Intelligent prediction method for temperature field of radio telescope structural components

Abstract

The present invention relates to the field of radio telescope structural temperature analysis and discloses a method for intelligently predicting the temperature field of radio telescope structural components. The method comprises the following steps: Step 1: establishing an LSTM-based meteorological data prediction model, where the meteorological data includes wind speed, air temperature, and solar radiation intensity; Step 2: calculating the temperature field of the radio telescope components during the observation period by combining the meteorological data, the rod shadow coefficient, and the rod geometric parameters; Step 3: constructing a rod temperature prediction model based on an LSTM neural network based on the meteorological data, the rod shadow coefficient, the rod geometric parameters, and the temperature field data; and Step 4: obtaining the rod temperature information at the next moment based on the predictions from the rod temperature prediction model. The multivariable input architecture of this solution integrates multi-dimensional influencing factors such as wind speed, radiation, and shadow coefficient, resulting in more accurate temperature prediction results. The prediction response time is short, and the temperature field distribution at the future moment can be directly output.

Description

Technical Field

[0001] The present invention relates to the field of radio telescope structural temperature analysis, and in particular to an intelligent temperature field prediction method for radio telescope structural components. Background Art

[0002] Radio telescopes are essential equipment for observing and studying radio waves from celestial bodies. To observe more distant, fainter, and subtler celestial structures, the aperture of a radio telescope must be increased. However, as the aperture of a radio telescope increases and the detection frequency increases, the requirements for its pointing accuracy become increasingly stringent. Structural deformation caused by sunlight can dramatically reduce antenna accuracy, impacting its observation performance.

[0003] Accurately calculating the temperature distribution of a radio telescope is the premise and key to controlling the temperature effects of the radio telescope. The temperature field of the radio telescope structure is affected by many factors and has complex boundary conditions. The finite element method is currently used to calculate the temperature field of the radio telescope structure, but the calculation efficiency is low and it is difficult to meet the needs of real-time compensation. Summary of the Invention

[0004] The present invention aims to provide an intelligent prediction method for the temperature field of radio telescope structural components to solve the problem of low computational efficiency in calculating the structural stability field using the finite element method.

[0005] To achieve the above-mentioned object, the present invention adopts the following technical solution: a method for intelligently predicting the temperature field of a radio telescope structural component, comprising:

[0006] Step 1: Establish a meteorological data prediction model based on LSTM, where the meteorological data includes wind speed, temperature, and solar radiation intensity;

[0007] Step 2: Calculate the temperature field of the radio telescope components during the observation period by combining meteorological data, the shadow coefficient of the rod and the geometric parameters of the rod;

[0008] Step 3: Based on meteorological data, the shadow coefficient of the pole, the geometric parameters of the pole and the temperature field data, a pole temperature prediction model based on the LSTM neural network is constructed;

[0009] Step 4: Input the predicted wind speed, predicted temperature, predicted solar radiation intensity, shadow coefficient and rod geometric parameters into the rod temperature prediction model to obtain the rod temperature information at the next moment.

[0010] In this scheme, meteorological data and component temperature information of the radio telescope structure over a period of time are collected to construct a training sample set. At the same time, the shading relationship between sunlight and the structure is considered to calculate the shadow coefficient of each component under different sun positions.

[0011] The temperature field during the radio telescope construction observation period (i.e., the observation period of the acquired meteorological data) is further calculated using meteorological data, the shadow coefficient of the rod, the geometric parameters of the rod, and the temperature field data. The temperature is further calculated using the shadow coefficient of the rod to expand the training sample set and obtain a multi-dimensional training sample set. The introduction of sunlight shadow analysis can accurately predict the temperature field distribution of space truss structures under sunlight conditions, which has important engineering significance for structural design, construction and maintenance; LSTM is used to establish meteorological data and temperature construction prediction models respectively, and the LSTM neural network is used as a proxy model for radio telescope temperature field calculation. The temperature distribution of the radio telescope can be calculated and predicted by fusing multi-dimensional meteorological data and time-series temperature data; in the rod temperature prediction model, the input data includes wind speed, air temperature, solar radiation intensity, shadow coefficient and rod geometric parameters, and the output data is the rod temperature information at the next moment. This scheme discretizes the space truss structure into a single rod, and then obtains the temperature information of a single component through a neural network algorithm. Finally, the corresponding integration method is used to complete the prediction of the sunlight temperature field of the entire space truss structure. This method not only simplifies the calculation process of the shadow coefficient and improves the calculation efficiency, but also facilitates the implementation of relevant real-time compensation measures for the construction structure according to the prediction results; it also provides a unified data format for the construction and selection of subsequent sample libraries.

[0012] Preferably, in order to simplify the training process, a space truss structure of the same form as the radio telescope back frame structure is adopted, the structural components are divided into the upper chord area, the lower chord area and the web area, and the solar position and the rod attitude are coupled to establish a shadow coefficient database.

