Flexible support pile foundation process for water photovoltaic power generation

Abstract

The invention relates to the technical field of water photovoltaic flexible support pile foundation construction data processing, in particular to a process for a water photovoltaic flexible support pile foundation, which comprises the following steps: fusing meteorological, hydrological and pile deformation multi-source data through a dynamic weight distribution mechanism, and generating a space-time associated dynamic load spectrum by adopting wavelet packet transformation and entropy feature extraction; a collaborative analysis architecture of a finite element simulation and machine learning agent model is constructed, an anchoring depth mapping relation and a real-time design parameter adjustment instruction are output in parallel, and macroscopic fluid parameters and microscopic material degradation characteristics are associated through a data cascade interaction correction mechanism; full-life-cycle data version management is achieved through the block chain technology, and finite element model re-calibration and protection strategy dynamic updating are triggered. The dynamic load response analysis precision and long-term stability of the pile foundation are improved, and the deformation coordination requirement of the flexible support in the complex water area environment is met.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing for flexible support pile foundations in floating solar power, and particularly to a flexible support pile foundation technology for floating solar power. Background Art

[0002] The flexible support pile of floating solar power is a support structure designed based on the dynamic load characteristics of the water area environment. Through the regulation of the material elastic modulus and the optimization of the structure topology, it realizes the adaptive deformation coordination of the photovoltaic modules under the action of water flow, wind waves and buoyancy. This type of pile foundation is composed of high-strength steel or composite materials. Combining the pile-soil-water coupling dynamics model, for different water depths, geological conditions and wave spectrum characteristics, it buffers the impact of external loads through elastic deformation and reduces the risk of stress concentration. During the construction process, based on the geological exploration data and the results of fluid mechanics simulation, vibration hammering, bored casting or precast pile technology is adopted to optimize the penetration trajectory of the pile body and reduce the disturbance of the foundation.

[0003] With the large-scale development of floating solar power systems, the data-driven decision-making ability of traditional pile foundation technologies faces significant bottlenecks. In the prior art, the pile foundation design relies on static load calculation models and fails to effectively integrate multi-source heterogeneous data of wind, snow and dynamic loads of photovoltaic modules, resulting in blind spots in the correlation analysis between the anchorage depth and the mechanical properties of the pile foundation under the combined action of loads. In addition, the pile spacing planning lacks multi-dimensional collaborative modeling of water area navigation data, water flow dynamic parameters and structural safety thresholds, resulting in limited accuracy in predicting local scour risks and evaluating navigation compatibility. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a flexible support pile foundation technology for floating solar power, which solves the problems that the existing pile foundation design relies on static load models and fails to integrate multi-source dynamic load data, resulting in blind spots in the correlation analysis between the anchorage depth and the mechanical properties of the pile foundation. At the same time, the pile spacing planning lacks collaborative modeling of navigation, hydrodynamic and structural safety data, resulting in insufficient accuracy in scour risk prediction and navigation assessment, and it is difficult to adapt to the dynamic deformation and long-term stability requirements of flexible supports in complex water area environments.

[0005] To solve the above technical problems, the specific technical solutions of the present invention are as follows: The present invention provides a flexible support pile foundation technology for floating solar power, including: Obtaining multi-source dynamic data of the water area environment, where the multi-source dynamic data includes meteorological parameters, hydrological parameters and pile body deformation parameters; Performing dynamic fusion processing on the multi-source dynamic data to generate a dynamic load spectrum, where the dynamic fusion processing includes reverse-correcting the fusion weight based on the pile foundation actual deformation data of the sensor confidence and construction feedback; Screen the feature combinations of multi-source dynamic data based on the adaptive Bayesian optimization algorithm to optimize the correlation between the dynamic load spectrum and the actual deformation data of the pile foundation; Parallelly input the dynamic load spectrum into the finite element simulation model and the machine learning proxy model, output the anchorage depth mapping relationship through the finite element simulation model, and generate real-time design parameter adjustment instructions through the machine learning proxy model; Based on the anchorage depth mapping relationship and the real-time design parameter adjustment instructions, use a hybrid optimization algorithm to perform multi-objective optimization on the pile foundation layout parameters, where the hybrid optimization algorithm integrates a reinforcement learning module and a genetic algorithm module, and generates dynamic constraint conditions based on the pile foundation degradation data; Execute pile foundation construction control according to the optimized pile foundation layout parameters and collect construction process data; Perform feature extraction and cloud collaborative analysis on the construction process data to generate formation classification labels, and feedback the formation classification labels to the construction control module to adjust construction parameters; Construct a scour and corrosion coupling prediction model based on the pile foundation stress data after construction, and generate a protection decision instruction through the data cascade interaction correction of the scour and corrosion coupling prediction model; Perform data collaborative update on the protection decision instruction, the pile foundation design parameters, and the construction control instruction through the blockchain, where the smart contract verifies the data version consistency to trigger parameter correction.

[0006] Further, for the flexible support pile foundation technology for floating solar power, perform dynamic fusion processing on the multi-source dynamic data to generate a dynamic load spectrum, where the dynamic fusion processing includes reverse correction of the fusion weight of the actual deformation data of the pile foundation based on the sensor confidence and construction feedback, including: calculating the initial confidence weight based on the sliding window statistical result of the sensor historical error rate; Adjust the initial confidence weight according to the dynamic deviation value between the actual deformation data of the pile foundation and the meteorological and hydrological data to obtain the adjusted weight value; Input the adjusted weight value into a preset spatio-temporal alignment algorithm to optimize the interpolation compensation parameters of the multi-source data.

[0007] Further, the flexible support pile foundation technology for floating solar power further includes: using the anchorage depth mapping relationship output by the finite element simulation model as the training sample of the machine learning proxy model; The machine learning proxy model generates adjustment instructions for the pile foundation construction inclination parameter through transfer learning; The pile foundation stress data collected during the construction stage is fed back to the machine learning proxy model to correct the loss function weight of the machine learning proxy model.

[0008] Further, for the flexible support pile foundation process for floating solar power, the adjustment instructions for the pile foundation construction inclination parameters generated by the machine learning agent model include: The reinforcement learning module explores the global solution space based on the pile group stability objective; The genetic algorithm module loads the historical engineering case library to generate the initial population and injects the navigation safety constraint conditions; The pile foundation degradation monitoring data dynamically constrains the feasible solution range of the hybrid optimization algorithm.

[0009] Further, for the flexible support pile foundation process for floating solar power, the pile foundation degradation monitoring data dynamically constraining the feasible solution range of the hybrid optimization algorithm includes: The cloud deep learning model analyzes the frequency domain characteristics of the construction process data and generates formation classification labels; The edge computing device adjusts the pile driving frequency and penetration speed parameters after receiving the formation classification labels; The updated construction parameters are input into the cloud deep learning model as training samples to optimize the cloud deep learning model.

[0010] Further, for the flexible support pile foundation process for floating solar power, it further includes: The scour and corrosion coupling prediction model consists of a macroscopic scour model and a microscopic corrosion model; The macroscopic scour model is applied in the macroscopic scour module; The microscopic corrosion model is applied in the microscopic corrosion rate prediction module; The water flow velocity gradient distribution data output by the macroscopic scour module is input into the microscopic corrosion rate prediction module; The microscopic corrosion rate prediction module reversely corrects the boundary layer thickness parameters of the macroscopic scour module based on the collected corrosion monitoring data to generate updated flow field distribution data; The updated flow field distribution data and the microscopic corrosion rate prediction results are aligned in space and time, input into a preset graph neural network, a non-linear mapping relationship of the filling material gradation parameters is established, a scour and corrosion coupling prediction result is generated, and a protection construction plan instruction set is generated to drive the construction equipment to adjust the pile driving frequency and material ratio.

[0011] Further, for the flexible support pile foundation process for floating solar power, the aligning the updated flow field distribution data and the microscopic corrosion rate prediction results in space and time, inputting them into a preset graph neural network, establishing a non-linear mapping relationship of the filling material gradation parameters, generating a scour and corrosion coupling prediction result, and generating a protection construction plan instruction set to drive the construction equipment to adjust the pile driving frequency and material ratio includes: Bind the construction control instruction version number to the protection decision timestamp and store it in the blockchain; The smart contract verifies the matching between the characteristic dimension of the dynamic load spectrum and the input layer of the machine learning agent model; When a version conflict is detected, trigger the re - calibration of parameters of the finite element simulation model.

[0012] Further, for the flexible support pile foundation technology for floating solar power, the generation of the dynamic load spectrum includes: Decompose the time series of the multi - source dynamic data into frequency - domain sub - bands through wavelet packet transform; Extract entropy value features from the frequency - domain sub - bands to generate a high - dimensional feature space.

[0013] Further, for the flexible support pile foundation technology for floating solar power, the adaptive Bayesian optimization algorithm includes: Construct a Gaussian process surrogate model, with the feature selection combination of multi - source dynamic data as the input and the correlation index of the coupled scour and corrosion prediction result as the output; Adopt a local penalty expected improvement function to balance the global exploration and local exploitation search strategies; Dynamically inject the actual pile foundation deformation data in the feedback construction parameters into the weight coefficient update process of the objective function.

[0014] Further, for the flexible support pile foundation technology for floating solar power, the construction of the Gaussian process surrogate model, with the feature selection combination of multi - source dynamic data as the input and the correlation index of the coupled scour and corrosion prediction result as the output includes: Input the coupled scour and corrosion prediction result into the material durability evaluation model; Generate a filling construction plan based on the grading optimization parameters output by the graph neural network; Verify the degradation trend of the historical protection decision data recorded in the blockchain and the prediction result of the material durability evaluation model.

[0015] Advantages of the present invention; The present invention integrates multi-source heterogeneous data through a dynamic weight allocation mechanism to improve the generation accuracy of dynamic load spectra, effectively solving the problem that traditional static models cannot capture the spatio-temporal correlation of dynamic loads in the water area environment; the collaborative optimization architecture of the finite element simulation model and the machine learning proxy model significantly shortens the optimization cycle of the pile foundation layout parameters while ensuring the calculation accuracy of the anchorage depth mapping relationship; the hybrid optimization algorithm integrates the global search ability of reinforcement learning and the historical experience guidance mechanism of the genetic algorithm to achieve multi-dimensional collaborative modeling of navigation safety constraints, pile group stability objectives, and full-life cycle degradation data, breaking through the limitations of traditional single-objective optimization; the edge-cloud collaborative control system reduces the construction trial-and-error cost under complex geological conditions through the closed-loop feedback of intelligent pre-judgment of formation characteristics and dynamic adjustment of construction parameters; the multi-scale scour-corrosion coupling prediction model reveals the interaction mechanism between material property degradation and hydrodynamic parameters, providing a decision-making basis for protection strategies across spatio-temporal scales; blockchain technology ensures the traceability and version consistency of data interaction between multiple modules, forming a trusted closed-loop of the technical chain covering the entire process of design, construction, and operation and maintenance, and finally achieving dynamic deformation coordination and long-term stable bearing of flexible support pile foundations in complex water area environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. Obviously, for those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0017] Figure 1 The flowchart of a flexible support pile foundation process for floating solar power provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the specific embodiments and corresponding accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention will be described in detail below with reference to the accompanying drawings. To better understand the objectives of the present invention, the present invention will be further described in detail below.

