A method and system for hydration heat cooling of high-cold high-altitude gravity dam mass concrete
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA RAILWAY CONSTR BRIDGE ENG BUREAU GRP CO LTD
- Filing Date
- 2025-11-07
- Publication Date
- 2026-07-14
Smart Images
Figure CN121451597B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of large-volume concrete construction technology for high-altitude and cold-climate water conservancy and hydropower projects, and in particular to a method and system for cooling the hydration heat of large-volume concrete in high-altitude and cold-climate gravity dams. Background Technology
[0002] In traditional large-volume concrete structure construction, temperature control and crack prevention technologies still face significant technical bottlenecks. Currently, there are two main processes, each with the following technical shortcomings: 1. Surface watering: While forced repeated watering can cool the surface, its heat conduction lag effect results in a temperature response delay of 12-24 hours, making real-time temperature field control difficult. This lag can easily create a temperature gradient exceeding 15℃ / m within the structure, inducing surface thermal stress concentration. 2. Aggregate pre-cooling technology: Although it can reduce the initial pouring temperature by 8-12℃, it is limited by the hydration heat release kinetics of concrete, only delaying the temperature rise curve by about 6-8 hours. Furthermore, it cannot effectively control the maximum temperature difference between the core area and the surface, causing temperature stress to exceed the critical threshold for concrete creep. These technical shortcomings collectively result in an early cracking rate of 38%-45% in large-volume concrete structures in current engineering practice, severely impacting structural durability. Summary of the Invention
[0003] This invention aims to at least solve the technical problem of high early cracking rate in large-volume concrete structures in the prior art, and innovatively proposes a method for cooling the hydration heat of large-volume concrete in high-altitude and cold-weather gravity dams.
[0004] To achieve the above-mentioned objectives of this invention, this invention provides a method for cooling the hydration heat of large-volume concrete in high-altitude and cold-climate gravity dams, the method comprising:
[0005] S1. An internal cooling pipe network is installed inside the concrete, an external cooling pipe network is installed on the outer surface of the concrete, and multi-dimensional sensing data is collected on the concrete.
[0006] S2. Preprocess the multi-dimensional perception data and extract features to obtain structured feature data;
[0007] S3. Construct a dual-network structure model based on the structured feature data, and use the evaluation network in the dual-network structure model to predict the distribution of hydration heat and temperature gradient changes inside the concrete in real time according to the structured feature data to obtain the prediction results;
[0008] S4. Based on the prediction results and the reinforcement learning reward function, an adaptive cooling strategy is generated using the control network in the dual-network structure model.
[0009] S5. Based on the adaptive cooling strategy, control the valves and pump station power of the internal cooling pipe network and the external cooling pipe network to dynamically adjust the cooling process;
[0010] S6. Continuously collect the multi-dimensional perception data, combine it with the output of the dual-network structure model for real-time feedback optimization, and update the model parameters of the dual-network structure model.
[0011] In another aspect, the present invention also provides a hydration heat cooling system for large-volume concrete in high-altitude and cold-climate gravity dams, the system comprising:
[0012] processor;
[0013] Memory used to store processor-executable instructions;
[0014] The processor is configured to implement the hydration heat cooling method for large-volume concrete in high-altitude and cold-climate gravity dams when executing the executable instructions.
[0015] The beneficial effects of this invention are as follows: By collecting multi-dimensional sensing data in real time and constructing a dual-network structure model, combined with the real-time prediction of the evaluation network and the adaptive cooling strategy of the control network, this invention achieves precise dynamic control of the hydration heat distribution and temperature gradient inside concrete. This effectively overcomes the shortcomings of traditional surface watering methods, such as large heat conduction time lag and difficulty in controlling the core of aggregate pre-cooling technology—the surface temperature difference. It greatly reduces the early cracking rate of large-volume concrete in high-altitude and cold-climate gravity dams. At the same time, the real-time feedback optimization mechanism ensures that the model parameters continuously adapt to environmental changes, comprehensively improving the structural durability and temperature control crack prevention effect.
[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:
[0018] Figure 1 This is a flowchart of a method for cooling the hydration heat of large-volume concrete in a high-altitude, cold-climate gravity dam according to the present invention. Detailed Implementation
[0019] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0020] Example 1
[0021] like Figure 1 As shown, a method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam is disclosed, the method comprising:
[0022] S1. An internal cooling pipe network is installed inside the concrete, an external cooling pipe network is installed on the outer surface of the concrete, and multi-dimensional sensing data is collected on the concrete.
[0023] In step S1, it should be noted that the internal cooling pipe network installed inside the concrete specifically includes stainless steel pipes, evenly distributed along the concrete pouring layer, with the pipe spacing controlled within the range of 0.8 meters to 1.2 meters, and connected to the external cooling pipe network through intelligent valves to ensure the efficiency of coolant circulation. Simultaneously, the external cooling pipe network installed on the outer surface of the concrete specifically includes an array of spray heads, spaced 0.5 meters apart, and integrated with temperature sensors to monitor surface temperature changes in real time. Furthermore, multi-dimensional sensing sensors, including temperature sensors, humidity sensors, and stress sensors, are embedded in key areas of the concrete (such as dam corners and joints between new and old concrete). The sensors are buried at a depth of 1 / 3 to 1 / 2 of the concrete thickness, and the collected data is uploaded to the central processor in real time via a wireless transmission module to ensure the comprehensiveness and real-time nature of the data collection.
[0024] S2. Preprocess the multi-dimensional perception data and extract features to obtain structured feature data;
[0025] S3. Construct a dual-network structure model based on the structured feature data, and use the evaluation network in the dual-network structure model to predict the distribution of hydration heat and temperature gradient changes inside the concrete in real time according to the structured feature data to obtain the prediction results;
[0026] S4. Based on the prediction results and the reinforcement learning reward function, an adaptive cooling strategy is generated using the control network in the dual-network structure model.
