Machine learning based energy pile green low carbon building design method

Abstract

This invention belongs to the field of geothermal energy development and utilization technology, and proposes a green and low-carbon building design method for energy piles based on machine learning. Since the sample attributes affecting thermal stress and heat transfer efficiency have opposite effects, the invention identifies energy pile feature parameters that simultaneously apply to both thermal stress and heat transfer efficiency. These parameters are input into a federated learning algorithm for green and low-carbon building design of energy piles, serving as thresholds and criteria for determining the optimal solution. The same sample attributes are input into the simulated annealing method for heat transfer efficiency and the K-nearest neighbor regression prediction model for thermal stress of energy piles. By selecting the energy pile feature parameters corresponding to heat transfer efficiency and thermal stress, a green and low-carbon building design method for energy piles is established, ultimately determining the unit equipment, energy pile structure, and building load. This invention has the advantages of low technical cost, simple operation, and strong universality, overcoming the subjective experience of traditional building design methods and improving computational efficiency and the scientific nature of design results.

Description

Technical Field

[0001] This invention relates to the field of geothermal energy development and utilization technology, and in particular to a green and low-carbon building design method for energy piles based on machine learning. Background Technology

[0002] Ground source heat pump technology extracts recyclable shallow geothermal energy to regulate indoor building temperatures, essentially functioning as a seasonal energy storage mechanism. Currently, ground source heat pumps are considered the most promising energy-saving and environmentally friendly HVAC technology for achieving green and low-carbon buildings. However, in practice, traditional ground source heat pump systems typically require a large land area to install buried pipe heat exchangers, resulting in high initial investment costs. Compared to traditional cooling and heating methods, their advantages are significantly reduced, thus becoming one of the main obstacles hindering the further development and application of ground source heat pump technology.

[0003] Energy piles are an economical, efficient, energy-saving, and emission-reducing new technology that combines ground source heat pump technology with traditional pile foundations. By embedding heat exchange pipes of various shapes within the pile foundation, shallow geothermal energy is converted. While fulfilling the conventional mechanical functions of pile foundations, the pile body also facilitates heat exchange with shallow geothermal energy, serving the dual purpose of simultaneous construction of the pile foundation and the buried pipe heat exchanger. Compared to traditional ground source heat pump technology, it offers significant advantages in investment cost, heat exchange efficiency, and land utilization. Existing field tests demonstrate that evaluating and predicting the heat exchange efficiency of energy piles and the impact of temperature loads on pile stress is the foundation and prerequisite for green and low-carbon building design using energy piles.

[0004] Chinese invention patent application number CN202110124290.0, entitled "A Test Platform for Energy Piles in Grain Depots and Its Test Operation Method", discloses a test platform for energy piles in grain depots and its test operation method, which can monitor the thermal stress changes of energy piles in grain depots through a thermo / mechanical monitoring system in an energy pile model, a grain depot model, and a raft model.

[0005] Chinese invention patent application number CN201410177624.0, entitled "Energy Pile-Soil Load and Temperature Transfer Mechanism Model Test Device and Test Method", discloses a visual energy pile-soil load and temperature transfer mechanism model test device and test method, which can observe the pile-soil interaction under temperature action in real time through PIV technology and fiber optic technology.

[0006] Chinese invention patent application number CN201410077507.7, entitled "A Test Device and Test Method for Friction Force of Energy Pile Pile-Soil Contact Surface", discloses a test device for friction force of energy pile pile-soil contact surface consisting of a circulation control system, a loading system, a pressurization system and a measurement system.

[0007] Chinese invention patent application number CN201810818755.0, entitled "Energy Pile Test Monitoring System and Test Method", discloses a test monitoring system including an energy pile body, a loading device, a pile top displacement testing device, a circulating temperature control device, a data acquisition device, and a geotechnical thermal response testing device, which can perform more comprehensive testing and analysis of the thermodynamic parameters of energy piles.

[0008] Chinese invention patent application number CN201611087188.3, ​​entitled "A Method for Calculating Axial Force under Thermal Coupling of Energy Piles", discloses a one-dimensional finite difference axial force calculation method for energy piles based on the load transfer method under thermal coupling.

[0009] The limited number of thermal response analyses of energy piles mentioned above primarily rely on model tests, field tests, or numerical simulations. Therefore, a simple and easily applicable calculation method is needed to accurately evaluate the impact of temperature loads on the bearing capacity of piles in practical engineering projects.

[0010] Currently, the traditional architectural design method involves designers performing design calculations based on specifications, and then conceptualizing and drawing up design drawings based on empirical safety factors. However, this method suffers from problems such as unreasonable design and a lack of scientific rigor. The green and low-carbon building design for energy piles requires accurate evaluation of heat exchange efficiency and thermal stress in the pile body, comprehensively considering the parameters of the energy pile's embedded pipes, building loads, building load characteristics, and the mutual constraints between various modules of unit equipment selection. Continuing to use the traditional design method based on safety factors can easily lead to excessive material strength and overly large equipment configurations, resulting in overly conservative design schemes that violate the original design principles of energy conservation, environmental protection, low carbon emissions, and economy.

[0011] Machine learning is a scientific technique that uses appropriate algorithms to enable computers to automatically analyze and obtain patterns from a type of data, and then use these patterns to predict unknown data.

[0012] Chinese invention patent application number CN202110380281.8, entitled "Evaluation Method, Storage Medium and Electronic Device for Pile Driving Process Based on Machine Learning", discloses a principal component evaluation model for pile driving process.

[0013] Chinese invention patent application number CN202010319454.0, entitled "Prediction and Analysis Method of Pile-Soil Interaction Based on Machine Learning", discloses a neural network model of pile bearing capacity based on numerical simulation samples.

[0014] Machine learning is a scientific technology that enables computers to automatically analyze patterns from a type of data through appropriate algorithms and use these patterns to predict unknown data. Training a large number of data samples with machine learning can be used to predict pile foundation construction and evaluate pile bearing capacity. However, the aforementioned machine learning-based pile evaluation methods only analyze the pile bearing capacity and evaluate pile driving construction. Furthermore, the two learning models used suffer from low readability in the calculation process and simplified input parameters, making them unsuitable for evaluating the heat exchange performance and thermal characteristics of energy piles, thus hindering the design and calculation of green and low-carbon building systems using energy piles. Summary of the Invention

[0015] To overcome the following shortcomings and problems of existing technologies: (1) the calculation cost of existing thermal analysis methods for energy piles is too high and the implementation process is complicated; (2) the traditional design method based on the concept of empirical safety factor is difficult to meet the green and low-carbon concept of energy pile green and low-carbon building system design. Therefore, this invention provides a machine learning-based energy pile green and low-carbon building design method to achieve the dual goals of improving the scientific rationality of energy pile green building design and reducing design calculation and economic costs.

[0016] The technical solution of this invention is as follows: A machine learning-based green and low-carbon building design method for energy piles. Since sample attributes that have the same influence on thermal stress and heat transfer efficiency have opposite effects on the two, the method seeks energy pile feature parameters that are simultaneously applied to the prediction results of thermal stress and heat transfer efficiency. These parameters serve as the basic data for training the energy pile buried pipe design parameters in the green building design of energy piles. The data is then input into the federated learning algorithm for green and low-carbon building design of energy piles, serving as the threshold and basis for judging the optimal solution. The machine learning-based green and low-carbon building design method for energy piles includes energy pile heat transfer efficiency prediction, energy pile thermal stress prediction, and a federated learning architecture-based green and low-carbon building design method for energy piles.

[0017] The prediction of the heat exchange efficiency of the energy pile specifically includes the following steps:

[0018] S1: Obtain training sample data and corresponding labels for the heat exchange efficiency of energy piles; the training sample data includes energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters; the label data includes information reflecting the heat exchange efficiency of the energy piles; the energy pile characteristic parameters include energy pile size parameters, energy pile material parameters, and energy pile embedded pipe geometric parameters.

[0019] S2: Preprocessing and feature analysis of training sample data for energy pile heat exchange efficiency; quality analysis, feature analysis, and cleaning of training sample data for energy pile heat exchange efficiency are performed to divide the training set and prediction set of energy pile heat exchange efficiency required for supervised learning.

[0020] S3: Training of the prediction model for heat exchange efficiency of energy piles. The prediction model for heat exchange efficiency of energy piles is established and trained based on the sample data information of the training set of heat exchange efficiency of energy piles.

[0021] S4: Input the sample attribute information of the energy pile heat exchange efficiency prediction set. The sample attributes of the energy pile heat exchange efficiency prediction set include: energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters.

[0022] S5: Calculate the heat exchange efficiency prediction results corresponding to the heat exchange efficiency prediction set of the energy pile based on the heat exchange efficiency prediction model of the energy pile.

