well as an edge computing unit for real-time processing of sensor data. The cloud-based digital twin platform is connected to the edge computing unit via a wireless transmission module and has a built-in nonlinear damage constitutive model for UHPC materials and a control model for the variable stiffness smart sandwich layer. The intelligent decision controller, deployed in the edge computing unit or the cloud-based digital twin platform, is configured as a data-driven intelligent decision model. It takes seismic characteristics, structural response, and damage indicators as inputs and outputs the optimal control command for the variable stiffness intelligent interlayer and the optimal driving force command for the top active control device.
[0008] Furthermore, the variable stiffness intelligent interlayer adopts magnetorheological elastomer material or electrorheological elastomer material; the excitation device embedded inside is uniformly arranged along the circumference of the tower and forms multiple independent control zones in the height direction of the tower; the intelligent decision controller is configured to independently adjust the excitation intensity of each of the independent control zones to achieve asymmetric adjustment of the lateral stiffness of the tower, thereby actively controlling the torsional response under seismic action.
[0009] Furthermore, the equivalent shear modulus of the variable stiffness intelligent sandwich layer can be continuously adjusted within the range of Gmin to Gmax, and the response time from Gmin to Gmax is less than 20 milliseconds; the overall fundamental frequency of the tower structure has an adjustable bandwidth of not less than 30% under the adjustment of the variable stiffness intelligent sandwich layer.
[0010] Furthermore, the mass of the movable mass block of the top active control device is 3% to 8% of the total mass of the tower top, and its maximum stroke is not less than ±400mm; the drive mechanism is a ball screw type or linear motor type servo drive system, which can provide active control acceleration of 0.1g to 1.0g.
[0011] Furthermore, the distributed optical fiber sensor array adopts a helical winding arrangement with a winding angle of 45°±5°, which is used to simultaneously calculate the bending modal strain and torsional modal strain of the tower; the accelerometer array includes at least three-dimensional accelerometers respectively set at the root, middle and top of the tower, which are used to identify the translational vibration mode and rocking vibration mode of the tower.
[0012] On the other hand, this application also claims protection for a seismic control method applied to the aforementioned adaptive UHPC-steel-concrete tower structure based on digital twins and intelligent decision-making, comprising the following steps: Step S1: Rapid identification of ground motion characteristics: The P-wave signal is captured by the accelerometer at the base of the tower. The time-frequency features of the P-wave signal are extracted by wavelet packet decomposition and input into a pre-trained lightweight convolutional neural network to quickly predict the dominant frequency range and potential destructive force level of the upcoming S-wave. Step S2: Real-time inversion of digital twin and damage field: Based on strain field data and acceleration data collected by distributed fiber optic sensors, combined with data assimilation algorithm, the instantaneous stiffness distribution and damping distribution of the tower are inverted in real time in the cloud digital twin platform, and the three-dimensional damage field index of each section is calculated according to the nonlinear damage constitutive model of UHPC material. Step S3: Multi-objective collaborative optimization and decision-making: The dominant frequency range of the shear wave predicted in step S1, the instantaneous structural state parameters inverted in step S2, and the real-time tower top displacement response are combined into a state space; an intelligent decision-making model based on deep reinforcement learning is adopted to output the optimal control action sequence under the optimization objectives of "minimizing the peak displacement at the tower top", "minimizing the overall structural damage index", and "minimizing control energy consumption". The action sequence includes: the target equivalent stiffness value of each independent control zone and the target control force time history of the top active control device; Step S4: Real-time coordinated adjustment of structural parameters and load path: The edge computing unit sends decision commands to each actuator; the variable stiffness intelligent interlayer changes its equivalent shear modulus according to the command, thereby reconstructing the lateral stiffness distribution of the tower, so that its instantaneous fundamental frequency deviates from the dominant seismic frequency in real time, achieving frequency avoidance; at the same time, the top active control device drives the mass block to move according to the command, generating a reverse inertial force, forming a time-domain active control force that complements the frequency domain of the variable stiffness adjustment; Step S5: Model self-evolution and strategy iteration: After the earthquake, the structural response data, control commands and external ground motion records of the entire process are uploaded to the cloud digital twin platform as experience samples. The deep reinforcement learning model is periodically retrained offline, the model parameters are updated and distributed to the edge computing unit to realize the continuous evolution of the control strategy.
[0013] Furthermore, the deep reinforcement learning model in step S3 adopts a multi-agent reinforcement learning framework, wherein each independent control partition is defined as an agent and the top active control device is defined as another agent; under the framework of centralized training and distributed execution, the agents form a collaborative control strategy through collaborative learning to solve the curse of dimensionality problem in high-dimensional action space.
[0014] Furthermore, after the real-time inversion of the three-dimensional damage field index of the tower in step S2, a predictive control step is also included: the currently inverted damage field is used as the initial condition and input into the damage evolution model built into the digital twin platform to predict the damage propagation path in the next 1-2 seconds under three conditions: no control, existing control, and ideal control; the intelligent decision model uses the predicted damage propagation path as one of its decision-making bases to achieve forward-looking suppression of structural damage.
