Control method of permafrost refrigeration system based on cooling load

Abstract

The invention discloses a permafrost refrigeration system control method based on a cold load, and belongs to the technical field of permafrost protection measures, and the method comprises the following steps: 1, predicting the cold load of a permafrost protection work site; secondly, a refrigerating system control module based on the cooling load is constructed. Intelligent control over the operation process of the refrigerating system is achieved through an intelligent algorithm, and the dual purposes of permafrost temperature control and refrigerating system energy saving are achieved. The permafrost cold load is predicted based on a machine learning model, a refrigeration system control module based on the cold load is constructed, the control module effectively integrates geological information and meteorological information of different time scales, the dynamic adaptability to key influence factors of permafrost degradation is improved, and the advantages of being high in generalization ability, energy-saving, efficient and the like are achieved; and according to the cold load prediction and feature recognition result, the control strategy of the refrigeration system is adjusted in time, a control instruction is generated, and therefore the permafrost intelligent protection method with the environment self-adaption function is achieved.

Description

Technical Field

[0001] This invention belongs to the technical field of permafrost protection measures, and specifically relates to a control method for a permafrost refrigeration system based on cold load. Background Technology

[0002] With global warming, permafrost faces the challenge of temperature rise and degradation, which can cause a series of engineering geological problems such as surface thawing subsidence and landslides. Permafrost degradation causes deformation and damage to surface structures, seriously affecting the normal use of various infrastructures. Therefore, protecting permafrost is a top priority. Given the insufficient timeliness of traditional measures such as boulders and air-cooled structures and heat pipes, the engineering field has begun to introduce refrigeration technology in recent years to provide forced cooling protection for permafrost foundations. For example, invention patent application number CN201711190185.7 discloses a compression refrigeration system for preventing permafrost degradation, which is a refrigeration system driven by a combination of solar and wind power; invention patent application number CN202311743133.3 discloses a year-round heat pipe refrigeration device and construction method for permafrost areas, which is a compression refrigeration system driven by solar photovoltaic and with multiple refrigeration sections. The refrigeration system is arranged such that the refrigeration sections are evenly distributed at certain intervals on the outside of the foundation of the structure, ensuring that the permafrost between the refrigeration sections can obtain cooling.

[0003] Most existing refrigeration systems employ the vapor compression thermodynamic cycle principle, consisting of a closed loop comprised of four main components: a compressor, a condenser, an expansion valve, and an evaporator. These components are filled with refrigerant, and heat absorption, transfer, and release are achieved through the phase changes of the refrigerant (gas-liquid). The evaporator is the heat-absorbing component, while the condenser is the heat-releasing component. The refrigeration cycle is driven by the compressor and dynamically regulated by the expansion valve; the relevant details are well-known and will not be elaborated upon here. The evaporator in the refrigeration system has a columnar structure, mechanically drilled and embedded into the permafrost layer. The other components (compressor, condenser, and expansion valve) are integrated into a single enclosure, located alongside various types of structures such as railway subgrades.

[0004] The main problems identified in the publicly available patent documents are as follows: the refrigeration system requires external electrical power. If traditional industrial and construction control strategies such as continuous operation, timed intermittent start-stop, or PID temperature tuning are blindly adopted without considering the actual needs of permafrost, it can lead to insufficient or excessive cooling of the permafrost and excessive energy consumption. Especially in the absence of grid power, a photovoltaic energy storage system needs to be installed, and excessive energy consumption will drastically increase power supply costs.

[0005] The balance between the effectiveness of permafrost protection and energy efficiency depends on the relative relationship between the cooling capacity of the refrigeration system and the cooling load of the permafrost. However, due to the complexity of permafrost geological conditions and the variability of meteorological environment, traditional refrigeration system control strategies lack parameter self-adaptation capabilities and cannot meet the requirements for permafrost cooling protection under the drastically fluctuating meteorological environment of the plateau. It is also difficult to identify and avoid phenomena such as no-load, light-load, and overload of the refrigeration system.

[0006] Currently, there is no technology to accurately quantify the cooling load of permafrost and the operating mode of refrigeration systems, which prevents refrigeration systems from achieving optimal operation. Therefore, in the application of permafrost refrigeration protection technology, how to balance the relationship between refrigeration system performance and energy consumption urgently needs to be addressed. Summary of the Invention

[0007] This invention provides a control method for a permafrost refrigeration system based on cold load, aiming to solve the technical problems of existing permafrost refrigeration protection technologies, such as the inability to quantify the cooling demand of permafrost, insufficient basis for operation schemes, and high energy consumption.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A control method for a permafrost refrigeration system based on cooling load includes the following steps:

[0010] The first step is to predict the cooling load at the construction site on the permafrost foundation.

[0011] The second step is to construct a refrigeration system control module based on the cooling load, with the following steps:

[0012] S1: Identify the cooling load characteristics of permafrost foundations based on the cooling load prediction results;

[0013] Cooling load characteristics refer to the magnitude and fluctuation pattern of cooling load, and the load factor is calculated. The load factor is the ratio of the cooling load at a certain moment or time interval of the day to the peak cooling load of the day. Based on the load factor, a time interval of 0.5 hours is set, and the average cooling load value within each time interval is calculated as the cooling load characteristic value of that time interval. A total of 48 cooling load characteristic values ​​are calculated each day. The 48 time intervals of the day are divided into peak intervals, flat intervals, and valley intervals. The load factor of the peak interval is ≥80%, the load factor of the flat interval is <80%, and the load factor of the valley interval is <30%.

[0014] S2: Develop the control strategy for the refrigeration system as follows:

[0015] When the time is in a low period, the cooling system shuts down;

[0016] When the time falls within the peak period, the refrigeration system is turned on, and the opening degree of the expansion valve of the refrigeration system is set to 100%.

[0017] When the time is within a flat period, the refrigeration system is turned on and the opening of the expansion valve is dynamically adjusted.

[0018] S3: Optimize the operation control scheme of the refrigeration system;

[0019] S4: Determine the adaptive environmental control method for the refrigeration system;

[0020] S5: Intelligent control module for designing refrigeration systems;

[0021] Based on the first and second steps, the design of the intelligent control module for the permafrost refrigeration system was completed using the Niagara software platform.

[0022] Furthermore, in the first step, the method for predicting the cooling load of permafrost is as follows:

[0023] S1: Collect multi-source data related to permafrost stability at the construction site of the permafrost foundation, including historical data of meteorological and geological parameters;

[0024] S2: Data preprocessing based on periodic registration for multi-source data;

[0025] S3: Establish a permafrost cooling load mechanism model, calculate the permafrost cooling load corresponding to each hour throughout the year; analyze the spatial distribution characteristics and temporal variation characteristics of permafrost cooling load;

[0026] S4: Construct a feature set of permafrost cooling load, meteorological parameters, and geological parameters;

[0027] The acquired hourly cooling load data for the entire year were preprocessed based on periodic registration to obtain a dataset of permafrost cooling load and local meteorological and geological parameters. Meteorological and geological parameters with significant impacts on permafrost stability were selected as input feature variables. The Pearson correlation coefficient method was used to analyze the correlation between different feature variables and the cooling load. Feature variables with a Pearson correlation coefficient ≥ 0.3 were selected and combined with the cooling load data to form a multi-source data feature set for subsequent steps. The formula for calculating the Pearson correlation coefficient P is:

[0028] (1)

[0029] In the formula, n is the number of data points in each feature variable; X i Let i be the value of the i-th feature variable; Y represents the average of the characteristic variable data. i This is the data for the i-th cooling load. This represents the average value of the cooling load data;

[0030] S5: Perform empirical mode decomposition of permafrost cooling load;

[0031] First, the multi-source data feature set obtained in step S4 is processed by signal processing methods to decompose the original data sequence of permafrost cold load into modal components imf and residual res, as shown in formula (2), to form modal data x(t) composed of imf and res;

[0032] (2)

[0033] In the formula; t is time; K is the total number of modal data components; imf i (t) represents the i-th intrinsic mode data component, which is the data sequence after subtracting the average envelope from the original data sequence; res(t) is the residual signal after subtracting the imf from the original data, which can reflect the long-term trend of cooling load.