[0027] S5. Based on the adaptive cooling strategy, control the valves and pump station power of the internal cooling pipe network and the external cooling pipe network to dynamically adjust the cooling process;
[0028] In step S5, it is necessary to explain in detail that after the control network generates the adaptive cooling strategy, the central processing unit immediately sends instructions to the intelligent valve controller and the pump station frequency converter. Specifically, for the internal cooling pipe network, the intelligent valves adjust their opening within 0.5 seconds according to the valve opening sequence instructions, with an adjustment range accurate to 0.1 degrees, ensuring that the coolant flow rate matches the internal heat load of the concrete in real time. Simultaneously, the pump station frequency converter adjusts the power to the target value within 0.3 seconds according to the power adjustment scheme, with an adjustment accuracy controlled within ±2%, achieving dynamic optimization of the coolant circulation speed. For the external cooling pipe network, the spray head array uses electromagnetic valve groups to achieve precise zoned spraying based on surface temperature monitoring data. The spray intensity and frequency are automatically adjusted according to the temperature gradient to ensure the uniformity of the concrete surface temperature. The entire adjustment process is fed back to the central processing unit in real time through the wireless transmission module, forming a closed-loop control to ensure that the cooling process is synchronized with the concrete hydration heat release rate, effectively avoiding local overheating or overcooling.
[0029] S6. Continuously collect the multi-dimensional perception data, combine it with the output of the dual-network structure model for real-time feedback optimization, and update the model parameters of the dual-network structure model.
[0030] The principle of a method for cooling the hydration heat of large-volume concrete in a high-altitude, cold-climate gravity dam, as described in this embodiment, is as follows: First, internal and external cooling pipe networks are set up inside and on the surface of the concrete, respectively, and multi-dimensional sensing data is collected. This data covers key indicators such as temperature, humidity, and stress, which can comprehensively reflect the hydration heat state of the concrete. Subsequently, the collected data is preprocessed and features are extracted to obtain structured feature data. Based on the structured feature data, a dual-network structure model is constructed. The evaluation network is responsible for predicting the distribution of hydration heat and temperature gradient changes inside the concrete in real time. The control network generates an adaptive cooling strategy based on the prediction results of the evaluation network and the reinforcement learning reward function. This strategy comprehensively considers multiple objectives such as minimizing the temperature gradient, suppressing crack risk, and optimizing energy consumption. When implementing the cooling strategy, the valve opening and pump power of the internal cooling pipe network are precisely controlled by intelligent valves and pump station frequency converters to achieve dynamic adjustment of the coolant flow rate and circulation speed. At the same time, the spray head array of the external cooling pipe network automatically adjusts the spray intensity and frequency according to the surface temperature monitoring data to ensure uniform concrete surface temperature. In addition, the system continuously collects multi-dimensional sensing data and combines the output of the dual-network structure model for real-time feedback optimization, constantly updating model parameters to adapt to environmental changes and ensure the accuracy and efficiency of the cooling process.
[0031] As an optional embodiment of the present invention, optionally, obtaining structured feature data in step S2 includes:
[0032] S201. Based on the concrete temperature dataset in the multi-dimensional sensing data, construct a special temperature spatiotemporal feature for cold environments;
[0033] In step S201, it is necessary to explain in detail that when constructing the spatiotemporal temperature features specifically for high-altitude and cold environments, the concrete temperature dataset is first spatially divided. The concrete pouring area is divided into multiple sub-regions according to volume and structural characteristics, and each sub-region is equipped with an independent temperature monitoring node to ensure that the spatial resolution meets the needs of local heat conduction analysis. In the temporal dimension, the sliding window method is used to extract short-term (10-minute) and long-term (hourly) temperature change trends. Combined with environmental factors such as large diurnal temperature differences and strong daytime radiation in high-altitude and cold regions, a feature vector containing spatial coordinates, timestamps, temperature values, and their first / second derivatives is constructed. Furthermore, the temperature signal is decomposed through wavelet transform to extract high-frequency thermal fluctuation components and low-frequency thermal accumulation trends, forming a multi-scale spatiotemporal feature matrix.
[0034] S202. Based on the humidity dataset in the multi-dimensional sensing data, construct humidity-stress coupled response features;
[0035] In step S202, it is necessary to explain in detail that when constructing the humidity-stress coupled response features, the concrete humidity dataset is first preprocessed to remove outliers and fill in missing data to ensure data continuity. Then, combining stress sensor data from the same monitoring node, the Pearson correlation coefficient is used to analyze the correlation between humidity and stress, and nodes with significant correlations are selected as key monitoring points. In the time dimension, the rates of change of humidity and stress are extracted simultaneously, constructing a feature vector containing humidity values, stress values, humidity change rate, stress change rate, and their intersection. Further dimensionality reduction is achieved through principal component analysis (PCA) to extract the main coupling modes, forming a humidity-stress coupled response feature matrix to reflect the impact of humidity changes on the stress state of concrete.
[0036] S203. Based on the stress dataset in the multi-dimensional sensing data, construct a high-altitude cold environment correction coefficient;
[0037] In step S203, it is necessary to explain in detail that when constructing the high-altitude cold environment correction coefficient, historical data on concrete stress in high-altitude cold regions must first be collected, covering stress changes under different seasons, day and night, and extreme weather conditions. Through statistical analysis, the sensitivity parameters of stress to temperature and humidity changes are determined, and an initial stress-environment parameter relationship model is established. Further considering the unique freeze-thaw cycle effect of high-altitude cold regions, a freeze-thaw damage factor is introduced. This factor is obtained through laboratory simulated freeze-thaw tests and reflects the stress attenuation law of concrete during repeated freeze-thaw processes. Simultaneously, combined with field monitoring data, the initial model is corrected, and the least squares method is used to fit the correction coefficient, minimizing the error between the model's predicted values and the measured values. The final high-altitude cold environment correction coefficient can accurately reflect the actual state of concrete stress under high-altitude cold conditions.
[0038] In this embodiment, the method for establishing the initial stress-environment parameter relationship model is as follows: First, historical stress data of concrete in high-altitude and cold regions under different environmental parameters (temperature and humidity) are collected. This data needs to cover various typical working conditions, such as day-night cycles, seasonal changes, and extreme weather. The collected data is cleaned to remove outliers and missing values to ensure data quality. Next, a multiple linear regression analysis method is used, with stress as the dependent variable and temperature and humidity as independent variables, to establish an initial stress-environment parameter linear relationship model. To verify the accuracy of the model, the dataset can be divided into a training set and a test set. The model is built using the training set, and the prediction accuracy of the model is evaluated using the test set. If the model's prediction error is large, a nonlinear term can be introduced or a more complex nonlinear regression method, such as support vector regression or neural network regression, can be used to improve the model's fitting effect. Finally, the sensitivity parameters in the model are determined through statistical analysis, completing the establishment of the initial stress-environment parameter relationship model.