[0015] Furthermore, after predicting the dominant frequency range of the shear wave in step S1, a risk level prediction and strategy pre-activation step is also included: if the difference between the predicted dominant frequency and the current fundamental frequency of the tower is less than a preset threshold, it is determined to be high risk, and the variable stiffness intelligent mezzanine is immediately pre-set to a high response standby state, while the servo driver of the top active control device is pre-charged to eliminate the response delay of the actuator.
[0016] Furthermore, the multi-objective optimization function in step S3 is expressed as:
[0017] Where max|y(t)| is the peak displacement at the top of the tower, y lim Permissible displacement limit; D max D represents the maximum damage index of the entire tower. crit E is the critical damage index. consump To control total energy consumption, E max To maximize the allowable energy consumption, the weighting coefficients w1, w2, and w3 are dynamically and nonlinearly adjusted based on the real-time earthquake risk level and structural importance coefficient. When the peak value of the earthquake P-wave acceleration exceeds the threshold, the weights of w1 and w2 are automatically increased.
[0018] Furthermore, it also includes self-diagnosis and fault-tolerant control steps: the edge computing unit monitors the health status of each sensor and actuator in real time; when a sensor failure is diagnosed, it automatically switches to a data-driven virtual sensing mode, and reconstructs the physical quantity of the failure location using data from adjacent sensors and a digital twin model; when a variable stiffness partition actuator failure is diagnosed, it automatically re-optimizes the stiffness distribution of the remaining normal partitions, and maintains the overall control performance of the tower through functional reconstruction.
[0019] Furthermore, the lightweight convolutional neural network pre-trained in step S1 takes as input the first 2-second P-wave signal captured by the root accelerometer, and its output layer uses the Softmax function. The output result is the probability distribution of the earthquake dominant frequency belonging to the predefined frequency bands [0.1Hz, 0.5Hz], [0.5Hz, 1.0Hz], and [1.0Hz, 2.0Hz], and the frequency band corresponding to the highest probability is used as the prediction result.
[0020] Furthermore, the frequency domain allocation strategy for the coordinated control between the variable stiffness intelligent interlayer and the top active control device is as follows: the variable stiffness intelligent interlayer is mainly responsible for adjusting the fundamental frequency of the tower to be more than 0.2Hz below the dominant seismic frequency, thereby achieving "softening avoidance"; the top active control device is mainly responsible for suppressing the residual vibration energy that still exists after adjustment, with a frequency range of 1.0Hz to 5.0Hz.
[0021] Compared with the prior art, the present invention achieves the following beneficial technical effects: This invention achieves millisecond-level active adjustment of tower stiffness and frequency avoidance under seismic loading by introducing variable stiffness intelligent sandwich and digital twin technology, fundamentally avoiding resonance damage. Combined with deep reinforcement learning and multi-objective optimization, it realizes coordinated control of structural displacement, damage and energy consumption, and the control strategy can self-evolve. At the same time, it integrates cloud-based digital twin and self-diagnostic fault-tolerant mechanisms, which significantly improves the structure's adaptive capability, post-earthquake recoverability and full life cycle economy. Attached Figure Description
[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 A flowchart illustrating the seismic control method for the adaptive UHPC-steel-concrete tower structure based on digital twins and intelligent decision-making, provided in an embodiment of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0025] It should be noted that traditional steel-concrete hybrid tower structures typically include a concrete foundation, a UHPC-steel-concrete transition section, a UHPC concrete section, and a top steel section, connected sequentially from bottom to top. Unlike these traditional structures, this invention creatively proposes an "intelligent variable stiffness UHPC functional gradient tower section." This section completely replaces and integrates the traditional UHPC-steel-concrete transition section and UHPC concrete section, integrating stiffness transition, load-bearing capacity, and energy dissipation into a single functional gradient tower section.
[0026] In one embodiment of this application, an adaptive UHPC-reinforced concrete tower structure based on digital twins and intelligent decision-making is provided. The structure includes, from bottom to top, a concrete foundation, an intelligent variable stiffness UHPC functional gradient tower section, and a top steel section.
[0027] Specifically, the concrete foundation is a cast-in-place reinforced concrete structure, embedded in the ground, providing a stable bottom constraint for the entire tower. The intelligent variable stiffness UHPC functional gradient tower section is directly connected to the concrete foundation. Its wall, from the outside to the inside, includes an outer UHPC main load-bearing layer, a variable stiffness intelligent interlayer, and an inner UHPC counterweight energy dissipation layer. The outer UHPC main load-bearing layer is approximately 300mm thick and uses ultra-high performance concrete incorporating high-strength steel fibers, with a design compressive strength of not less than 150MPa. It is used to bear the vertical pressure and bending moment generated by the wind turbine's self-weight, wind load, and seismic action. The variable stiffness intelligent interlayer is approximately 80mm thick and is composed of controllable stiffness materials (such as magnetorheological elastomers or electrorheological elastomers). An excitation device (electromagnetic coil or electric field plate) is embedded inside. By adjusting the excitation intensity, the equivalent shear modulus of the interlayer can be changed, thereby achieving continuous and adjustable overall wall stiffness. The inner UHPC counterweight energy dissipation layer is approximately 120mm thick and also uses ultra-high performance concrete. Its high mass density increases the inertial mass of the tower, thus aiding in energy dissipation under seismic loads.