[0023] S6: Evaluation of the prediction results of the heat exchange efficiency of the energy pile; compare the prediction results obtained in S5 with the labels corresponding to the samples of the heat exchange efficiency prediction set of the energy pile, calculate the prediction error of the heat exchange efficiency prediction model of the energy pile, and evaluate its accuracy.

[0024] S7: The characteristic parameters of the energy pile, the parameters of the soil and rock mass, and the state parameters of the heat exchange circulating fluid to be predicted are used as the attribute space of the energy pile heat exchange efficiency prediction set. The heat transfer characteristics of the energy pile are predicted and calculated through the energy pile heat exchange efficiency prediction model.

[0025] The prediction of thermally induced stress in energy piles specifically includes the following steps:

[0026] T1: Obtain thermal stress training sample data and corresponding labels; thermal stress training sample data includes energy pile characteristic parameters and soil and rock parameters, and label data includes relevant information reflecting the thermal stress of energy piles, including thermally induced pile side friction, thermally induced pile stress, and thermally induced pile strain.

[0027] T2: Preprocessing of thermal stress training sample data, defining state vectors; performing quality analysis, feature analysis, and cleaning on the thermal stress training sample data to obtain thermal stress training set and thermal stress prediction set; sorting sample attribute variables according to correlation through sensitivity analysis, and taking the sample attribute variables with the highest ranking as elements of state parameters, the state vector at time t is represented as X(t)=[x1(t),x1(t-1),x2(t),x2(t-1),x3(t),x3(t-1),…,x n (t),x n (t-1)];

[0028] T3: Select a distance metric; the correlation between each state vector X and the target vector Y in the state space is represented by the distance in the state space.

[0029] T4: The sample attributes of the thermal stress prediction set include: energy pile characteristic parameters and soil and rock parameters; the sample attributes of the thermal stress training set are used to form a state vector, and the K nearest neighbors of the current state vector are searched in the thermal stress training set to form a K nearest neighbor prediction model; the sample attribute information of the thermal stress prediction set is input into the K nearest neighbor prediction model; the predicted mean of the K nearest neighbors. Calculate the thermal stress prediction results corresponding to the thermal stress prediction set for the K-nearest neighbor prediction model;

[0030] T5: Evaluation of the thermal stress prediction results of energy piles. The prediction results obtained in T4 are compared with the labels corresponding to the thermal stress prediction set samples to form a confusion matrix. Based on the confusion matrix, the sensitivity, accuracy, and precision of the K-nearest neighbor prediction model are calculated to further evaluate the prediction error and model accuracy of the K-nearest neighbor prediction model.

[0031] Sensitivity The closer the value is to 1, the better the model fit. Accuracy The closer the value is to 1, the stronger the model's predictive ability. The closer the value is to 1, the higher the accuracy of the model's predictions.

[0032] Table 1 Confusion Matrix

[0033]

[0034] T6: The characteristic parameters of the energy pile and the parameters of the soil and rock mass that need to be predicted are used as the attribute space of the thermal stress prediction set, and the thermal stress prediction calculation of the energy pile is performed by the K-nearest neighbor prediction model.

[0035] The energy pile green and low-carbon building design method based on a federated learning architecture establishes a client-server architecture-based federated learning algorithm framework for energy pile green and low-carbon building design. The client includes four main modules: building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module, which are used for learning calculations of all sub-modules. After the anonymized parameters calculated by the client are aggregated to the central server for calculation, they are distributed to each client to update their local model until the global model is robust. The specific steps are as follows:

[0036] D1: Collect and organize building and environmental information, perform data preprocessing, analyze sample data characteristics, and assign sample data characteristics to the corresponding computing module client; sample data is obtained through one or more of the following technologies: on-site survey, on-site monitoring, questionnaire survey, and intelligent image recognition; sample data information includes building information parameters, environmental information parameters, foundation information parameters, unit equipment parameters, and geometric parameters of energy pile buried pipes;

[0037] D2: The four clients—building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module—update their local models.

[0038] D3: Four clients—building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module—upload encrypted and de-identified parameters to the central server using public keys; the relevant data for client i is represented as follows: Can be encrypted as

[0039] D4: The central server uses its private key to decrypt the encrypted and anonymized parameters uploaded by the four clients, performs secure aggregation, and then updates the global shared model; the central server then... Decode the encrypted and anonymized parameters; the central server-side federated learning obtains the globally shared model using at least one of the following synchronization methods: gradient averaging, federated averaging, or knowledge distillation; aggregate all client models based on knowledge distillation using federated learning, according to the formula... Perform global model weight updates for four clients, where W n+1 These are the global model parameters for the nth round. The client sub-model weights are uploaded to the server by client i in the nth round. After each round of model weight updates, the central server calculates the global model error and accuracy. The central server can also control the transmission speed and the shutdown of model training.

[0040] D5: The central server generates a public key for encrypting data transmission from the global shared model and distributes it to each client; based on the iteration results of each local client transmitted by the global shared model, it updates the attribute parameters of other related clients as new sample data; for example, the random forest model client for the energy pile buried pipe scheme updates the building load calculated by the building load prediction client and the circulating pump parameters and heat pump parameters calculated by the unit selection client based on the global shared model.

[0041] D6: Repeat steps D2 to D5 iteratively until the global model is robust. Finally, the client calculates the corresponding results based on the global shared model. The design parameters that match each other in the green and low-carbon building of energy piles are calculated, including building load, heat pump unit model, circulation pump model, and geometric parameters of energy pile buried pipes.

[0042] The geometric parameters of the energy pile buried pipes include heat exchange pipe length, pipe spacing, pipe diameter, and pipe layout classification; the material parameters of the energy pile include thermal conductivity of the heat exchange medium, elastic modulus of the pile material, coefficient of expansion, thermal conductivity of concrete, and concrete temperature; the parameters of the soil and rock mass include elastic modulus of the soil and rock mass, thermal conductivity of the soil and rock mass, groundwater depth, groundwater flow velocity, and soil and rock mass temperature; the state parameters of the heat exchange circulating fluid include circulating fluid flow velocity, density, specific heat, and inlet temperature; the heat exchange efficiency of the energy pile includes comprehensive thermal conductivity, heat exchange per linear meter, heat extraction, and heat exchange power; the building information parameters include indoor human behavior, indoor equipment behavior, and building envelope structure. The data includes structural materials, enclosure structure dimensions, and enclosure structure thermal properties; environmental information parameters include ambient temperature, ambient humidity, and solar radiation; foundation information parameters include foundation bearing capacity and stratum parameters; building load is hourly load and compliance indicators; unit equipment parameters include heat pump unit parameters and circulating pump parameters; the sample data label data is classified according to field test data, model test data, and numerical simulation data; the energy pile thermally induced stress includes thermally induced strain, pile side friction, and pile end resistance; sample attribute variables include energy pile size parameters, energy pile material parameters, and energy pile buried pipe geometric parameters.

[0043] The data feature analysis is based on at least one of the following: extracting the distribution pattern of the data, normalizing the data, and evaluating the correlation between sample data of different labels.

[0044] The data cleaning process targets incomplete, erroneous, duplicate, and outlier data. It includes identifying and processing invalid, outlier, and missing values. Methods for identifying invalid, outlier, and missing values ​​can be based on one of the following: 1. Box plots to calculate the maximum, minimum, median, and upper and lower quartiles of the dataset. Outliers or invalid values ​​can be identified based on the upper and lower quartiles of the box plot; 2. The 3σ rule is commonly used when the data follows a normal distribution. In this case, outliers are defined as values ​​in a set of measurements that deviate from the mean by more than three standard deviations; 3. Scatter plots visually identify outliers, which are then considered invalid or outliers, by showing the positional relationship between two sets of data. Methods for processing invalid, outlier, and missing values ​​can be based on one of the following: 1. Direct deletion; 2. Filling in invalid, outlier, and missing values ​​using statistical data characteristics based on the mean, median, and mode; 3. Fitting invalid, outlier, and missing values ​​using a regression model or maximum likelihood estimation.

[0045] The energy pile heat exchange efficiency prediction model is implemented based on one or more of the following algorithms: simulated annealing, regression tree algorithm, random forest regression algorithm, support vector regression algorithm, multiple linear regression algorithm, simulated annealing improved support vector regression algorithm, and clustering regression algorithm.

[0046] When the energy pile heat transfer efficiency prediction model is trained using the simulated annealing method to improve the support vector regression algorithm, the specific steps are as follows:

[0047] S3.1: To address the relationship between sample attribute information and sample labels, a slack variable ξ is introduced. i , A nonlinear segmentation support vector classifier considering soft margins is constructed using the penalty coefficient C; the prediction accuracy and stability of the energy pile heat transfer efficiency prediction model are obtained through a conditional extremum function.