[0019] Please refer to Figure 1 , the present invention provides a flexible support pile foundation process for floating solar power, including: Step S101, obtaining multi-source dynamic data of the water area environment, where the multi-source dynamic data includes meteorological parameters, hydrological parameters, and pile body deformation parameters; In the process of obtaining multi-source dynamic data of the water area environment, first, meteorological parameters, hydrological parameters, and pile deformation parameters are collected in real time through a sensor array deployed in the pile structure and the surrounding water area. Meteorological parameters include wind speed, wave height, and the vibration frequency of photovoltaic modules, which are collected by combining a Doppler radar and a piezoelectric accelerometer; hydrological parameters cover water flow velocity, water depth, and suspended sediment concentration, and are synchronously monitored by an acoustic Doppler current profiler and an optical turbidity sensor; pile deformation parameters are obtained in real time by a distributed fiber optic sensor and a micro-electromechanical system displacement sensor. Among them, the fiber optic sensor captures the axial strain gradient of the pile through a Bragg grating array, and the displacement sensor records the three-dimensional displacement trajectory of the pile top based on the capacitive measurement principle. The data collection frequency of all sensors is dynamically adjusted according to the dynamic change rate of the water area environment. For example, the sampling rate is automatically increased during typhoon or strong tide periods to match the monitoring requirements of high-frequency load changes.

[0020] The collected raw data goes through a preprocessing process, including outlier filtering and timestamp alignment. Outlier filtering uses the sliding window interquartile range method, and a dynamic threshold is set in combination with the historical error distribution of the sensor to eliminate noise data beyond the confidence interval; timestamp alignment synchronizes the clock signals of each sensor through the Beidou satellite timing system, and uses a linear interpolation algorithm to compensate for the sampling time deviation between different sensors to generate a multi-source data sequence with a unified time baseline. The preprocessed data is stored in the circular buffer of the edge computing node, and the data compression algorithm is used to reduce the transmission bandwidth requirement, and a data quality label is marked to distinguish data packets with different confidence levels.

[0021] The acquisition frequency adjustment mechanism of multi-source dynamic data is realized based on the water area environment state recognition model. This model judges whether the current water area environment belongs to a steady state, a transition state, or a transient condition by analyzing the time-varying characteristics of wind speed, water flow velocity, and wave spectrum in real time. For example, when it is detected that the main wave frequency exceeds the preset threshold, the high-frequency sampling mode is triggered, and the power consumption allocation strategy of the sensor array is synchronously adjusted to give priority to ensuring the acquisition accuracy of key parameters. During the dynamic adjustment process, a redundant sensor cross-check mechanism is adopted. Multiple types of sensors are deployed for the same monitoring parameter, and the systematic error of a single sensor is eliminated through a weighted fusion algorithm to improve the robustness of data acquisition.

[0022] The preprocessed multi-source dynamic data is uploaded to the cloud data management platform through an encrypted transmission protocol and is associated and mapped with the water area environment characteristic data in the historical engineering case library. The data management platform constructs a multi-dimensional database based on a spatio-temporal index, classifies and stores meteorological, hydrological, and pile deformation parameters according to time series and spatial coordinates, and supports the rapid retrieval and feature extraction of the subsequent dynamic fusion processing module. At the same time, the data management platform integrates a data integrity verification module, which verifies the consistency of the data during transmission and storage through hash value comparison and cyclic redundancy check codes to prevent data tampering or loss.

[0023] In the above steps, the deployment of the sensor array covers the key stress concentration areas and water dynamic sensitive areas of the pile body, ensuring the comprehensiveness and representativeness of the monitoring data through spatial topology optimization; the preprocessing process eliminates the noise interference and time mismatch problems in the data acquisition stage through outlier filtering and timestamp alignment, providing high-quality input for the generation of the dynamic load spectrum; the dynamic acquisition frequency adjustment mechanism adapts to the monitoring requirements of different water environment conditions, balancing data accuracy and system energy consumption; the cloud data management platform realizes the structured storage and secure transmission of multi-source data, providing a reliable data basis for subsequent data processing and analysis. The collaborative effect of each link forms a complete technical link from data acquisition, preprocessing to storage, supporting the efficient execution of subsequent dynamic fusion processing and model optimization processes.

[0024] Step S102: Perform dynamic fusion processing on the multi-source dynamic data to generate a dynamic load spectrum, where the dynamic fusion processing includes reverse-correcting the fusion weights based on the actual deformation data of the pile foundation with sensor confidence and construction feedback; During the dynamic fusion processing, first calculate the confidence weight based on the sensor historical error data set. Analyze the error mean and variance of each sensor within a preset time interval through a sliding window statistical method, and correct the initial weight allocation in combination with the temperature drift compensation factor. For example, in view of the sensitivity attenuation characteristics of the piezoelectric accelerometer under high-temperature working conditions, introduce ambient temperature monitoring data to dynamically adjust the weight coefficient, reducing the influence of external interference on the fusion accuracy of meteorological parameters. The confidence weight calculation module outputs a multi-dimensional weight matrix, characterizing the credibility distribution of different sensor data in the spatio-temporal dimension.

[0025] In the dynamic weight adjustment stage, the actual deformation data of the pile foundation fed back by the construction control module is received in real time, and the deviation degree between the sensor monitoring data and the physical response of the pile body is quantified through the root mean square error algorithm. When the deviation value exceeds the preset threshold, trigger the adaptive weight attenuation mechanism, and use the exponential smoothing algorithm to reduce the proportion of the fusion weight of the abnormal data source. At the same time, the hammer penetration resistance data and the pile body inclination angle data obtained during the construction process are used as auxiliary constraint conditions to reverse-correct the sensor data lag deviation caused by sudden changes in geological conditions. The adjusted weight matrix contains spatio-temporal dynamic characteristics, reflecting the change of data reliability at different monitoring points during the construction process.

[0026] The optimized weight value is input into the spatio-temporal alignment algorithm module to perform data synchronization processing for the spatio-temporal heterogeneity of multi-source data. In the time dimension, the dynamic time warping algorithm is used to align the sensor time series data with different sampling frequencies, and the phase difference is eliminated by matching the non-linear time warping path. In the space dimension, a three-dimensional interpolation grid is constructed based on the pile foundation topology coordinates, and the spatial attenuation coefficient adjustment mechanism in the inverse distance weighted interpolation algorithm is combined to compensate for the spatial data blind area caused by uneven sensor distribution. During the optimization process of the interpolation compensation parameters, the local deformation monitoring data of the pile foundation is used as the verification benchmark to evaluate the physical rationality of the interpolation results in real time, forming a closed-loop feedback link for weight adjustment and spatio-temporal alignment.

[0027] The dynamic fusion processing module finally generates a dynamic load spectrum with spatio-temporal correlation characteristics. The load spectrum decomposes multi-source data into different frequency domain sub-bands through wavelet packet transform, and combines the entropy value feature extraction method to quantify the energy distribution and non-linear characteristics of each sub-band. The load fluctuation characteristics of the key frequency bands are retained during the fusion process to suppress the interference of high-frequency noise on the analysis of the mechanical response of the pile foundation. The spatio-temporal resolution of the load spectrum is dynamically adjusted according to the water area environment conditions. For example, in strong water flow areas, the extraction accuracy of the spatial gradient characteristics is enhanced to match the analysis requirements of the local flow field details for the optimization of the pile group layout.

[0028] In the above processing flow, the calculation of the sensor confidence weight lays the foundation for data credibility for dynamic fusion; the weight adjustment mechanism driven by construction feedback enhances the adaptability of the fusion model to complex working conditions; the spatio-temporal alignment algorithm eliminates the spatio-temporal mismatch problem of multi-source data; the closed-loop verification mechanism ensures the physical rationality and engineering applicability of the load spectrum. Each link realizes the transformation of multi-source heterogeneous data into a high-precision dynamic load spectrum through the coordinated action of the data flow and the control flow, providing a standardized input for subsequent finite element simulation and machine learning analysis.

[0029] Step S103: Based on the adaptive Bayesian optimization algorithm, screen the feature combinations of multi-source dynamic data to optimize the correlation between the dynamic load spectrum and the actual deformation data of the pile foundation; In the feature combination screening stage, first, a Gaussian process surrogate model is constructed based on the high-dimensional feature space of the dynamic load spectrum. This model takes the entropy value features of the frequency domain sub-bands of multi-source dynamic data, the geometric parameters of the pile body, and the water flow gradient parameters as the input domain, and captures the non-linear correlation and linear trend between features through a composite kernel function. In the initialization stage of the surrogate model, the pile foundation strain response data set in historical engineering cases is loaded to establish a prior knowledge base of the correlation between feature combinations and pile body deformation, providing a search direction guidance for Bayesian optimization.

[0030] The construction of the objective function integrates multi-dimensional evaluation indicators, including the Pearson correlation coefficient between the dynamic load spectrum characteristics and the actual deformation data of the pile foundation, the KL divergence distribution difference, and the scour risk prediction deviation. Among them, the correlation coefficient reflects the strength of linear correlation, the KL divergence quantifies the matching degree of non-linear distribution, and the scour risk deviation is calculated based on the simulation results of the flow field around the pile. The objective function fuses the above indicators through weighted summation, and the weight coefficients are dynamically adjusted according to the construction stage. For example, when the inclination angle of the pile changes, the weight ratio of enhancing the correlation of bending strain is increased, enhancing the adaptability of the optimization process to the construction conditions.

[0031] Adaptive Bayesian optimization uses a local penalty expected improvement function to balance the global exploration and local exploitation strategies. In the global exploration stage, the Latin hypercube sampling is used to cover the low-density regions of the feature space to identify potential high-correlation feature subsets; in the local exploitation stage, the gradient descent algorithm is used to lock the high-probability optimal solution region, and a dynamic penalty radius mechanism is adopted to suppress redundant sampling. The adjustment of the penalty radius is linked with the pile foundation degradation monitoring data. When a sudden change in material properties is detected, the search range is expanded to prevent the algorithm from falling into a local optimal solution. During the optimization iteration process, the real-time hammering resistance data fed back by the construction control module is injected into the objective function to dynamically correct the evaluation criteria of the feature combination.

[0032] The feedback mechanism of construction data is realized through edge computing nodes. The pile foundation deformation data collected in real time is processed by sliding window normalization and then input into the objective function update module of the surrogate model. This module adjusts the weight coefficient allocation strategy according to the spatio-temporal distribution characteristics of the strain data. For example, the weight of the axial strain dynamically decays as the penetration depth of the pile increases, and the weight of the bending strain triggers non-linear growth based on the inclination angle threshold. A regularization constraint term is introduced during the feedback process to limit the oscillation amplitude of the weight coefficients and maintain the numerical stability of the optimization process. The updated objective function drives the Bayesian optimization to generate a new candidate feature combination, forming a data closed-loop optimization link.

[0033] After the optimized feature combination is reduced in dimension by principal component analysis, a refined dynamic load spectrum is generated, which is used as the collaborative input of the finite element simulation and the machine learning model. The load spectrum retains the frequency domain sub-band features strongly correlated with the pile foundation deformation, eliminates redundant noise components, and improves the convergence efficiency of the subsequent pile group layout optimization model. The feature screening results are synchronously fed back to the sensor array control module to dynamically adjust the acquisition frequency and preprocessing strategy of key parameters. For example, the sampling density of highly correlated hydrological parameters is enhanced to optimize the system resource allocation. This process realizes the dynamic matching of monitoring resource configuration and engineering requirements through a data-driven approach, forming a reverse control closed-loop from feature optimization to data acquisition.