[0028] The top steel section connects to the upper end of the intelligent variable stiffness (UHPC) functionally graded tower section, supporting the nacelle and rotor. A top active control device is installed between the top steel section and the nacelle. This device includes a movable mass block capable of bidirectional movement in the horizontal plane and its drive mechanism. The movable mass block is typically a steel or cast iron counterweight, accounting for 3% to 8% of the total mass of the tower top. The drive mechanism is a ball screw or linear motor servo drive system capable of providing active control acceleration of 0.1g to 1.0g, with a maximum stroke of not less than ±400mm. When an earthquake or strong wind causes vibration at the tower top, the drive mechanism, according to instructions, drives the mass block to accelerate, generating a reverse inertial force to counteract the input energy and suppress tower top displacement.
[0029] To achieve real-time structural status monitoring, this embodiment incorporates a multi-source sensing and edge computing unit. This unit includes a distributed fiber optic sensor array, an accelerometer array, and environmental monitoring sensors positioned along the tower's height. The distributed fiber optic sensor array employs a helical winding arrangement with a winding angle of 45°±5°, enabling simultaneous calculation of the tower's bending and torsional modal strains. The accelerometer array includes at least three-dimensional accelerometers positioned at the tower's root, middle, and top, used to identify the tower's translational and sway modes. Environmental monitoring sensors include anemometers and temperature sensors. The edge computing unit utilizes a high-performance embedded computing module (such as NVIDIA Jetson AGX Orin), deployed within a control cabinet at the tower's base, for real-time processing and feature extraction of sensor data, and for running a lightweight intelligent decision-making inference model, keeping end-to-end inference latency below 10 milliseconds.
[0030] The cloud-based digital twin platform is deployed on a remote cloud server or industrial cloud platform, maintaining bidirectional data communication with the edge computing unit via a 4G / 5G wireless transmission module. The platform incorporates nonlinear damage constitutive models for UHPC materials (e.g., concrete damage plasticity models based on continuous damage mechanics) and control models for variable stiffness intelligent sandwich layers (e.g., magnetic-force coupling models). Utilizing data assimilation algorithms, the platform uses real-time acquired strain and acceleration data as boundary conditions to invert the instantaneous stiffness distribution, damping distribution, and three-dimensional damage field of the tower, achieving synchronous mapping between the physical entity and the virtual model.
[0031] The intelligent decision controller can be deployed on edge computing units or on cloud-based digital twin platforms. This controller employs a data-driven intelligent decision-making model, specifically a multi-objective optimization model based on deep reinforcement learning. Its inputs include seismic motion characteristics (such as the dominant S-wave frequency predicted by P-waves), structural response states (such as tower top displacement, velocity, and acceleration), and damage indices (damage indices for each section). Outputs include optimal control commands for each zone of the variable stiffness intelligent mezzanine (such as electromagnetic excitation current values) and optimal driving force commands for the top active control device (such as servo motor voltage commands). Through the above structural design, this application constructs a complete closed-loop intelligent seismic resistance system encompassing "perception-inversion-decision-execution."
[0032] In one embodiment of this application, an optimized design of the variable stiffness intelligent sandwich layer was implemented to further improve control accuracy and address the torsional component of seismic vibrations. The variable stiffness intelligent sandwich layer utilizes magnetorheological elastomer (MRE) or electrorheological elastomer (ERE) materials. Taking the magnetorheological elastomer as an example, its matrix is a polymer elastomer with micron-sized soft magnetic particles uniformly dispersed within it. Under the influence of an external magnetic field, these magnetic particles align in a chain-like structure along the magnetic field direction, causing the overall shear modulus of the material to increase significantly with increasing magnetic induction intensity. This change is continuous, rapid (millisecond-level), and reversible. To achieve refined control, excitation devices (electromagnetic coils or electric field plates) are uniformly arranged circumferentially along the tower, with an independent control zone set every 5 meters along the tower height, forming multiple independent control zones. The intelligent decision controller is configured to independently adjust the excitation intensity of each independent control zone. For example, when the torsional component of local vibration is large, the controller can increase the excitation intensity of one side zone while decreasing the excitation intensity of the opposite side zone, thereby achieving asymmetric adjustment of the lateral stiffness of the tower and actively generating a reaction force to resist the torque, effectively suppressing the torsional response of the tower.