[0013] The spatial location of the bottom chord at the very bottom of the structural system creates a significantly complex solar environment. As the sun's altitude changes throughout the year, the top chord creates a continuous shading zone, forming a composite shading system with the staggered web members. This multi-layered shading effect results in significant temporal and spatial variations in the solar radiation received by the bottom chord surface. The top chord region, due to its spatial position at the very top of the structure and direct exposure to solar radiation, exhibits a high shading coefficient. However, the shading coefficient is significantly reduced in the bottom chord and web lower chord regions, given their relatively low position and shading by upper structural members. Zonal statistics provide the data foundation for training the neural network model. By incorporating the geometric characteristics of the spatial truss structure, a statistical analysis of the lighting conditions at different locations is performed to extract the shading patterns of each member at different time points and to determine the appropriate shading coefficients. This provides accurate input parameters for subsequent temperature field calculations, ensuring the model's suitability for actual project requirements.

[0014] Preferably, the establishment of the LSTM-based meteorological data prediction model in step 1 specifically includes:

[0015] Step 11: Obtain historical meteorological data and use the sliding window technology to restructure the time series data of each type of meteorological data to construct a training sample set. The historical observation time parameter L is introduced, and the meteorological data values ​​of L consecutive historical moments are used as input feature vectors, and the measurement values ​​at subsequent moments are used as prediction targets.

[0016] Step 12: Divide the dataset into training, validation, and test sets, and build a meteorological data prediction model based on LSTM.

[0017] The sliding window technology is used to restructure the time series data and construct a training sample set suitable for deep learning models. The original hourly wind speed data is divided into 365 independent time series segments according to the natural day, and each segment contains 24 consecutive observations. Based on the time-dependent modeling requirements, the historical observation duration parameter L is introduced. The wind speed observation values ​​of L consecutive historical moments are used as input feature vectors, and the measured values ​​at subsequent moments are used as prediction targets, thereby generating input-output mapping pairs. For a single-day data segment, A set of valid samples is constructed and spliced ​​together in time series to form a global training dataset, strictly maintaining temporal coherence between samples. The LSTM model can effectively capture the temporal correlation between data and demonstrates superior performance in wind speed forecasting.

[0018] Preferably, the optimized value of the historical observation duration L in step 11 is determined by Bayesian hyperparameter search. Bayesian hyperparameter optimization is an automated parameter adjustment method based on Bayes' theorem. By constructing a probabilistic model of the objective function (model performance), it intelligently selects the parameter combination most likely to improve the effect, significantly reducing the computing resources required for parameter adjustment.

[0019] Preferably, the dataset is segmented in a time series ratio of 6:2:2 in step 12, with the validation set starting 30 days after the training set ending, and a 15-day gap between the test and validation sets. A strict time barrier strategy is implemented during the segmentation process: the validation set starts 30 days after the training set ending to eliminate overlapping interference from seasonal cycle features; the test and validation sets are separated by 15 days, serving as independent time domains for the final model evaluation. This progressive segmentation approach effectively captures the seasonal evolution of meteorological elements while ensuring the statistical independence of each data subset.

[0020] Preferably, the number of model training cycles in step 1 and step 3 is dynamically adjusted using an early stopping strategy. The early stopping strategy is an effective method for preventing model overfitting. Its principle is to continuously monitor the validation set loss during model training. If the validation set loss does not decrease for multiple consecutive times (controlled by the Patience parameter), training is terminated early, and the model parameters at the time of the last loss decrease are output.

[0021] Preferably, the geometric parameters of the rod include a rod azimuth and a rod surface inclination. The rod azimuth indicates the orientation of the rod itself, and the rod surface inclination indicates the relationship between the rod surface unit and the incident direction of the sun.

[0022] Preferably, before constructing the model in step 3, the meteorological data, the shadow coefficient of the rod and the geometric parameters of the rod are normalized. Input parameters such as wind speed, temperature, solar radiation intensity, shadow coefficient and geometric parameters of the rod have significant differences in dimensions and numerical ranges. For example, the unit of wind speed is m / s, the unit of temperature is ℃, and the shadow coefficient is a dimensionless value. This difference in dimensions and numerical ranges will lead to inconsistent sensitivity of the model to different features during training, thereby affecting the convergence speed and prediction accuracy of the model. In this solution, the data is normalized and mapped to a unified scale range, thereby eliminating the impact of dimension and numerical range differences on model training, ensuring that the model can treat all types of input and output data equally, and improving the learning efficiency of the model.