[0048]

[0049] Where w is the normal vector of the hyperplane, C is the penalty coefficient, and ξ, ξ * y is the relaxation factor, ∈ is the hyperparameter that determines the boundary width; i The actual test results for the training samples; f(X) i )=w·Φ(X)+b is the classification hyperplane of the energy pile heat transfer efficiency prediction model; i is the training sample number; N is the number of training set samples; Φ(X) is the nonlinear mapping function; b represents the intercept of the hyperplane;

[0050] S3.2: Transform the conditional extremum function of step S3.1 into a multivariate function using the Lagrangian function for solution. Set the partial derivative of the Lagrangian function with respect to the optimization objective to 0. The optimization objective includes the normal vector w of the hyperplane, the intercept b of the hyperplane, and the relaxation factors ξ and ξ'. * Transform the original conditional extremum function into a dual function to find the minimum value within the prediction boundary.

[0051]

[0052] Where α, α * μ, μ * Represents the Lagrange multipliers;

[0053] S3.3: The nonlinear mapping function Φ(X) contained in the classification hyperplane of the energy pile heat transfer efficiency prediction model, and its inner product Φ(X) i ) T Φ(X j The processing of ) uses a combination of kernel functions: the combination kernel function is one or more of the following: Gaussian kernel, linear kernel, polynomial kernel, and Sigmoid kernel; Gaussian kernel: Linear kernel: φ(X) i ) T φ(X j )=κ(X i ,X j ) = X i T Xj Polynomial kernel: φ(X) i ) T φ(X j )=κ(X i X j )=(eX i T X j +c) d Sigmoid kernel: φ(X) i ) T φ(X j )=κ(X i ,X j ) = tanh(eX) i T X j +c); Combined kernel: φ(X) i ) T φ(X j )=λκ1(X i ,X j )+(λ-1)κ2(X i ,X j ).

[0054] S3.4: Optimize model parameters in support vector regression algorithm using piecewise simulated annealing: representing the insensitive loss function ∈, the penalty coefficient C, and the hyperparameters γ, λ, e, c, and d in the kernel function;

[0055] Step S3.4.1: Randomly generate an initial model parameter set for interactive verification. Record the error value EEP as the current annealing system state E0, the initial temperature T0, the temperature of annealing stage one T1, and the annealing end temperature as T2.

[0056] Step S3.4.2: According to m′ i =m i +s·(τ-0.5)(B i -A i The model parameters are perturbed to form a new set of model parameters, and the current annealing system state parameters E are obtained through interactive verification. n Calculate ΔE = E n -E n-1 ;

[0057] Where: m′ i Let m be the variable after perturbation. i Let be the current variable, s be the perturbation ratio, τ be a random number in [0,1], and B be the variable. i A i For the current variable m i Scope;

[0058] Step S3.4.3: When ΔE < 0, accept the new model parameter set and jump to step S3.4.5; otherwise, accept the corresponding model parameter set according to the Metropolis criterion exp(ΔE / kT)-τ > 0 and jump to step S3.4.4; when none of the above conditions are met, reject the critical state, execute step S3.4.2, re-perturb to generate a new model parameter set, perform interactive verification, until the parameter set acceptance condition in step S3.4.3 is met;

[0059] Step S3.4.4: Cooling down. Upon obtaining the new state, proceed according to the cooling plan. The temperature is lowered. If the set temperature T1 is not reached, the process returns to step S3.4.2. Once the set temperature T1 is reached, a new annealing plan is initiated.

[0060] Step S3.4.5: According to the new perturbation method and annealing plan, continue to perturb the model parameter set of the first-stage annealing, and calculate the corresponding annealing system state parameters E. n ;

[0061] Step S3.4.6: When ΔE < 0, accept the new parameter set and jump to step S3.4.7; otherwise, accept the corresponding parameter set according to the Metropolis criterion exp(ΔE / kT)-τ > 0 and jump to step S3.4.7; when none of the above conditions are met, reject the parameter set and return to step S3.4.5, re-perturb to generate a new model parameter set, perform interactive verification, until the model parameter set acceptance condition in step S3.4.7 is met;

[0062] Step S3.4.7: Set the end temperature T2 as the algorithm exit point, and set the global maximum number of EEP calculations to N; stop annealing when T2 or N is reached, and accept the critical state cross-validation error E at this time. n For the lowest E n If the corresponding parameter is the optimal prediction parameter, then return to step S3.4.5.

[0063] The distance metric is one or more of the following: Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, standardized Euclidean distance, Mahalanobis distance, and cosine similarity. Euclidean distance: Manhattan distance: Chebyshev distance: d xy =max i (x i -y i Minkowski distance: Standardized Euclidean distance: Mahalanobis distance: Cosine of included angle: Combined distance: d xy =λd xy1 +(1-λ)d xy2 .

[0064] The building load calculation module uses environmental and building information parameters as sample attribute data to establish a support vector regression local model for building energy consumption, predicts and calculates the building load, and uploads it to the central processing unit in encrypted form; it also calculates the building's cooling / heating load based on the environmental and building information parameters; the specific steps are as follows:

[0065] D2.1.1: Decrypt the public key of the encrypted data using the private key, and update the iteration results of the building load module, unit selection module, and energy pile buried pipe module clients according to the global shared model of the central server as new sample data attribute parameters; based on the sample data, use environmental information parameters and building information parameters as sample attribute space sets, and use building load as the learning target;

[0066] D2.1.2: To address the relationship between sample attribute parameters and building load, slack variables are introduced. A nonlinear segmentation support vector classifier considering soft margins is constructed using the penalty coefficient β; the prediction accuracy and model stability of the local support vector regression model for building energy consumption are represented by the loss function.

[0067] D2.1.3: The loss function needs to be reduced in dimension to a multivariate function using the Lagrangian function for solution. Let the Lagrangian function be such that it applies to the optimization objective ζ, δ, The partial derivatives are 0, so we obtain the Lagrange multipliers, which transform the original conditional extremum function into the dual function, thereby finding the minimum value within the predicted boundary.

[0068] D2.1.4: The nonlinear mapping function Φ(X) contained in the classification hyperplane in the local model of building energy consumption support vector regression, and its inner product Φ(X) i ) T Φ(X j The kernel function selected can be one or more of the following: Gaussian kernel, linear kernel, polynomial kernel, and Sigmoid kernel.

[0069] D2.1.5: Optimize the model parameters in the support vector regression local model for building energy consumption: insensitive loss function ∈, penalty coefficient C, hyperparameters γ, λ, e, c, d in the kernel function; select one or more optimization methods from simulated annealing, grid search, particle swarm optimization, PSO algorithm, and genetic algorithm;

[0070] D2.1.6: Input the feature parameters of the target building, use the trained support vector regression model to predict and calculate the annual hourly environmental control load of the target building, and use the public key to convert the building load results into encrypted parameters and upload them to the central server for global shared model iteration in federated learning.

[0071] The building load calculation module establishes a support vector regression local model of building load based on basic building information to predict and calculate building load, and calculates building mechanical load based on building information parameters and foundation information parameters; when the number of samples is insufficient, numerical methods such as finite element, finite difference or discrete element are used to calculate building load; finally, the load results are encrypted and uploaded to the central processing unit.

[0072] The unit selection module establishes a decision tree-based local model for unit selection based on building load results and unit equipment parameter samples. It determines the parameters of the heat pump and circulating pump and uploads them to the central processing unit with encryption. The unit equipment parameters include: the cooling method of the heat pump unit, the cooling capacity of the heat pump unit, the flow rate of the heat pump unit, the flow rate of the chilled water circulation pump, and the head information of the chilled water circulation pump. The decision tree model is generated based on one or more of the following algorithms: ID3, C4.5, and CART. The specific steps are as follows:

[0073] D2.3.1: Decrypt the public key of the encrypted data using the private key, update the iteration results of the relevant clients of the building load module, building load module, and energy pile buried pipe module according to the global shared model of the central server as the new sample data attribute parameters; calculate the information gain of all attributes of the training samples according to the information entropy, and sort all attributes according to the information gain.

[0074] Total information entropy of a given sample:

[0075] Sample subset {s 1j ,s 2j ,…,s mj Total information entropy: P ij For sample subsets s j Category C i The sample probability;

[0076] Information entropy value of the samples based on attribute A: The corresponding information gain is: Gain(A) = I(s1,s2,…,s) m )-E(A);

[0077] D2.3.2: Select the attribute with the largest information gain as the optimal attribute, use the optimal attribute as the basis for sample partitioning, and form the root node by taking the same sample set as the same sample set of the optimal attribute.

[0078] D2.3.3: Each root node is then treated as a complete dataset. Samples are divided according to the suboptimal attribute. Samples with the same suboptimal attribute value are grouped into leaf nodes. The decision tree is pruned using the REP, PEP, or MEP methods. The decision tree is formed iteratively. The unit parameters are converted into encrypted parameters using the public key and uploaded to the central server for global shared model iteration in federated learning.