[0039] S204. Based on the weather dataset in the multi-dimensional sensing data, construct meteorological impact characteristics;
[0040] In step S204, it is necessary to elaborate on the following: When constructing the meteorological impact features, the weather dataset is first preprocessed, including data cleaning to remove noise and outliers, and interpolation to fill in missing data to ensure the continuity of the time series. Subsequently, key meteorological indicators are extracted, such as ambient temperature, humidity, wind speed, solar radiation intensity, and precipitation type. These indicators directly affect the hydration heat release rate and surface temperature changes of concrete. In the time dimension, the diurnal and seasonal variations of meteorological indicators are analyzed. Combined with the concrete pouring and curing cycle, a feature vector containing instantaneous values, daily averages, extreme values, and rates of change of meteorological elements is constructed. Further consideration is given to the interactions between meteorological elements, such as the combined effect of temperature and humidity on the evaporation rate of the concrete surface, and the coupling effect of wind speed and solar radiation on the regulation of the thermal balance of the concrete surface. These complex relationships are reflected by constructing cross-term features. Simultaneously, weather type classification (such as sunny, cloudy, rainy, snowy, etc.) is introduced. Combining historical meteorological data with concrete performance monitoring results, the differences in concrete temperature gradient and stress state under different weather types are analyzed to form a weather type impact feature matrix. Finally, the feature set is optimized by feature selection algorithms (such as feature filtering based on information gain or mutual information) to ensure that the meteorological impact features can not only fully reflect the effect of meteorological conditions on the hydration heat process of concrete, but also have low redundancy and high computational efficiency.
[0041] S205. By integrating the specific temperature and spatiotemporal characteristics of the high-altitude cold environment, the low-temperature brittle stress response characteristics, the high-altitude cold environment correction coefficient, and the meteorological influence characteristics, the structured feature data is generated through principal component analysis for dimensionality reduction.
[0042] In step S205, it is necessary to explain in detail that when fusing various features to generate structured feature data, the spatiotemporal temperature features specific to high-altitude and cold environments, humidity-stress coupled response features, high-altitude and cold environment correction coefficients, and meteorological influence features must first be standardized. This is because the data dimensions and value ranges of different features may vary significantly, and standardization can eliminate these differences, making various features comparable in subsequent analyses. Specifically, the Z-score standardization method can be used, subtracting the mean of each feature value and then dividing by the standard deviation of that feature to obtain the standardized feature value.
[0043] After standardization, these features are combined into a high-dimensional feature vector. This high-dimensional feature vector contains information about concrete in terms of temperature, humidity, stress, and weather conditions, but it also suffers from data redundancy and high computational complexity. To address these issues, principal component analysis (PCA) is used for dimensionality reduction.
[0044] The PCA method projects the original high-dimensional feature vectors onto a new set of orthogonal coordinate axes through a linear transformation. These new coordinate axes are called principal components. The principal components are arranged in order of variance; the larger the variance, the more information the principal component contains. During dimensionality reduction, the first few principal components with larger variances are selected to represent the original high-dimensional feature vectors, thus significantly reducing the dimensionality of the data while retaining most of the important information.
[0045] In practice, the covariance matrix of the standardized eigenvectors is first calculated, reflecting the correlation between different features. Then, the eigenvalues and eigenvectors of the covariance matrix are solved; the eigenvectors corresponding to the eigenvalues represent the directions of the principal components. Based on the magnitude of the eigenvalues, the eigenvectors corresponding to the k largest eigenvalues are selected to form the projection matrix. Finally, the standardized eigenvectors are multiplied by the projection matrix to obtain the dimensionality-reduced structured feature data.
[0046] Through the above steps, the structured feature data generated by integrating various features and performing dimensionality reduction through principal component analysis can more effectively reflect the hydration heat state and performance changes of concrete.
[0047] As an optional embodiment of the present invention, optionally, obtaining the prediction result in step S3 includes:
[0048] S301. Construct a dual-network structure model that includes the collaboration of an evaluation network and a control network, wherein the evaluation network includes a convolutional neural network and a long short-term memory network.
[0049] In step S301, it is necessary to explain in detail that when constructing the dual-network structure model, a combined architecture of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) is adopted to fully utilize the advantages of CNN in spatial feature extraction and LSTM in time series analysis. Specifically, the CNN part is responsible for processing spatial information from multi-dimensional perceptual data, automatically extracting spatial features of temperature, humidity, and stress distribution inside the concrete through structures such as convolutional layers and pooling layers. These spatial features are flattened and then input into the LSTM part. The LSTM uses its unique gating mechanism and memory units to capture the dynamic changes in the time series of the concrete hydration heat process, such as the evolution of temperature gradient and the adjustment of stress state. Through the collaborative work of CNN and LSTM, the evaluation network can comprehensively and accurately predict the distribution of hydration heat and temperature gradient changes inside the concrete, providing reliable prediction results for subsequent adaptive cooling strategies. At the same time, the control network works closely with the evaluation network, generating and adjusting the cooling strategy in real time based on the prediction results of the evaluation network and the reinforcement learning reward function, ensuring the efficiency and accuracy of the cooling process.
[0050] S302. Train and optimize the dual-network structure model;
[0051] In step S302, it is necessary to explain in detail that when training and optimizing the dual-network structure model, a training dataset is collected and prepared. The dataset covers multi-dimensional sensing data of concrete under various working conditions, including data on temperature, humidity, stress, and meteorology. The training dataset is divided into a training set, a validation set, and a test set. The training set is used for learning the model's parameters, the validation set is used to adjust the model's hyperparameters to prevent overfitting, and the test set is used to evaluate the final performance of the model.
[0052] During training, backpropagation and gradient descent optimization algorithms are used to update the model parameters. For the convolutional neural network (CNN) part of the evaluation network, the weights of the CNN kernels are gradually adjusted using gradient descent by calculating the gradient of the loss function with respect to the kernel parameters. This allows the CNN to better extract the spatial characteristics of temperature, humidity, and stress distribution within the concrete. For the long short-term memory (LSTM) network part, parameters are also updated based on the gradient of the loss function with respect to its internal control mechanism and memory unit parameters to enhance its ability to capture the dynamic changes in the time series of concrete hydration heat.