[0023] Preferably, in step 2, a finite element model is established to calculate the temperature field of the radio telescope components during the observation period, and a pre-written macro operation script is used to dynamically adjust the wind speed, air temperature, and solar radiation intensity parameters to complete the update of the finite element model, enter the temperature field simulation calculation, and use a pre-written batch processing script to batch process data and extract the temperature field data of the components. Through the combination of macro operations and batch processing files, the entire calculation process from parameter input to result extraction is realized, significantly improving the efficiency and reliability of the temperature field calculation.

[0024] Preferably, the historical observation time L of the rod temperature prediction model is 4 hours, the number of LSTM network layers is 2, the optimizer is Adam, the number of training times adopts the Earlystopping strategy, the Dropout is 0.1, the activation function is SELU, the Batch-size is 64, the loss function is MSEloss, and the learning rate is 0.002.

[0025] Preferably, in step 2, the overall geometric modeling of the space truss structure of the radio telescope is performed to extract the geometric parameter information of the rods; a single rod is simulated to extract the surface unit information to obtain the shadow coefficient and geometric parameters of the rod.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] 1. An efficient and reliable meteorological data preprocessing system uses authoritative NASA meteorological data sources to ensure data integrity and scientificity. It uses sliding window technology to effectively extract time series features and preserve the temporal dependence of meteorological parameters. The LSTM network design conforms to the temporal characteristics of meteorological data, and the historical observation duration parameter L enhances the ability to model temporal correlations.

[0028] 2. Innovative temperature field simulation system: The batch processing system realizes the full process calculation of "data-simulation-results". The daily sequence number drive mechanism automatically matches meteorological parameters to ensure temporal and spatial consistency. The parallel computing capability supports large-scale temperature field simulation, which greatly improves efficiency. It can quickly and efficiently obtain data samples covering multiple dimensions. Only the required observation period and the main working posture need to be input. The system will automatically obtain the meteorological data (temperature, wind speed and solar radiation intensity data) within the corresponding period and perform automatic thermal analysis for different working postures, avoiding the high cost and error risks of manual operation.

[0029] 3. Intelligent prediction model: Grid search method optimizes LSTM hyperparameters to determine the optimal historical observation duration and network depth. The multivariable input architecture integrates multi-dimensional influencing factors such as wind speed, radiation, and shadow coefficient. The prediction response time is short and the temperature field distribution at the future moment can be directly output. The temperature of the rod at the next moment can be predicted, and the temperature field of the structure at the future moment can be constructed. The temperature compensation of the components can be performed in a timely manner, providing a reference for precision control and guaranteeing future observation missions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention.

[0031] Figure 2 Schematic diagram of calculating pole occlusion using a depth buffer according to an embodiment of the present invention.

[0032] Figure 3 Schematic diagram of the flow of the temperature field calculation method according to an embodiment of the present invention.

[0033] Figure 4 Schematic diagram of wind speed data processing according to an embodiment of the present invention.

[0034] Figure 5 This is a graph of the average RMSE of the model under different combinations of historical observation duration and number of LSTM layers in an embodiment of the present invention.

[0035] Figure 6 This is the average R2 graph of the model under different combinations of historical observation time and number of LSTM layers in an embodiment of the present invention.

[0036] Figure 7This is a comparison chart of the actual and predicted wind speed values ​​according to an embodiment of the present invention.

[0037] Figure 8 This is a relative error diagram between the actual and predicted wind speed values ​​according to an embodiment of the present invention.

[0038] Figure 9 This is a diagram showing the relative error range between the actual and predicted wind speed values ​​according to an embodiment of the present invention.

[0039] Figure 10 This is the Ansys simulation result diagram of the upper chord member of an embodiment of the present invention.

[0040] Figure 11 This is the Ansys simulation result diagram of the lower chord member of an embodiment of the present invention.

[0041] Figure 12 This is the Ansys simulation result diagram of the web member of an embodiment of the present invention.

[0042] Figure 13 This is a schematic diagram of sample construction using the rod unit temperature according to an embodiment of the present invention.

[0043] Figure 14 This figure shows the comparison of the prediction results of the LSTM neural network model of the upper chord according to the embodiment of the present invention.

[0044] Figure 15 This figure shows the comparison of the prediction results of the LSTM neural network model of the lower chord according to the embodiment of the present invention.

[0045] Figure 16 This figure shows the comparison of the prediction results of the LSTM neural network model of the belly rod of the embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following is further described in detail through specific implementation methods:

[0047] Example:

[0048] The present invention provides a specific implementation method for intelligent prediction of temperature field of radio telescope structural components, such as Figure 1 Shown, including:

[0049] Step 11: Acquisition and preprocessing of meteorological data to obtain meteorological data on wind speed, temperature and solar radiation intensity;

[0050] The meteorological data on wind speed, temperature and solar radiation intensity in this plan are all derived from the NASA meteorological database. This plan selected the temperature (two meters above the ground), wind speed (ten meters above the ground) and ground solar radiation intensity in Weihai area in 2023 as the research data set, and the sampling frequency of each meteorological data is 1 hour.