[0034] S6: Perform sample entropy preprocessing on the modal data (modal components IMF and residuals Res) of permafrost cold load;

[0035] S7: Construct a neural network model of permafrost cooling load to obtain the predicted value of permafrost cooling load;

[0036] S8: Conduct verification and cyclic training of the results of permafrost cold load prediction;

[0037] S9: Predicting permafrost cooling load: Based on the permafrost foundation conditions in different scenarios, input the corresponding feature variable values ​​into the neural network prediction model to obtain the predicted value of permafrost cooling load.

[0038] Furthermore, in step S1 of the first step, the permafrost foundation construction site includes different types of structures such as roadbeds, building pile foundations, and power tower foundations. The multi-source data includes meteorological data (average annual temperature, average winter temperature, average annual solar irradiance, average annual wind speed, and average annual rainfall) and geological data (vegetation coverage, permafrost ice content, permafrost upper limit, permafrost average annual temperature, engineering structure design dimensions, and engineering structure orientation).

[0039] Furthermore, in step S2 of the first step, the time series data from various data sources obtained in step S1 are aligned to a unified time period, synchronizing and aligning data from different time scales or different time periods.

[0040] Furthermore, in step S3 of the first step, the DeST software is used to construct a cold load mechanism model for permafrost subgrade, and hourly cold load data for the permafrost subgrade target protection area for 8760 hours throughout the year are obtained.

[0041] Furthermore, in step S6 of the first step, the modal data after modal decomposition of the cooling load data undergoes sample entropy preprocessing. The sample entropy preprocessing steps are as follows:

[0042] S61. Let N be the total number of data points included in the components imf and res after empirical mode decomposition of the cooling load data, and let X be the calculation vector corresponding to the sample entropy preprocessing tool (sliding data window). X is a calculation vector consisting of m consecutive data points starting from the i-th data point, where m is the dimension of the sliding data window. Then, N-m+1 m-dimensional calculation vectors X can be constructed. m X m The expression is shown in formula (3):

[0043] (3)

[0044] Where i represents the data window X m The number, ; u represents the time series data of the cooling load.

[0045] S62, Calculate Vectors and The maximum distance L between them m :

[0046] (4)

[0047] In the formula, 0 ≤ k ≤ m-1, j represents the number of another sliding data window (used for comparison with the i-th sliding data window), and k represents the element position index within the sliding data window. Since... and Both sliding data windows are m-dimensional computational vectors, so the value of k ranges from 0 to m-1, corresponding to the positions of the 1st to mth elements in the window.

[0048] S63, Set a given threshold (Distance threshold for similarity judgment), calculate the distance threshold that satisfies and The number of each is denoted as . and Then the matching rate between the m-dimensional vector and the m+1-dimensional vector is calculated using formulas (5) and (6), respectively.

[0049] (5)

[0050] In the formula, Nm is the total number of (m+1)-dimensional sliding data windows, and N-m+1 is the total number of (m)-dimensional sliding data windows. This refers to the number of m-dimensional vectors in the data sequence whose distance to the i-th m-dimensional sliding data window is less than r.

[0051] (6)

[0052] In the formula, This refers to the number of m+1 dimensional vectors in the data sequence whose distance to the i-th m+1 dimensional sliding data window is less than r.

[0053] S64. The expression for sample entropy is:

[0054] (7)

[0055] The sample entropy SampEn(m,r) of the cooling load modal data (composed of components imf and res) is calculated by formula (7), and the overall stability of the cooling load data sequence is judged accordingly. Then, the cooling load data whose sample entropy SampEn(m,r) fluctuation range conforms to the normal pattern are selected, and the neural network model construction in the next step S7 is carried out.

[0056] Furthermore, in step S7 of the first step, the cold load data filtered in step S6 and the meteorological and geological parameters in the multi-source data feature set filtered in step S4 are used to train and construct a neural network model of permafrost cold load, thereby predicting the cold load of permafrost under other different scenario conditions.

[0057] The neural network model for permafrost cooling load is a CNN-BiLSTM (Convolutional-Bidirectional Long Short-Term Memory) neural network model. This model combines the advantages of CNN in extracting local temporal features with BiLSTM in capturing bidirectional long-term temporal dependencies. The model structure includes an input layer, a CNN feature extraction module (convolutional layer + pooling layer), a BiLSTM modeling layer, and an output layer, all composed of neurons. The modeling approach is as follows:

[0058] First, the input layer incorporates the cooling load data selected in step S6 and the meteorological and geological parameters from the multi-source data feature set selected in step S4. Second, the input data's local correlation features are extracted using a CNN feature extraction module. Then, data meeting the local correlation feature requirements are combined into a feature matrix and input into a BiLSTM modeling layer to capture the bidirectional long-term temporal variation patterns of the computation vector. Finally, the cooling load prediction result is obtained through the output layer. The specific method is as follows:

[0059] S71. Using a convolutional layer, feature convolution is performed on the permafrost cold load data and multi-source data feature sets (meteorological parameters and geological parameters) from step S6. The convolution kernel (i.e., the computational tool of the convolutional layer) slides on the input data and performs convolution operations to generate the feature matrix of the input data. The formula for calculating this feature matrix is ​​as follows:

[0060] (8)

[0061] In the formula: and These represent the number of input data to the convolutional layer and the number of data points in the output feature matrix of the convolutional layer, respectively; a is the kernel size, i.e., the number of data points covered by the convolution operation; q is the number of zero-padding layers, used to maintain the matching degree between the output length and the input length of the feature matrix and avoid the loss of edge features; s is the stride, the step size of each slide of the convolutional kernel;

[0062] Pooling layers perform max pooling by downsampling the feature matrix output by convolutional layers. This is achieved by using a sliding data window to scan the feature matrix and taking the maximum value within the window as the output data. This reduces the dimensionality of the feature matrix, lowers the computational cost, and enhances the robustness of the output data.

[0063] S72. The one-dimensional data output from the CNN feature extraction module is then fed into the BiLSTM modeling layer. The BiLSTM modeling layer consists of two LSTM neural networks operating in opposite directions. The BiLSTM modeling layer simultaneously activates both the forward and backward branches of the LSTM neural network: the forward LSTM neural network learns the influence of historical local features on the current moment by moving from the earliest to the latest input data sequence; the backward LSTM neural network captures the correlation between future local features and the current moment by moving from the latest to the earliest input data sequence.

[0064] After the bidirectional LSTM neural network completes the full sequence traversal, the BiLSTM modeling layer will concatenate and fuse the forward and backward hidden states corresponding to each input data to generate a high-dimensional feature matrix containing bidirectional temporal dependencies.

[0065] Subsequently, the high-dimensional feature matrix is ​​input into the fully connected layer (output layer), and the high-dimensional features are converted into scalar results corresponding to the cold load data through linear mapping, finally obtaining the predicted value of the permafrost cold load.