[0053] The control network is trained based on the predictions of the evaluation network and the reinforcement learning reward function. The design of the reinforcement learning reward function needs to comprehensively consider multiple objectives, such as minimizing the temperature gradient, suppressing crack risk, and optimizing energy consumption. For example, a higher reward value is given when the cooling strategy reduces the internal temperature gradient of the concrete, lowers the crack risk, and reduces energy consumption; conversely, a lower reward value is given. The control network continuously tries different cooling strategies and adjusts its parameters based on the feedback from the reward function to generate the optimal adaptive cooling strategy.
[0054] During model training, several optimization methods are needed to improve model performance and training efficiency. For example, batch training can be employed, inputting a batch of data into the model for training each time, rather than individual data points. This leverages the advantages of matrix operations to accelerate the training process. Simultaneously, a learning rate decay strategy can be introduced, gradually reducing the learning rate as training progresses. This allows the model to converge quickly in the early stages of training and enables more precise parameter adjustment in the later stages.
[0055] After a certain number of training iterations, the model is evaluated using a validation set. Based on the evaluation results, the model's hyperparameters are adjusted, such as the kernel size and number of convolutional neural networks, and the number of hidden layer nodes in long short-term memory networks, to further improve the model's performance. Once the model's performance on the validation set reaches a satisfactory level, it is finally evaluated using a test set to ensure that the model has good generalization ability in practical applications. Through this training and optimization process, a high-performance dual-network structure model can be obtained.
[0056] S303. Input the structured feature data and historical time series information into the evaluation network in the dual network structure model to obtain the preliminary prediction results of the hydration heat distribution and temperature gradient change in n(1) hours;
[0057] In step S303, it is necessary to explain in detail that the structured feature data includes comprehensive information on concrete in terms of temperature, humidity, stress, and weather conditions. After standardization and principal component analysis for dimensionality reduction, this information can more effectively reflect the hydration heat state of concrete. Historical time series information provides the state changes of concrete at different points in time.
[0058] The convolutional neural network portion of the evaluation network first extracts spatial features from the structured feature data. Through processing by convolutional and pooling layers, the network can automatically identify the temperature, humidity, and stress distribution patterns inside the concrete, capturing both local and global spatial features. These features are then flattened and transformed into a format suitable for processing by long short-term memory networks.
[0059] The Long Short-Term Memory (LSTM) network utilizes its powerful time series analysis capabilities, combined with historical time series information, to model the dynamic changes in the heat of hydration of concrete. The gating mechanism and memory units of LSTM enable it to process long-sequence data, capturing temporal dependencies and trends. Through LSTM processing, the network can predict the changes in the heat of hydration and temperature gradient of concrete over the next n (1) hours.
[0060] The preliminary predictions are compared with the actual monitoring data to assess the accuracy and reliability of the predictions. If there is a significant deviation between the predictions and the actual data, the model needs further training and optimization, adjusting network parameters or improving feature extraction methods to enhance prediction accuracy.
[0061] S304. The preliminary prediction results are corrected, and a three-dimensional thermal field reconstruction map is generated, showing the hydration heat distribution and temperature gradient changes at key measuring points, to obtain the prediction results.
[0062] In step S304, it is necessary to explain in detail that when correcting the preliminary prediction results, the sources of deviation between the prediction results and the actual monitoring data are first analyzed. These deviations may stem from limitations in the model training data, incomplete feature extraction, or the inherent similarity of the dual-network structure model itself. To address these sources of deviation, data augmentation techniques are used to expand the training dataset. For example, generative adversarial networks (GANs) are used to simulate multi-dimensional perception data of concrete under more working conditions to improve the model's adaptability to different scenarios. Simultaneously, an attention mechanism is introduced to improve the evaluation network, enabling the model to dynamically focus on feature dimensions that have a greater impact on the prediction results and reduce interference from irrelevant information.
[0063] During the correction process, the preliminary prediction results are corrected online by incorporating real-time monitoring data. Specifically, the latest collected data on temperature, humidity, stress, and meteorological conditions are fused with the preliminary prediction results, and the predicted values are dynamically adjusted using a Kalman filter algorithm. The Kalman filter algorithm can provide the optimal state estimate by utilizing the system state equation and observation equation, taking into account prediction uncertainties and observation noise. This online correction method can promptly correct prediction biases caused by environmental changes or data drift, improving the real-time performance and accuracy of the prediction results.
[0064] When generating the 3D thermal field reconstruction map, the corrected prediction results and the 3D geometric model of the concrete structure are used. An interpolation algorithm maps the discrete prediction data to a continuous 3D space. The radial basis function interpolation method is chosen, which can construct a smooth function based on the spatial distribution and magnitude of known data points to estimate the values of unknown points, thus generating a detailed and accurate 3D thermal field distribution. The 3D thermal field reconstruction map can intuitively display the hydration heat distribution at different locations within the concrete, helping engineers quickly locate high-temperature regions and areas with drastic temperature gradient changes.
[0065] For the hydration heat distribution and temperature gradient changes at key measuring points, data at corresponding locations were extracted from the three-dimensional thermal field reconstruction map, and detailed analysis and visualization were performed. By plotting the temperature-time curves and temperature gradient-time curves at key measuring points, the hydration heat development process and temperature change trend of concrete at different time points were clearly presented. At the same time, the crack risk and structural safety at key measuring points were assessed by combining the thermodynamic performance parameters of concrete materials. The final prediction results not only include quantitative values of the hydration heat distribution and temperature gradient changes of concrete in the next n (1) hours, but also include the three-dimensional thermal field reconstruction map and detailed analysis reports of key measuring points.
[0066] As an optional embodiment of the present invention, the expression of the evaluation network may be:
[0067]
[0068] in, Indicates the future Hourly forecast results This represents the output layer weight matrix. Represents the gated timing modeling unit. Indicates the hidden state of the LSTM history. Indicates the historical cell state of LSTM. This represents a convolutional feature extractor. express Input features at all times, This represents the weight matrix of the convolutional layer. Represents the set of LSTM weight matrices. This represents the LSTM bias vector set. This represents the output layer bias vector. This represents the correction factor for high-altitude and cold environments. This represents the uncertainty covariance matrix.