[0051] This solution uses sliding window technology to restructure time series data and construct a training sample set suitable for deep learning models. The original hourly wind speed data is divided into 365 independent time series segments according to natural days, such as Figure 4 As shown, each segment contains 24 consecutive observations. Based on the requirements of time-dependent modeling, a historical observation duration parameter, L, is introduced. Wind speed observations at L consecutive historical moments are used as input feature vectors, and the measured values ​​at subsequent moments are used as prediction targets. This generates input-output mapping pairs. For a single-day data segment, 24-L+1 sets of valid samples are generated. These samples are then spliced ​​together to form a global training dataset, strictly maintaining temporal coherence between samples.

[0052] The initial L-1 time points of the first day's data serve only as historical information carriers and do not constitute independent training samples. The resulting sample database contains a multidimensional tensor structure, providing the LSTM network with an input format that meets the requirements of time series modeling. The optimal value of the historical observation period L is determined through Bayesian hyperparameter search.

[0053] Bayesian Hyperparameter Optimization is an automated parameter tuning method based on Bayes' theorem. By constructing a probabilistic model of the objective function (model performance), it intelligently selects the parameter combination most likely to improve the effect, significantly reducing the computing resources required for parameter tuning.

[0054] Step 12: Divide the dataset into training, validation, and test sets; and build a meteorological data prediction model based on LSTM.

[0055] During deep learning model development, the dataset partitioning strategy directly impacts the reliability and generalization capabilities of model performance evaluation. This embodiment strictly adheres to temporal patterns in constructing training, validation, and test sets. In this solution, the training set (data from January to August 2023) is used for gradient updates of model parameters, the validation set (data from September to October 2023) is used for hyperparameter optimization and early stopping, and the test set (data from November to December 2023) is completely isolated from the training process and used to evaluate the model's generalization performance in unknown time domains. This partitioning method effectively avoids the temporal leakage problem caused by traditional random partitioning, ensuring that model evaluation meets the timing constraints of actual prediction scenarios.

[0056] The dataset was split into time series with a ratio of 6:2:2. A strict time barrier strategy was implemented during the splitting process: the validation set started 30 days after the training set to eliminate overlap in seasonal cyclical features. A 15-day interval separated the test and validation sets, serving as independent time domains for the final model evaluation. This progressive splitting approach effectively captures the seasonal evolution of meteorological elements while ensuring statistical independence between the data subsets. All data preprocessing operations (including normalization and feature engineering) were performed on the training set before being applied to the validation and test sets via parameter transfer.

[0057] LSTM (Long Short-Term Memory) is a recurrent neural network (RNN) variant specifically designed for processing sequential data. It addresses the long-term dependency issues of traditional RNNs by introducing a gating mechanism. Through its unique gating mechanism (forget gate, input gate, and output gate), LSTM selectively accepts historical information and automatically adjusts its transmission. In the field of meteorological time series forecasting, changes in wind speed, temperature, and solar radiation intensity typically exhibit time series characteristics, with certain periodicities and trends. LSTM, through its unique gating mechanism, can remember and utilize historical information over long time spans, not just the current state. Furthermore, forecasts of wind speed, temperature, and solar radiation intensity are not based solely on input data at a single moment; rather, they typically use a series of historical data to predict future values.

[0058] Among them, in the process of building the LSTM model of this solution, the hyperparameters that need to be set can be divided into the following three categories:

[0059] (1) Parameters related to the sample set: including the type of input features, the type of output features, and the length of historical observations L;

[0060] (2) Parameters related to the network model structure: including the number of LSTM network layers and the number of hidden layer neurons;

[0061] (3) Parameters related to the model training process: including activation function, batch size, loss function, learning rate, optimizer type, number of training times, and dropout rate.

[0062] This example uses wind speed as an example to illustrate model hyperparameter settings. LSTM networks typically require fewer than three layers, so this method uses values ​​from one, two, and three layers. Nine candidate model configurations are constructed through cross-combination, and a grid search method is used to find the optimal parameter combination.

[0063] Based on the above steps, the present invention obtains the root mean square error (RMSE) and determination coefficient (R) of the wind speed prediction model under different historical observation time.2 ), the specific results are as follows Figure 5 and Figure 6 As shown. The analysis results show that the optimal historical observation time is 4 hours and the number of LSTM layers is 1. On this basis, the optimal values ​​of hyperparameters such as the number of hidden layer neurons, batch size, learning rate, and dropout rate are determined through the Bayesian optimization method. In addition, the activation function uses the SeLU function with self-normalization characteristics, the optimizer uses the Adam optimizer, and the number of training times is dynamically adjusted using the early stopping strategy. The early stopping strategy is an effective method to prevent model overfitting. Its principle is to continuously monitor the loss value of the validation set during the model training process. When the validation set loss value does not decrease for multiple consecutive times (controlled by the Patience parameter), the training will be terminated early and the model parameters at the time of the last loss decrease will be output. In the present invention, the number of training steps (epoch) is set to 2000 and the number of early stopping steps (Patience) is set to 50. The loss function uses the mean square error loss function (MSEloss) to further optimize the prediction performance of the model.