[0079] The energy pile installation design module, based on building load, building load, and unit equipment parameters, establishes a local random forest model for energy pile installation on the client side, using energy pile structural parameters as sample attributes. It then calculates the energy pile installation parameters as the optimal installation scheme and uploads it to the central processing unit in encrypted form. The specific steps are as follows:

[0080] D2.4.1: Decrypt the public key of the encrypted data using the private key, update the iteration results of the relevant clients of the building load module, building load module, and unit selection module according to the global shared model of the central server as the new sample data attribute parameters; use the Bootstrap sampling method to randomly generate a subset of samples from the samples as the training samples of one of the decision tree models, and repeat the sampling k times to form k decision tree training samples.

[0081] D2.4.2: Train decision trees based on a subset of attributes from k training samples to form k independent random decision trees;

[0082] D2.4.3: The random forest randomly selects a classifier to vote on the pipe laying schemes predicted by k decision trees, and the voting result is taken as the optimal pipe laying scheme of the random forest.

[0083] D2.4.4: Convert the energy pile buried pipe parameters into encrypted parameters using a public key and upload them to the central server for global shared model iteration in federated learning.

[0084] Beneficial effects of the present invention: Compared with existing energy pile technology, the present invention has the following technical advantages:

[0085] (1) Existing heat transfer calculation models for energy piles mostly adopt buried pipe heat transfer models or complex numerical models, which have the drawbacks of large calculation errors and low calculation efficiency. This invention trains a heat transfer model suitable for energy piles themselves through support vector regression algorithm, and the calculation efficiency and prediction error are greatly improved.

[0086] (2) The thermal stress of energy piles is obtained through various forms of field tests and numerical modeling methods, which are complex, time-consuming and labor-intensive in practice, and lack mature mathematical models. This invention trains a thermal stress model of energy piles using the K-nearest neighbor regression algorithm, which has the advantages of low technical cost and simple operation.

[0087] (3) The energy pile green low-carbon building federated learning design method proposed in this invention realizes the integrated design of the three parts of green low-carbon building user end, unit selection and buried pipe end, which improves design efficiency, scientific rationality of design parameters and reduces design risk. Attached Figure Description

[0088] Figure 1 This is a schematic diagram of an optional support vector regression calculation process for the heat exchange efficiency of an energy pile according to an embodiment of the present invention;

[0089] Figure 2 This is a schematic diagram of an optional support vector regression structure for the heat exchange efficiency of an energy pile according to an embodiment of the present invention;

[0090] Figure 3 This is a schematic diagram of an optional simulated annealing method improved support vector regression algorithm according to an embodiment of the present invention;

[0091] Figure 4 This is a schematic diagram of the calculation process of an optional energy pile thermal stress K-nearest neighbor prediction model according to an embodiment of the present invention;

[0092] Figure 5 This is a schematic diagram of a conceptual design method for energy pile green and low-carbon buildings based on a federated learning architecture, according to an embodiment of the present invention.

[0093] Figure 6 This is a schematic diagram of the federated learning framework structure of an optional energy pile green low-carbon building design method based on a federated learning architecture according to an embodiment of the present invention.

[0094] Figure 7 This is a schematic diagram of the support vector regression prediction model structure of an optional building load calculation module according to an embodiment of the present invention;

[0095] Figure 8 This is a schematic diagram of the support vector regression prediction model flow of an optional building load calculation module according to an embodiment of the present invention;

[0096] Figure 9 This is a schematic diagram of the support vector regression prediction model structure of an optional building load calculation module according to an embodiment of the present invention;

[0097] Figure 10 This is a schematic diagram of an optional unit selection module decision tree model structure according to an embodiment of the present invention;

[0098] Figure 11 This is a schematic diagram of an optional unit selection module decision tree model flow according to an embodiment of the present invention;

[0099] Figure 12This is a schematic diagram of a random forest model structure for an optional energy pile buried pipe design module according to an embodiment of the present invention;

[0100] Figure 13 This is a schematic diagram of the random forest model process for an optional energy pile buried pipe design module according to an embodiment of the present invention. Detailed Implementation

[0101] The specific embodiments of this invention are described in detail below with reference to the accompanying drawings. However, the scope of protection of this invention is not limited to the description of these embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0102] Figure 1 This is a flowchart of an embodiment of the energy pile heat exchange efficiency prediction method of this patent, which includes the following steps:

[0103] Step S1: Collection of training sample data and corresponding labels for energy pile heat exchange efficiency. The training sample dataset includes energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters. The label data includes relevant information reflecting the heat exchange efficiency of the energy pile. The sample data and corresponding label data need to be classified according to field test data, model test data, and numerical simulation data. Energy pile characteristic parameters include energy pile size parameters, energy pile material parameters, and energy pile buried pipe geometric parameters. Among them, the energy pile buried pipe geometric parameters are heat exchange pipe length, pipe spacing, pipe diameter, and pipe layout classification; the energy pile material parameters are heat exchange medium thermal conductivity, heat exchange medium temperature, heat exchange medium flow velocity, concrete thermal conductivity, and concrete temperature; the soil and rock parameters are soil and rock thermal conductivity, groundwater depth, groundwater flow velocity, and soil and rock temperature; the heat exchange circulating fluid state parameters are circulating fluid flow velocity, density, specific heat, and inlet temperature; the energy pile heat exchange efficiency parameters are comprehensive thermal conductivity, heat exchange per meter, heat extraction, and heat exchange power.

[0104] Step S2: Sample Data Preprocessing and Feature Analysis. This involves quality analysis, feature analysis, data cleaning, and partitioning the sample data into training and prediction sets for supervised learning. The main contents include: cleaning the sample data, analyzing data features, and partitioning the training and prediction sets for supervised learning. The specific operation process is as follows: 1. Analyze the data distribution according to the 3σ principle. In this case, values ​​whose deviation from the mean exceeds three times the standard deviation are considered invalid or outliers; 2. Fit invalid, outlier, and missing values ​​using a regression model or maximum likelihood estimation; 3. Extract the distribution pattern of the cleaned data, perform data normalization, and evaluate the correlation between sample data with different labels. Table 1 shows the sample space and sample quality for predicting the heat exchange efficiency of energy piles in a green low-carbon building in Dalian.

[0105] Table 1. Sample Space and Sample Quality for Predicting the Heat Exchange Efficiency of Energy Piles in a Green and Low-Carbon Building in Dalian City

[0106]

[0107]

[0108] Step S3: Heat transfer efficiency model training. An energy pile heat transfer efficiency prediction model is established and trained based on the sample attribute information of the training set. The energy pile heat transfer efficiency prediction model is trained using simulated annealing to improve support vector regression. Figure 2 and Figure 3 The following are the structural diagram and flowchart of an optional simulated annealing method improved support vector regression training model for predicting the heat transfer efficiency of energy piles, according to a patent embodiment. The specific steps are as follows:

[0109] Step S3.1: To resolve the relationship between sample attribute data and sample label data, a slack variable ξ is introduced. i , A nonlinear segmentation support vector classifier considering soft margins is constructed using the penalty coefficient C. The model's prediction accuracy and stability can be represented by the loss function (L = max(0, |z|-∈)).

[0110] Step S3.2: The above conditions need to be converted into a multivariate function for solution using the Lagrange function. By setting the partial derivatives of the Lagrange function with respect to the optimization objectives w, b, ξ to 0, the Lagrange multipliers are obtained. The original condition extremum function can be transformed into a dual function, thereby finding the minimum value within the constraint region.

[0111] Step S3.3: The nonlinear mapping function Φ(X) contained in the classification hyperplane of the energy pile heat transfer efficiency prediction model, and its inner product Φ(X) i ) T Φ(X j The processing of ) can be achieved by using a combined kernel function φ(X). i ) T φ(X j )=λκ1(X i ,X j )+(λ-1)κ2(X i ,X j ) to process.

[0112] Step S3.4: The training process of the support vector regression model mainly involves optimizing the model parameters: the insensitive loss function ∈, the penalty coefficient C, and the hyperparameters γ, λ, c, e, and d in the kernel function. Preferably, piecewise simulated annealing is used for model parameter optimization, and the steps are as follows:

[0113] Step S3.4.1: Randomly generate an initial parameter set for interactive verification. Record the error value EEP as the current annealing system state E0, the initial temperature T0, the temperature of annealing stage one T1, and the annealing end temperature as T2.

[0114] Step S3.4.2: According to m′ i ≤=m i +s·(τ-0.5)(B i -A i The parameters are perturbed to form a new parameter set, and the current annealing system state E is obtained through interactive verification. n Calculate ΔE = E n -E n-1 .

[0115] Step S3.4.3: When ΔE < 0, accept the new model parameter set and jump to step S3.4.5; otherwise, accept the corresponding model parameter set according to the Metropolis criterion exp(ΔE / kT)-τ > 0 and jump to step S3.4.4; when none of the above conditions are met, reject the critical state, execute step S3.4.2, re-perturb to generate a new model parameter set, perform interactive verification, until the parameter set acceptance condition in step S3.4.3 is met;

[0116] Step S3.4.4: Cooling down. Upon obtaining the new state, proceed according to the cooling plan. The temperature is lowered. If the set temperature T1 is not reached, the process returns to step S3.4.2. Once the set temperature T1 is reached, a new annealing plan is initiated.