[0034] Step S104: Input the dynamic load spectrum into the finite element simulation model and the machine learning proxy model in parallel. Output the mapping relationship of the anchorage depth through the finite element simulation model, and generate real-time design parameter adjustment instructions through the machine learning proxy model. In the parallel processing stage of the dynamic load spectrum, the finite element simulation model is constructed based on the pile-soil-water coupling dynamics equation, and the spatio-temporal distribution load parameters and pile body geometric topology data in the dynamic load spectrum are input. The model outputs the mapping relationship between the anchorage depth and the pile foundation stress response by solving the coupled equations of the stress field and displacement field under nonlinear contact conditions, including the correlation characteristics between the critical instability load and the pile penetration resistance under different geological conditions. During the simulation process, the adaptive mesh refinement technology is used to capture the gradient change of the shear stress at the pile-soil interface, and at the same time, a material degradation parameter correction module is introduced to dynamically update the boundary condition parameter set according to the pile body corrosion monitoring data fed back during construction, improving the engineering applicability of the simulation results.

[0035] The machine learning proxy model adopts a transfer learning architecture. In the pre-training stage, the anchorage depth data set of historical engineering cases is loaded, and the implicit correlation between the pile foundation geometric parameters and load characteristics is learned through the feature extraction network. In the fine-tuning stage, the finite element simulation results of the current project are fused, and a dynamic curriculum learning strategy is constructed to gradually increase the weight ratio of high-confidence samples, generating a real-time adjustment instruction set for the pile foundation construction inclination and penetration speed. The instruction set is dynamically corrected according to the real-time collected water flow velocity gradient data. For example, the edge computing node analyzes the change trend of the flow field and triggers the adjustment of the guide frame angle of the hydraulic control system to suppress the asymmetric stress concentration during the pile penetration process.

[0036] In the model collaborative working mechanism, the mapping relationship of the anchorage depth output by the finite element simulation model is used as the incremental training sample of the machine learning proxy model, and the network weight parameters are updated through the online learning module. The pile foundation axial stress and bending stress data collected during the construction stage are transmitted to the feedback control module of the proxy model in real time by the distributed optical fiber sensor. The deviation between the calculated prediction instruction and the actual response is calculated, and the adaptive momentum optimization algorithm is used to dynamically adjust the weight coefficients of different stress components in the loss function. During the adjustment process, the inversion result of the elastic modulus of the soil layer around the pile is injected into the optimization process as a regularization constraint term to reduce the risk of model overfitting, forming a data closed-loop optimization link between simulation and machine learning.

[0037] The generation process of real-time design parameter adjustment instructions integrates the collaborative optimization capabilities of a reinforcement learning module and a genetic algorithm module. The reinforcement learning module constructs a state space and a reward function based on the pile group stability objective to explore the optimal pile position layout in the global solution space; the genetic algorithm module loads the historical engineering case library to generate an initial population, injects navigation safety constraints and pile foundation degradation monitoring data, and dynamically adjusts the crossover and mutation probabilities. The optimization results are synchronized to the construction control terminal through the blockchain network, driving the pile driver to adjust the hammering frequency and penetration trajectory, and at the same time transmitting the updated construction parameters back to the cloud model training module to form a positive feedback loop for iterative optimization of the design parameters.

[0038] In the above process, the finite element simulation model provides the anchorage depth mapping relationship driven by physical mechanisms, and the machine learning surrogate model realizes data-driven real-time parameter adjustment. The two balance the calculation accuracy and response speed through a collaborative optimization architecture. The dynamic curriculum learning strategy enhances the model's adaptability to complex working conditions, and the closed-loop feedback mechanism ensures the dynamic matching of construction parameters and the actual response of the pile foundation. Through the interaction of data flow and control flow among various modules, the efficiency bottleneck of the traditional serial processing mode is broken through, and multi-dimensional collaborative optimization of pile foundation design parameters among the goals of construction safety, navigation compatibility, and long-term stability is realized.

[0039] Step S105, based on the anchorage depth mapping relationship and the real-time design parameter adjustment instructions, use a hybrid optimization algorithm to perform multi-objective optimization on the pile foundation layout parameters, where the hybrid optimization algorithm integrates a reinforcement learning module and a genetic algorithm module, and generates dynamic constraint conditions based on the pile foundation degradation data; In the initialization stage of the hybrid optimization algorithm, the reinforcement learning module constructs a state space including pile foundation layout parameters, the interaction force between adjacent piles, and the hydrodynamic characteristics of the water flow, and defines the combined action of the pile spacing adjustment amount and the inclination correction amount as the action space. The reward function comprehensively considers the natural frequency offset of the pile group, the stress distribution uniformity index, and the navigation compatibility evaluation result, and explores the global optimal solution in the solution space through the deep deterministic policy gradient algorithm to avoid the risk of local pile overload. For example, when it is detected that the stress difference between adjacent piles exceeds the threshold, the reward function triggers a negative penalty mechanism to guide the algorithm to optimize the uniformity of the pile position distribution.

[0040] The genetic algorithm module loads the pile foundation layout schemes in the historical engineering case library, encodes the pile position coordinates, anchorage depth, and material parameters into chromosome gene sequences to construct an initial population. The navigation safety constraints are constructed into a constraint matrix through the channel width threshold and the ship draft depth limit parameters, and the adaptive penalty function method is used to handle infeasible solutions, and the population individuals that meet the navigation rules are preferentially retained in the crossover and mutation operations. The pile foundation degradation monitoring data is collected in real time through corrosion rate sensors and scour depth monitoring devices, and the degradation index is generated through data fusion to dynamically adjust the minimum safety distance constraint boundary between adjacent piles in the pile spacing optimization model.

[0041] The generation process of dynamic constraint conditions analyzes the frequency domain characteristics of construction process data through a cloud deep learning model, extracts the energy distribution characteristics of vibration waveforms and formation reflection signals, and generates formation classification labels including the distribution probability of soft interlayers and the elastic modulus grade of rock and soil masses. After receiving the classification labels, the edge computing device adjusts the pile driving frequency and penetration speed parameters based on the fuzzy control algorithm. For example, in the area of soft interlayers, the hammering energy is reduced and the penetration interval time is extended to suppress the risk of pile body deflection. The adjusted construction parameters are synchronously uploaded to the cloud model, and the diversity of training samples is enhanced through the random time masking and frequency domain channel dropout strategy to optimize the accuracy of formation feature recognition.

[0042] During the optimization process, the blockchain network records the timestamps of the pile foundation design parameter versions and construction control instructions in real time, and verifies the data consistency through smart contracts. When it is detected that the version sequence of the pile foundation degradation index does not match the layout parameters, the recalibration process of the finite element simulation model is triggered, and the latest material property parameters and boundary condition datasets are loaded to iteratively update the pile-soil contact stiffness coefficient and the turbulent boundary layer calculation model. The updated simulation parameters are digitally signed and then transmitted back to the optimization algorithm module to dynamically correct the fitness function weight allocation strategy of the genetic algorithm.

[0043] In the above process, the reinforcement learning module realizes the efficient exploration of the global solution space, and the genetic algorithm module injects historical experience and safety constraint conditions. The two share optimization information through the population mixing strategy; the dynamic constraint conditions are dynamically adjusted based on real-time degradation data and formation characteristics to enhance the adaptability of the optimization model to complex working conditions; the blockchain technology supports the collaborative management and traceability verification of multi-version data to ensure the dynamic matching between the optimization process and the actual engineering state. Each link realizes the multi-dimensional collaborative optimization of the pile foundation layout parameters among the goals of navigation safety, construction efficiency, and long-term stability through the closed-loop interaction of data flow and control flow.

[0044] Step S106, perform pile foundation construction control according to the optimized pile foundation layout parameters, collect construction process data, extract features from the construction process data and conduct collaborative analysis in the cloud, generate formation classification labels, and feedback the formation classification labels to the construction control module to adjust the construction parameters; During the execution stage of pile foundation construction control, the optimized pile foundation layout parameters are sent to the hydraulic drive system through the edge computing node, driving the pile driver to adjust the hammering frequency, penetration speed, and the pitching angle of the guide frame. During the process of adjusting construction parameters, the PID control algorithm is used to dynamically match the preset penetration trajectory. For example, when the inclination angle of the pile body exceeds the threshold, a reverse compensation mechanism is triggered, and the pile body attitude is corrected in real time by adjusting the pressure of the hydraulic cylinder. A multi-axis vibration sensor array is deployed at the connection between the pile hammer and the pile top to collect the hammering vibration waveform at a sampling rate of not less than 1 kHz, and synchronously record the time-varying characteristics of the penetration depth and the soil layer reaction force.

[0045] The construction process data is transmitted to the edge gateway through the 5G narrowband Internet of Things for denoising and resampling preprocessing. The denoising process uses the wavelet threshold denoising algorithm to separate the high-frequency impact noise and the effective vibration signal; the resampling module dynamically adjusts the sampling interval based on the penetration speed of the pile body to maintain the integrity of the time-frequency characteristics of the vibration waveform. After preprocessing, the data stream generates a feature vector packet through the data compression algorithm. After marking the timestamp and spatial coordinates, it is uploaded to the cloud distributed storage cluster through an encrypted tunnel to establish a spatio-temporal association index with the historical engineering database.

[0046] The cloud collaborative analysis module deploys a convolutional neural network model to perform multi-scale analysis on the frequency-domain energy distribution characteristics of the vibration waveform. The input layer of the model receives the spectrogram generated by the fast Fourier transform, extracts the energy gradient characteristics of different frequency bands through the residual connection structure, and combines the cross-water formation feature library loaded by transfer learning to output the formation classification label including the distribution probability of soft interlayers, the elastic modulus grade of rock and soil masses, and the presence identification of reefs. For example, when it detects the characteristics of sudden increase in low-frequency energy accompanied by high-frequency attenuation, the model determines the existence of a dense gravel layer and generates the corresponding label.

[0047] The formation classification label is fed back to the edge construction control terminal in real time through the blockchain network, triggering the dynamic adjustment mechanism of construction parameters. The label parsing module adjusts the hammering energy output curve of the hydraulic system according to the elastic modulus grade of the rock and soil mass. For example, in the area of soft interlayers, it reduces the peak impact force and extends the interval time; the presence identification of reefs triggers the three-dimensional path planning module to recalculate the penetration trajectory and adopts the spiral drilling mode to bypass obstacles. The updated construction parameters are synchronously injected into the incremental training process of the cloud model, and the training sample set is expanded through random time window slicing and spectrum enhancement technology to improve the recognition and generalization ability of the model for complex formation characteristics.

[0048] In the above process, the edge execution system realizes millisecond-level dynamic adjustment of construction parameters to ensure the accuracy of pile foundation construction; the multi-modal data acquisition and preprocessing technology eliminates the influence of environmental interference on feature extraction; the cloud collaborative analysis establishes the mapping relationship between vibration characteristics and formation attributes through a deep learning model; the blockchain feedback mechanism ensures data traceability and instruction integrity. Each link forms a closed-loop control link of "parameter execution - data acquisition - feature analysis - decision feedback", and through the adaptive optimization of construction parameters driven by formation characteristics, the construction trial-and-error cost under complex geological conditions is reduced.

[0049] Step S107: Construct a scour and corrosion coupling prediction model based on the post-construction pile foundation stress data, and generate a protection decision instruction through the data cascade interaction correction of the scour and corrosion coupling prediction model. In the process of constructing the scour and corrosion coupling prediction model, first, the axial stress and bending stress distribution data of the post-construction pile foundation are collected through a distributed fiber optic sensor array, and the corrosion current density and potential gradient parameters are obtained by combining the electrochemical sensor array deployed on the pile surface. After the stress data are aligned by time stamps and standardized by a sliding window, they are input into the macroscopic scour analysis module. Based on the Navier-Stokes equation, the velocity gradient distribution of the flow field around the pile is solved, and the predicted value of the local scour depth is output; the corrosion monitoring data are used by the microscopic corrosion rate prediction module to calculate the dynamic corrosion rate distribution, and combined with the performance degradation curve in the material accelerated aging test database, a material durability evaluation model is established.