[0069] As an optional embodiment of the present invention, optionally, in step S4, generating an adaptive cooling strategy using the control network in the dual-network structure model based on the reinforcement learning reward function of the prediction result includes:
[0070] S401. Based on the prediction results, the action space is defined as the valve opening adjustment range and the pump station power adjustment value. The action space corresponds to the control parameters of the external cooling pipe network and the internal cooling pipe network.
[0071] In this embodiment, the method for defining the action space is as follows: First, based on the requirements of temperature gradient changes during the hydration heat of concrete and the actual capacity of the cooling system, the adjustment range of the valve opening is determined. The size of the valve opening directly affects the flow rate of cooling water, and thus the cooling effect. By reasonably setting the adjustment range of the valve opening, it can be ensured that the cooling water can be dynamically adjusted according to the changes in the internal temperature gradient of the concrete, achieving precise cooling. For example, when the internal temperature gradient of the concrete is large, the valve opening is increased to increase the flow rate of cooling water to accelerate heat dissipation; when the temperature gradient is small, the valve opening is decreased to reduce the flow rate of cooling water and avoid over-cooling.
[0072] Secondly, determine the adjustment value of the pump station power. The pump station power determines the circulation speed and pressure of the cooling water in the pipe network. Different pump station power adjustment values correspond to different cooling water circulation states, thus affecting cooling efficiency. Based on the heat generation and dissipation during the concrete hydration heat process, rationally setting the pump station power adjustment value can ensure efficient circulation of cooling water in the pipe network, promptly removing heat from the concrete. For example, during the peak of concrete hydration heat, increasing the pump station power accelerates the cooling water circulation speed and enhances the cooling effect; during the later stages of hydration heat, reducing the pump station power reduces energy consumption.
[0073] Simultaneously, the action space is mapped to the control parameters of the external and internal cooling pipe networks. The external cooling pipe network is primarily responsible for heat exchange with the external environment; its control parameters, such as valve opening and pump power, affect the supply and circulation of external cooling water. The internal cooling pipe network, on the other hand, is in direct contact with the concrete; adjusting its control parameters allows for more precise control of the cooling process within the concrete. By mapping the action space to the control parameters of the external and internal cooling pipe networks, the control network can generate appropriate control commands based on prediction results, achieving precise regulation of the cooling system. For example, when the prediction result indicates that the temperature in a certain area inside the concrete is high, the control network can generate commands to increase the valve opening of the corresponding internal cooling pipe network in that area and appropriately increase the pump power to enhance cooling in that area.
[0074] S402. Construct a reinforcement learning reward function based on the action space. The reward function is constructed based on the objectives of minimizing temperature gradient, suppressing crack risk, and optimizing energy consumption.
[0075] In step S402, it is necessary to explain in detail that when constructing the reinforcement learning reward function, three core objectives must be comprehensively considered: minimizing the temperature gradient, suppressing crack risk, and optimizing energy consumption. Minimizing the temperature gradient aims to ensure a uniform temperature distribution inside the concrete, avoiding structural problems caused by excessive temperature differences. To this end, the reward function can be set to give a positive reward when the internal temperature gradient of the concrete decreases and a negative reward when the temperature gradient increases, thereby guiding the control network to generate cooling strategies that can effectively reduce the temperature gradient.
[0076] During the hydration process, concrete is prone to cracking due to temperature changes and stress. A reward function can incorporate crack risk assessment indicators, such as predicting the likelihood of crack formation based on the internal stress state and temperature history of the concrete. A higher reward value is awarded when cooling strategies reduce crack risk, and a lower reward value is awarded when cooling strategies reduce crack risk, thus prompting the control network to take measures to suppress crack risk.
[0077] In practical engineering, the operation of cooling systems consumes a significant amount of energy, thus energy consumption reduction is a crucial consideration. The reward function calculates the energy consumption of the cooling system based on control parameters such as pump power and valve opening. A positive reward is given when the cooling strategy reduces energy consumption while minimizing the temperature gradient and suppressing crack risk; conversely, a negative reward is given if energy consumption increases, incentivizing the control network to generate energy-saving cooling strategies.
[0078] Combining the above three objectives, a multi-objective optimization reward function is constructed. This reward function can be calculated using a weighted summation method, integrating the reward values of the three objectives: minimizing the temperature gradient, suppressing crack risk, and optimizing energy consumption. The weight coefficients can be set according to actual engineering needs and importance. For example, if structural safety requirements are high, the weight of the crack risk suppression objective can be appropriately increased; if energy costs are more sensitive, the weight of the energy consumption optimization objective can be increased. Through this multi-objective optimization reward function, the control network is guided to comprehensively consider each objective when generating adaptive cooling strategies, achieving high efficiency, accuracy, and energy saving in the cooling process. The expression for the reinforcement learning reward function is:
[0079]
[0080] in, express Constantly reinforce learning reward values. , and All represent weighting coefficients. Indicates the future Hourly temperature gradient prediction Indicates the future Hourly cracking risk probability Represents numerically stable terms, preventing The mathematical definition is undefined, which guarantees the continuity of the reward function. express Real-time measured power of the pumping station This indicates the maximum power threshold of the pumping station;
[0081] S403. Based on the reinforcement learning reward function, the optimal action policy is calculated using the forward propagation of the control network, wherein the control network adopts a deep Q-network structure; the expression for calculating the optimal action policy is:
[0082]
[0083]
[0084] in, This indicates control over network output. This represents the control network weight matrix. Indicates the future Hourly temperature gradient prediction Indicates the future Hourly cracking risk probability express Real-time measured power of the pumping station This represents the action vector at time step 1. This represents the bias vector of the control network. Indicates time The optimal action vector, Indicates the safety threshold. This indicates the safety control gain.