[0033] In one embodiment of this application, specific requirements are set for the performance parameters of the variable stiffness smart sandwich layer to ensure the effectiveness of the "frequency avoidance" strategy. The equivalent shear modulus of the variable stiffness smart sandwich layer can reach G... min To G maxThe stiffness can be continuously adjusted within a certain range, and the response time from the minimum to the maximum modulus is less than 20 milliseconds. This means that the controller can complete multiple stiffness adjustments within seconds before the arrival of a destructive S-wave. Simultaneously, the overall fundamental frequency of the tower structure, under the adjustment of the variable stiffness intelligent interlayer, has an adjustable bandwidth of no less than 30%. For example, a tower with an original fundamental frequency of 0.9Hz can have its fundamental frequency continuously adjusted between 0.7Hz and 1.2Hz by adjusting the variable stiffness interlayer, thus enabling it to avoid the dominant frequencies of most seismic vibrations. In this embodiment, experimental measurements show that using a magnetorheological elastomer interlayer, under current excitation from 0A to 5A, the shear modulus can vary from 0.8MPa to 2.5MPa, corresponding to a tower fundamental frequency variation range of 35%, meeting the above requirements.
[0034] In one embodiment of this application, the specific parameters of the top active control device are preferably defined. The mass of the movable mass block is 3% to 8% of the total mass of the tower top. Too small a mass will not generate sufficient control force, while too large a mass will increase the static load on the tower. The maximum stroke is not less than ±400mm. Combined with the mass block's mass, the maximum available reverse inertial force can be calculated. The drive mechanism adopts a ball screw or linear motor servo drive system, capable of providing 0.1g to 1.0g of active control acceleration. Here, "g" refers to gravitational acceleration, meaning the maximum acceleration generated by the active control device is approximately 0.1 to 1 times the gravitational acceleration, which is sufficient to effectively suppress earthquake responses of moderate to strong intensity. In this embodiment, a 20-ton mass block can generate approximately 100kN of control force at an acceleration of 0.5g, and simulation verification shows that it can reduce the peak displacement at the tower top by more than 40%.
[0035] In one embodiment of this application, the sensor deployment method is optimized. The distributed fiber optic sensor array adopts a helical winding deployment method with a winding angle of 45°±5°. A significant advantage of this deployment method is that a single fiber can simultaneously measure axial strain and circumferential strain. By calculating the axial strain difference between the 0° and 180° positions, the bending curvature of the cross-section can be obtained; by calculating the circumferential strain difference between the 90° and 270° positions, the torsional shear strain of the cross-section can be obtained, thereby simultaneously calculating the bending modal strain and torsional modal strain of the tower. The accelerometer array includes at least three-dimensional accelerometers respectively located at the root, middle, and top of the tower. The accelerometer at the root is used to capture seismic P-wave signals; the accelerometer in the middle is used to monitor the deformation modes of the structure along its height; and the accelerometer at the top is used to provide real-time feedback on the tower top displacement and acceleration response, and to identify the translational and rocking modes of the tower.
[0036] In one embodiment of this application, a seismic control method for the above-mentioned structure is provided. This method includes the following five core steps, which are detailed below. Figure 1 .
[0037] Step S1: Rapid Identification of Seismic Motion Characteristics. When an earthquake occurs, the faster-propagating P-wave first reaches the tower location. Accelerometers at the tower base capture the P-wave signal, and edge computing units immediately perform wavelet packet decomposition on the signal to extract its time-frequency features (such as peak ground acceleration, dominant frequency, and energy distribution). These features are input into a pre-trained lightweight convolutional neural network. This network, trained on a large number of seismic records, can quickly predict the dominant frequency range of upcoming S-waves (e.g., predicting S-wave energy concentration in the low-frequency band of 0.5Hz-1.2Hz) and their potential destructive power level. This prediction provides a valuable lead time for subsequent decision-making, typically from a few seconds to tens of seconds.
[0038] Step S2: Real-time Inversion of Digital Twin and Damage Field. During the P-wave and throughout the S-wave, distributed fiber optic sensors and accelerometers continuously collect data at a sampling rate of up to 200Hz. The edge computing unit preprocesses the raw data and uploads it to the cloud-based digital twin platform. The platform uses data assimilation algorithms (such as Kalman filtering) to use discrete strain and acceleration data as boundary conditions to drive the built-in UHPC nonlinear damage constitutive model, inverting and updating the stiffness, damping, and damage index of each element in the virtual model in real time. For example, when the fiber optic sensor at a certain cross-section detects a nonlinear strain increment, the model automatically determines that the UHPC material at that cross-section has been damaged and updates its damage index D from 0 (no damage) to 1 (complete failure), thereby obtaining the three-dimensional damage field indices for each cross-section.
[0039] Step S3: Multi-objective Cooperative Optimization and Decision Making. At each control time step (e.g., 20 milliseconds), the intelligent decision controller acquires the current state space, including: the dominant shear wave frequency range predicted in Step S1, the instantaneous structural state parameters (stiffness distribution, damping distribution, and damage indices of each section) retrieved in Step S2, and the real-time tower top displacement response. The controller employs an intelligent decision model based on deep reinforcement learning to output the optimal control action sequence under preset optimization objectives. The optimization objectives include three aspects: minimizing the peak displacement at the tower top, minimizing the overall structural damage index, and minimizing control energy consumption. The output action sequence includes: the target equivalent stiffness value for each independent control zone and the target control force time history of the top active control device.