[0066] S73. Optimize the hyperparameters of the CNN-BiLSTM neural network model using the GA / PSO optimization algorithm (genetic algorithm / particle swarm optimization algorithm);

[0067] A combination of particle swarm optimization and genetic algorithm was used to optimize the hyperparameters of the CNN-BiLSTM neural network model. Hyperparameters are the configuration parameters of the CNN-BiLSTM neural network model before training, including the number of neurons in the BiLSTM modeling layer, the learning rate, etc. The steps are as follows:

[0068] 1) Initialize GA parameters: Initialize hybridization probability , The probability of performing a crossover operation (exchanging some parameters) between two different hyperparameter combinations; initial mutation probability. , The probability of performing a mutation operation (fine-tuning the value) on a single parameter in a combination of hyperparameters;

[0069] 2) Initialize PSO parameters: Initialize the inertia weight w, which refers to the degree to which the combined scheme continues the previous search direction; initialize the learning factors c1 and c2, where c1 refers to the degree to which the combined scheme learns from its own historical best scheme, and c2 refers to the degree to which the combined scheme learns from the group's global best scheme.

[0070] 3) Initialize the range of hyperparameters to be optimized: Initialize the number of neurons in the BiLSTM modeling layer, the learning rate, and the size of the CNN convolutional kernel;

[0071] 4) Treat each hyperparameter combination as an optimization particle, and search for the globally optimal and individually optimal particles using the PSO optimization algorithm. The objective function is fitness (the fitness value of the optimization particle; a higher fitness value indicates a better hyperparameter combination). The objective function formula is:

[0072] (9)

[0073] Where: n is the number of cooling load samples; This is the predicted cooling load value; This represents historical (actual) cooling load data.

[0074] 5) Based on particle fitness, particles are divided into two categories: dominant particles and suboptimal particles. Dominant particles are retained for the next generation.

[0075] 6) Randomly select two particles from the dominant particles and cross them to update the particle velocity and position;

[0076] 7) Randomly select two particles from the crossed particles to mutate;

[0077] 8) Update the fitness value, global optimum, and individual optimum of the mutated particles;

[0078] 9) Determine if the requirements are met. If the optimal parameters are achieved, output them. Otherwise, return to step 3) Initialize the range of hyperparameters to be optimized.

[0079] Furthermore, in step S8 of the first step, the prediction results of step S7 are tested, and the index for evaluating the prediction accuracy is the coefficient of determination R. 2 Mean absolute percentage error E MAPE :

[0080] (10)

[0081] In the formula, The average of the true values; This represents the true value of the k-th sample; This represents the model's prediction for the k-th sample;

[0082] (11)

[0083] Furthermore, in step S3 of the second step, the optimization method for the refrigeration system operation control scheme is as follows: First, specify an objective function with the minimum daily power consumption minP as the objective and the controlled operation state of the refrigeration system and the opening degree of the expansion valve as variables, i.e., formula (12); then, specify two constraint function functions, i.e., formulas (13) and (14), corresponding to the cooling capacity and power consumption of the refrigeration system, respectively; finally, solve formula (12) to obtain the controlled operation state of the refrigeration system that satisfies the constraints. The opening degree φ of the expansion valve is the optimized operation control scheme for the refrigeration system.

[0084] S31, Objective Function

[0085] The primary task of refrigeration system operation control is to meet the cooling load of permafrost roadbeds, with the objective being the minimum daily power consumption minP. The objective function is:

[0086] (12)

[0087] In the formula, P is the total daily power consumption of the refrigeration system, kW·h; t is the time period, h; p is the power consumption rate of the refrigeration system during the time period t, kW; φ is the opening degree of the expansion valve, which is adjusted between 0 and 100%; for a fixed model of refrigeration system, the power consumption rate of the refrigeration system depends on the opening degree φ of the expansion valve; t is the time period, h. This represents the controlled operating state of the refrigeration system during time period t, with 1 for on and 0 for off.

[0088] S32, Constraints

[0089] 1) Cooling capacity constraints

[0090] The relationship between the cooling load of permafrost and the required cooling capacity of the refrigeration system is as follows:

[0091] (13)

[0092] In the formula, COP is the coefficient of performance of the refrigeration system, and y(t) is the permafrost cooling load during time period t, in kW.

[0093] 2) Power consumption constraints

[0094] The power consumption of the cooling system must not exceed the power supply capacity of the photovoltaic system, and the minimum daily power consumption (minP) must not exceed the power supply capacity of the photovoltaic system.

[0095] (14)

[0096] In the formula, The maximum daily power supply of the photovoltaic system is expressed in kW•h.

[0097] Furthermore, in step S4 of the second step, the adaptive environmental control method for the refrigeration system follows these steps:

[0098] First, the association rule mining method is used to deeply mine the historical data of the refrigeration system operation to find the correspondence between the expansion valve opening parameter of the refrigeration system and the cooling output power of the refrigeration system.

[0099] Then, based on the objective function and constraint requirements in step S3, from the perspective of global optimization, the optimal control of the refrigeration system's start-stop state and expansion valve opening parameters is achieved, forming an adaptive environmental control method for permafrost refrigeration systems.

[0100] The technological advancements achieved by this invention compared to existing technologies are as follows:

[0101] This invention predicts the cooling load at permafrost foundation construction sites using a machine learning model and constructs a refrigeration system control module based on the cooling load. This control module effectively integrates meteorological and geological information, enhancing its dynamic adaptability to key influencing factors of permafrost degradation. It adapts to permafrost engineering scenarios with significant uncertainties and external dynamic disturbances, exhibiting advantages such as strong generalization ability and energy efficiency, thereby improving the accuracy of permafrost cooling load prediction. Furthermore, based on the cooling load prediction and feature recognition results, it promptly adjusts the refrigeration system's control strategy and generates control commands. This invention achieves autonomous perception, decision-making, execution, learning, and feedback capabilities during the refrigeration system's operation through intelligent algorithms, achieving the dual goals of permafrost temperature control and energy saving in the refrigeration system. Attached Figure Description

[0102] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0103] In the attached diagram:

[0104] Figure 1 This is a schematic diagram of the structural composition of the refrigeration system in an embodiment of the present invention;

[0105] Figure 2 This is a schematic diagram of the field application scheme of the refrigeration system in an embodiment of the present invention;

[0106] Figure 3 A flowchart illustrating a control method for a permafrost refrigeration system based on cold load, provided as an embodiment of the present invention;

[0107] Figure 4 This is a flowchart of the first step in an embodiment of the present invention;

[0108] Figure 5 This is a flowchart of the second step in an embodiment of the present invention;

[0109] Figure 6 This is a schematic diagram of the calculation process for cold load on permafrost subgrade based on DeST software in step S3 of the first step.

[0110] Figure 7 The flowchart shows the Empirical Mode Decomposition (EMD) method in step S5 of the first step.

[0111] Figure 8 A schematic diagram of the algorithm flow for optimizing CNN-BiLSTM using GA / PSO in step (3) of S7 in the first step;

[0112] In the diagram: 1-compressor, 2-condenser, 3-expansion valve, 4-evaporator, 5-chassis, 6-roadbed, 7-permafrost layer. Detailed Implementation

[0113] The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0114] like Figure 1 As shown, the existing refrigeration system adopts the principle of vapor compression thermodynamic cycle, consisting of a closed loop composed of four main components: compressor 1, condenser 2, expansion valve 3, and evaporator 4. The system is filled with refrigerant, and heat absorption, transfer, and release are achieved through the phase changes of the refrigerant (gas-liquid). The evaporator is the heat-absorbing component, and the condenser is the heat-releasing component. The refrigeration cycle is driven by the compressor and dynamically regulated by the expansion valve. The relevant details are well-known and will not be elaborated further.

[0115] like Figure 2 As shown, the evaporator 4 in the existing refrigeration system adopts a columnar structure and is implanted into the permafrost layer 7 through mechanical drilling. Other components (compressor 1, condenser 2, expansion valve 3) are integrated and installed in a single casing 5, and distributed along the sides of various types of structures such as railway subgrade 6.

[0116] In this invention, cooling load refers to the amount of cooling the refrigeration system needs to supply per unit time to maintain the target temperature of the object (permafrost foundation). It directly reflects the end-user demand for cooling capacity from the permafrost foundation. Cooling load is a key basis for the stability assessment of permafrost foundations and the design and operation control optimization of refrigeration systems.