[0064] Table 1 details the specific hyperparameter settings for the network model. The historical observation period L is determined by balancing the ability to capture long-term patterns with the ability to respond to short-term fluctuations. An adaptive dropout mechanism is employed, with an initial dropout rate of 0.5 to enhance the model's exploration capabilities, and then dynamically adjusted to 0.2 over the training cycle to maintain effective feature transfer. The learning rate is set to 0.05, combined with the momentum adaptation of the Adam optimizer to achieve stable convergence. Properly configuring these hyperparameters can significantly improve the model's predictive performance.

[0065] Table 1 Parameter settings of the wind speed prediction LSTM network model

[0066] Hyperparameter Types Setting situation Input feature type Local historical wind speed Output feature type Wind speed at the next moment Historical observation duration 4 hours Number of LSTM network layers 1 Activation Function SELU Batch-size 16 Loss Function MSEloss Learning rate 0.05 Optimizer Adam Number of training sessions Early stopping Dropout 0.2

[0067] Based on the above parameter settings, the present invention predicts and analyzes the typical summer wind speed sequence in Weihai from July 10 to 20, 2024. The results are as follows: Figure 7 、 Figure 8 and Figure 9As shown in the figure, the LSTM neural network model accurately captures the characteristics of the alternating sea and land breezes and the fluctuation patterns within the daily cycle. The coefficient of determination between the wind speed forecast curve and the actual NASA monitoring value reaches 0.97. A quantitative evaluation using a historical observation period of 4 hours and a single LSTM layer shows that the model's prediction root mean square error (RMSE) is only 0.36 m / s, the mean absolute error (MAE) is 0.27 m / s, and the maximum relative error is 1.5 m / s. Furthermore, error statistical analysis shows that 84.2% of the wind speed forecasts have a relative error of less than 10% from the actual values, further verifying the model's high accuracy and reliability.

[0068] In this solution, the LSTM models for temperature and solar radiation intensity are constructed using the same approach as the wind speed model. The input and output sample construction methods for temperature data are the same as for the wind speed model. The hyperparameter settings for the temperature prediction model are the same as for the wind speed prediction model, except that the input feature type is local historical temperature. The input and output sample construction methods for solar radiation intensity data are the same as for the wind speed model. The hyperparameter settings for the solar radiation intensity prediction model are the same as for the wind speed prediction model, except that the input feature type is local historical solar radiation intensity.

[0069] Step 2: Calculate the temperature field of the radio telescope components during the observation period by combining meteorological data, the shadow coefficient of the rod and the geometric parameters of the rod.

[0070] The geometric parameters of a rod include its azimuth and surface inclination. The azimuth indicates the rod's orientation, while the surface inclination represents the relationship between the rod's surface unit and the solar incidence direction. In this step, the overall geometry of the radio telescope's spatial truss structure is modeled and its geometric parameters extracted. A single rod is simulated, its surface unit information extracted, and its shadow coefficient and geometric parameters are obtained.

[0071] Current temperature field simulations fail to fully account for the dynamic changes in solar shadow distribution. This proposal introduces solar shadow analysis to accurately predict the temperature field distribution of spatial truss structures under solar conditions, which has important engineering implications for structural design, construction, and maintenance. This proposal draws on the depth buffer algorithm (Z-Buffer) in computer graphics to develop a solar shadow analysis algorithm for spatial truss structures and calculate the solar shadow coefficient for each component. The depth buffer algorithm simulates the occlusion relationship between sunlight and the structure, calculating the shadow coefficient for each component at different sun positions, thereby providing precise input parameters for temperature field analysis.

[0072] Figure 2A schematic diagram of calculating member occlusion using a depth buffer is shown. This method, through visibility determination and occlusion culling logic in the depth buffer, enables rapid calculation of solar radiation exposure in complex structures. Compared to traditional ray tracing or Boolean determination methods, this method offers higher computational efficiency and scalability for large structural systems, providing effective technical support for solar radiation analysis of spatial trusses. The distance between different member elements and the sun is visualized using color coding: the distance between member 1 and the sun is represented in red, while the distance between member 2 and the sun is represented in cyan. The results clearly show that, because member 1 is closer to the sun, all pixels covered by its projected outline in the depth buffer appear red. However, since member 2 is partially occluded by member 1, the occluded pixels appear red in the depth buffer, leaving only the unobstructed area in cyan. By counting the member numbers corresponding to each pixel in the depth buffer, the actual visible projected area of ​​each member element in the direction of solar radiation can be obtained. In other words, the number of pixels that appear in the depth buffer of each unit directly reflects its effective area illuminated by solar radiation.