[0050] The data cascade interaction correction mechanism realizes the parameter linkage between the macroscopic and microscopic models through a two-way data interface. The water flow velocity gradient distribution data output by the macroscopic scour module are input into the microscopic corrosion model as the environmental driving factor for corrosion rate calculation; the extreme point distribution data of the microscopic corrosion rate are used to inversely correct the turbulent boundary layer thickness parameter of the macroscopic scour module, and the shear stress calculation model at the pile-water interface is iteratively updated. During the correction process, the dynamic time warping algorithm is used to align the macroscopic flow field data and microscopic corrosion data with different sampling frequencies, and the key spatio-temporal correlation features are extracted through principal component analysis to form a multi-dimensional coupling feature matrix.

[0051] The generation process of the protection decision instruction is constructed based on a graph neural network. The nodes are defined as the discretized grid units of the pile body, and the attributes include water flow gradient, corrosion rate, and material parameters. The graph convolutional layer aggregates the feature vectors of adjacent nodes, learns the interaction mechanism between scour and corrosion, and outputs an optimization scheme for the grading parameters of the filling material. The grading parameters match the performance curve of the anti-scour composite material in the preset material library, screen the candidate material types that meet the permeability threshold and corrosion inhibition efficiency, and generate a set of protection instructions including the corrected value of the pile foundation anchorage depth, the construction thickness of the anti-corrosion coating, and the aggregate ratio parameters.

[0052] The execution process of the protection instruction drives the adjustment of construction equipment parameters through edge computing nodes. The concrete mixing plant receives the grading parameter instruction and dynamically adjusts the aggregate particle size distribution and the proportion of porous medium incorporated; the anti-corrosion spraying equipment optimizes the coating coverage rate according to the corrosion rate distribution data by using an adaptive path planning algorithm. During the construction parameter adjustment process, the flow field monitoring data and material performance data collected in real time are fed back to the erosion-corrosion model, triggering the secondary correction of the boundary layer thickness parameter, and forming a closed-loop optimization link between model prediction and construction verification.

[0053] The blockchain network records the version number of the protection decision instruction and the time stamp of the construction parameters, and verifies the data consistency through a smart contract. When the deviation between the actual degradation rate of the material and the model prediction value is detected to exceed the threshold, the model recalibration process is triggered, and the latest monitoring data is loaded to iteratively update the weight parameters of the corrosion rate prediction module. The updated model parameters are synchronized to the distributed nodes after digital signature, ensuring the traceability of the data version in the whole life cycle and the transparency of the decision-making logic.

[0054] Step S108, the protection decision instruction, the pile foundation design parameters, and the construction control instruction are updated through data collaboration on the blockchain, and the smart contract verifies the data version consistency to trigger parameter correction.

[0055] In the data collaboration update stage, the protection decision instruction, the pile foundation design parameters, and the construction control instruction generate a unique data fingerprint through the hash algorithm and are bound to the time stamp provided by the Beidou time service system to form an immutable data block. The data block uses the Merkle tree structure to store the multi-level parameter relationships, such as the spatial correlation between the pile position coordinates and the anchorage depth, and the temporal dependence relationship between the construction instruction version and the protection decision logic, and ensures the security of the data transmission process through the asymmetric encryption algorithm. The blockchain nodes are deployed on the cloud server and the edge computing device, and a lightweight consensus mechanism is used to synchronously update the distributed ledger, reducing the network transmission bandwidth requirements.

[0056] The smart contract verification process obtains the time series of the pile foundation degradation monitoring data through an oracle and analyzes the matching relationship between the version number of the protection decision instruction and the time stamp of the construction parameters. When it is detected that the version of the pile body corrosion rate prediction model is incompatible with the current construction stage, the contract automatically triggers the recalibration task of the finite element simulation model. The recalibration process loads the latest material performance parameters and flow field monitoring data stored on the blockchain, screens the key variables through parameter sensitivity analysis, and uses the Latin hypercube sampling method to generate the boundary condition update set, iteratively optimizing the pile-soil contact stiffness coefficient and the turbulent boundary layer calculation model.

[0057] The generation process of the parameter correction instruction aligns the updated simulation model parameters with the protection decision instruction in terms of time and space, and establishes the mapping relationship between the material grading parameters and the pile foundation stress distribution through a graph neural network. The correction instruction includes the adjustment amount of the anchoring depth, the construction thickness threshold of the anti-corrosion coating, and the optimized value of the aggregate permeability. After being digitally signed, it is written into a newly added block of the blockchain. After the smart contract verifies the validity of the digital signature of the correction instruction, it drives the construction equipment control terminal to update the parameter set, such as adjusting the aggregate ratio of the concrete mixing plant or the hammering energy output curve of the hydraulic system.

[0058] The blockchain network realizes the interconnection with the material supplier database through the cross-chain protocol, and verifies the matching degree between the grading parameters and the anti-scour composite material performance curve in real time. When the deviation between the actual performance index of the material and the design parameter is detected, a reverse traceability mechanism is triggered to locate the abnormal data node in the supply chain link, and a supplier credit assessment report is generated. After being encrypted and stored, the assessment report is associated with the pile foundation full life cycle database to provide data support for subsequent project selection.

[0059] In the above process, the blockchain data storage structure ensures the spatio-temporal correlation and traceability of multi-source parameters; the smart contract verification mechanism realizes the adaptive matching of data versions and engineering stages; the generation and execution of the parameter correction instruction form a cross-model collaborative optimization link; the interconnection of supply chain data enhances the controllability of material performance. Each link forms a full-process trusted technology chain covering design, construction, and operation and maintenance through the closed-loop interaction of data fingerprint verification, model recalibration, and equipment control instructions, improving the dynamic adaptation ability and engineering reliability of the pile foundation protection strategy.

[0060] In the flexible support pile foundation process for floating solar power provided by the present invention, first, a sensor array deployed on the pile body and the water area environment is used to collect meteorological parameters, hydrological parameters, and pile body deformation parameters. The meteorological parameters include wind speed, wave height, and the vibration frequency of the photovoltaic module. The hydrological parameters cover water flow velocity, water depth, and suspended solid concentration. The pile body deformation parameters are monitored in real time by an accelerometer and a displacement sensor. The data collection frequency matches the dynamic change rate of the water area, and preprocessing such as outlier filtering and timestamp alignment is performed. During the dynamic fusion processing, the initial confidence weight is calculated based on the historical error rate of the sensor, combined with the dynamic deviation value of the actual deformation data and the meteorological and hydrological data feedback during the pile foundation construction stage. The sliding window statistical method is used to update the weight coefficient, and the adjusted weight is input into the spatio-temporal alignment algorithm to optimize the interpolation compensation parameters of multi-source data and generate a dynamic load spectrum integrating spatio-temporal characteristics.

[0061] A parallel input finite element simulation model for dynamic load spectra and a lightweight machine learning proxy model. The finite element simulation model is constructed based on the pile - soil - water coupling dynamics equation and outputs the mapping relationship between the anchorage depth and the pile foundation stress distribution. The machine learning proxy model generates real - time adjustment instructions for the construction inclination angle and penetration speed of the pile foundation by transferring and learning the anchorage depth characteristics of historical engineering cases. The hybrid optimization algorithm integrates a reinforcement learning module and a genetic algorithm module. The reinforcement learning module globally optimizes within the solution space based on the pile group stability objective, and the genetic algorithm module initializes the population distribution by loading the historical case library, injects navigation safety constraints and pile foundation degradation monitoring data, dynamically limits the range of feasible solutions, and generates pile foundation layout parameters that take into account construction - period safety and long - term stability.

[0062] The optimized pile foundation layout parameters are transmitted to the edge computing device to drive the hydraulic control system to adjust the hammering frequency and penetration trajectory of the pile driver. At the same time, high - sampling - rate vibration sensor is used to collect high - frequency vibration waveform data of the pile - soil interaction. The construction process data is compressed by wavelet transform and then uploaded to the cloud server. The deep - learning model deployed on the cloud analyzes the frequency - domain energy distribution characteristics of the vibration waveform, identifies the distribution of soft interlayers and reefs in the formation, generates formation classification labels and feeds them back to the edge device, triggering pre - adjustment of the hammering parameters, and the updated construction parameters simultaneously enhance the generalization ability of the cloud model.

[0063] After construction is completed, pile foundation stress data is obtained based on distributed fiber optic sensors to construct a coupled prediction model for scour and corrosion. The macroscopic scour model predicts the local scour depth around the pile according to the distribution of the water flow velocity gradient. The microscopic corrosion model receives real - time data from corrosion monitoring electrodes and reversely corrects the boundary layer thickness parameters of the scour model through data cascading to generate protection decision instructions. The protection decision instructions, pile foundation design parameters, and construction control instructions are stored in the blockchain network. After the smart contract verifies the consistency of the data version, it triggers the recalibration of the finite element simulation model parameters and the update of the material gradation optimization model, forming a closed - loop of data collaboration throughout the life cycle.

[0064] Specifically, for the flexible support pile foundation technology for floating solar power, the multi - source dynamic data is dynamically fused to generate a dynamic load spectrum. The dynamic fusion process includes: reversely correcting the fusion weight based on the actual deformation data of the pile foundation considering sensor confidence and construction feedback: calculating the initial confidence weight based on the sliding - window statistical results of the sensor historical error rate; Adjusting the initial confidence weight according to the dynamic deviation value between the actual deformation data of the pile foundation and the meteorological and hydrological data to obtain an adjusted weight value; Inputting the adjusted weight value into a preset spatio - temporal alignment algorithm to optimize the interpolation compensation parameters of the multi - source data.

[0065] In the dynamic weight allocation mechanism described in the present invention, first, the historical error rate of the sensor is statistically analyzed using a sliding window. The window length is dynamically adjusted according to the requirements of water area environment monitoring, and is specifically set based on the sensor sampling frequency and data update period. For the historical error data sets of meteorological sensors, hydrological sensors, and pile deformation sensors, the initial confidence weight is calculated using the error mean and variance within the sliding window. Among them, high-frequency abnormal data points are filtered by the interquartile range method, and the error statistical results in the stable monitoring interval are retained. During the calculation process of the initial confidence weight, an environmental temperature compensation factor is introduced for sensors sensitive to temperature drift to reduce the influence of external interference on the initial weight value.

[0066] In the dynamic weight adjustment stage, the dynamic deviation values of the pile foundation deformation monitoring data and meteorological and hydrological monitoring data during the construction process are obtained in real time. The deviation value quantifies the deviation degree between the sensor data and the actual response of the pile foundation through the root mean square error algorithm. When the deviation value exceeds the preset threshold, the confidence weight attenuation mechanism is triggered, and the exponential smoothing algorithm is used to dynamically reduce the fusion weight of abnormal sensors. At the same time, the hammer penetration resistance data and pile inclination angle data feedback by the pile foundation construction module are injected into the weight adjustment process as additional constraint conditions to correct the sensor data lag deviation caused by geological mutations.