[0117] Step S3.4.5: According to the new perturbation method and annealing plan, continue to perturb the model parameter set of the first-stage annealing, and calculate the corresponding annealing system state parameters E. n ;

[0118] Step S3.4.6: When ΔE < 0, accept the new parameter set and jump to step S3.4.7; otherwise, accept the corresponding parameter set according to the Metropolis criterion exp(ΔE / kT)-τ > 0 and jump to step S3.4.7; when none of the above conditions are met, reject the parameter set and return to step S3.4.5, re-perturb to generate a new model parameter set, perform interactive verification, until the model parameter set acceptance condition in step S3.4.7 is met;

[0119] Step S3.4.7: Set the end temperature T2 as the algorithm exit point, and set the global maximum number of EEP calculations to N; stop annealing when T2 or N is reached, and accept the critical state cross-validation error E at this time. n For the lowest E nIf the corresponding parameter is the optimal prediction parameter, then return to step S3.4.5.

[0120] Step S4: Input the sample attribute information of the prediction set. The sample attributes of the energy pile heat exchange efficiency prediction set include: energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters. Specifically, the sample attributes include: energy pile buried pipe parameters, soil and rock thermal property parameters, and heat exchange circulating fluid state parameters. Among these, the energy pile buried pipe geometric parameters are heat exchange pipe length, pipe spacing, pipe diameter, and pipe layout classification; the energy pile material parameters are heat exchange medium thermal conductivity, pile material elastic modulus, expansion coefficient, concrete thermal conductivity, and concrete temperature; the soil and rock parameters are soil and rock elastic modulus, soil and rock thermal conductivity, groundwater depth, groundwater flow velocity, and soil and rock temperature; and the heat exchange circulating fluid state parameters are circulating fluid flow velocity, density, specific heat, and inlet temperature.

[0121] Step S5: Calculate the comprehensive thermal conductivity, heat exchange efficiency, and heat exchange results corresponding to the prediction set based on the energy pile heat exchange efficiency prediction model.

[0122] Step S6: Evaluate the predicted results of the energy pile heat exchange efficiency. Input the attribute information of the prediction set into the training model to calculate the corresponding prediction results. Compare the prediction results with the labels corresponding to the samples in the prediction set, calculate the model prediction error, and evaluate the model accuracy. The prediction error of the energy pile heat exchange efficiency prediction model for a green low-carbon building in Dalian is approximately 76.8%.

[0123] Step S7: Input 50 sets of basic attribute parameters of the energy piles to be predicted, and calculate the heat exchange efficiency of the energy piles. The distribution characteristics of the prediction results are shown in Table 2. Table 2: Distribution Characteristics of 50 Sets of Energy Pile Heat Exchange Efficiency Prediction Results

[0124]

[0125]

[0126] Figure 4 This is a flowchart of an embodiment of a method for predicting thermal stress in energy piles, including the following steps:

[0127] Step T1: Collection of training sample data and corresponding labels for thermal stress in energy piles. The training sample dataset includes energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters. Label data includes relevant information reflecting the thermal stress of the energy piles. The sample data and corresponding label data need to be categorized according to field test data, model test data, and numerical simulation data. Energy pile characteristic parameters include energy pile size parameters, energy pile material parameters, and energy pile embedded pipe geometric parameters. Specifically, the embedded pipe geometric parameters include heat exchange pipe length, pipe spacing, pipe diameter, and pipe layout classification; the energy pile material parameters include the thermal conductivity of the heat exchange medium, the elastic modulus of the pile material, the coefficient of thermal expansion, the thermal conductivity of concrete, and the concrete temperature; the soil and rock parameters include the elastic modulus of the soil and rock, the thermal conductivity of the soil and rock, the groundwater depth, the groundwater flow velocity, and the soil and rock temperature; the heat exchange circulating fluid state parameters include the circulating fluid flow velocity, density, specific heat, and inlet temperature; and the thermal stress of the energy pile includes thermal strain, pile side friction, and pile end resistance.

[0128] Step T2: Sample data preprocessing, defining the state vector. Quality analysis, feature analysis, and cleaning are performed on the sample data. The specific data preprocessing process is as follows: 1. Calculate the maximum, minimum, median, and upper and lower quartiles of the dataset using box plots. Outliers or invalid values ​​can be identified based on the upper and lower quartiles of the box plots. 2. Fill in the identified invalid, outlier, and missing values ​​according to the mean, median, and mode. 3. The cleaned data can be further processed to extract the data distribution pattern and perform data normalization. Sensitivity analysis can be used to sort the sample attribute variables according to their correlation, and the attributes with the highest ranking can be used as elements of the state parameters. The state vector at time t can be represented as X(t)=[x1(t),x1(t-1),x2(t),x2(t-1),x3(t),x3(t-1),…,x n (t),x n (t-1)].

[0129] Step T3: Select the distance metric. The correlation between each state vector X and the target vector Y in the state space can be represented by the distance in the state space; therefore, the "combined distance: d" is selected. xy =λd xy1 +(1-λ)d xy2 "Distance metric is the key to K-nearest neighbor prediction models."

[0130] Step T4: Input the sample attribute information of the prediction set, including: the buried pipe parameters of the energy pile and the mechanical parameters of the soil and rock mass. Form a state vector from the sample attributes of the training set, and search for the K nearest neighbors of the current vector in the training set to form a K-nearest neighbor prediction model. The K nearest neighbors correspond to... Calculate the predicted thermal stress results corresponding to the prediction set for the model.

[0131] Step T5: Evaluation of the thermal stress prediction results of the energy piles. The prediction results of the prediction set are compared with the labels corresponding to the samples in the prediction set to form a confusion matrix. Based on the confusion matrix, three indicators—sensitivity (TPR), accuracy (ACC), and precision—are calculated to evaluate the model's prediction error and model precision. The sensitivity (TPR), accuracy (ACC), and precision of the energy pile heat exchange efficiency prediction model for a green low-carbon building in Dalian are approximately 0.658, 0.729, and 0.821, respectively.

[0132] Step T6: Input 50 sets of characteristic parameters of the energy piles to be predicted as the prediction set attribute space, and perform thermal stress prediction calculation for the energy piles. Table 3 shows the distribution characteristics of the 50 sets of thermal stress prediction results for the energy piles.

[0133] Table 3. Distribution characteristics of the predicted thermal stress of 50 energy piles

[0134] Distribution index Pile side skin friction (kPa) Thermal stress (MPa) Thermally induced strain (με) Pile end resistance (kPa) Maximum value 45.61 3.48 305.8 33.38 Minimum value 10.24 1.55 145.9 9.45 average value 33.75 2.24 228.4 19.44 Standard deviation 5.32 0.64 34.8 2.38

[0135] Another aspect of this invention provides a green and low-carbon building design method for energy piles. Figure 5 A conceptual diagram for green building design using energy piles. A federated learning algorithm framework for green, low-carbon building design using energy piles, based on a client-server architecture, is established. Figure 6 This is a structural diagram of a green and low-carbon building design methodology for energy piles based on a federated learning architecture. The learning and calculation of all sub-modules are implemented on the client side, including four main modules: building load calculation module, building load calculation module, generator selection module, and energy pile buried pipe design module. Finally, the anonymized parameters calculated by the client are aggregated to the central server for calculation, and then distributed to each data holder (client) to update their local model until the global model is robust. The specific steps are as follows:

[0136] Step D1: Collect and organize building information, environmental information, unit equipment information, energy pile information, and foundation information using on-site survey technology, questionnaire survey technology, and intelligent image recognition technology. Perform simple data preprocessing, analyze the characteristics of 50 sets of sample data, and allocate them to the corresponding computing module client. The building information parameters include building enclosure structural materials, building dimensions, indoor human behavior, indoor equipment behavior, indoor lighting behavior, indoor air conditioning design parameters, building materials, beam and column reinforcement, and building load; the environmental information parameters include ambient temperature, ambient humidity, and solar radiation; the system equipment parameters include heat pump cooling capacity, heat pump heating capacity, heat pump flow rate, circulating pump flow rate, and circulating pump head; the energy pile parameters include pile foundation dimensions, buried pipe type, buried pipe length, heat exchange medium flow rate, and heat exchange capacity.