[0117] This invention provides a control method for a permafrost refrigeration system based on cooling load, the process of which is as follows: Figure 3 As shown, it includes the following steps:

[0118] The first step is to predict the cooling load at the permafrost foundation construction site. The specific process is as follows: Figure 4 As shown, the prediction method is as follows:

[0119] S1: Collect multi-source data related to the stability of permafrost foundations at the construction site and obtain cold load data of permafrost subgrade;

[0120] The permafrost foundation construction sites include different types of structures such as roadbeds, building pile foundations, and power tower foundations. The multi-source data includes meteorological parameters (average annual temperature, average winter temperature, average annual solar irradiance, average annual wind speed, and average annual rainfall) and geological parameters (vegetation coverage, permafrost ice content, permafrost upper limit, permafrost average annual temperature, engineering structure design dimensions, and engineering structure orientation), totaling 11 parameters.

[0121] Meteorological parameters can be obtained through meteorological platforms or simulation software. Geological parameters can be obtained through geological surveys, literature reviews, and field monitoring.

[0122] Compared with traditional prediction methods based on a single data source or a single scale, this invention can more comprehensively consider the combined impact of various factors on the stability of permafrost.

[0123] S2: Data preprocessing based on periodic registration for multi-source data;

[0124] By using a data preprocessing method based on periodic registration, the time series data from various data sources obtained in step S1 are aligned to a unified time period, thus synchronizing and aligning data from different time scales or different time periods.

[0125] In addition, missing values ​​are filled using interpolation or mean imputation methods; outliers are identified and removed; and for data of different magnitudes, scaling is performed using min-max normalization to make the data comparable on the same scale. It is ensured that time series data from all data sources have consistent time periods and timestamps, providing a data foundation for subsequent modeling and analysis.

[0126] S3: Optimize the operation control scheme of the refrigeration system;

[0127] An optimization model is established with the objective function of minimizing daily power consumption, and the controlled operating state of the refrigeration system and the expansion valve opening as variables. Cooling capacity and power consumption are set as constraints. Finally, by solving this model, the optimal operating state and expansion valve opening that satisfy the constraints are obtained, thus completing the optimization of the operation control scheme. The specific process is as follows: Figure 5 As shown.

[0128] Establish a mechanistic model for the cooling load on permafrost, the specific process of which is as follows: Figure 6 As shown, the permafrost cooling load for each hour throughout the year is obtained; the spatial distribution characteristics and temporal variation characteristics of the permafrost cooling load are analyzed.

[0129] A permafrost cooling load mechanism model was constructed using DeST (Designer's Simulation Toolkit) cooling load simulation software. This included establishing a micro-element combination model of the permafrost foundation and structures, setting thermal property parameters, and calculating hourly cooling load data for the target protection area of ​​the permafrost foundation over 8760 hours throughout the year. The method for constructing the permafrost cooling load mechanism model is as follows: Figure 6 As shown, the specific steps are as follows:

[0130] (1) Establish a calculation model for permafrost foundations in DeST software. The steps include: In the graphical user interface, through functions such as drawing straight walls and modifying slanted walls, the working conditions of the structure are input to establish a micro-element combination model of permafrost foundation and engineering structure; In the global settings menu, you can set the orientation and geographical location of the engineering structures; In the building components menu, set the thermal properties parameters for different parts of the foundation and engineering structure; In the properties interface, input the environmental conditions, adjust the thermal disturbance to set the initial temperature of the engineering structure; and edit the default ground to set the initial temperature of the permafrost foundation.

[0131] (2) Based on the geographical coordinates, call the meteorological model, and set the external terrain conditions through the shading model and the illumination model, and set the boundary conditions of the engineering structures and the permafrost foundation.

[0132] (3) Using the BAS thermal characteristics module of DeST software, a permafrost calculation model based on the BAS module was established, and the temperature field of the permafrost was calculated to obtain the predicted value of the micro-element temperature.

[0133] (4) Compare the similarity threshold between the predicted value and the measured value of the micro-element temperature. If the threshold condition is not met, correct the geothermal parameters (including heat capacity, soil thermal conductivity, etc.), that is, adjust the geothermal disturbance until the threshold condition is met (error ≤ 5%).

[0134] (5) Output accurate permafrost temperature field (predicted temperature distribution data of permafrost interior output by DeST software simulation). (6) Set the target temperature of permafrost foundation, calculate the cooling load of each micro-element, calculate the maximum cooling load and average cooling load and other characteristic values ​​through algebraic summation, and analyze the spatial distribution characteristics and temporal variation characteristics of permafrost cooling load. The above calculation methods are all common knowledge in the construction field and will not be elaborated here.

[0135] S4: Construct a feature set of permafrost cooling load, meteorological parameters, and geological parameters;

[0136] The acquired hourly cooling load data for the entire year were preprocessed using periodic registration to obtain a dataset of permafrost cooling load and its local meteorological and geological parameters (characteristic variables). Due to the significant differences in meteorological and geological conditions within permafrost regions, and the varying key elements, different permafrost degradation patterns can occur. To reduce the complexity of prediction, the Pearson correlation coefficient method was used to analyze the correlation between different characteristic variables and the cooling load, thus selecting and determining the input characteristic variables. The Pearson correlation coefficient measures the strength of the linear correlation between variables; when the absolute value of the correlation coefficient is less than 0.3, the correlation is considered weak, and characteristic variables with weak or no correlation are excluded, reducing the prediction dimensionality and determining the optimal characteristic variables for the prediction model.

[0137] The formula for calculating the Pearson correlation coefficient P is:

[0138] (1)

[0139] In the formula, n is the number of data points in each feature variable; X i Let i be the value of the i-th feature variable; Y represents the average of the characteristic variable data. i This is the data for the i-th cooling load. This represents the average value of the cooling load data;

[0140] S5: Perform empirical mode decomposition (EMD decomposition method) on multi-year permafrost cooling load data. The specific process is as follows: Figure 4 As shown;

[0141] To improve the accuracy of cold load prediction for permafrost, this invention employs a prediction model that combines signal processing and machine learning algorithms.

[0142] First, the feature set obtained in step S4 is processed using signal processing methods. Empirical mode decomposition (EMD) is a nonlinear stationary time-frequency data analysis method and a common signal processing method that can transform non-stationary cold load data into stationary modal components. The purpose of using the EMD method in this invention is to decompose the original data sequence x(t) of permafrost cold load into intrinsic modal components imf and residuals res, in order to better characterize the trend and fluctuation characteristics of cold load changes.

[0143] This invention divides the decomposed components and residuals into high-frequency and low-frequency components, and uses the EMD algorithm to decompose the original data into different modal components and residuals carrying time characteristics, such as... Figure 7 As shown in formula (2), the entire structure consists of imf and res;

[0144] (2)

[0145] In the formula; t is time; K is the total number of modal data components; imf i (t) represents the i-th intrinsic modal data component, i.e., the data sequence after subtracting the average envelope from the original data sequence; res(t) is the residual signal after subtracting the imf from the original data, which can reflect the long-term variation trend of the cooling load. This method can decompose the original data sequence into modal components from high frequency to low frequency, i.e., modal components (imf) and residuals (res), such as... Figure 4 The intermediate mode component (IMF) feature selection step, based on sample entropy, aims to remove interfering signals and retain only valid signals as input to the prediction model. Its advantages include strong adaptability and efficiency; it can extract data trends and eliminate nonlinearity and non-stationarity of data sequences, thereby reducing the complexity of cold-loaded data and uncovering latent features within the data.