[0040] Step S4: Real-time coordinated adjustment of structural parameters and load paths. The edge computing unit sends decision commands to each actuator. For the variable stiffness smart mezzanine, the command changes the excitation current of the electromagnetic coils in each zone, thereby changing the equivalent shear modulus of the magnetorheological elastic body, reconstructing the lateral stiffness distribution of the tower, and causing its instantaneous fundamental frequency to deviate from the dominant seismic frequency in real time, achieving "frequency avoidance". For example, if the predicted dominant S-wave frequency is 0.8Hz, while the current fundamental frequency of the tower is 0.9Hz, the controller will increase the current in each zone in the middle and upper parts, raising the fundamental frequency of the tower to 1.2Hz, successfully avoiding resonance. At the same time, the top active control device drives the mass block to move according to the command, generating a reverse inertial force to directly offset the seismic input energy. Variable stiffness adjustment mainly changes the structural dynamic characteristics in the frequency domain, while active control mainly provides direct damping force in the time domain. The two form a complementary and coordinated control effect in the frequency domain.
[0041] Step S5: Model Self-Evolution and Policy Iteration. After the earthquake, the edge computing unit completely packages all data recorded throughout the entire earthquake process (original ground motion, all sensor responses, and control commands at each step) and uploads it to the cloud-based digital twin platform. This data is stored in a historical sample library as experience samples. Monthly or quarterly, the cloud platform uses the accumulated large number of new samples to retrain the deep reinforcement learning model offline (e.g., using a deep Q-network algorithm with priority experience replay or a near-end policy optimization algorithm). After training, the updated model parameters are distributed and deployed to the edge computing units of all wind turbines, thereby enabling continuous evolution of the control strategy and making the entire intelligent earthquake-resistant system smarter with use.
[0042] In one embodiment of this application, to address the "curse of dimensionality" problem in high-dimensional action spaces, the deep reinforcement learning model employs a multi-agent reinforcement learning framework. Specifically, each independent control partition is defined as an agent, and the top active control device is defined as another agent. The agents collaborate in a framework of "centralized training and distributed execution." During the training phase, a centralized evaluator (Critic) acquires the state and action information of all agents, guiding the policy network (Actor) of each agent to update. During the execution phase, each agent makes decisions based solely on its local observations, but through collaborative learning, they form a cooperative control policy, thereby significantly reducing the complexity of the individual agent's policy network while ensuring control effectiveness.
[0043] In one embodiment of this application, to achieve proactive suppression of structural damage, a predictive control step is added after the real-time inversion of the three-dimensional damage field index in step S2. Specifically, the currently inverted damage field is used as the initial condition and input into the damage evolution model (such as a crack propagation model based on fracture mechanics) built into the digital twin platform. This model can quickly simulate and predict the damage propagation path in the next 1-2 seconds under three different conditions: condition one, no control; condition two, only the current control strategy; and condition three, a certain ideal control strategy. The intelligent decision-making model uses the predicted damage propagation path as one of its decision-making bases, prioritizing the control action that minimizes future damage accumulation, thereby achieving an upgrade from "reactive control" to "proactive control".
[0044] In one embodiment of this application, to eliminate the response delay of the actuator, a risk level prediction and strategy pre-activation step is added after predicting the dominant frequency range of the shear wave in step S1. Specifically, the predicted dominant frequency of the shear wave is compared with the current fundamental frequency of the tower, and the difference between the two is calculated. If the difference is less than a preset threshold (e.g., 15%), it is determined to be high-risk, and resonance is about to occur. At this time, the edge computing unit immediately sends a pre-activation command to each partition driver of the variable stiffness intelligent mezzanine, pre-setting the electromagnetic coil to a "high response standby state" (e.g., passing a small current to pre-magnetize the magnetorheological elastomer), and pre-charging the servo driver of the top active control device to raise the DC bus capacitor voltage to the rated value. In this way, when the destructive S-wave actually arrives, the actuator can respond within zero delay or an extremely short delay, achieving "instantaneous avoidance".
[0045] In one embodiment of this application, the multi-objective optimization function in step S3 is specifically expressed as:
[0046] Where max|y(t)| is the peak displacement at the top of the tower, y lim This is the allowable displacement limit for the structure (usually determined as 1 / 100 of the tower height). D max D represents the maximum damage index (ranging from 0 to 1) across all UHPC sections. crit The critical damage index (e.g., 0.8; exceeding this index indicates the structure is nearing failure); E consump E represents the total energy consumption of this control action (including electromagnetic excitation energy consumption and drive motor energy consumption). maxThis represents the system's maximum allowable energy consumption (determined by the backup power capacity). The weighting coefficients w1, w2, and w3 are not fixed values but are dynamically and non-linearly adjusted based on the real-time seismic risk level and structural importance coefficient. For example, when the peak P-wave acceleration detected by the root accelerometer exceeds 0.2g, the weights of w1 and w2 are automatically increased to prioritize structural and personnel safety, sacrificing some energy-saving targets.