[0085] In step S403, it is necessary to explain in detail that the control network adopts a deep Q-network structure. This structure has strong nonlinear mapping capabilities and adaptive learning capabilities, and can effectively handle complex cooling strategy generation problems. When calculating the optimal action strategy, the deep Q-network, through a forward propagation process, incorporates the input information such as the predicted temperature gradient for the next n hours, the cracking risk probability for the next n hours, the measured pump station power at time t, and the action vector from the previous time step, and performs calculations with the control network's weight matrix and bias vector. During the calculation process, the network gradually extracts key feature representations based on the characteristics and weight parameters of these input information, performs nonlinear transformations through activation functions, and finally outputs the optimal action vector at time t.
[0086] A safety threshold sets a safe range. When the predicted result or control action exceeds this range, the safety control gain adjusts it to ensure that the generated cooling strategy does not damage the concrete structure. For example, if the prediction shows that the internal temperature of the concrete is too high, which may lead to cracking, the safety threshold will limit the adjustment range of valve opening and pump power to avoid excessive cooling that could cause excessive changes in internal stress in the concrete. The safety control gain will then appropriately modify the action vector according to the actual situation, making the cooling strategy safer and more reliable.
[0087] By employing forward propagation calculations based on a deep Q-network structure, an optimal action strategy can be obtained that satisfies the requirements of minimizing temperature gradients and suppressing crack risk while also optimizing energy consumption. This strategy can precisely guide the adjustment of control parameters for both external and internal cooling pipe networks, such as valve opening and pump station power. This enables precise control of the hydration heat cooling process of large-volume concrete in high-altitude and cold-climate gravity dams, improving the quality and safety of the concrete structure while reducing the operating costs of the cooling system.
[0088] S404. Optimize the optimal action strategy using a strategy gradient algorithm (such as the PPO algorithm) to generate an adaptive cooling strategy, which includes a valve opening sequence and a pump station power adjustment scheme.
[0089] In step S404, it is necessary to explain in detail that the policy gradient algorithm (such as the PPO algorithm) maximizes the cumulative reward by continuously adjusting the policy parameters. In this embodiment, the PPO algorithm is used to optimize the optimal action policy calculated based on the deep Q-network structure to generate an adaptive cooling strategy that better meets the actual engineering needs.
[0090] Specifically, in each iteration, the PPO algorithm generates a series of actions based on the current policy and obtains corresponding reward signals by interacting with the environment. Then, the algorithm calculates the advantage function of these actions, which is the additional reward of the action relative to the average level. Next, the PPO algorithm uses the advantage function to update the policy parameters, increasing (if the advantage function is positive) or decreasing (if the advantage function is negative) the probability of generating similar actions in the future.
[0091] During optimization, the PPO algorithm introduces a pruning mechanism to prevent excessive policy updates, thus ensuring training stability. The pruning mechanism limits the magnitude of policy parameter updates, preventing excessive differences between the new and old policies. This way, even if a poor reward is obtained under the current policy, the algorithm will not blindly and drastically adjust the policy parameters, but will instead engage in careful exploration and optimization.
[0092] As an optional embodiment of the present invention, optionally, in step S6, the multi-dimensional perception data is continuously collected, and real-time feedback optimization is performed in combination with the output of the dual-network structure model, including updating the model parameters of the dual-network structure model;
[0093] S601. Continuously collect and preprocess multi-dimensional sensing data of concrete to obtain real-time structured feature vectors;
[0094] In step S601, it is important to explain in detail that continuously collecting multi-dimensional sensing data of the concrete is a crucial step in ensuring that the cooling system can make real-time adjustments based on actual conditions. This multi-dimensional sensing data includes, but is not limited to, the internal temperature of the concrete, temperature gradient, stress state, cooling water flow rate, valve opening, and pump station power. This is achieved by deploying various types of sensors, such as temperature sensors, stress sensors, and flow sensors, inside the concrete and in the cooling pipe network.
[0095] The collected multi-dimensional sensory data is preprocessed. The preprocessing process mainly includes data cleaning, data normalization, and feature extraction.
[0096] Feature extraction involves extracting features from preprocessed data that significantly influence the generation of cooling strategies. For example, temperature gradient features are extracted from temperature data, and stress change rate features are extracted from stress data. These features can more accurately reflect the thermo-mechanical coupling state inside concrete and the operation of the cooling system.
[0097] S602. Input the real-time structured feature vector into the evaluation network to obtain the prediction result. Compare the prediction result with the measured value, calculate the mean square error and the crack risk probability deviation, and simultaneously evaluate the energy consumption reduction rate and temperature gradient suppression rate of the adaptive cooling strategy to form a quantitative feedback index.
[0098] In step S602, it is necessary to explain in detail that,
[0099] The pre-processed real-time structured feature vectors are input into the evaluation network. The evaluation network, using its pre-trained model structure and parameters, performs in-depth analysis and processing of these input features, and then outputs predictions about the future state of the concrete. These predictions cover key information such as the temperature gradient changes and cracking probability within the concrete over a future period.
[0100] Subsequently, the predicted results output by the evaluation network were compared in detail with the actual measured values. By calculating the mean square error, the degree of difference between the predicted results and the measured values can be intuitively understood. The smaller the mean square error, the closer the predicted results are to the actual situation, and the higher the prediction accuracy of the evaluation network. At the same time, the cracking risk probability bias was calculated. This indicator reflects the accuracy of the evaluation network in predicting the cracking risk of concrete. The smaller the bias, the more reliable the prediction of cracking risk.
[0101] In addition to the aforementioned indicators related to concrete condition, the energy consumption reduction rate and temperature gradient suppression rate of the adaptive cooling strategy also need to be evaluated simultaneously. The energy consumption reduction rate reflects the energy-saving effect of the currently implemented cooling strategy and is calculated by comparing energy consumption before and after implementing the cooling strategy. The temperature gradient suppression rate reflects the effectiveness of the cooling strategy in reducing the internal temperature gradient of the concrete and can be calculated based on the changes in the internal temperature gradient of the concrete before and after implementing the cooling strategy.
[0102] By combining the calculated mean square error, crack risk probability deviation, energy consumption reduction rate, and temperature gradient suppression rate, a complete set of quantitative feedback indicators is formed. These quantitative feedback indicators enable the entire cooling system to make precise dynamic adjustments based on actual conditions, further improving the cooling effect and the quality and safety of the concrete structure.
[0103] S603. Update the evaluation network parameters based on the quantized feedback index using mini-batch gradient descent.