[0137] Step D2: The four clients—building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module—update their local models. The local models for the four clients are as follows:

[0138] 1. Building Load Calculation Module: Based on basic building information, a support vector regression local model of building energy consumption is established to predict and calculate the building load, which is then encrypted and uploaded to the central processing unit. Figure 7 This invention provides a support vector regression prediction model structure diagram for a building load calculation module. Its key feature is the calculation of the building's cooling / heating load based on environmental information parameters and building characteristics. Figure 8 The flowchart for the support vector regression prediction model in the building load calculation module is shown below. The specific steps of the model in round s are as follows:

[0139] Step D2.1.1: Decrypt the public key used to encrypt the transmitted data using the private key. Update the iteration results of the building load module, unit selection module, and energy pile installation module clients based on the global shared model of the central server, and use these as the new sample data attribute parameters. Use the environmental information parameters and building information parameters as sample attribute X. i This forms sample data X = {X1, X2, ..., X...} containing multiple features. n}, taking building load as the learning objective y={y1,y2,…y n}

[0140] Step D2.1.2: To resolve the relationship between the sample attribute space and the building load, slack variables are introduced. A nonlinear segmentation support vector classifier considering soft margins is constructed using the penalty coefficient β; the prediction accuracy and model stability of the local support vector regression model for building energy consumption are represented by the loss function L=max(0,|z|-∈).

[0141] Step D2.1.3: The above conditions need to be transformed into a multivariate function for solution using the Lagrangian function. Let the Lagrangian function be such that it applies to the optimization objective ζ, δ, The partial derivatives are 0, so we obtain the Lagrange multipliers, which can transform the original conditional extremum function into the dual function, thereby finding the minimum value within the predicted boundary.

[0142] Step D2.1.4: In the formula, the inner product φ(X) of the mapping function i ) T φ(X j The processing of ) can be done using a Gaussian kernel: "To be processed."

[0143] Step D2.1.5: Optimize the support vector regression model parameters based on the training set data: insensitive loss function ∈, penalty coefficient C, and hyperparameters γ, λ, e, c, and d in the kernel function. At least one of the following optimization methods can be used: simulated annealing, grid search, particle swarm optimization, PSO algorithm, or genetic algorithm.

[0144] Step D2.1.6: Input the characteristic parameters of a green and low-carbon building in Dalian City, and use the trained support vector regression model to predict and calculate the hourly environmental control load of the green and low-carbon building in Dalian City throughout the year. Encrypt the building load results using a public key and upload them to the central server.

[0145] 2. Building Load Calculation Module: Figure 9 This is a structural diagram of the support vector regression model for building loads. A local support vector regression model for building loads is established based on basic building information to predict and calculate building loads, which are then encrypted and uploaded to the central processing unit. Its key feature is that the building mechanical loads are calculated based on the building's structural characteristics. The learning sample attributes in the building load support vector regression model include building materials, building dimensions, material properties, frame structure, beam and column reinforcement, and load conditions. Additionally, a public key for encrypting data transmission is generated based on the globally shared model on the central server, and the iteration results of relevant clients at the building load end, generator end, and energy pile buried pipe end are updated as new sample data attribute parameters. When training samples are lacking, finite element, finite difference, and discrete element numerical simulation methods can also be used to calculate building loads. The specific steps of the s-th round of this model are as follows:

[0146] Step D2.2.1: Establish the building geometric model, define material properties, and determine the constitutive relations of the materials.

[0147] Step D2.2.2: Perform finite mesh generation, dividing the building physics model into a finite number of interconnected element bodies of various shapes. Data information between element bodies is transmitted through nodes to determine the shape function of each element body.

[0148] Step D2.2.3: Define the contact connection relationship of each component: Define the boundary conditions and load conditions.

[0149] Step D2.2.4: Define the element stiffness matrix and nodal displacement matrix.

[0150] Step D2.2.5: Assemble the elements into a total matrix equation (joint equation system) for the entire discrete domain. The simultaneous equation system can be solved using direct methods or iterative methods. The building load results are encrypted using a private key and uploaded to the central server.

[0151] 3. Unit Selection Module: Based on the building load and unit parameter samples, a decision tree local model for unit selection is established to determine the parameters of the heat pump and circulating pump and upload them to the central processor in encrypted form. Figure 10 This invention provides a unit selection decision tree model structure diagram, characterized in that the unit parameters include: heat pump cooling capacity, heat pump unit power, heat pump flow rate, refrigeration cycle pump flow rate, and refrigeration cycle pump head information. The ID3 algorithm is used as an example of this invention. Figure 11 Here is a flowchart of a unit selection decision tree model for an embodiment. The specific steps of the client in the s-th round are as follows:

[0152] Step D2.3.1: Decrypt the public key of the encrypted data using the private key. Update the iteration results of the relevant clients of the building load module, building load module, and energy pile buried pipe module of a green and low-carbon building in Dalian City according to the global shared model of the central server, and use them as the new sample data attribute parameters. Calculate the information gain of all attributes (various unit parameters) of the training samples according to the information entropy, and sort all attributes according to the information gain.

[0153] Step D2.3.2: Select the attribute with the largest information gain as the optimal attribute, use the optimal attribute as the basis for sample partitioning, and form the root node by taking samples with the same attribute value as the same sample set.

[0154] Step D2.3.3: Each root node is then treated as a complete dataset. Samples are partitioned based on the suboptimal attribute, with samples having the same attribute value forming the same leaf node. Pruning is performed using the REP method to address the overfitting problem of the decision tree model. This process iterates to form a decision tree. The calculated optimal unit parameters are encrypted and uploaded to the central server for iterative federated learning of the globally shared model.

[0155] 4. Energy Pile Burial Pipe Design Module: Figure 12 This invention provides a random forest model structure diagram for an energy pile buried pipe scheme. Based on building load, building load, and unit equipment parameter information, a local random forest model of the energy pile buried pipe is established on the client side, using energy pile parameters as sample attributes. The calculated energy pile parameters are encrypted and uploaded to the central processing unit. The energy pile parameters include: heat exchange pipe material, buried pipe length, buried pipe spacing, energy pile size, and heat exchange velocity information.

[0156] Figure 13 The flowchart for this model shows the specific steps of the s-th round of the client-side local model as follows:

[0157] Step D2.4.1: Decrypt the public key of the encrypted data using the private key. Update the iteration results of the relevant clients of the building load module, building load module, and unit equipment module of a green and low-carbon building in Dalian City according to the global shared model of the central server, and use them as the new sample data attribute parameters. Use the Bootstrap sampling method to randomly generate a subset of samples from the samples, which will be used as training samples for one of the decision tree models. Repeat the sampling k times to form k decision tree training samples.

[0158] Step D2.4.2: Train decision trees based on the attribute subsets from the k training samples to form k independent random decision trees.

[0159] Step D2.4.3: Vote on the pipe laying schemes predicted by the k decision trees, and use the voting result as the optimal pipe laying scheme of the random forest.

[0160] Step D2.4.4: Convert the calculated optimal pipe-laying scheme into encrypted parameters using the public key and upload it to the central server for iteration of the globally shared model in federated learning.

[0161] Step D3: The four clients—building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module—upload encrypted and de-identified parameters to the server in the form of public keys.

[0162] Step D4: The server uses its private key to decrypt the encrypted and anonymized parameters uploaded by the four clients, aggregates the decrypted parameters of all client models using federated learning based on knowledge distillation, and updates the global model weights of the M clients according to the following formula: Among them W s+1 For the global model parameters in round s, The client sub-model weights are uploaded to the server by client i in round s. After each round of model weight updates, the central server calculates the global model error and accuracy. The central server can also control the transmission speed and the shutdown of model training.

[0163] Step D5: The central server generates a public key for encrypting transmitted data using the globally shared model and distributes it to each client. Based on the globally shared model, each local client updates the iteration results of other relevant clients as the new sample data attribute parameters. For example, the building load forecast client updates the building load amount calculated by the global shared model, and the circulating pump parameters and heat pump parameters calculated by the unit selection client, so that the random forest model client for energy pile installation schemes can calculate energy pile installation schemes that match the building load and heat pump unit requirements.

[0164] Step D6: Repeat steps D2 to D5 iteratively until the global model is robust. Finally, the client calculates the corresponding results based on the global model. The matching building load end, heat pump unit model, circulation pump model, energy pile structural characteristics, and buried pipe design parameters for a green low-carbon building with energy piles in Dalian are calculated, as shown in Table 4.

[0165] Table 4. Design parameters of a green and low-carbon building with energy piles in Dalian.