[0047] In one embodiment of this application, to improve the robustness and reliability of the system, self-diagnosis and fault-tolerant control steps are also included. The edge computing unit monitors the health status of each sensor and actuator in real time at a high frequency (e.g., 10 times per second), including whether sensor signals are interrupted, whether values exceed reasonable ranges, and whether actuator drivers report faults. When a sensor failure is diagnosed, the system automatically switches to a "data-driven virtual sensing mode": using sensor data from adjacent locations, combined with the physical correlation of the digital twin model (e.g., the linear relationship between strain and curvature), the physical quantities at the failure location are reconstructed through interpolation or extrapolation algorithms. When a failure of a variable stiffness zone actuator (such as an electromagnetic coil or driver) is diagnosed, the system automatically re-optimizes the stiffness distribution of the remaining normal zones, for example, by increasing the current in adjacent zones to compensate for the stiffness contribution lost by the failed zone, maintaining the overall control performance of the tower through "functional reconfiguration."
[0048] In one embodiment of this application, the specific input and output design of the lightweight convolutional neural network pre-trained in step S1 is as follows. The input is the first 2-second P-wave signal captured by the root accelerometer (sampling rate 100Hz, 200 sampling points in total). The network structure consists of two convolutional layers (each followed by a pooling layer) and a fully connected output layer. The output layer uses the Softmax activation function, and the output result is the probability distribution of the dominant seismic frequency belonging to a predefined frequency band. The predefined frequency bands include three: [0.1Hz, 0.5Hz] (low frequency band), [0.5Hz, 1.0Hz] (mid frequency band), and [1.0Hz, 2.0Hz] (high frequency band). The network finally outputs the frequency band corresponding to the highest probability as the prediction result. For example, if the output probability is [0.1, 0.8, 0.1], then the predicted dominant shear wave frequency belongs to the mid frequency band of [0.5Hz, 1.0Hz]. This classification prediction method is more robust than direct regression prediction of frequency values and requires less computation, making it suitable for real-time operation on edge computing units.
[0049] In one embodiment of this application, a clear allocation strategy is implemented in the frequency domain to further optimize the synergistic efficiency between the variable stiffness intelligent mezzanine and the top active control device. The variable stiffness intelligent mezzanine is primarily responsible for low-frequency avoidance: adjusting the fundamental frequency of the tower to be at least 0.2 Hz below the dominant seismic frequency. This strategy is called "softening avoidance," which involves reducing stiffness (decreasing current) or increasing stiffness (increasing current) to keep the structural fundamental frequency away from the dominant seismic frequency. Since variable stiffness adjustment does not consume much energy (electromagnetic excitation only needs to maintain the magnetic field), this low-frequency avoidance method is very energy-efficient. The top active control device is primarily responsible for suppressing residual vibration energy remaining after adjustment, especially mid-to-high frequency vibrations in the range of 1.0 Hz to 5.0 Hz. This residual vibration is often caused by high-frequency components of wind loads or seismic motions. The active mass block has a fast response speed and can effectively suppress these high-frequency components. Through this frequency domain division of labor, where "variable stiffness is responsible for low-frequency avoidance and active mass block is responsible for high-frequency energy consumption," the two complement each other's advantages, achieving the overall optimal control effect.
[0050] To more clearly illustrate the embodiments of the present invention, a specific example of a 160m UHPC-concrete composite tower will be described below.
[0051] This example focuses on a wind turbine tower with a total height of 160 meters and a rated power of 5 MW. The concrete foundation section ranges in elevation from 0 to 40 meters; the intelligent variable stiffness UHPC functional gradient tower section ranges in elevation from 40 to 120 meters, with a total wall thickness of 500 mm. This includes a 300 mm thick outer UHPC main load-bearing layer, an 80 mm thick variable stiffness intelligent interlayer (using magnetorheological elastomer), and a 120 mm thick inner UHPC counterweight energy dissipation layer. Sixteen independent control zones are set up every 5 meters along the height direction, with eight sets of electromagnetic coils evenly distributed circumferentially in each zone. The top steel section ranges in elevation from 120 to 160 meters, and the top active control device has a mass block weighing 20 tons and a maximum stroke of ±500 mm.
[0052] Sensor deployment: Fiber grating sensor rings are deployed at heights of 40m, 60m, 80m, 100m, and 120m, with 4 measuring points (0°, 90°, 180°, 270°) in each ring direction; triaxial accelerometers are deployed at the base of the tower, at 80m, and at the top of the tower.