[0104] In step S603, it is necessary to explain in detail that, based on the formed quantized feedback index, the parameters of the evaluation network are updated using a mini-batch gradient descent algorithm. In the specific implementation, a mini-batch of data is first randomly selected from the collected multi-dimensional sensing data, and the corresponding real-time structured feature vector is input into the evaluation network to obtain the prediction result. Then, based on the mean square error, crack risk probability deviation, and other indicators calculated from the prediction result and the measured value, a loss function is constructed. The loss function is used to measure the difference between the current prediction result of the evaluation network and the actual value.
[0105] Next, the mini-batch gradient descent algorithm is used to calculate the gradient of the loss function with respect to each parameter of the evaluation network. The gradient represents the direction and rate of change of the loss function at the current parameter point. Through the backpropagation algorithm, the gradient is passed back from the output layer to the input layer layer by layer, thereby obtaining the gradient value corresponding to each parameter.
[0106] Finally, the parameters of the evaluation network are updated based on the calculated gradient values and the preset learning rate.
[0107] S604. Monitoring environmental mutation information triggers transfer learning, and the evaluation network parameters are updated again.
[0108] As an optional embodiment of the present invention, optionally, in step S604, monitoring environmental mutation information to trigger transfer learning and updating the evaluation network parameters again includes:
[0109] S6041. When environmental mutation information is detected, freeze some of the underlying parameters of the evaluation network;
[0110] In step S6041, it is important to explain in detail that when the monitoring system detects sudden environmental changes, such as extreme weather changes or fluctuations in the properties of concrete raw materials, which may significantly affect the cooling process, it immediately triggers the transfer learning mechanism. Under this mechanism, some low-level parameters of the evaluation network are first frozen. These low-level parameters are usually formed after training with a large amount of data, such as filter parameters in convolutional layers. The purpose of freezing these parameters is to preserve the network's ability to stably extract fundamental characteristics such as the thermo-mechanical coupling state of concrete, which it has already learned, and to avoid loss of accuracy due to over-adjustment of these fundamental perceptions caused by sudden environmental changes.
[0111] S6042. Extract environmental feature data from the environmental mutation information, construct a fine-tuning dataset, and define a transfer learning loss function. This loss function combines the historical distribution before the mutation with the difference in real-time data distribution after the mutation. The expression for the transfer learning loss function is:
[0112]
[0113] in, This represents the total loss from transfer learning. This represents the historical-to-new data weighting coefficient. Indicates the Kullback-Leibler divergence. This represents the predicted distribution of the evaluation network on data prior to the environmental abrupt change. Represents the predicted result vector, This represents the input feature vector. This indicates the evaluation network's real-time predicted distribution on post-mutation data. This represents the mean square error in the new environment. Represents the regularization coefficient. This indicates the evaluation of the top-level parameters of the network. This represents the top-level parameters before the mutation.
[0114] In step S6042, it is necessary to explain in detail that after capturing information about sudden environmental changes, the system accurately extracts key environmental feature data from this information. This feature data covers the magnitude of temperature changes, the rate of humidity change, and specific indicators of raw material performance fluctuations, comprehensively reflecting the potential impact of environmental changes on the concrete cooling process. Subsequently, using this extracted environmental feature data, a fine-tuning dataset is constructed specifically for fine-tuning the evaluation network. This dataset not only includes real-time data after the change but also incorporates historical data before the change to ensure that the model can fully understand the continuity and scope of environmental changes.
[0115] By using the transfer learning loss function, the evaluation network can be guided to quickly adapt to new challenges brought about by sudden environmental changes while retaining its original stable feature extraction capabilities, thereby achieving a more accurate update of the evaluation network parameters.
[0116] S6043. Using the fine-tuning dataset and the transfer learning loss function, quickly fine-tune the unfrozen parameters in the evaluation network;
[0117] During fine-tuning, the input feature vectors from the fine-tuning dataset are sequentially fed into the evaluation network. At this stage, only the unfrozen parameters participate in the computation. The network performs forward propagation based on these input feature vectors to obtain the corresponding prediction result vectors.
[0118] Then, based on the expression of the transfer learning loss function, and combining information such as the evaluation network's predicted distribution on data before the environmental change, the real-time predicted distribution on data after the change, and the mean squared error of the new environment, the current total transfer learning loss is calculated. This total loss reflects the evaluation network's adaptability to environmental changes under the current parameters.
[0119] Next, using the backpropagation algorithm, the gradients of the loss function with respect to the unfrozen parameters are calculated layer by layer, starting from the output layer, based on the calculated total transfer learning loss. These gradients indicate the direction and magnitude in which the parameters should be adjusted to reduce the value of the loss function.
[0120] Finally, based on the calculated gradient values and the preset learning rate, the unfrozen parameters in the evaluation network are updated. By continuously iterating this process, the values of the unfrozen parameters are gradually adjusted, enabling the evaluation network to better adapt to new situations after sudden environmental changes and improving the accuracy and reliability of predicting the relevant states of the concrete cooling process.
[0121] S6044. Synchronize the fine-tuned evaluation network parameters to the control network and perform transfer learning update of the dual-network structure model.
[0122] In step S6044, it is necessary to explain in detail that the evaluation network parameters obtained through the above fine-tuning process are accurately synchronized to the control network through a specific data transmission and synchronization mechanism. This ensures that the evaluation network and the control network in the dual-network structure model maintain a consistent parameter state, thereby guaranteeing the collaborative working capability of the entire model in response to sudden environmental changes.
[0123] During parameter synchronization, an efficient and reliable data transmission protocol is employed to prevent data loss or errors during transmission. Simultaneously, to ensure real-time parameter synchronization, a reasonable synchronization cycle is set, enabling the control network to promptly acquire the latest parameters from the evaluation network and adjust its control strategy accordingly.
[0124] After parameter synchronization is complete, the dual-network structure model completes the transfer learning update. At this point, the control network will recalculate the optimal action policy based on the updated parameters to adapt to new situations after sudden environmental changes. This transfer learning update mechanism makes the dual-network structure model more adaptable and robust, enabling it to work stably and continuously in complex and ever-changing environments.