[0073] The relevant calculation method is shown in formula (1)

[0074] (1)

[0075] Where n is the number of pixels in the unit, Pixel pitch (m).

[0076] In order to measure the shadow size of each component of the space truss structure under sunlight, the shadow coefficient of the component is defined as the ratio of the surface area of ​​the component actually receiving direct solar radiation to the total area of ​​its outline on the projection plane on a virtual projection plane perpendicular to the direction of solar incidence. The formula is shown in 2

[0077] (2)

[0078] Where, The effective area of ​​the component exposed to solar radiation in the projection surface, The total area of ​​the component in the projection surface. This study uses the deep buffer algorithm and combines the data of the sun's position change in Weihai area throughout the year to carry out a detailed calculation of the sunlight shadow coefficient of each part of the structure, and accurately calculates the shadow distribution of the space truss structure under sunlight conditions.

[0079] Among them, this scheme divides the space truss structure of the same form as the radio telescope back frame structure into the upper chord area, the lower chord area and the web area, and couples the solar position and the rod attitude to establish a shadow coefficient database to facilitate the rapid acquisition of the shadow coefficient.

[0080] In this solution, the shadow coefficient database was built by coupling the sun's position with the member's attitude. The database included 864 typical operating conditions (the 5th, 15th, and 25th of each month were selected as typical operating conditions throughout the year) and 324 member attitude geometric parameters, including azimuth attitude (18 conditions) and pitch attitude (18 conditions). After coupling, a total of 279,936 conditions were identified. This approach relied on the ANSYS-PyMAPDL co-simulation platform for batch calculations, involving massive data processing. This ultimately resulted in the creation of a shadow coefficient feature database.

[0081] Figure 3 The flow chart of the full-process calculation system based on macro operations is presented. The macro operation function provided by PyMAPDL is an efficient solution that can record, edit, and replay all operations in interactive Ansys, and manually edit the operation log through simple programming. Macro operations usually include macro recording and macro playback, which are used to record Ansys operation records and automatically play Ansys operation records respectively. Therefore, operations such as "parameter input to finite element software", "temperature field simulation calculation", and "temperature result output" can be recorded in the macro recording file .ipynb through macro recording. Then, by modifying the meteorological data that needs to be modified in the macro recording file, macro playback can realize the calculation of the entire operation process.

[0082] Macro operation mainly includes two steps: macro recording and macro playback:

[0083] (1) Macro recording: Record the operations such as “parameter input into finite element software”, “temperature field simulation calculation” and “temperature result output” in a macro recording file (.ipynb).

[0084] (2) Macro playback: By adjusting the meteorological data that needs to be modified in the macro recording file, the recorded operations are automatically played to achieve full-process calculation.

[0085] A batch script (typically with a .bat extension) is an executable file consisting of a series of command lines. It uses flow control logic (such as while and for) and conditional control mechanisms (such as if) to control the execution flow of commands and call other programs. In this application, batch scripts are used to automatically call Ansys and macro recording files. First, the software environment is automatically loaded by calling the Ansys Mechanical APDL startup interface; then, pre-written macro operation scripts are scheduled in an orderly manner.

[0086] Specifically, after the system activates the Ansys kernel service through the "launch_mapdl" command, it immediately executes the parameter reading macro file "Parameter.ipynb" - this file completes the update of the finite element model by dynamically adjusting thermal boundary conditions such as solar radiation intensity and ambient wind speed. After the thermal design parameters are modified, the system enters the temperature field simulation calculation stage. Since the finite element calculation process is relatively time-consuming, the system determines whether the temperature field calculation is completed by monitoring the running status of the process. Once the calculation is completed, the system automatically enters the next stage of operation, the extraction of temperature field result data. The extraction of temperature field result data is also achieved through macro recording, and its process is similar to parameter modification.

[0087] This solution realizes the whole process calculation from parameter input to result extraction by combining Python scripts and batch files, which significantly improves the efficiency and reliability of temperature field calculation. Figure 10 、 Figure 11 、 Figure 12 The calculation results of the upper chord, lower chord and web members on July 15 are shown. Figure 10 、 Figure 11 and Figure 12 In the table, step represents a load step, sub represents a substep within a load step, TIME represents time, temp(avg) represents the average temperature, RSYS represents the result coordinate system, smn represents the minimum result value, and smx represents the maximum result value. The horizontal axis below represents the element temperature in °C.