[0053] Operational Process: During the P-wave arrival phase, the root accelerometer is triggered, and the edge computing unit completes P-wave feature extraction and spectrum prediction within milliseconds. The predicted dominant frequency of the shear wave is 0.7Hz, which is less than 15% different from the tower's current fundamental frequency of 0.9Hz, indicating a high risk. The variable stiffness interlayer and active control device are immediately pre-activated. During the decision-making phase, the intelligent decision model outputs instructions to increase the electromagnetic current in the eight sections of the upper and middle part from 1A to 4A, increasing the shear modulus of the magnetorheological elastomer by 50% and raising the tower's fundamental frequency from 0.9Hz to 1.2Hz, successfully avoiding the 0.7Hz resonant frequency. Simultaneously, the top active control device, based on tower top displacement feedback, drives a 20-ton mass block with an acceleration of 0.3g, generating a reverse inertial force of approximately 60kN. During the S-wave action phase, the digital twin platform shows in real-time inversion that the damage index of each section is below 0.05, indicating that the structure maintains its elasticity. In the post-earthquake phase, the cloud uses the earthquake data to retrain the decision model offline and optimize the control strategy.
[0054] Comparative test results show that, compared with the traditional uncontrolled scheme, the peak displacement at the top of the tower is reduced by more than 40%; compared with the scheme that only uses the active control device at the top, the peak displacement at the top of the tower is further reduced by about 15%, and the control power consumption is reduced by about 20%; the damage index of the UHPC layer is significantly reduced, and the residual displacement is greatly reduced.
[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. An adaptive UHPC-steel hybrid tower structure based on digital twin and intelligent decision making, characterized in that, Including those connected from bottom to top: Concrete foundation; The intelligent variable stiffness UHPC functional gradient tower section has a wall consisting of an outer UHPC main load-bearing layer, a variable stiffness intelligent interlayer, and an inner UHPC counterweight energy dissipation layer from the outside to the inside. The variable stiffness intelligent interlayer is made of a controllable stiffness material and has an excitation device embedded inside to achieve continuous adjustment of the overall stiffness of the wall. Top steel section; as well as The top active control device, installed between the top steel section and the nacelle, includes a movable mass block capable of bidirectional movement in the horizontal plane and its driving mechanism. The multi-source sensing and edge computing unit includes a distributed optical fiber sensor array, an acceleration sensor array, and an environmental monitoring sensor located at the base of the tower, all arranged along the tower height, as well as an edge computing unit for real-time processing of sensor data. The cloud-based digital twin platform is connected to the edge computing unit via a wireless transmission module and has a built-in nonlinear damage constitutive model for UHPC materials and a control model for the variable stiffness smart sandwich layer. The intelligent decision controller, deployed in the edge computing unit or the cloud-based digital twin platform, is configured as a data-driven intelligent decision model. It takes seismic characteristics, structural response, and damage indicators as inputs and outputs the optimal control command for the variable stiffness intelligent interlayer and the optimal driving force command for the top active control device.
2. The adaptive UHPC-steel hybrid tower structure based on digital twin and intelligent decision making according to claim 1, characterized in that, The variable stiffness smart sandwich layer is made of magnetorheological elastomer material or electrorheological elastomer material. The internally embedded excitation devices are uniformly arranged along the circumference of the tower and form multiple independent control zones in the tower height direction; the intelligent decision controller is configured to independently adjust the excitation intensity of each independent control zone to achieve asymmetric adjustment of the tower lateral stiffness, thereby actively controlling the torsional response under seismic action.
3. The adaptive UHPC-steel hybrid tower structure based on digital twin and intelligent decision making according to claim 1, characterized in that, The equivalent shear modulus of the variable stiffness smart sandwich can be continuously adjusted in the range of G min to G max , and the response time of adjusting from G min to G max is less than 20 milliseconds; the overall fundamental frequency of the tower structure under the adjustment of the variable stiffness smart sandwich is not less than 30% of the adjustable bandwidth.
4. The adaptive UHPC-steel hybrid tower structure based on digital twin and intelligent decision making according to claim 1, characterized in that, The movable mass block of the top active control device has a mass of 3% to 8% of the total mass of the tower top, and its maximum stroke is not less than ±400mm; the drive mechanism is a ball screw or linear motor servo drive system that can provide active control acceleration of 0.1g to 1.0g.
5. The adaptive UHPC-steel hybrid tower structure based on digital twin and intelligent decision making according to claim 1, characterized in that, The distributed optical fiber sensor array is arranged in a spiral winding manner with a winding angle of 45°±5°, and is used to simultaneously calculate the bending modal strain and torsional modal strain of the tower. The accelerometer array includes at least three-dimensional accelerometers respectively set at the root, middle and top of the tower, and is used to identify the translational vibration mode and rocking vibration mode of the tower.