[0146] S6: Perform sample entropy preprocessing on permafrost cold load modal data;

[0147] Given that the permafrost cooling load data, even after empirical mode decomposition, still exhibits massive sample complexity and characteristics such as non-stationarity, nonlinearity, and time lag, this invention continues to employ the sample entropy method to measure the complexity of its time series and identify and remove meaningless data. This reduces the complexity of the cooling load sequence, minimizes time loss due to excessive decomposition, and improves the accuracy and reliability of data analysis. The higher the sample entropy value, the greater the complexity of the time series. The sample entropy preprocessing steps are as follows:

[0148] S61. Let N be the total number of data points included in the components imf and res after empirical mode decomposition of the cooling load data, and let X be the calculation vector corresponding to the sample entropy preprocessing tool (sliding data window). X is a calculation vector consisting of m consecutive data points starting from the i-th data point, where m is the dimension of the sliding data window. Then, N-m+1 m-dimensional calculation vectors X can be constructed. m X m The expression is shown in formula (3):

[0149] (3)

[0150] Where i represents the data window X m The number, ; u represents the time series data of the cooling load.

[0151] S62, Calculate Vectors and The maximum distance L between them m :

[0152] (4)

[0153] In the formula, 0 ≤ k ≤ m-1, j represents the number of another sliding data window (used for comparison with the i-th sliding data window), and k represents the element position index within the sliding data window. Since... and Both sliding data windows are m-dimensional computational vectors, so the value of k ranges from 0 to m-1, corresponding to the positions of the 1st to mth elements in the window.

[0154] S63, Set a given threshold (Distance threshold for similarity judgment), calculate the distance threshold that satisfies and The number of each is denoted as . and Then the matching rate between the m-dimensional vector and the m+1-dimensional vector is calculated using formulas (5) and (6), respectively.

[0155] (5)

[0156] In the formula, Nm is the total number of (m+1)-dimensional sliding data windows, and N-m+1 is the total number of (m)-dimensional sliding data windows. This refers to the number of m-dimensional vectors in the data sequence whose distance to the i-th m-dimensional sliding data window is less than r.

[0157] (6)

[0158] In the formula, This refers to the number of m+1 dimensional vectors in the data sequence whose distance to the i-th m+1 dimensional sliding data window is less than r.

[0159] S64. The expression for sample entropy is:

[0160] (7)

[0161] The values ​​of the sample entropy parameters m and r are determined through trial and error. Based on repeated experiments, the sample entropy parameters of this invention... , When std is the standard deviation of the cooling load data sequence, the calculated sample entropy has better characteristics. The sample entropy SampEn(m,r) of each cooling load component imf and res is calculated by formula (7), and the overall stability of the cooling load data sequence is judged accordingly. Then, the allowable fluctuation range of SampEn(m,r) is set, which is generally set to 95% of the mean of SampEn(m,r). Based on this, cooling load data with sample entropy SampEn(m,r) fluctuation amplitude conforming to normal rules and without abnormal deviation are selected, and then proceed to the next step of neural network model construction.

[0162] It is important to note that the modal data components IMF and RES are intermediate analysis data, not the final objects to be modeled and predicted. That is, the neural network model ultimately predicts the cooling load data. Therefore, if the sample entropy calculation results of the IMF and RES components show stability, the corresponding cooling load data will be retained; if the sample entropy calculation results of the IMF and RES components show anomalies, the corresponding cooling load data will be discarded. In other words, the IMF and RES components are only tools used to check the stability of the cooling load data, not the targets for selection.

[0163] S7: Construct a neural network model of permafrost cooling load to obtain the predicted value of permafrost cooling load;

[0164] A neural network model of permafrost cooling load was constructed. The model was trained using the cooling load data selected in step S6 and the multi-source data feature set (meteorological and geological parameters) selected in step S4. Finally, the prediction results of each group of imf and res were superimposed and reconstructed to obtain the predicted cooling load values ​​of permafrost foundation under different scenario conditions (meteorological and geological conditions are different from the calculation conditions of the mechanistic model), such as... Figure 7 As shown.

[0165] The neural network model for calculating the cold load of permafrost is a CNN-BiLSTM (Convolutional-Bidirectional Long Short-Term Memory) neural network model. This model combines the advantages of CNN in extracting local temporal features with BiLSTM in capturing bidirectional long-term temporal dependencies. The model structure includes an input layer, a CNN feature extraction module (convolutional layer + pooling layer), a BiLSTM modeling layer, and an output layer, all composed of neurons. The modeling approach is as follows: First, the cold load data x(t) from step S6 and meteorological and geological factor data are incorporated into the input layer; second, the local correlation features of the input data are extracted by the CNN feature extraction module; then, the data that meet the requirements of local correlation features are combined into a feature matrix and input into the BiLSTM modeling layer to capture the bidirectional long-term temporal variation patterns of the computational vector; finally, the cold load prediction result is obtained through the output layer. The main steps and methods include:

[0166] S71. Use convolutional layers to perform feature convolution on the cold load data and feature variable dataset (meteorological parameters and geological parameters) in step S6. The purpose is to aggregate high-dimensional data into one-dimensional data. Then, use pooling layers to further reduce the complexity of the data and increase the diversity of data features, so as to capture short-term inherent features in cold load prediction.

[0167] A convolutional kernel (the computational tool of a convolutional layer, also known as a filter or feature detector) slides across the input data and performs convolution operations, thereby generating a feature matrix of the input data. The calculation process of this feature matrix can be summarized as follows: a weighted summation of local regions of the input data using the convolutional kernel, with a bias term added, followed by a nonlinear transformation using an activation function. The formula for calculating the feature matrix is:

[0168] (8)

[0169] In the formula: and These represent the number of input data to the convolutional layer and the number of data points in the output feature matrix of the convolutional layer, respectively; a is the kernel size, i.e., the number of data points covered by the convolution operation; q is the number of zero-padding layers, used to maintain the matching degree between the output length and the input length of the feature matrix and avoid the loss of edge features; s is the stride, the step size of each slide of the convolutional kernel;

[0170] Through convolutional layers, CNN models can automatically learn useful features from the input data. Pooling layers downsample the feature matrix output by the convolutional layers, i.e., perform max pooling, to reduce the spatial size and computational cost of the data while retaining important feature information, thereby improving the model's robustness and generalization ability. The pooling method uses a sliding data window to scan the feature matrix and takes the maximum value within the window as the output data, thus achieving feature matrix dimensionality compression, reducing computational cost, and enhancing output robustness.

[0171] The beneficial effects of CNN models include sharing convolutional kernels, capturing high-dimensional data features without requiring manual feature selection, and exploring linear and nonlinear relationships between different variables through weighted feature matrices. They efficiently extract data feature information through local perception, parameter sharing, pooling operations, and spatial hierarchical structures.

[0172] S72. The one-dimensional data output from the CNN feature extraction module is then fed into the BiLSTM modeling layer. The BiLSTM modeling layer consists of two LSTM neural networks operating in opposite directions. The BiLSTM modeling layer simultaneously activates both the forward and backward branches of the LSTM neural network: the forward LSTM neural network learns the influence of historical local features on the current moment by moving from the earliest to the latest input data sequence; the backward LSTM neural network captures the correlation between future local features and the current moment by moving from the latest to the earliest input data sequence.

[0173] After the bidirectional LSTM neural network completes the full sequence traversal, the BiLSTM modeling layer will concatenate and fuse the forward and backward hidden states corresponding to each input data to generate a high-dimensional feature matrix containing bidirectional temporal dependencies.

[0174] Subsequently, the high-dimensional feature matrix is ​​input into the fully connected layer (output layer), and the high-dimensional features are converted into scalar results corresponding to the cold load data through linear mapping, finally obtaining the predicted value of the permafrost cold load.