[0166]

Claims

1. A green and low-carbon building design method for energy piles based on machine learning, characterized in that, Since the same sample attributes affecting thermal stress and heat transfer efficiency have opposite effects on both, we search for energy pile feature parameters that are simultaneously applied to the prediction results of thermal stress and heat transfer efficiency. These parameters serve as the basic data for training energy pile buried pipe design parameters in the green building design of energy piles. They are then input into the federated learning algorithm for green and low-carbon building design of energy piles to serve as the threshold and basis for judging the optimal solution. The machine learning-based green and low-carbon building design methods for energy piles include energy pile heat exchange efficiency prediction, energy pile thermal stress prediction, and energy pile green and low-carbon building design methods based on federated learning architecture. The prediction of the heat exchange efficiency of the energy pile specifically includes the following steps: S1: Obtain training sample data on the heat exchange efficiency of energy piles and corresponding labels; the training sample data on the heat exchange efficiency of energy piles includes energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters; the label data includes information reflecting the heat exchange efficiency of energy piles; the characteristic parameters of energy piles include energy pile size parameters, energy pile material parameters, and energy pile buried pipe geometric parameters. S2: Preprocessing and feature analysis of training sample data for energy pile heat exchange efficiency; quality analysis, feature analysis, and cleaning of training sample data for energy pile heat exchange efficiency are performed to divide the training set and prediction set of energy pile heat exchange efficiency required for supervised learning. S3: Training of the prediction model for heat exchange efficiency of energy piles. The prediction model for heat exchange efficiency of energy piles is established and trained based on the sample data information of the training set of heat exchange efficiency of energy piles. S4: Input the sample attribute information of the energy pile heat exchange efficiency prediction set. The sample attributes of the energy pile heat exchange efficiency prediction set include: energy pile characteristic parameters, soil and rock parameters, and heat exchange circulating fluid state parameters. S5: Calculate the heat exchange efficiency prediction results corresponding to the heat exchange efficiency prediction set of the energy pile based on the heat exchange efficiency prediction model of the energy pile. S6: Evaluation of the prediction results of the heat exchange efficiency of the energy pile; compare the prediction results obtained in S5 with the labels corresponding to the samples of the heat exchange efficiency prediction set of the energy pile, calculate the prediction error of the heat exchange efficiency prediction model of the energy pile, and evaluate its accuracy. S7: The characteristic parameters of the energy pile, the parameters of the soil and rock mass, and the state parameters of the heat exchange circulating fluid to be predicted are used as the attribute space of the energy pile heat exchange efficiency prediction set. The heat transfer characteristics of the energy pile are predicted and calculated through the energy pile heat exchange efficiency prediction model. The prediction of thermally induced stress in energy piles specifically includes the following steps: T1: Obtain thermal stress training sample data and corresponding labels; thermal stress training sample data includes energy pile characteristic parameters and soil and rock parameters, and label data includes relevant information reflecting the thermal stress of energy piles, including thermally induced pile side friction, thermally induced pile stress, and thermally induced pile strain. T2: Preprocessing of thermal stress training sample data, defining state vectors; performing quality analysis, feature analysis, and cleaning on the thermal stress training sample data to obtain thermal stress training set and thermal stress prediction set; sorting sample attribute variables according to correlation through sensitivity analysis, and taking the sample attribute variables with the highest ranking as elements of state parameters, the state vector at time t is represented as X(t)=[x1(t),x1(t-1),x2(t),x2(t-1),x3(t),x3(t-1),…,x n (t),x n (t-1)]; T3: Select a distance metric; the correlation between each state vector X and the target vector Y in the state space is represented by the distance in the state space. T4: The sample attributes of the thermal stress prediction set include: energy pile characteristic parameters and soil and rock parameters; the sample attributes of the thermal stress training set are used to form a state vector, and the K nearest neighbors of the current state vector are searched in the thermal stress training set to form a K nearest neighbor prediction model; the sample attribute information of the thermal stress prediction set is input into the K nearest neighbor prediction model; the predicted mean of the K nearest neighbors. Calculate the thermal stress prediction results corresponding to the thermal stress prediction set for the K-nearest neighbor prediction model; T5: Evaluation of the thermal stress prediction results of energy piles. The prediction results obtained in T4 are compared with the labels corresponding to the thermal stress prediction set samples to form a confusion matrix. Based on the confusion matrix, the sensitivity, accuracy, and precision of the K-nearest neighbor prediction model are calculated to further evaluate the prediction error and model accuracy of the K-nearest neighbor prediction model. T6: The characteristic parameters of the energy pile and the parameters of the soil and rock mass that need to be predicted are used as the attribute space of the thermal stress prediction set, and the thermal stress prediction calculation of the energy pile is performed by the K-nearest neighbor prediction model. The energy pile green and low-carbon building design method based on a federated learning architecture establishes a client-server architecture-based federated learning algorithm framework for energy pile green and low-carbon building design. The client includes four main modules: building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module, which are used for learning calculations of all sub-modules. After the anonymized parameters calculated by the client are aggregated to the central server for calculation, they are distributed to each client to update their local model until the global model is robust. The specific steps are as follows: D1: Collect and organize building and environmental information, perform data preprocessing, analyze sample data characteristics, and assign sample data characteristics to the corresponding computing module client; sample data is obtained through one or more of the following technologies: on-site survey, on-site monitoring, questionnaire survey, and intelligent image recognition; sample data information includes building information parameters, environmental information parameters, foundation information parameters, unit equipment parameters, and geometric parameters of energy pile buried pipes; D2: The four clients—building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module—update their local models. D3: The four clients—building load calculation module, building load calculation module, unit selection module, and energy pile buried pipe design module—upload encrypted and de-identified parameters to the central server in the form of public keys. D4: The central server uses its private key to decrypt the encrypted and anonymized parameters uploaded by the four clients, performs secure aggregation, and then updates the globally shared model; the central server decodes the encrypted and anonymized parameters; the federated learning on the central server side obtains the globally shared model in at least one of the following synchronous methods: gradient averaging, federated averaging, or knowledge distillation. D5: The central server generates a public key for encrypting data transmission using the global shared model and distributes it to each client; based on the iteration results of each local client transmitted by the global shared model, it updates the parameters of other relevant clients as new sample data attributes. D6: Repeat steps D2 to D5 iteratively until the global model is robust. Finally, the client calculates the corresponding results based on the global shared model. The design parameters that match each other in the green and low-carbon building of energy piles are calculated, including building load, heat pump unit model, circulation pump model, and geometric parameters of energy pile buried pipes.

2. The green and low-carbon building design method for energy piles based on machine learning according to claim 1, characterized in that, The geometric parameters of the energy pile's buried pipes include heat exchange pipe length, pipe spacing, pipe diameter, and pipe layout classification; the material parameters of the energy pile include the thermal conductivity of the heat exchange medium, the elastic modulus of the pile material, the coefficient of expansion, the thermal conductivity of the concrete, and the concrete temperature; the parameters of the soil and rock mass include the elastic modulus of the soil and rock mass, the thermal conductivity of the soil and rock mass, the groundwater depth, the groundwater flow velocity, and the soil and rock mass temperature; the state parameters of the heat exchange circulating fluid include the circulating fluid flow velocity, density, specific heat, and inlet temperature; the heat exchange efficiency of the energy pile includes the comprehensive thermal conductivity, heat exchange per meter, heat extraction, and heat exchange power; and the building information parameters... The data includes indoor human behavior, indoor equipment behavior, building envelope materials, building envelope dimensions, and building envelope thermal properties. Environmental information parameters include ambient temperature, ambient humidity, and solar radiation. Foundation information parameters include foundation bearing capacity and stratum parameters. Building loads are hourly loads and compliance indicators. Unit equipment parameters include heat pump unit parameters and circulating pump parameters. The sample data's label data is categorized based on field test data, model test data, and numerical simulation data. The energy pile's thermally induced stress includes thermally induced strain, pile side friction, and pile end resistance. The sample attribute variables include energy pile size parameters, energy pile material parameters, and energy pile buried pipe geometric parameters.

3. The green and low-carbon building design method for energy piles based on machine learning according to claim 1 or 2, characterized in that, The objects of the cleaning process include incomplete data, erroneous data, duplicate data, and abnormal data; the content of the cleaning process includes identifying invalid values, outliers, and missing values, and processing invalid values, outliers, and missing values.

4. The green and low-carbon building design method for energy piles based on machine learning according to claim 3, characterized in that, The energy pile heat exchange efficiency prediction model is implemented based on one or more of the following algorithms: simulated annealing, regression tree algorithm, random forest regression algorithm, support vector regression algorithm, multiple linear regression algorithm, simulated annealing improved support vector regression algorithm, and clustering regression algorithm.