[0067] After the optimized weight value is input into the spatio-temporal alignment algorithm, for the spatio-temporal heterogeneity characteristics of multi-source data, the dynamic time warping algorithm is used to align the sensor time series with different sampling frequencies. In the spatial dimension, a spatial interpolation grid is established based on the pile foundation topology coordinates, and the spatial attenuation coefficient in the inverse distance weighted interpolation algorithm is adjusted according to the weight value to compensate for the data blind area caused by uneven sensor distribution. During the optimization process of the interpolation compensation parameters, the local deformation data feedback by the pile foundation construction is used to verify the effectiveness of the interpolation results, forming a closed-loop optimization link for sensor weight adjustment and spatio-temporal alignment.

[0068] Specifically, the flexible support pile foundation technology for floating solar power described in the present invention further includes: using the anchoring depth mapping relationship output by the finite element simulation model as the training sample of the machine learning proxy model; The machine learning proxy model generates adjustment instructions for the pile foundation construction inclination parameters through transfer learning; The pile foundation stress data collected during the construction stage is fed back to the machine learning proxy model to correct the weight of the loss function of the machine learning proxy model.

[0069] In the collaborative analysis of the finite element simulation model and the machine learning surrogate model described in the present invention, the finite element simulation model is constructed based on the pile-soil-water coupling dynamics equation. By inputting the spatio-temporal distribution load parameters in the dynamic load spectrum, the coupled equation of the pile body stress field and displacement field is solved, and the non-linear mapping relationship between the anchorage depth and the pile foundation stress response is output. The mapping relationship of the anchorage depth includes the correlation characteristics between the pile penetration resistance and the critical instability load under different geological conditions, which is used as the training sample set of the machine learning surrogate model to initialize the network weight parameters of the surrogate model.

[0070] The machine learning surrogate model adopts a transfer learning architecture. In the pre-training stage, the anchorage depth data sets in different water environments in historical engineering cases are loaded, and the implicit correlation between the pile foundation geometric parameters and the anchorage depth is learned through the feature extraction network. In the fine-tuning stage, the finite element simulation results of the current project are introduced, and a dynamic curriculum learning strategy is adopted to gradually fuse high-confidence samples to generate an adjustment instruction set for the pile foundation construction inclination parameter. The adjustment instruction set is dynamically corrected according to the water flow velocity gradient data collected in real time, driving the hydraulic control system to adjust the pitching angle of the pile driver guide frame to suppress the asymmetric stress concentration during the pile penetration process.

[0071] During the construction stage, the axial stress and bending stress data of the pile foundation are collected by distributed fiber optic sensors and transmitted to the feedback control module of the machine learning surrogate model in real time. After the feedback data is statistically analyzed by a sliding time window, the deviation between the predicted inclination parameter and the actual stress distribution is calculated, and an adaptive momentum optimization algorithm is used to dynamically adjust the weight coefficients of different stress components in the loss function of the surrogate model. During the process of correcting the loss function weights, the inversion result of the elastic modulus of the soil layer around the pile is injected into the optimization process as a regularization constraint term to reduce the risk of model overfitting. The updated surrogate model parameters are synchronized to the edge computing node to generate the inclination adjustment instruction for the next construction stage, forming a data closed-loop optimization link between the finite element simulation and machine learning.

[0072] Specifically, for the flexible support pile foundation technology for floating solar power described in the present invention, the adjustment instructions for the pile foundation construction inclination parameter generated by the machine learning surrogate model through transfer learning include: The reinforcement learning module explores the global solution space based on the pile group stability goal; The genetic algorithm module loads the historical engineering case library to generate the initial population and injects the navigation safety constraint conditions; The dynamic monitoring data of the pile foundation degradation restricts the feasible solution range of the hybrid optimization algorithm.

[0073] In the multi-objective optimization process of the hybrid optimization algorithm described in the present invention, the reinforcement learning module constructs a state space and an action space based on the pile group stability objective. The state space includes real-time monitoring data of pile foundation layout parameters, hydrodynamic parameters of water flow, and mutual acting forces between adjacent piles. The action space is defined as a combined operation of pile spacing adjustment amount and pile inclination angle correction amount. The reward function of reinforcement learning is jointly constructed by the uniformity index of pile stress distribution, the deviation of the natural frequency of the pile group, and the evaluation result of navigation compatibility. The global optimal solution is explored in the solution space through the deep deterministic policy gradient algorithm to suppress the risk of local pile overload.

[0074] In the initialization stage of the genetic algorithm module, the pile foundation layout scheme of a similar water area environment is extracted from the historical engineering case database. The pile position coordinates, anchoring depth, and material parameters are encoded into a chromosome gene sequence to construct an initial population. The navigation safety constraint conditions construct a constraint matrix through the channel width threshold, ship draft depth limit, and visible distance parameter of pile foundation warning signs, and the adaptive penalty function method is used to handle infeasible solutions, and the population individuals that meet the navigation rules are retained for crossover and mutation operations.

[0075] The pile foundation degradation monitoring data is collected in real time by corrosion rate sensors, material fatigue detectors, and scour depth monitoring devices, and a degradation index is generated after data fusion. The degradation index is coupled with the optimization objective function to dynamically adjust the constraint boundary of the minimum safety distance between adjacent piles in the pile spacing optimization model. In each generation of evolution of the genetic algorithm, the pile position coordinate variables with the degradation index exceeding the limit are locked as immutable gene loci, forcing the optimization algorithm to search for feasible solutions in the remaining solution space, forming a multi-stage collaborative optimization mechanism for pile foundation layout parameters during the construction period and the operation and maintenance period.

[0076] Specifically, for the flexible support pile foundation technology for floating solar power described in the present invention, the feasible solution range of the hybrid optimization algorithm dynamically constrained by the pile foundation degradation monitoring data includes: The cloud deep learning model analyzes the frequency domain characteristics of the construction process data to generate formation classification labels; After receiving the formation classification labels, the edge computing device adjusts the pile driving frequency and penetration speed parameters; The updated construction parameters are input into the cloud deep learning model as training samples to optimize the generalization ability of the cloud deep learning model.

[0077] In the process of generating and feedback of formation classification labels according to the present invention, the construction process data is preprocessed by edge computing nodes, including denoising and resampling of vibration waveform signals, and the time-domain signals are decomposed into multi-scale frequency-domain subbands by wavelet packet transform. The deep learning model deployed in the cloud receives the preprocessed frequency-domain feature matrix, extracts the frequency-domain energy distribution features through a convolutional neural network, and combines the cross-water formation feature library loaded by transfer learning to output a formation classification label set, and the label set includes the distribution probability of soft interlayers, the grade of elastic modulus of rock and soil masses, and the identification of the existence of reefs.

[0078] After receiving the formation classification labels, the edge computing device dynamically adjusts the hammering frequency and penetration speed parameters of the hydraulic drive system based on the elastic modulus grade of the rock and soil mass and the distribution probability of soft interlayers in the labels. During the adjustment process, the real-time feedback data of the vibration sensor and the deviation amount of the target parameters are input into the fuzzy controller to generate a hammering energy compensation coefficient to suppress the pile body deflection vibration caused by formation mutation. The adjusted construction parameters are synchronously recorded in the local cache queue to form a mapping relationship data set of parameters - formation labels.

[0079] After the updated construction parameter data set is processed by data augmentation, it is input into the incremental training module of the deep learning model in the cloud. The data augmentation adopts random time masking and frequency-domain channel dropping strategies to expand the sample diversity. During the incremental training process, the model measures the similarity of new and old formation features through a contrast loss function, dynamically adjusts the convolution kernel weight allocation of the feature extraction network, and optimizes the generalization and recognition ability of the model for cross-water formation features. The trained model parameters are encrypted and transmitted to the edge node to update the formation classification rule library of the local inference engine, forming a closed-loop feedback link for construction parameter adjustment and model optimization.

[0080] Specifically, the flexible support pile foundation technology for floating solar power described in the present invention further includes: The scouring and corrosion coupling prediction model consists of a macroscopic scouring model and a microscopic corrosion model; The macroscopic scouring model is applied to the macroscopic scouring module; The microscopic corrosion model is applied to the microscopic corrosion rate prediction module; The water flow velocity gradient distribution data output by the macroscopic scouring module is input into the microscopic corrosion rate prediction module; The microscopic corrosion rate prediction module reversely corrects the boundary layer thickness parameter of the macroscopic scouring module based on the collected corrosion monitoring data to generate updated flow field distribution data; Spatially and temporally align the updated flow field distribution data with the microscopic corrosion rate prediction results, input them into a preset graph neural network, establish a non-linear mapping relationship of the grading parameters of the filling material, generate a prediction result of the coupling of scouring and corrosion, and generate a set of instruction sets for the protection construction plan to drive the construction equipment to adjust the hammering frequency and material ratio.

[0081] In the flexible support pile foundation technology for floating solar power described in the present invention, the coupling prediction model of scouring and corrosion consists of a macroscopic scouring module and a microscopic corrosion rate prediction module. The macroscopic scouring module simulates the flow field distribution around the pile based on the hydrodynamic equation, inputs the water flow velocity, water depth and pile geometry parameters monitored in real time, and outputs the water flow velocity gradient distribution data by solving the Navier-Stokes equation. The microscopic corrosion rate prediction module receives the water flow velocity gradient data provided by the macroscopic scouring module, combines the corrosion current density and potential distribution collected by the electrochemical sensor array on the pile surface, calculates the local mass transfer coefficient, and predicts the dynamic corrosion rate distribution.

[0082] The water flow velocity gradient distribution data output by the macroscopic scouring module is transmitted to the microscopic corrosion rate prediction module through a data interface as an environmental driving factor for corrosion rate calculation. The microscopic corrosion rate prediction module extracts the gradient extreme points of the corrosion rate distribution, reversely corrects the turbulent boundary layer thickness parameter of the macroscopic scouring module, and updates the calculation conditions of the shear stress distribution at the pile-water interface. The corrected macroscopic scouring module recalculates the flow field distribution, generates updated flow field data, and spatially and temporally aligns it with the microscopic corrosion rate prediction results. The spatio-temporal alignment process uses the dynamic time warping algorithm to align the macroscopic and microscopic data with different sampling frequencies, and extracts the key spatio-temporal correlation features through principal component analysis to form a multi-dimensional coupling feature matrix.

[0083] The updated flow field data and the microscopic corrosion rate prediction results are input into a preset graph neural network. The nodes are defined as the discretized grid units of the pile, and the attributes include the water flow gradient, corrosion rate and material parameters. The edge weights are determined by the correlation of the water flow velocity gradient and the difference in corrosion rate between adjacent units. The graph neural network aggregates the feature vectors of adjacent nodes through the graph convolution layer, learns the coupling mechanism of scouring and corrosion, and outputs the non-linear mapping relationship of the grading parameters of the filling material. The grading parameters include the permeability of the porous medium, the aggregate particle size distribution and the type of anti-scouring composite material. After matching the performance curves in the preset material library, a set of instruction sets for the protection construction plan is generated.

[0084] The instruction set of the protection construction plan drives the construction equipment through the edge computing device to adjust the process parameters, specifically including optimizing the aggregate ratio of the concrete mixing plant according to the grading parameters, adjusting the pile foundation anchorage depth based on the predicted scour depth, and controlling the construction thickness and coverage rate of the anti-corrosion coating based on the corrosion rate distribution. The adjusted construction parameters are synchronously fed back to the blockchain network and updated in coordination with the pile foundation design parameters and real-time monitoring data to form a dynamic optimization closed-loop for the entire life cycle.