[0175] S73. Optimize the hyperparameters of the CNN-BiLSTM neural network model using the GA / PSO optimization algorithm. The specific process is as follows: Figure 8 As shown. To avoid the optimization process of the CNN-BiLSTM neural network model getting stuck in local optima and premature convergence, this invention uses a combination of particle swarm optimization (PSO) and genetic algorithm (GA) to optimize the hyperparameters of the neural network model. The advantage of PSO is that it can retain local and global optimum information by continuously updating local and global optimum particles, resulting in fast convergence, but it is prone to getting stuck in local optima. The advantage of GA is that it can enrich the population diversity by performing crossover and mutation operations on particles, avoiding premature convergence. The advantages of GA and PSO are complementary, achieving a co-evolution effect. This invention uses GA and PSO algorithms to optimize the update strategy of the weights and thresholds of the neurons in the neural network model. The steps of GA / PSO optimization of CNN-BiLSTM are as follows: Figure 8 As shown, the specific steps are as follows:

[0176] 1) Initialize GA parameters: Initialize hybridization probability , The probability of performing a crossover operation (exchanging some parameters) between two different hyperparameter combinations; initial mutation probability. , The probability of performing a mutation operation (fine-tuning the value) on a single parameter in a combination of hyperparameters;

[0177] 2) Initialize PSO parameters: Initialize the inertia weight w, which controls the degree to which the combination scheme continues the previous search direction; initialize the learning factor c, which controls the degree to which the combination scheme learns from the better combination scheme.

[0178] 3) Initialize the range of hyperparameters to be optimized: Initialize the number of neurons in the BiLSTM modeling layer, the learning rate, and the size of the CNN convolutional kernel;

[0179] 4) Treat each hyperparameter combination as an optimization particle, and search for the globally optimal and individually optimal particles using the PSO optimization algorithm. The objective function is fitness (the fitness value of the optimization particle; a higher fitness value indicates a better hyperparameter combination). The objective function formula is:

[0180] (9)

[0181] Where: n is the number of cooling load samples; This is the predicted cooling load value; This represents historical (actual) cooling load data.

[0182] 5) Based on particle fitness, particles are divided into two categories: dominant particles and suboptimal particles. Dominant particles are retained for the next generation.

[0183] 6) Randomly select two particles from the dominant particles and cross them to update the particle velocity and position;

[0184] 7) Randomly select two particles from the crossed particles to mutate;

[0185] 8) Update the fitness value, global optimum, and individual optimum of the mutated particles;

[0186] 9) Determine if the requirements are met. If the optimal parameters are achieved, output them. Otherwise, return to step 3.

[0187] The beneficial effects of the CNN-BiLSTM model optimized based on the GA / PSO algorithm include: Firstly, the GA / PSO optimization algorithm can optimize the network structure parameters of the CNN-BiLSTM, reducing the bias of subjective human judgment. Secondly, using the CNN-BiLSTM network model can extract the spatial features of factors influencing cooling load, as well as the time series features of cooling load, thereby improving the accuracy of cooling load prediction.

[0188] S8: Prediction result verification and iterative training;

[0189] The prediction results of step S7 are tested, and the coefficient of determination R is used to evaluate the accuracy of the prediction. 2 Mean absolute percentage error E MAPE Among them, R 2 This reflects the goodness of fit between the predicted and actual values; the closer it is to 1, the more accurate the prediction. E MAPE The closer it is to 0, the better the prediction effect.

[0190] (10)

[0191] In the formula, The average of the true values; This represents the true value of the k-th sample; This represents the model's prediction for the k-th sample;

[0192] (11)

[0193] S9: Predicting permafrost cooling load: Based on other different permafrost foundation scenarios, input the corresponding feature variables (meteorological parameters and geological parameters) into the neural network model to obtain the predicted value of permafrost cooling load.

[0194] For specific permafrost foundation scenarios and working conditions, input the values ​​of the corresponding characteristic variables (meteorological parameters and geological parameters) to obtain the predicted value of permafrost cooling load.

[0195] The second step is to construct a refrigeration system control module based on the cooling load. The specific process is as follows: Figure 5 As shown.

[0196] To ensure the refrigeration system meets the cooling load requirements of permafrost and operates at its optimal state while minimizing energy consumption, optimized control is a crucial issue. The optimization objective of this invention's refrigeration system operation control is to precisely and rapidly adjust the cooling performance of the refrigeration system by controlling its start-stop state and expansion valve opening. This aims to maintain the cooling performance of the refrigeration system and its effect on permafrost temperature regulation while minimizing energy consumption.

[0197] S1: Identify the cold load characteristics of permafrost subgrade based on the cold load prediction results;

[0198] Cooling load characteristics refer to the magnitude and fluctuation pattern of cooling load, and the load factor is calculated. The load factor is the ratio of the cooling load at a certain moment or time interval of the day to the peak cooling load of the day. For specific permafrost protection sites, after obtaining the cooling load prediction results, the predicted values ​​are first reduced for noise using a limited-amplitude recursive averaging filtering method. Specifically, the average cooling load value within each 0.5-hour time interval is calculated as the cooling load characteristic value for that time interval, resulting in a total of 48 cooling load characteristic values ​​per day.

[0199] Based on the load factor, the 48 time segments of each day (each time segment lasting 0.5 hours) are divided into peak, off-peak, and valley segments; the load factor of the peak segment is ≥80%, the load factor of the off-peak segment is ≤30%, the load factor of the valley segment is <80%, and the load factor of the valley segment is <30%.

[0200] S2: Develop the control strategy for the refrigeration system as follows:

[0201] When the time is in a low period, the cooling system shuts down;

[0202] When the time falls within the peak period, the refrigeration system is turned on, and the opening degree of the expansion valve of the refrigeration system is set to 100%.

[0203] When the time is within a flat period, the refrigeration system is turned on and the opening of the expansion valve is dynamically adjusted.

[0204] S3: Optimize the operation control scheme of the refrigeration system;

[0205] Cooling load is a key factor in the operation and control of refrigeration systems; therefore, cooling load prediction and its characteristic identification are the foundation for optimizing the operation and control of refrigeration systems.

[0206] S31, Objective Function

[0207] The primary task of refrigeration system operation control is to meet the cooling load of permafrost roadbeds, with the target being the minimum daily power consumption (minP).

[0208]

[0209] In the formula, P is the total daily power consumption of the refrigeration system, kW·h; t is the time period, h; and p(t) is the power consumption rate of the refrigeration system during the time period t, kW.

[0210] For a fixed-model refrigeration system, the power consumption rate depends on the opening degree φ of the expansion valve. The power consumption rate p(t) of the refrigeration system during time period t is calculated as follows:

[0211]

[0212] In the formula, φ is the opening degree of the expansion valve, which can be adjusted between 0 and 100%. This represents the controlled operating state of the refrigeration system during time period t, with 1 for on and 0 for off.

[0213] The objective function of the refrigeration system is derived as follows:

[0214] (12)

[0215] S32, Constraints

[0216] 1) Cooling capacity constraints

[0217] The relationship between the cooling load of permafrost and the required cooling capacity of the refrigeration system is as follows:

[0218]

[0219] In the formula, q(t) is the cooling output power of the refrigeration system, kW; y(t) is the permafrost cooling load during time period t, kW;

[0220] The relationship between the power consumption and cooling capacity of a refrigeration system is as follows:

[0221]

[0222] In the formula, COP is the coefficient of performance of the refrigeration system;

[0223] The cooling capacity constraint condition is derived as follows:

[0224] (13)

[0225] 2) Power consumption constraints

[0226] The power consumption of the cooling system must not exceed the power supply capacity of the photovoltaic system, and the minimum daily power consumption (minP) must not exceed the power supply capacity of the photovoltaic system.

[0227] (14)

[0228] In the formula, The maximum daily power supply of the photovoltaic system is expressed in kW•h.