5. The green and low-carbon building design method for energy piles based on machine learning according to claim 4, characterized in that, When the energy pile heat transfer efficiency prediction model is trained using the simulated annealing method to improve the support vector regression algorithm, the specific steps are as follows: S3.1: To address the relationship between sample attribute information and sample labels, a slack variable ξ is introduced. i , A nonlinear segmentation support vector classifier considering soft margins is constructed using the penalty coefficient C; the prediction accuracy and stability of the energy pile heat transfer efficiency prediction model are obtained through a conditional extremum function. Where w is the normal vector of the hyperplane, C is the penalty coefficient, and ξ, ξ * y is the relaxation factor, ∈ is the hyperparameter that determines the boundary width; i The actual test results for the training samples; f(X) i )=w·Φ(X)+b is the classification hyperplane of the energy pile heat transfer efficiency prediction model; i is the training sample number; N is the number of training set samples; Φ(X) is the nonlinear mapping function; b represents the intercept of the hyperplane; S3.2: Transform the conditional extremum function of step S3.1 into a multivariate function using the Lagrangian function for solution. Set the partial derivative of the Lagrangian function with respect to the optimization objective to 0. The optimization objective includes the normal vector w of the hyperplane, the intercept b of the hyperplane, and the relaxation factors ξ and ξ'. * Transform the original conditional extremum function into a dual function to find the minimum value within the prediction boundary. Where α, α * μ, μ * Represents the Lagrange multipliers; S3.3: The nonlinear mapping function Φ(X) contained in the classification hyperplane of the energy pile heat transfer efficiency prediction model, and its inner product Φ(X) i ) T Φ(X j The processing of ) uses a combination kernel function: the combination kernel function is one or more of the following: Gaussian kernel, linear kernel, polynomial kernel, and Sigmoid kernel; S3.4: Optimize model parameters in the support vector regression algorithm using piecewise simulated annealing: insensitive loss function ∈, penalty coefficient C, and hyperparameters γ, λ, e, c, d in the kernel function; Step S3.4.1: Randomly generate an initial model parameter set for interactive verification. Record the error value EEP as the current annealing system state E0, the initial temperature T0, the temperature of annealing stage one T1, and the annealing end temperature as T2. Step S3.4.2: According to m′ i =m i +s·(τ-0.5)(B i -A i The model parameters are perturbed to form a new set of model parameters, and the current annealing system state parameters E are obtained through interactive verification. n Calculate ΔE = E n -E n-1 ; Where: m′ i Let m be the variable after perturbation. i Let be the current variable, s be the perturbation ratio, τ be a random number in [0,1], and B be the variable. i A i For the current variable m i Scope; Step S3.4.3: When ΔE < 0, accept the new model parameter set and jump to step S3.4.5; otherwise, accept the corresponding model parameter set according to the Metropolis criterion exp(ΔE / kT)-τ > 0 and jump to step S3.4.4; when none of the above conditions are met, reject the critical state, execute step S3.4.2, re-perturb to generate a new model parameter set, perform interactive verification, until the parameter set acceptance condition in step S3.4.3 is met; Step S3.4.4: Cooling down. Upon obtaining the new state, proceed according to the cooling plan. The temperature is lowered. If the set temperature T1 is not reached, the process returns to step S3.4.

2. Once the set temperature T1 is reached, a new annealing plan is initiated. Step S3.4.5: According to the new perturbation method and annealing plan, continue to perturb the model parameter set of the first-stage annealing, and calculate the corresponding annealing system state parameters E. n ; Step S3.4.6: When ΔE < 0, accept the new parameter set and jump to step S3.4.7; otherwise, accept the corresponding parameter set according to the Metropolis criterion exp(ΔE / kT)-τ > 0 and jump to step S3.4.7; when none of the above conditions are met, reject the parameter set and return to step S3.4.5, re-perturb to generate a new model parameter set, perform interactive verification, until the model parameter set acceptance condition in step S3.4.7 is met; Step S3.4.7: Set the end temperature T2 as the algorithm exit point, and set the global maximum number of EEP calculations to N; stop annealing when T2 or N is reached, and accept the critical state cross-validation error E at this time. n For the lowest E n If the corresponding parameter is the optimal prediction parameter, then return to step S3.4.

5.

6. The green and low-carbon building design method for energy piles based on machine learning according to claim 1 or 2, characterized in that, The distance metric is one or more of the following: Euclidean distance, Manhattan distance, Chebyshev distance, Minkowski distance, standardized Euclidean distance, Mahalanobis distance, and cosine of the included angle.

7. The green and low-carbon building design method for energy piles based on machine learning according to claim 1 or 2, characterized in that, The building load calculation module uses environmental and building information parameters as sample attribute data to establish a support vector regression local model for building energy consumption, predicts and calculates the building load, and uploads it to the central processing unit in encrypted form; it also calculates the building's cooling / heating load based on the environmental and building information parameters; the specific steps are as follows: D2.1.1: Decrypt the public key of the encrypted data using the private key, and update the iteration results of the building load module, unit selection module, and energy pile buried pipe module clients according to the global shared model of the central server as new sample data attribute parameters; based on the sample data, use environmental information parameters and building information parameters as sample attribute space sets, and use building load as the learning target; D2.1.2: To address the relationship between sample attribute parameters and building load, slack variables are introduced. A nonlinear segmentation support vector classifier considering soft margins is constructed using the penalty coefficient β; the prediction accuracy and model stability of the local support vector regression model for building energy consumption are represented by the loss function. D2.1.3: The loss function needs to be reduced in dimension to a multivariate function using the Lagrangian function for solution. Let the Lagrangian function be such that it applies to the optimization objective ζ, δ, The partial derivatives are 0, so we obtain the Lagrange multipliers, which transform the original conditional extremum function into the dual function, thereby finding the minimum value within the predicted boundary. D2.1.4: The nonlinear mapping function Φ(X) contained in the classification hyperplane in the local model of building energy consumption support vector regression, and its inner product Φ(X) i ) T Φ(X j The kernel function selected can be one or more of the following: Gaussian kernel, linear kernel, polynomial kernel, and Sigmoid kernel. D2.1.5: Optimize the model parameters in the support vector regression local model for building energy consumption: insensitive loss function ∈, penalty coefficient C, hyperparameters γ, λ, e, c, d in the kernel function; select one or more optimization methods from simulated annealing, grid search, particle swarm optimization, PSO algorithm, and genetic algorithm; D2.1.6: Input the feature parameters of the target building, use the trained support vector regression model to predict and calculate the annual hourly environmental control load of the target building, and use the public key to convert the building load results into encrypted parameters and upload them to the central server for global shared model iteration in federated learning. The building load calculation module establishes a support vector regression local model of building load based on basic building information to predict and calculate building load, and calculates building mechanical load based on building information parameters and foundation information parameters; when the number of samples is insufficient, numerical methods such as finite element, finite difference or discrete element are used to calculate building load; finally, the load results are encrypted and uploaded to the central processing unit. The unit selection module establishes a decision tree-based local model for unit selection based on building load results and unit equipment parameter samples. It determines the parameters of the heat pump and circulating pump and uploads them to the central processing unit with encryption. The unit equipment parameters include: the cooling method of the heat pump unit, the cooling capacity of the heat pump unit, the flow rate of the heat pump unit, the flow rate of the chilled water circulation pump, and the head information of the chilled water circulation pump. The decision tree model is generated based on one or more of the following algorithms: ID3, C4.5, and CART. The specific steps are as follows: D2.3.1: Decrypt the public key of the encrypted data using the private key, update the iteration results of the relevant clients of the building load module, building load module, and energy pile buried pipe module according to the global shared model of the central server as the new sample data attribute parameters; calculate the information gain of all attributes of the training samples according to the information entropy, and sort all attributes according to the information gain. Total information entropy of a given sample: Sample subset {s 1j ,s 2j ,…,s mj Total information entropy: P ij For sample subsets s j Category C i The sample probability; Information entropy value of the samples based on attribute A: The corresponding information gain is: Gain(A) = I(s1,s2,…,s) m )-E(A); D2.3.2: Select the attribute with the largest information gain as the optimal attribute, use the optimal attribute as the basis for sample partitioning, and form the root node by taking the same value of the optimal attribute as the same sample set. D2.3.3: Each root node is then treated as a complete dataset. Samples are divided according to the suboptimal attribute. Samples with the same suboptimal attribute value are grouped into leaf nodes. The decision tree is pruned using the REP, PEP, or MEP methods. The decision tree is formed iteratively. The unit parameters are converted into encrypted parameters using the public key and uploaded to the central server for global shared model iteration in federated learning. The energy pile installation design module, based on building load, building load, and unit equipment parameters, establishes a local random forest model for energy pile installation on the client side, using energy pile structural parameters as sample attributes. It then calculates the energy pile installation parameters as the optimal installation scheme and uploads it to the central processing unit in encrypted form. The specific steps are as follows: D2.4.1: Decrypt the public key of the encrypted data using the private key, update the iteration results of the relevant clients of the building load module, building load module, and unit selection module according to the global shared model of the central server as the new sample data attribute parameters; use the Bootstrap sampling method to randomly generate a subset of samples from the samples as the training samples of one of the decision tree models, and repeat the sampling k times to form k decision tree training samples. D2.4.2: Train decision trees based on a subset of attributes from k training samples to form k independent random decision trees; D2.4.3: The random forest randomly selects a classifier to vote on the pipe laying schemes predicted by k decision trees, and the voting result is taken as the optimal pipe laying scheme of the random forest. D2.4.4: Convert the energy pile buried pipe parameters into encrypted parameters using a public key and upload them to the central server for global shared model iteration in federated learning.

Images