[0088] This step calculates the temperature field of radio telescope components during the observation period by building a finite element model. A pre-programmed macro script dynamically adjusts wind speed, air temperature, and solar radiation intensity parameters. This updates the finite element model and then proceeds to the temperature field simulation. A pre-programmed batch script processes the data and extracts component temperature field data. This combination of macros and batch files enables a complete calculation process, from parameter input to result extraction, significantly improving the efficiency and reliability of temperature field calculations.

[0089] Step 3: Based on meteorological data, the shadow coefficient of the rod, the geometric parameters of the rod and the temperature field data, a rod temperature prediction model based on the LSTM neural network is constructed.

[0090] Taking the radio telescope as the research object, each unit in the structure has a unified structural form. All rods with independent spatial postures in the structure are modeled. In terms of geometric parameters, the angle between the rod and the ground ranges from -90° to 90°, and the azimuth angle of the rod ranges from -90° to 90°. A single rod is divided into 48 shell units in the circumferential direction and used as the basic analysis unit. In order to achieve accurate prediction of the rod temperature, the present invention uses 24 hours a day as the time series unit and calculates the rod unit temperature for the whole year of 2023 as a sample. An input sample database containing characteristic points such as thermal boundary parameters (wind speed, air temperature, solar radiation intensity), geometric features (rod azimuth, surface inclination), and shadow coefficient is constructed, and the rod unit temperature at the next moment is used as the output sample. The specific construction method is as follows Figure 13 Compared with the wind speed and air temperature prediction models, the input characteristics of the bar unit temperature prediction model are more complex. It includes not only meteorological data but also the geometric information of the bar unit and the sunshine conditions, thus more comprehensively reflecting the variation pattern of the bar unit temperature.

[0091] Due to the high complexity of the rod temperature prediction model, this paper uses a grid search method to optimize the historical observation duration and the number of LSTM neural network layers to find the optimal hyperparameter combination. Through systematic parameter tuning, the hyperparameter settings for the LSTM neural network model for rod temperature prediction were ultimately determined, as shown in Table 2.

[0092] Table 2 Parameter settings of the LSTM network model for rod temperature prediction

[0093] Hyperparameter Types Setting situation Input feature type Wind speed, temperature, pole information Output feature type Rod temperature at the next moment Historical observation duration 4 hours Number of LSTM network layers 2 Optimizer Adam Number of training sessions Early stopping Dropout 0.1 Activation Function SELU Batch-size 64 Loss Function MSEloss Learning rate 0.002

[0094] In order to verify the prediction performance of the model, the present invention selected the meteorological data of July 15 and predicted the sunshine temperature changes of the components in the back frame structure from 8:00 to 17:00. Figure 14 、 Figure 15 and Figure 16 Comparisons of the prediction results from the LSTM neural network model for the top chord, bottom chord, and web members are presented. The mean absolute error (MAE) for the top chord is 0.515°C, the MAE for the bottom chord is 0.985°C, and the MAE for the bottom chord is 1.742°C, further confirming the accuracy of the model. Comparative analysis of the predicted curves with simulated data demonstrates that the model not only accurately captures the overall trend of temperature variation but also exhibits good fit at temperature extremes and inflection points, without exhibiting significant systematic deviations or outliers.

[0095] Before building the model in this step, the meteorological data, the shadow coefficient of the pole, and the geometric parameters of the pole are normalized. Input parameters such as wind speed, temperature, solar radiation intensity, shadow coefficient, and geometric parameters of the pole have significant differences in dimensions and numerical ranges. For example, the unit of wind speed is m / s, the unit of temperature is ℃, and the shadow coefficient is a dimensionless value. This difference in dimensions and numerical ranges will lead to inconsistent sensitivity of the model to different features during training, thereby affecting the convergence speed and prediction accuracy of the model. In this solution, the data is normalized and mapped to a unified scale range, thereby eliminating the impact of dimension and numerical range differences on model training, ensuring that the model can treat all types of input and output data equally, and improving the learning efficiency of the model.

[0096] In this scheme, the historical observation time L of the rod temperature prediction model is 4h, the number of LSTM network layers is 2, the optimizer is Adam, the number of training times adopts the Earlystopping strategy, the Dropout is 0.1, the activation function is SELU, the Batch-size is 64, the loss function is MSEloss, and the learning rate is 0.002.

[0097] Step 4: Input the predicted wind speed, predicted temperature, predicted solar radiation intensity, shadow coefficient and rod geometric parameters into the rod temperature prediction model to obtain the rod temperature information at the next moment.

[0098] The number of model training iterations in step 1 or step 3 is dynamically adjusted using the early stopping strategy. The early stopping strategy is an effective method for preventing model overfitting. It works by continuously monitoring the validation set loss during model training. If the validation set loss fails to decrease for multiple consecutive times (controlled by the Patience parameter), training is terminated early, and the model parameters at the time of the last loss decrease are output.