[0229] Based on the objective function (12) and constraints (13) and (14) constructed above, optimization algorithms (genetic algorithm and particle swarm optimization algorithm) are used to solve the problem. Under the conditions of satisfying the cooling requirement of formula (13) and the power limit of formula (14), a set of solution vectors that minimizes the P value in formula (12) is sought. and The solution vector represents the optimized refrigeration system operation control scheme, specifying the start / stop status and optimal valve opening at each moment.

[0230] S4: Determine the adaptive environmental control method for the refrigeration system;

[0231] First, association rule mining is used to deeply mine historical data of the refrigeration system to find the correspondence between the expansion valve opening parameters and the system's cooling output power. Then, based on the objective function and constraints in step S3, from a global optimization perspective, the start-up and shutdown of the compressor and the expansion valve opening parameters of the refrigeration system are optimized, forming an adaptive environmental control method for permafrost refrigeration systems.

[0232] S5: Intelligent control module for designing refrigeration systems;

[0233] An intelligent control system was developed using the Niagara software platform. It incorporates the aforementioned intelligent prediction algorithm for permafrost cooling load and an optimization control algorithm based on cooling load feature identification and bidirectional constraints of cooling capacity and power consumption. This led to the design of the intelligent control module for the permafrost refrigeration system.

[0234] In summary, this invention effectively integrates geological and meteorological information across multiple time scales, enhancing its dynamic adaptability to key influencing factors of permafrost degradation. It is suitable for permafrost foundation scenarios with significant uncertainties and external dynamic disturbances, possessing advantages such as strong generalization ability and high computational efficiency, thereby improving the accuracy of cooling load prediction. Furthermore, based on the cooling load prediction and feature recognition results, the control strategy of the refrigeration system is adjusted in a timely manner, generating control commands.

[0235] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A control method for a permafrost refrigeration system based on cooling load, characterized in that, Includes the following steps: The first step is to predict the cooling load at the construction site on the permafrost foundation. The second step is to construct a refrigeration system control module based on the cooling load, with the following steps: S1: Identify the cold load characteristics of permafrost subgrade based on the cold load prediction results; Cooling load characteristics refer to the magnitude and fluctuation pattern of the cooling load, and the load factor is calculated. The load factor is the ratio of the cooling load at a certain moment or time period of the day to the peak cooling load of the day. Based on the load factor, the day is divided into peak period, flat period and valley period. The load factor of the peak period is ≥80%, the load factor of the flat period is <80% and the load factor of the valley period is <30%. S2: Develop the control strategy for the refrigeration system as follows: When the time is in a low period, the cooling system shuts down; When the time falls within the peak period, the refrigeration system is turned on, and the opening degree of the expansion valve of the refrigeration system is set to 100%. When the time is within a flat period, the refrigeration system is turned on and the opening of the expansion valve is dynamically adjusted. S3: Optimize the operation control scheme of the refrigeration system; S4: Determine the adaptive environmental control method for the refrigeration system; S5: Intelligent control module for designing refrigeration systems; Based on the first and second steps, the design of the intelligent control module for the permafrost refrigeration system was completed using the Niagara software platform.

2. The control method for a permafrost refrigeration system based on cold load according to claim 1, characterized in that: In the first step, the method for predicting the cooling load of permafrost is as follows: S1: Collect multi-source data related to permafrost stability at the construction site of the permafrost foundation, including historical data of meteorological and geological parameters; S2: Data preprocessing based on periodic registration for multi-source data; S3: Establish a permafrost cooling load mechanism model and calculate the permafrost cooling load corresponding to each hour throughout the year; S4: Construct a feature set of permafrost cooling load, meteorological parameters, and geological parameters; The acquired hourly cooling load data for the whole year was preprocessed based on periodic registration to obtain a permafrost cooling load dataset. Meteorological and geological parameters were used as input feature variables, and the correlation between different feature variables and cooling load was analyzed using the Pearson correlation coefficient method. Feature variables with a Pearson correlation coefficient ≥ 0.3 were selected and combined with the cooling load dataset to form a multi-source data feature set for subsequent steps. The formula for calculating the Pearson correlation coefficient P is: （1） In the formula, n is the number of data points in each feature variable; X i Let i be the value of the i-th feature variable; Y represents the average of the characteristic variable data. i This is the data for the i-th cooling load. This represents the average value of the cooling load data; S5: Perform empirical mode decomposition of permafrost cooling load; First, the multi-source data feature set obtained in step S4 is processed by signal processing methods to decompose the original data sequence of permafrost cold load into modal components imf and residual res, as shown in formula (2), to form modal data x(t) composed of imf and res; （2） In the formula; t is time; K is the total number of modal data components; imf i (t) represents the i-th intrinsic mode data component, which is the data sequence after subtracting the average envelope from the original data sequence; res(t) is the residual signal after subtracting the imf from the original data, which can reflect the long-term trend of cooling load. S6: Perform sample entropy preprocessing on permafrost cold load modal data; S7: Construct a neural network model of the cooling load on permafrost; S8: Conduct verification and cyclic training of the results of permafrost cold load prediction; S9: Predicting the cold load on permafrost: Based on the conditions of permafrost foundation sites in different scenarios, input the corresponding feature variable values ​​into the neural network prediction model to obtain the predicted cold load value.

3. The control method for a permafrost refrigeration system based on cold load according to claim 2, characterized in that: In step S1 of the first step, the multi-source data of the permafrost foundation construction site includes meteorological parameters and geological parameters. The meteorological parameters include average annual temperature, average winter temperature, average annual solar irradiance, average annual wind speed, and average annual rainfall. The geological parameters include vegetation coverage, permafrost ice content, permafrost upper limit, permafrost average annual temperature, engineering structure design dimensions, and engineering structure orientation.

4. The control method for a permafrost refrigeration system based on cold load according to claim 3, characterized in that: In step S2 of the first step, the time series data from various data sources obtained in step S1 are aligned to a unified time period, synchronizing and aligning data from different time scales or different time periods.

5. The control method for a permafrost refrigeration system based on cold load according to claim 4, characterized in that: In step S3 of the first step, the DeST software was used to construct a cold load mechanism model for permafrost subgrade and obtain hourly cold load data for the target protection area of ​​permafrost foundation for 8760 hours throughout the year.

6. The control method for a permafrost refrigeration system based on cold load according to claim 5, characterized in that: In step S6 of the first step, the modal data of the cooling load data after empirical mode decomposition are preprocessed with sample entropy. The sample entropy preprocessing steps are as follows: S61. Let N be the total number of data points included in the components imf and res after empirical mode decomposition of the cooling load data, and let X be the calculation vector corresponding to the sample entropy preprocessing tool. X is a calculation vector consisting of m consecutive data points starting from the i-th data point, where m is the dimension of the sliding data window. Then, N-m+1 m-dimensional calculation vectors X can be constructed. m X m The expression is shown in formula (3): （3） Where i represents the data window X m The number, N represents the total number of data points included in the components imf and res after empirical mode decomposition of the cooling load data; m represents the dimension of the sliding data window; u is the time series data of the cooling load. S62, Calculate Vectors and The maximum distance L between them m : （4） In the formula, 0≤k≤m-1, j represents the number of another sliding data window, and k represents the element position index inside the sliding data window; because and Both sliding data windows are m-dimensional computational vectors, so the value of k ranges from 0 to m-1, corresponding to the positions of the 1st to the mth elements in the window; S63, Set a given threshold The calculation satisfies and The number of each is denoted as . and Then the matching rate between the m-dimensional vector and the m+1-dimensional vector is calculated using formulas (5) and (6), respectively. （5） In the formula, Nm is the total number of (m+1)-dimensional sliding data windows, and N-m+1 is the total number of (m)-dimensional sliding data windows. This refers to the number of m-dimensional vectors in the data sequence whose distance to the i-th m-dimensional sliding data window is less than r. （6） In the formula, This refers to the number of m+1 dimensional vectors in the data sequence whose distance to the i-th m+1 dimensional sliding data window is less than r. S64. The expression for sample entropy is: （7） The sample entropy SampEn(m,r) of the cooling load modal data components imf and res is calculated using formula (7), and the overall stability of the cooling load data sequence is judged accordingly. Then, cooling load data with sample entropy SampEn(m,r) fluctuation amplitude conforming to normal rules are selected and proceeded to the next step S7 for neural network model construction.