[0085] In the flexible support pile foundation technology for floating solar power of the present invention, the coupled prediction results of scour and corrosion include the following data: the water flow velocity gradient distribution data output by the macroscopic scour module, the predicted value of the local scour depth around the pile, the corrosion rate distribution data output by the microscopic corrosion rate prediction module, the predicted parameters of the material degradation trend, and the grading parameters of the filling material generated by the graph neural network. Among them, the water flow velocity gradient distribution data reflects the dynamic characteristics of the flow field around the pile body, the predicted value of the local scour depth characterizes the degree of hydraulic erosion of the soil around the pile foundation, the corrosion rate distribution data quantifies the dynamic corrosion rate of different regions on the pile body surface, the predicted parameters of the material degradation trend are generated based on the electrochemical sensor data and the historical aging curve, and the grading parameters of the filling material include the selection suggestions for the permeability of the porous medium, the aggregate particle size distribution, and the type of anti-scour composite material.

[0086] The generation process of the instruction set of the protection construction plan is as follows: the graph neural network matches the performance curves of the anti-scour composite materials in the preset material library according to the grading parameters in the coupled prediction results of scour and corrosion, and screens the candidate material types that meet the permeability threshold and corrosion inhibition efficiency. Combining the predicted value of the local scour depth and the corrosion rate distribution data, the adjustment amount of the pile foundation spacing, the correction value of the anchorage depth, and the construction thickness of the anti-corrosion coating are calculated through the constraint satisfaction algorithm. The construction parameter optimization module converts the above calculation results into an instruction set executable by the construction equipment, including the instruction for adjusting the aggregate ratio of the concrete mixing plant, the instruction for controlling the hammering frequency of the pile driver, and the instruction for setting the operation parameters of the anti-corrosion spraying equipment.

[0087] The logical relationship between the prediction results of erosion-corrosion coupling and the instruction set of the protection construction plan is as follows: Through multi-scale data fusion and machine learning modeling, the prediction results of erosion-corrosion coupling identify the dynamic risk distribution of pile foundations in complex water environments, providing a decision-making basis for protection strategies. Based on the quantitative parameters in the prediction results, the instruction set of the protection construction plan transforms the decision-making basis into specific construction actions, forming a closed-loop control link of "risk prediction - strategy matching - parameter adjustment". For example, when the prediction results show that the corrosion rate in a certain area of the pile body exceeds the safety threshold, the instruction set drives the anti-corrosion spraying equipment to increase the coating coverage rate in this area; when the predicted value of the local scour depth exceeds the design allowable range, the instruction set triggers the dynamic correction of the pile foundation anchorage depth and synchronously updates the pile position coordinate data through the blockchain network. Through the data cascading and reverse correction mechanism, this logical relationship realizes the cross-scale collaborative optimization of macroscopic fluid parameters and microscopic material properties, ensuring the dynamic adaptation of construction parameter adjustment and long-term stability goals.

[0088] Specifically, for the flexible support pile foundation process for floating solar power, the preset graph neural network establishes a non-linear mapping relationship of the grading parameters of the filling material based on the boundary layer thickness parameters of the corrected macroscopic scour module, including: Bind the construction control instruction version number with the protection decision timestamp and store it in the blockchain; The smart contract verifies the matching of the characteristic dimensions of the dynamic load spectrum with the input layer of the machine learning agent model; When a version conflict is detected, trigger the recalibration of parameters of the finite element simulation model.

[0089] During the blockchain data collaborative update process of the present invention, the construction control instruction version number generates a unique identifier using the hash algorithm, and is bound to the timestamp generated by the protection decision module through the Merkle tree structure and then written into the blockchain distributed ledger. The timestamp is synchronized based on the Beidou satellite timing system, includes the pile foundation construction stage identifier and the data operation type code, and ensures the integrity and traceability of the stored data through the asymmetric encryption algorithm. The data block after binding is verified by the consensus node to generate a block hash value, which is associated with the index directory of the pile foundation full life cycle database.

[0090] The smart contract is deployed on the blockchain node, obtains the characteristic dimension data of the dynamic load spectrum through the oracle. The characteristic dimensions include the quantity and type of time-domain statistics, frequency-domain energy distribution, and spatial gradient parameters. The smart contract calls the metadata interface of the machine learning agent model, parses the tensor dimension and data type constraints of its input layer, and uses the feature vector similarity matching algorithm to verify the compatibility between the load spectrum characteristics and the model input layer. When a dimension mismatch or data type conflict is detected, trigger the exception handling process, pause the model inference task and generate an audit log event.

[0091] The version conflict detection mechanism monitors the version serial numbers of the pile foundation design parameters, construction control instructions, and protection decision-making data in the blockchain ledger in real time. When the smart contract verifies that the time series of the parameter version and the load spectrum version is discontinuous, it sends a model recalibration instruction to the finite element simulation task scheduling system. During the recalibration process, the latest version of the pile body material parameters and boundary condition datasets are loaded. Key variables are screened through parameter sensitivity analysis, and the Latin hypercube sampling method is used to generate a calibration sample set to drive the finite element simulation model to iteratively update the turbulent boundary layer calculation parameters and the pile-soil contact stiffness coefficient. The updated simulation parameters are digitally signed and then transmitted back to the blockchain network, triggering an incremental update of the data version numbers of relevant modules.

[0092] Specifically, for the flexible support pile foundation process for floating solar power, the generation of the dynamic load spectrum includes: Decompose the time series of the multi-source dynamic data into frequency domain sub-bands through wavelet packet transform; Extract entropy value features from the frequency domain sub-bands to generate a high-dimensional feature space.

[0093] During the generation process of the dynamic load spectrum of the present invention, the time series of the multi-source dynamic data is first processed by the timestamp alignment method to eliminate the sampling time delay difference between sensors, and the Daubechies wavelet basis function is used to perform wavelet packet transform decomposition on the aligned time series. The decomposition level is adaptively selected according to the spectral characteristics of the dynamic load in the water area environment. A higher resolution decomposition is adopted for the main frequency band of the wind and wave load to generate a set of frequency domain sub-bands containing time-frequency energy. After the frequency domain sub-bands are filtered by a Butterworth band-pass filter to remove high-frequency noise interference, the effective frequency band reflecting the dynamic response of the pile foundation is retained.

[0094] Perform multi-scale entropy value feature extraction on the preprocessed frequency domain sub-bands, calculate the approximate entropy, sample entropy, and permutation entropy indexes of each sub-band to quantify the complexity and non-linear characteristics of the load waveform. The entropy value feature extraction adopts a sliding time window strategy, and the window length is synchronized with the natural vibration period of the pile foundation to capture the time-varying load characteristics. The extracted entropy value feature vector is normalized by the maximum and minimum values to eliminate the influence of the energy scale difference of different frequency bands and construct a high-dimensional feature space matrix.

[0095] The adaptive Bayesian optimization algorithm uses the high-dimensional feature space matrix as the input domain. In the initialization stage, the historical strain monitoring data set of the pile foundation is loaded as prior knowledge to construct a Gaussian process surrogate model. The kernel function of the surrogate model adopts a composite structure of the Matérn kernel and the linear kernel to capture the non-linear relationship between features and the energy trend of frequency domain sub-bands respectively. During the optimization process, the expected improvement function combined with the local penalty mechanism is used to dynamically balance between global feature space exploration and local high-correlation region development. The strain data collected in real time from the pile foundation constructs an objective function through the joint use of the Pearson correlation coefficient and the KL divergence, driving the Bayesian optimization to screen for a subset of feature combinations that maximizes the correlation between the load spectrum and the strain response. The selected feature combination is dimensionally reduced by principal component analysis to generate a dynamic load spectrum, which is used as the collaborative input for finite element simulation and machine learning models.

[0096] Specifically, for the flexible support pile foundation process for floating solar power in the present invention, the adaptive Bayesian optimization algorithm includes: Construct a Gaussian process surrogate model with the feature selection combination of multi-source dynamic data as the input and the correlation index of the scour and corrosion coupling prediction result as the output; Adopt the local penalty expected improvement function to balance the global exploration and local exploitation search strategies; Dynamically inject the actual deformation data of the pile foundation in the feedback construction parameters into the weight coefficient update process of the objective function.

[0097] During the implementation of the adaptive Bayesian optimization algorithm of the present invention, the feature selection combination consists of the frequency domain sub-band entropy value features output by the dynamic load spectrum generation module and the pile body geometric parameters, and is input into the Gaussian process surrogate model. The kernel function of the surrogate model adopts a composite structure. The Matérn kernel captures the non-linear correlation between features, and the linear kernel models the linear trend of the boundary layer thickness parameter and the corrosion rate in the scour and corrosion coupling prediction result. The output correlation index measures the linear correlation between the load spectrum features and the actual deformation data of the pile foundation through the Pearson correlation coefficient, and combines the KL divergence to quantify the non-linear distribution difference to form a multi-dimensional evaluation index.

[0098] The local penalty expected improvement function adopts a large-range parameter sampling strategy in the global exploration stage, covering the feature combinations in the high-uncertainty region of the scour-corrosion coupling prediction model; in the local exploitation stage, it locks the high-correlation feature subspace based on the gradient descent algorithm, and suppresses the redundant search of adjacent sampling points through the dynamic adjustment mechanism of the penalty radius. The penalty radius is associated with the update frequency of the pile foundation degradation monitoring data. When a sudden change in material properties is detected, the penalty radius is enlarged to force the exploration of a new solution space to avoid the optimization falling into a local optimum.

[0099] The actual deformation data of the pile foundation in the construction parameters are collected in real time by the edge computing node. After being processed by the sliding window normalization, it is injected into the target function weight update module. The weight coefficient is dynamically allocated according to the spatio-temporal distribution characteristics of the strain data, and the ratio of the axial strain weight to the bending strain weight is adaptively adjusted with the inclination angle of the pile body. During the weight update process, the boundary layer thickness parameter corrected by the scour model is used as a regularization constraint term to limit the oscillation amplitude of the weight coefficient and maintain the numerical stability of the optimization process. The updated target function drives the Bayesian optimization to iteratively screen the feature combinations, generates a subset of the load spectrum that is strongly correlated with the dynamic response of the pile foundation, and synchronously feeds it back to the boundary condition input interface of the finite element simulation model to form a data closed-loop of load optimization and pile foundation performance prediction.

[0100] Specifically, for the flexible support pile foundation technology for floating solar power, the construction of the Gaussian process surrogate model, with the feature selection combination of multi-source dynamic data as the input and the correlation index of the coupled scour and corrosion prediction result as the output, includes: Input the coupled scour and corrosion prediction result into the material durability evaluation model; Generate a filling construction plan based on the grading optimization parameters output by the graph neural network; Verify the degradation trend by comparing the historical data of the protection decision recorded in the blockchain with the prediction result of the material durability evaluation model.

[0101] During the generation process of the protection dynamic decision instruction of the present invention, the coupled scour and corrosion prediction result is input into the material durability evaluation model. Based on the chemical composition of the pile body material and the microscopic corrosion rate distribution data, combined with the material performance degradation curve in the accelerated aging test database, the model predicts the tensile strength attenuation rate and fatigue life threshold under different protection periods. The output of the evaluation model includes a set of durability indicators containing the critical corrosion depth and the allowable stress relaxation amount, which are used as the constraint conditions for the selection of the filling material.

[0102] The graph neural network receives the durability evaluation index set and the pile perimeter flow field distribution data, aggregates the scour-corrosion feature vectors of adjacent pile body nodes through the graph attention mechanism, and generates the grading optimization parameters of the porous medium permeability and the aggregate particle size. The parameter set matches the composite material anti-scour performance curve in the preset material library, and the constraint satisfaction algorithm is used to screen the candidate materials that simultaneously meet the permeability threshold and the corrosion inhibition efficiency, and generate a set of filling plan instructions including the construction ratio and process parameters.