[0099] This solution addresses the issues of traditional temperature field simulation methods, which often lack empirical consideration of thermal boundary parameters and the time-varying characteristics of solar radiation and shadows. By using a long short-term memory (LSTM) network to establish a time-series prediction model for meteorological parameters, this model dynamically predicts temperature, wind speed, and solar radiation intensity, improving the accuracy of thermal boundary condition inputs. A solar radiation and shadow analysis algorithm, combined with a deep buffering algorithm (Z-Buffer), is developed to accurately calculate the solar radiation and shadow coefficients of components. Statistical analysis provides reliable data support for deriving shadow coefficients at construction sites. Using a single member as the basic unit, a proxy model for numerical simulation of the solar radiation temperature field based on an LSTM neural network is constructed. This model uses member geometry, thermal boundary parameters (temperature, wind speed, and solar radiation intensity), and shadow coefficients as input features to rapidly predict the temperature of a single member. An integrated approach is then used to reconstruct the overall structural temperature field. The accuracy of the proxy model is verified using ANSYS numerical simulations.

[0100] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are not described in detail here. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the technical solution of the present invention. In the present invention, unless otherwise clearly specified and limited, the terms "install", "connect", "connect", "fix" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected, or indirectly connected through an intermediate medium, or it can be the internal connection of two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

Claims

1. An intelligent prediction method for the temperature field of radio telescope structural components, characterized in that: include: Step 1: Establish a meteorological data prediction model based on LSTM, where the meteorological data includes wind speed, temperature, and solar radiation intensity; Step 2: Calculate the temperature field of the radio telescope components during the observation period by combining meteorological data, the shadow coefficient of the rod and the geometric parameters of the rod; Step 3: Based on meteorological data, the shadow coefficient of the pole, the geometric parameters of the pole and the temperature field data, a pole temperature prediction model based on the LSTM neural network is constructed; Step 4: Input the predicted wind speed, predicted temperature, predicted solar radiation intensity, shadow coefficient, and rod geometric parameters into the rod temperature prediction model to obtain the rod temperature information at the next moment; In step 2, a finite element model is established to calculate the temperature field of the radio telescope components during the observation period, and a pre-written macro operation script is used to dynamically adjust the wind speed, air temperature and solar radiation intensity parameters to complete the update of the finite element model, enter the temperature field simulation calculation, and use a pre-written batch processing script to batch process data and extract the temperature field data of the components; In step 3, the grid search method is used to optimize the historical observation time and the number of LSTM neural network layers to determine the optimal hyperparameter combination.

2. The method for intelligently predicting the temperature field of a radio telescope structural component according to claim 1, characterized in that: The establishment of the LSTM-based meteorological data prediction model in step 1 specifically includes: Step 11: Obtain historical meteorological data, and use sliding window technology to structure the time series data of each type of meteorological data, construct a training sample set, introduce the historical observation time parameter L, and use the meteorological data values ​​of L consecutive historical moments as input feature vectors, and the measurement values ​​at subsequent moments as prediction targets; Step 12: Divide the training set, validation set and test set, and construct a LSTM-based meteorological data prediction model.

3. The method for intelligently predicting the temperature field of radio telescope structural components according to claim 2, characterized in that: The optimized value of the historical observation time L in step 11 is determined by Bayesian hyperparameter search.

4. The method for intelligently predicting the temperature field of radio telescope structural components according to claim 2, characterized in that: In step 12, the time series segmentation ratio of the data set is set to 6:2:2, the start time point of the validation set is 30 days later than the end time of the training set, and the interval between the test set and the validation set is 15 days.

5. The method for intelligently predicting the temperature field of radio telescope structural components according to claim 1, characterized in that: The number of model training times in step 1 and the number of model training times in step 3 are both dynamically adjusted using the early stopping strategy.

6. The method for intelligently predicting the temperature field of radio telescope structural components according to claim 1, characterized in that: The geometric parameters of the rod include the rod azimuth angle and the rod surface inclination angle.

7. The method for intelligently predicting the temperature field of radio telescope structural components according to claim 1, characterized in that: Before constructing the model in step 3, the meteorological data, the shadow coefficient of the rod and the geometric parameters of the rod are normalized.

8. The method for intelligently predicting the temperature field of radio telescope structural components according to claim 1, characterized in that: The historical observation time L of the rod temperature prediction model is 4 hours, the number of LSTM network layers is 2, the optimizer is Adam, the training times adopt the Earlystopping strategy, the Dropout is 0.1, the activation function is SELU, the Batch-size is 64, the loss function is MSEloss, and the learning rate is 0.

002.

9. The method for intelligently predicting the temperature field of a radio telescope structural component according to claim 1, characterized in that: In step 2, the overall geometric modeling of the space truss structure of the radio telescope is performed to extract the geometric parameter information of the rods; A single rod is simulated, surface unit information is extracted, and the shadow coefficient and geometric parameters of the rod are obtained.

Images