7. A control method for a permafrost refrigeration system based on cold load according to claim 6, characterized in that: In step S7 of the first step, the meteorological and geological parameters of the cold load data filtered in step S6 and the multi-source data feature set filtered in step S4 are used to train and construct a neural network model of permafrost cold load, and then predict the cold load of permafrost in other different scenarios. The neural network model for permafrost cooling load is a CNN-BiLSTM neural network model. The model structure includes an input layer, a CNN feature extraction module, a BiLSTM modeling layer, and an output layer, all composed of neurons. The CNN feature extraction module includes convolutional layers and pooling layers. The training and construction steps include: S71. A convolutional layer is used to perform feature convolution on meteorological and geological parameters from permafrost cold load data and multi-source data feature sets. The convolution kernel slides across the input data and performs convolution operations, thereby generating a feature matrix of the input data. The formula for calculating this feature matrix is ​​as follows: （8） In the formula: and These represent the number of input data to the convolutional layer and the number of data points in the output feature matrix of the convolutional layer, respectively; a is the kernel size, i.e., the number of data points covered by the convolution operation; q is the number of zero-padding layers, used to maintain the matching degree between the output length and the input length of the feature matrix and avoid the loss of edge features; s is the stride, the step size of each slide of the convolutional kernel; Pooling layers perform max pooling by downsampling the feature matrix output by convolutional layers. This is achieved by using a sliding data window to scan the feature matrix and taking the maximum value within the window as the output data. This reduces the dimensionality of the feature matrix, lowers the computational cost, and enhances the robustness of the output data. S72. The one-dimensional data output from the CNN feature extraction module is then fed into the BiLSTM modeling layer. The BiLSTM modeling layer simultaneously activates two branches of the forward and backward LSTM neural networks: the forward LSTM neural network learns the influence of historical local features on the current moment along the input data time sequence from early to late; the backward LSTM neural network captures the correlation between future local features and the current moment along the input data time sequence from late to early. After the bidirectional LSTM neural network completes the full sequence traversal, the BiLSTM modeling layer concatenates and fuses the forward and backward hidden states corresponding to each input data to generate a high-dimensional feature matrix containing bidirectional temporal dependencies. Subsequently, this high-dimensional feature matrix is ​​input into the fully connected layer, which converts the high-dimensional features into scalar results corresponding to the cold load data through linear mapping, ultimately obtaining the predicted value of the permafrost cold load. S73. Optimize the hyperparameters of the CNN-BiLSTM neural network model using the GA / PSO optimization algorithm; The GA / PSO optimization algorithm combines particle swarm optimization and genetic algorithms to optimize the hyperparameters of the CNN-BiLSTM neural network model. Hyperparameters are the configuration parameters of the CNN-BiLSTM neural network model before training, including the number of neurons in the BiLSTM modeling layer and the learning rate. The steps are as follows: 1) Initialize GA parameters: Initialize hybridization probability , The probability of performing a crossover operation between two different hyperparameter combinations; initial mutation probability. , The probability of a single parameter in a combination of hyperparameters performing a mutation operation; 2) Initialize PSO parameters: Initialize the inertia weight w, which refers to the degree to which the combined scheme continues the previous search direction; initialize the learning factors c1 and c2, where c1 refers to the degree to which the combined scheme learns from its own historical best scheme, and c2 refers to the degree to which the combined scheme learns from the group's global best scheme. 3) Initialize the range of hyperparameters to be optimized: Initialize the number of neurons in the BiLSTM modeling layer, the learning rate, and the size of the CNN convolutional kernel; 4) Treat each hyperparameter combination as an optimization particle, and search for the global optimum and individual optimum particles using the PSO optimization algorithm. Use the reciprocal of the root mean square error (RMSE) of the predicted cooling load as the objective function, fitness. The objective function formula is: （9） Where: n is the number of cooling load samples; This is the predicted cooling load value; The values ​​represent historical cooling load data; the objective function `fitness` represents the fitness value of the optimized particles, with a higher `fitness` value indicating a better combination of hyperparameters. 5) Based on particle fitness, particles are divided into two categories: dominant particles and suboptimal particles. Dominant particles are retained for the next generation. 6) Randomly select two particles from the dominant particles and cross them to update the particle velocity and position; 7) Randomly select two particles from the crossed particles to mutate; 8) Update the fitness value, global optimum, and individual optimum of the mutated particles; 9) Determine if the requirements are met. If the optimal parameters are achieved, output them. Otherwise, return to step 3.

8. A control method for a permafrost refrigeration system based on cold load according to claim 7, characterized in that: In step S8 of the first step, the prediction results of step S7 are tested, and the index for evaluating the accuracy of the prediction is the coefficient of determination R. 2 Mean absolute percentage error E MAPE : （10） In the formula, This represents the average value of the actual cooling load. This represents the true value of the k-th sample; This represents the model's prediction for the k-th sample; （11）。 9. A control method for a permafrost refrigeration system based on cold load according to any one of claims 2-8, characterized in that: In step S3 of the second step, the optimization method for the refrigeration system operation control scheme is as follows: First, specify an objective function with the minimum daily power consumption minP as the objective and the controlled operation state of the refrigeration system and the opening degree of the expansion valve as variables, i.e., formula (12); then, specify two constraint function functions, formula (13) corresponding to the cooling capacity of the refrigeration system and formula (14) corresponding to the power consumption of the refrigeration system; finally, solve formula (12) to obtain the controlled operation state of the refrigeration system that satisfies the constraints. The opening degree φ of the expansion valve is the optimized refrigeration system operation control scheme; S31, Objective Function The primary task of refrigeration system operation control is to meet the cooling load of permafrost roadbeds, with the objective being the minimum daily power consumption minP. The objective function is: （12） In the formula, P is the total daily power consumption of the refrigeration system, kW·h; p is the power consumption rate of the refrigeration system during time period t, kW; φ is the opening degree of the expansion valve, which is adjustable between 0 and 100%; for a fixed model of refrigeration system, the power consumption rate of the refrigeration system depends on the opening degree φ of the expansion valve. t represents a time period, h; This represents the controlled operating state of the refrigeration system during time period t, with 1 for on and 0 for off. S32, Constraints 1) Cooling capacity constraints The relationship between the cooling load of permafrost and the required cooling capacity of the refrigeration system is as follows: （13） In the formula, COP is the coefficient of performance of the refrigeration system, and y(t) is the cooling load of the permafrost during time period t, in kW; 2) Power consumption constraints The power consumption of the cooling system must not exceed the power supply capacity of the photovoltaic system, and the minimum daily power consumption (minP) must not exceed the power supply capacity of the photovoltaic system. （14） In the formula, The maximum daily power supply of the photovoltaic system is expressed in kW•h.

10. The control method for a permafrost refrigeration system based on cold load according to claim 9, characterized in that: In step S4 of the second step, the adaptive environmental control method for the refrigeration system follows these steps: First, the association rule mining method is used to mine the historical data of the refrigeration system operation to find the correspondence between the expansion valve opening parameter of the refrigeration system and the cooling output power of the refrigeration system. Then, based on the objective function and constraints in step S3, from a global optimization perspective, the start-stop state of the refrigeration system and the opening degree of the expansion valve are optimized, thus forming an adaptive environmental control method for permafrost refrigeration systems.

Images