[0103] The blockchain network retrieves historical protection decision data, extracts the deviation sequence between the actual degradation rate of the material and the predicted value, and constructs a degradation trend verification model through time series similarity analysis. The verification model captures the temporal dependence relationship of material performance degradation based on a long short-term memory network, compares the difference in the decay mode between the current prediction result and historical data, and triggers a protection plan backtracking mechanism when an abnormal deviation is detected, updates the threshold parameters of the material durability evaluation model, and generates a correction instruction. The corrected parameters are synchronized to the blockchain ledger after digital signature, associated with the decision-making process of the next construction cycle, and form a dynamic optimization closed-loop of the protection strategy.

[0104] The explanations of the technical feature terms in the technical solution of the present invention are as follows: Multi-source dynamic data: It refers to the real-time monitoring data set collected by meteorological sensors, hydrological sensors, and pile body deformation sensors, including parameters such as wind speed, wave height, water flow velocity, pile body strain, and displacement, which are used to reflect the dynamic characteristics of the water area environment and the real-time response of the pile foundation structure.

[0105] Dynamic weight allocation mechanism: Dynamically adjusts the weight ratio of different sensor data during the fusion process based on the historical error rate of the sensor and construction feedback data. For example, the initial weight is calculated by statistically averaging the error mean and variance through a sliding window, and the weight value is reversely corrected in combination with the actual deformation deviation of the pile body to improve the reliability of data fusion.

[0106] Wavelet packet transform: A signal decomposition method that decomposes time series data into different frequency domain sub-bands, which is used to extract time-frequency characteristics in the dynamic load spectrum. For example, the Daubechies wavelet basis is used for multi-scale frequency domain analysis of wind speed and water flow velocity data.

[0107] Entropy value feature extraction: Quantifies the complexity and non-linear characteristics of the load waveform by calculating indicators such as approximate entropy and sample entropy of the frequency domain sub-bands, and is used to construct a high-dimensional feature space to screen the load feature combinations strongly related to pile foundation deformation.

[0108] Finite element simulation model: A numerical model constructed based on the pile-soil-water coupling dynamics equation. By solving the coupling equation of the stress field and displacement field, it simulates the mechanical response of the pile foundation under different anchorage depths and outputs the mapping relationship between the anchorage depth and the critical instability load.

[0109] Machine learning proxy model: A lightweight model using a transfer learning architecture. In the pre-training stage, historical engineering case data is loaded, and in the fine-tuning stage, real-time simulation results are fused to generate adjustment instructions for pile inclination and penetration speed. For example, the pile group layout parameters are optimized through a reinforcement learning module.

[0110] Hybrid Optimization Algorithm: A multi-objective optimization method integrating reinforcement learning and genetic algorithm. Reinforcement learning explores the global solution space based on the pile group stability objective, and the genetic algorithm loads the historical case library and injects navigation safety constraints to dynamically adjust the constraint boundaries of pile spacing and material degradation monitoring data.

[0111] Blockchain Technology: It uses a distributed ledger to store versioned data of pile foundation design parameters, construction instructions, and protection decisions, and verifies data consistency through smart contracts. For example, when the deviation between the detected material degradation rate and the predicted value is detected, it triggers the recalibration of the finite element model and parameter correction.

[0112] Scour and Corrosion Coupling Prediction Model: It consists of a macroscopic hydrodynamic model and a microscopic corrosion prediction module. The macroscopic model calculates the velocity gradient distribution of the flow field around the pile, and the microscopic model predicts the corrosion rate based on electrochemical sensor data, and establishes a grading parameter optimization scheme through a graph neural network.

[0113] Edge-Cloud Collaborative Control: Edge devices execute real-time adjustment of construction parameters driven by formation classification labels (such as adjusting the hammering frequency), and the cloud analyzes the frequency domain characteristics of the vibration waveform to generate formation classification labels, forming a closed-loop feedback link between construction parameters and model training.

[0114] Spatio-Temporal Alignment Algorithm: Aiming at the spatio-temporal heterogeneity of multi-source data, it uses dynamic time warping to align time series data with different sampling frequencies, and constructs a spatial interpolation grid in combination with the pile foundation topology coordinates to eliminate the spatio-temporal mismatch problem in data acquisition.

[0115] Material Durability Evaluation Model: Based on the accelerated aging test database and real-time corrosion monitoring data, it predicts the tensile strength attenuation rate and fatigue life threshold of pile materials, providing a basis for material selection in the protection strategy.

[0116] Dynamic Weight Allocation Model: Based on the historical error data set of sensors and the actual deformation data of the pile foundation from construction feedback, it calculates the initial confidence weights through a sliding window statistical method, and combines the environmental temperature compensation factor to correct the weight allocation. The deviation of the pile body deformation collected in real time triggers the exponential smoothing algorithm to dynamically adjust the weights, optimize the spatio-temporal interpolation compensation parameters of multi-source data, eliminate the data deviation caused by uneven sensor distribution or environmental interference, and provide a high-confidence data fusion basis for generating dynamic load spectra.

[0117] Finite Element Simulation Model: A physical model constructed based on the pile-soil-water coupling dynamics equation. It inputs the spatio-temporal distribution load parameters in the dynamic load spectrum, solves the nonlinear equations of the pile body stress field and displacement field, and outputs the mapping relationship between the anchorage depth and the critical instability load. Through the adaptive grid encryption technology, it captures the gradient change of the shear stress at the pile-soil interface, introduces a material degradation parameter correction module, and dynamically updates the boundary conditions according to the corrosion monitoring data to improve the adaptability of the simulation results to complex working conditions.

[0118] Machine learning proxy model: A lightweight model using a transfer learning architecture. In the pre-training stage, it loads the anchorage depth feature dataset of historical engineering cases. In the fine-tuning stage, it fuses the finite element simulation results of the current project to generate real-time adjustment instructions for the pile inclination and penetration speed. The model explores the global solution space through a reinforcement learning module to optimize the pile group stability, and combines a genetic algorithm module to inject navigation safety constraints and pile foundation degradation monitoring data to form a multi-dimensional collaborative optimization mechanism for construction parameters.

[0119] Hybrid optimization algorithm model: A multi-objective optimization framework integrating reinforcement learning and genetic algorithms. The reinforcement learning module constructs a state space and a reward function based on the pile group stability objective, and explores the global optimal solution through the deep deterministic policy gradient algorithm. The genetic algorithm module loads the historical case library to generate the initial population, uses an adaptive penalty function to handle navigation constraints, and dynamically adjusts the pile spacing to optimize the safety distance boundary in the model. The model dynamically corrects the weight allocation strategy of the optimization objective through the linkage of the pile foundation degradation index and construction feedback data.

[0120] Scour and corrosion coupling prediction model: A cross-scale model containing a macroscopic scour module and a microscopic corrosion rate prediction module. The macroscopic module simulates the velocity gradient distribution of the flow field around the pile based on the Navier-Stokes equation. The microscopic module combines electrochemical sensor data to predict the dynamic corrosion rate, and through a data cascade interaction correction mechanism, reversely adjusts the turbulent boundary layer thickness parameter. The graph neural network aligns the corrected flow field data and the corrosion rate prediction results in space and time, establishes a non-linear mapping relationship of the filling material gradation parameters, and generates protection decision instructions such as anchorage depth correction and aggregate ratio optimization.

[0121] Blockchain data collaboration model: It uses a distributed ledger to store the versioned data of pile foundation design parameters, construction control instructions, and protection decisions, and verifies the consistency of data timestamps through smart contracts. For example, when it detects that the version of the material corrosion rate prediction model does not match the construction stage, it triggers the finite element model parameter recalibration process, loads the latest monitoring data, and iteratively updates the turbulent boundary layer calculation conditions to ensure the dynamic synchronization of the model prediction and the actual engineering state.

[0122] Edge-cloud collaborative analysis model: The edge computing node executes real-time adjustment of construction parameters driven by formation classification labels (such as dynamic adjustment of the hammering frequency). The convolutional neural network deployed in the cloud analyzes the frequency domain energy characteristics of the vibration waveform and generates formation classification labels including the elastic modulus level of the rock and soil mass. The labels are fed back to the edge device to form a closed-loop control. At the same time, the updated construction parameters are transmitted back to the cloud through a data augmentation strategy to optimize the generalization and recognition ability of the model.

[0123] Material Durability Evaluation Model: Based on the material performance degradation curves in the accelerated aging test database, combined with real-time corrosion current density and potential gradient data, predict the attenuation rate of the pile body's tensile strength and the fatigue life threshold. The output results of the model are input into a graph neural network to match the performance curves of erosion-resistant composite materials in the preset material library, screen the grading parameters that meet the permeability and corrosion inhibition efficiency, and support the generation of the protection construction plan.

[0124] Adaptive Bayesian Optimization Algorithm: An intelligent optimization method based on a probability model that establishes the mapping relationship between input features (such as the frequency-domain sub-band entropy value of the dynamic load spectrum and the pile body geometric parameters) and the target output (such as the pile foundation deformation correlation index) by constructing a Gaussian process surrogate model. This algorithm uses a local penalty expected improvement function to balance the global exploration and local exploitation strategies. In the global stage, the Latin hypercube sampling is used to cover the potentially high-value regions of the feature space, and in the local stage, the gradient descent is based on to lock in the high-probability optimal solution region. The dynamic adjustment mechanism dynamically updates the weights of the objective function according to the pile foundation deformation data feedback during construction. For example, when a sudden change in material performance is detected, the search range is expanded to prevent local convergence, and at the same time, the feasible solution space is dynamically constrained by combining the pile foundation degradation monitoring data. By iteratively screening the feature combinations strongly correlated with the pile foundation response, the engineering applicability of the dynamic load spectrum is effectively improved.

[0125] The above technical features are linked through dynamic data fusion, model collaborative optimization, and a closed-loop feedback mechanism to achieve improved accuracy in pile foundation load analysis, dynamic adaptation of construction parameters, and optimization of the full-life cycle protection strategy, solving the problems that traditional static models cannot integrate multi-source dynamic data and the lack of coordination between pile foundation layout and navigation and degradation constraints, and ultimately adapting to the dynamic deformation coordination requirements of flexible supports in complex water environments.

[0126] The specific implementation of the present invention relates to the flexible support pile foundation technology for floating solar power plants, achieving precise control of pile foundation design through multi-source data dynamic fusion and intelligent optimization. During the implementation process, Doppler radars, piezoelectric accelerometers, and distributed fiber optic sensor arrays are first deployed to collect wind speed, wave height, water flow velocity, and pile body deformation parameters in real time. After the sensor data is synchronized by the Beidou timing system, the sliding window interquartile range method is used to filter out outliers, and the dynamic time warping algorithm is used to align the time series data with different sampling frequencies to generate a multi-source data set with a unified spatio-temporal baseline. The preprocessed data is input into the dynamic fusion module, the initial confidence weights are calculated based on the historical error rate of the sensors, and the weights are corrected in reverse by combining the actual pile foundation deformation data feedback during construction to optimize the spatio-temporal interpolation compensation parameters and generate a high-dimensional dynamic load spectrum containing frequency-domain sub-band energy distribution and entropy value features.