6. A seismic control method applied to the adaptive UHPC-steel hybrid tower structure based on digital twin and intelligent decision according to any one of claims 1 to 5, characterized in that, Includes the following steps: Step S1: Rapid identification of ground motion characteristics: The P-wave signal is captured by the accelerometer at the base of the tower. The time-frequency features of the P-wave signal are extracted by wavelet packet decomposition and input into a pre-trained lightweight convolutional neural network to quickly predict the dominant frequency range and potential destructive force level of the upcoming S-wave. Step S2: Real-time inversion of digital twin and damage field: Based on strain field data and acceleration data collected by distributed fiber optic sensors, combined with data assimilation algorithm, the instantaneous stiffness distribution and damping distribution of the tower are inverted in real time in the cloud digital twin platform, and the three-dimensional damage field index of each section is calculated according to the nonlinear damage constitutive model of UHPC material. Step S3: Multi-objective collaborative optimization and decision-making: The dominant frequency range of the shear wave predicted in step S1, the instantaneous structural state parameters inverted in step S2, and the real-time tower top displacement response are combined into a state space; an intelligent decision-making model based on deep reinforcement learning is adopted to output the optimal control action sequence under the optimization objectives of "minimizing the peak displacement at the tower top", "minimizing the overall structural damage index", and "minimizing control energy consumption". The action sequence includes: the target equivalent stiffness value of each independent control zone and the target control force time history of the top active control device; Step S4: Real-time coordinated adjustment of structural parameters and load path: The edge computing unit sends decision commands to each actuator; the variable stiffness intelligent interlayer changes its equivalent shear modulus according to the command, thereby reconstructing the lateral stiffness distribution of the tower, so that its instantaneous fundamental frequency deviates from the dominant seismic frequency in real time, achieving frequency avoidance; at the same time, the top active control device drives the mass block to move according to the command, generating a reverse inertial force, forming a time-domain active control force that complements the frequency domain of the variable stiffness adjustment; Step S5: Model self-evolution and strategy iteration: After the earthquake, the structural response data, control commands and external ground motion records of the entire process are uploaded to the cloud digital twin platform as experience samples. The deep reinforcement learning model is periodically retrained offline, the model parameters are updated and distributed to the edge computing unit to realize the continuous evolution of the control strategy.
7. The seismic control method according to claim 6, characterized by, The deep reinforcement learning model in step S3 adopts a multi-agent reinforcement learning framework, in which each independent control partition is defined as an agent and the top active control device is defined as another agent; under the framework of centralized training and distributed execution, the agents form a collaborative control strategy through collaborative learning to solve the curse of dimensionality problem in high-dimensional action space.
8. The seismic control method according to claim 6, characterized by, After the real-time inversion of the three-dimensional damage field index of the tower in step S2, a predictive control step is also included: the currently inverted damage field is used as the initial condition and input into the damage evolution model built into the digital twin platform to predict the damage propagation path in the next 1-2 seconds under three conditions: no control, existing control, and ideal control; the intelligent decision model uses the predicted damage propagation path as one of its decision-making bases to achieve forward-looking suppression of structural damage.
9. The seismic control method according to claim 6, characterized by, After predicting the dominant frequency range of the shear wave in step S1, a risk level prediction and strategy pre-activation step is also included: if the difference between the predicted dominant frequency and the current fundamental frequency of the tower is less than a preset threshold, it is determined to be high risk, and the variable stiffness intelligent interlayer is immediately pre-set to a high response standby state, while the servo drive of the top active control device is pre-charged to eliminate the response delay of the actuator.
10. The seismic control method according to claim 6, characterized by, The multi-objective optimization function in step S3 is expressed as follows: Wherein, max | y (t) | is the peak displacement at the top, y lim To allow displacement limit; D max D is the maximum damage index of the tower tube as a whole, D crit E is the critical damage index; E consump E is the total energy consumption, E max E is the maximum allowable energy consumption; the weight coefficients w1, w2 and w3 are dynamically and nonlinearly adjusted according to the real-time seismic risk level and the structure importance coefficient, and the weights of w1 and w2 are automatically increased when the peak value of the P wave acceleration of the earthquake exceeds the threshold value.
11. The seismic control method according to claim 6, wherein It also includes self-diagnosis and fault-tolerant control steps: the edge computing unit monitors the health status of each sensor and actuator in real time; when a sensor failure is diagnosed, it automatically switches to a data-driven virtual sensing mode and reconstructs the physical quantity of the failure location using data from adjacent sensors and a digital twin model; when a variable stiffness partition actuator failure is diagnosed, it automatically re-optimizes the stiffness distribution of the remaining normal partitions to maintain the overall control performance of the tower through functional reconstruction.
12. The seismic control method according to claim 6, characterized in that, The lightweight convolutional neural network pre-trained in step S1 takes as input the first 2-second P-wave signal captured by the root accelerometer, and its output layer uses the Softmax function. The output result is the probability distribution of the earthquake dominant frequency belonging to the predefined frequency bands [0.1Hz, 0.5Hz], [0.5Hz, 1.0Hz], and [1.0Hz, 2.0Hz], and the frequency band corresponding to the highest probability is used as the prediction result.
13. The seismic control method according to claim 6, characterized in that, The coordinated control between the variable stiffness intelligent interlayer and the top active control device has the following frequency domain allocation strategy: the variable stiffness intelligent interlayer is mainly responsible for adjusting the fundamental frequency of the tower to be more than 0.2Hz below the dominant seismic frequency, thereby achieving "softening avoidance"; the top active control device is mainly responsible for suppressing the residual vibration energy that still exists after adjustment, with a frequency range of 1.0Hz to 5.0Hz.