[0127] Parallel input finite element simulation model and machine learning surrogate model for dynamic load spectra. The finite element model is constructed based on the pile-soil-water coupling dynamics equation, and the adaptive mesh refinement technology is used to capture the stress gradient at the pile-soil interface, outputting the mapping relationship between the anchorage depth and the pile foundation stress response. The machine learning surrogate model loads the anchorage depth characteristics of historical engineering cases through transfer learning, and combines real-time water flow gradient data to generate instructions for adjusting the pile inclination and penetration speed. The hybrid optimization algorithm integrates a reinforcement learning module and a genetic algorithm module. The reinforcement learning constructs the state space and reward function based on the pile group stability goal, and the genetic algorithm loads the historical case library to initialize the population and injects navigation safety constraints, dynamically adjusting the pile spacing and material degradation constraints to generate pile foundation layout parameters that take into account construction safety and long-term stability.

[0128] During the construction stage, the optimized parameters are sent to the hydraulic control system through the edge computing node, driving the pile driver to adjust the hammering frequency and penetration trajectory. The vibration sensor array collects the pile-soil interaction waveform at a sampling rate not lower than 1 kHz, and after wavelet threshold denoising and resampling, it is uploaded to the cloud. The convolutional neural network in the cloud analyzes the frequency domain energy distribution characteristics of the vibration waveform, outputs the formation classification label and feeds it back to the edge device, triggering the dynamic adjustment of the hammering energy and penetration speed. After construction, the pile foundation stress data is input into the scour and corrosion coupling prediction model. The macroscopic scour module calculates the flow field distribution around the pile based on the Navier-Stokes equation, and the microscopic corrosion module combines the electrochemical sensor data to predict the material degradation trend. A grading parameter optimization scheme is established through the graph neural network, driving the concrete mixing station and anti-corrosion spraying equipment to adjust the construction parameters.

[0129] The blockchain network uses the Merkle tree structure to store the versioned data of design parameters, construction instructions, and protection decisions. The smart contract triggers the model recalibration after verifying the timestamp consistency. For example, when it is detected that the deviation between the material corrosion rate and the predicted value exceeds the threshold, the latest monitoring data is loaded to iteratively update the finite element boundary conditions and optimize the turbulent boundary layer thickness parameters. The corrected instructions are written into the new block after digital signature, driving the construction equipment to adjust the grading parameters and coating coverage. The supply chain data verifies the material performance indicators through the cross-chain protocol, and the abnormal data triggers reverse traceability and generates a supplier credit assessment report, which is associated with the pile foundation full life cycle database. The above process forms a data closed-loop for design, construction, and operation and maintenance, realizing the dynamic deformation coordination and long-term stable bearing of the pile foundation in the complex water area environment.

[0130] Through the fusion processing of multi-source dynamic data and the closed-loop feedback mechanism, the present invention solves the problem that traditional pile foundation design relies on static load models. First, a sensor array is deployed to collect meteorological, hydrological, and pile deformation parameters in real time. Based on the dynamic weight allocation mechanism, the sensor fusion weights are adjusted. Combining the actual pile deformation data feedback from construction, the parameters of the spatio-temporal alignment algorithm are corrected in reverse to generate a high-precision dynamic load spectrum. The load spectrum is decomposed by wavelet packet transform and entropy value feature extraction to construct a spatio-temporal correlated load feature space, effectively eliminating the dynamic load coupling effect ignored by the static model and filling the blind spot in the correlation analysis of the anchorage depth and the mechanical properties of the pile foundation.

[0131] Furthermore, a collaborative architecture of a finite element simulation model and a machine learning surrogate model is adopted to process the dynamic load spectrum data in parallel. The finite element model outputs the mapping relationship of the anchorage depth based on the pile-soil-water coupling dynamics equation. The machine learning surrogate model generates real-time adjustment instructions for the pile inclination and penetration speed through transfer learning. Combining the hybrid optimization algorithm to integrate the global optimization of reinforcement learning and the historical case guidance of the genetic algorithm, injecting navigation safety constraints and pile foundation degradation monitoring data, a multi-dimensional collaborative optimization model is constructed. This architecture breaks through the limitations of traditional single-objective optimization and realizes the collaborative modeling of pile foundation layout parameters in terms of navigation compatibility, hydrodynamic stability, and long-term safety, improving the spatial resolution of scour risk prediction.

[0132] In addition, an edge-cloud collaborative control and blockchain data management mechanism is established to solve the problem of dynamic adaptation in complex water environments. The edge device executes millisecond-level adjustment of construction parameters driven by formation classification labels. The cloud deep learning model analyzes high-frequency vibration data to identify abnormal formation features, forming a construction control closed-loop; the multi-scale scour-corrosion coupling model correlates macroscopic fluid parameters with microscopic material degradation characteristics through data cascading and reverse correction mechanisms; the blockchain stores the full-life cycle data and verifies the version consistency, triggering the re-calibration of the finite element model parameters and the dynamic update of the protection strategy, ultimately achieving the dual goals of dynamic deformation coordination and long-term stable load-bearing of the flexible support pile foundation.

Claims

1. A flexible support pile foundation process for floating solar power, characterized in that, Including: Obtaining multi-source dynamic data of the water area environment, where the multi-source dynamic data includes meteorological parameters, hydrological parameters, and pile body deformation parameters; Performing dynamic fusion processing on the multi-source dynamic data to generate a dynamic load spectrum, where the dynamic fusion processing includes reverse-correcting the fusion weight based on the actual deformation data of the pile foundation with sensor confidence and construction feedback; Screening the feature combinations of the multi-source dynamic data based on the adaptive Bayesian optimization algorithm to optimize the correlation between the dynamic load spectrum and the actual deformation data of the pile foundation; Parallelly inputting the dynamic load spectrum into a finite element simulation model and a machine learning proxy model, outputting the anchorage depth mapping relationship through the finite element simulation model, and generating a real-time design parameter adjustment instruction through the machine learning proxy model; Based on the anchorage depth mapping relationship and the real-time design parameter adjustment instruction, performing multi-objective optimization on the pile foundation layout parameters using a hybrid optimization algorithm, where the hybrid optimization algorithm integrates a reinforcement learning module and a genetic algorithm module, and generates dynamic constraint conditions based on the pile foundation degradation data; Executing pile foundation construction control according to the optimized pile foundation layout parameters and collecting construction process data; Performing feature extraction and cloud collaborative analysis on the construction process data to generate a formation classification label, and feeding back the formation classification label to the construction control module to adjust construction parameters; Constructing a scour and corrosion coupling prediction model based on the pile foundation stress data after construction, and generating a protection decision instruction through data cascade interaction correction of the scour and corrosion coupling prediction model; Data collaboratively updates the protection decision instruction, the pile foundation design parameters, and the construction control instruction through the blockchain, where the smart contract verifies the data version consistency to trigger parameter correction.

2. The flexible support pile foundation process for floating PV according to claim 1, wherein, Performing dynamic fusion processing on the multi-source dynamic data to generate a dynamic load spectrum, where the dynamic fusion processing includes reverse-correcting the fusion weight based on the actual deformation data of the pile foundation with sensor confidence and construction feedback includes: calculating the initial confidence weight based on the sliding window statistical result of the sensor historical error rate; Adjusting the initial confidence weight according to the dynamic deviation value between the actual deformation data of the pile foundation and the meteorological and hydrological data to obtain an adjusted weight value; Inputting the adjusted weight value into a preset spatio-temporal alignment algorithm to optimize the interpolation compensation parameters of the multi-source data.

3. The flexible support pile foundation process for floating PV according to claim 2, wherein, Also including: The anchorage depth mapping relationship output by the finite element simulation model is used as the training sample of the machine learning proxy model; The machine learning proxy model generates an adjustment instruction for the pile foundation construction inclination parameter through transfer learning; The pile foundation stress data collected during the construction stage is fed back to the machine learning proxy model to correct the loss function weight of the machine learning proxy model.

4. The flexible support pile foundation process for floating PV according to claim 3, wherein, The machine learning proxy model generates an adjustment instruction for the pile foundation construction inclination parameter through transfer learning includes: The reinforcement learning module explores the global solution space based on the pile group stability target; The genetic algorithm module loads the historical engineering case library to generate an initial population and injects navigation safety constraint conditions; The pile foundation degradation monitoring data dynamically constrains the feasible solution range of the hybrid optimization algorithm.

5. The flexible support pile foundation process for floating PV according to claim 4, wherein, The pile foundation degradation monitoring data dynamically constrains the feasible solution range of the hybrid optimization algorithm includes: A cloud deep learning model analyzes the frequency domain characteristics of construction process data and generates formation classification labels. An edge computing device adjusts the pile driving hammering frequency and penetration speed parameters after receiving the formation classification labels. The updated construction parameters are input into the cloud deep learning model as training samples to optimize the cloud deep learning model.

6. The flexible support pile foundation process for floating PV according to claim 5, wherein, It further includes: The erosion and corrosion coupling prediction model consists of a macroscopic erosion model and a microscopic corrosion model. The macroscopic erosion model is applied in the macroscopic erosion module. The microscopic corrosion model is applied in the microscopic corrosion rate prediction module. The water flow velocity gradient distribution data output by the macroscopic erosion module is input into the microscopic corrosion rate prediction module. The microscopic corrosion rate prediction module reversely corrects the boundary layer thickness parameter of the macroscopic erosion module based on the collected corrosion monitoring data and generates updated flow field distribution data. The updated flow field distribution data and the microscopic corrosion rate prediction results are aligned in space and time, input into a preset graph neural network, a nonlinear mapping relationship of the filling material gradation parameters is established, an erosion and corrosion coupling prediction result is generated, and a protection construction plan instruction set is generated to drive the construction equipment to adjust the hammering frequency and material ratio.

7. The flexible support pile foundation process for floating PV according to claim 6, characterized in that, The step of aligning the updated flow field distribution data and the microscopic corrosion rate prediction results in space and time, inputting them into a preset graph neural network, establishing a nonlinear mapping relationship of the filling material gradation parameters, generating an erosion and corrosion coupling prediction result, and generating a protection construction plan instruction set to drive the construction equipment to adjust the hammering frequency and material ratio includes: Binding the construction control instruction version number and the protection decision timestamp and storing them in the blockchain. The intelligent contract verifies the matching of the feature dimension of the dynamic load spectrum and the input layer of the machine learning agent model. When a version conflict is detected, the finite element simulation model is triggered to recalibrate the parameters.

8. The flexible support pile foundation process for floating PV according to any one of claims 1-7, characterized in that, The generation of the dynamic load spectrum includes: Decomposing the time series of the multi-source dynamic data into frequency domain sub-bands through wavelet packet transform. Performing entropy value feature extraction on the frequency domain sub-bands to generate a high-dimensional feature space.

9. The flexible support pile foundation process for floating PV according to claim 8, wherein The adaptive Bayesian optimization algorithm includes: Constructing a Gaussian process surrogate model with the feature selection combination of multi-source dynamic data as the input and the correlation index of the erosion and corrosion coupling prediction result as the output. Adopting a local penalty expected improvement function to balance the global exploration and local exploitation search strategies. Dynamically injecting the actual pile deformation data in the feedback construction parameters into the weight coefficient update process of the objective function.

10. The flexible support pile foundation process for floating PV according to claim 9, characterized in that, The step of constructing a Gaussian process surrogate model with the feature selection combination of multi-source dynamic data as the input and the correlation index of the erosion and corrosion coupling prediction result as the output includes: Inputting the erosion and corrosion coupling prediction result into the material durability evaluation model. Generating a filling construction plan based on the gradation optimization parameters output by the graph neural network. Verifying the degradation trend of the protection decision historical data recorded in the blockchain and the prediction result of the material durability evaluation model.

Images