[0125] Example 2
[0126] A hydration heat cooling system for large-volume concrete in high-altitude and cold-climate gravity dams includes a processor and a memory for storing processor-executable instructions; wherein the processor is configured to implement a hydration heat cooling method for large-volume concrete in high-altitude and cold-climate gravity dams when executing the executable instructions.
[0127] It should be noted that the computer device includes a processor, memory, and may also include one or more of the following: multimedia components, input / output (I / O) interfaces, and communication components.
[0128] The processor controls the overall operation of the computer equipment to complete all or part of the steps of the hydration heat cooling method for large-volume concrete in high-altitude and cold-climate gravity dams.
[0129] Memory stores various types of data to support device operation, including instructions and related data for application programs or methods. It can be implemented by volatile or non-volatile storage devices or combinations thereof, such as SRAM, EEPROM, etc.
[0130] Multimedia components include a screen (such as a touchscreen) and audio components. The audio components are used for outputting and / or inputting audio signals, have a microphone to receive external audio signals, and can store or transmit them. They also have at least one speaker to output audio signals. I / O interfaces provide interfaces for the processor and other interface modules (such as keyboards, mice, buttons, etc., which can be virtual or physical buttons). Communication components are used for wired or wireless communication between the computer device and other devices. Wireless communication includes Wi-Fi, Bluetooth, etc., and corresponding components include Wi-Fi modules, Bluetooth modules, etc.
[0131] As a preferred option, the computer device can be implemented using electronic components such as ASIC and DSP to perform a method for cooling the hydration heat of large-volume concrete in high-altitude and cold-climate gravity dams.
[0132] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
Claims
1. A method for cooling the heat of hydration of large-volume concrete in high-altitude, cold-climate gravity dams, characterized in that, The method includes: S1. An internal cooling pipe network is installed inside the concrete, and an external cooling pipe network is installed on the outer surface of the concrete. Multi-dimensional sensing data is collected on the concrete. The multi-dimensional sensing data includes concrete temperature dataset, humidity dataset, stress dataset, and weather dataset. S2. Preprocess the multi-dimensional perception data and extract features to obtain structured feature data; S3. Construct a dual-network structure model based on the structured feature data, and use the evaluation network in the dual-network structure model to predict the distribution of hydration heat and temperature gradient changes inside the concrete in real time according to the structured feature data to obtain the prediction results; S4. Based on the prediction results and the reinforcement learning reward function, an adaptive cooling strategy is generated using the control network in the dual-network structure model. S5. Based on the adaptive cooling strategy, control the valves and pump station power of the internal cooling pipe network and the external cooling pipe network to dynamically adjust the cooling process; S6. Continuously collect the multi-dimensional perception data, combine it with the output of the dual-network structure model to perform real-time feedback optimization, and update the model parameters of the dual-network structure model. In step S4, based on the prediction results and the reinforcement learning reward function, an adaptive cooling strategy is generated using the control network in the dual-network structure model, including: S401. Based on the prediction results, the action space is defined as the valve opening adjustment range and the pump station power adjustment value. The action space corresponds to the control parameters of the external cooling pipe network and the internal cooling pipe network. S402. Construct a reinforcement learning reward function based on the action space. The reward function is constructed based on the objectives of minimizing temperature gradient, suppressing crack risk, and optimizing energy consumption. S403. Based on the reinforcement learning reward function, the optimal action policy is calculated using the forward propagation of the control network, wherein the control network adopts a deep Q-network structure. S404. Optimize the optimal action strategy through the strategy gradient algorithm to generate an adaptive cooling strategy, which includes a valve opening sequence and a pump station power adjustment scheme. The structured feature data obtained in step S2 includes: S201. Based on the concrete temperature dataset in the multi-dimensional sensing data, construct a special temperature spatiotemporal feature for cold environments; S202. Based on the humidity dataset in the multi-dimensional sensing data, construct humidity-stress coupled response features; S203. Based on the stress dataset in the multi-dimensional sensing data, construct a high-altitude cold environment correction coefficient; S204. Based on the weather dataset in the multi-dimensional sensing data, construct meteorological impact characteristics; S205. By integrating the specific temperature and spatiotemporal characteristics of the high-altitude cold environment, the low-temperature brittle stress response characteristics, the high-altitude cold environment correction coefficient, and the meteorological influence characteristics, the structured feature data is generated through principal component analysis for dimensionality reduction.
2. The method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam as described in claim 1, characterized in that, The prediction results obtained in step S3 include: S301. Construct a dual-network structure model that includes the collaboration of an evaluation network and a control network, wherein the evaluation network includes a convolutional neural network and a long short-term memory network. S302. Train and optimize the dual-network structure model; S303. Input the structured feature data and historical time series information into the evaluation network in the dual-network structure model to obtain preliminary prediction results of the hydration heat distribution and temperature gradient change over n hours. S304. The preliminary prediction results are corrected, and a three-dimensional thermal field reconstruction map is generated, showing the hydration heat distribution and temperature gradient changes at key measuring points, to obtain the prediction results.
3. A method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam as described in claim 1 or 2, characterized in that, The expression for the evaluation network is: in, Indicates the future Hourly forecast results This represents the output layer weight matrix. Represents the gated timing modeling unit. Indicates the hidden state of the LSTM history. Indicates the historical cell state of LSTM. This represents a convolutional feature extractor. express Input features at all times, Represents the convolutional layer weight matrix. Represents the set of LSTM weight matrices. This represents the LSTM bias vector set. This represents the output layer bias vector. This represents the correction factor for high-altitude and cold environments. This represents the uncertainty covariance matrix.
4. The method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam as described in claim 1, characterized in that, The expression for calculating the optimal action policy is: in, This indicates control over network output. This represents the control network weight matrix. Indicates the future Hourly temperature gradient prediction Indicates the future Hourly cracking risk probability express Real-time measured power of the pumping station This represents the action vector at time step 1. This represents the bias vector of the control network. Indicates time The optimal action vector, Indicates the safety threshold. This indicates the safety control gain.
5. The method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam as described in claim 1, characterized in that, In step S6, the multi-dimensional perception data is continuously collected, and real-time feedback optimization is performed based on the output of the dual-network structure model. The model parameters of the dual-network structure model are updated, including: S601. Continuously collect and preprocess multi-dimensional sensing data of concrete to obtain real-time structured feature vectors; S602. Input the real-time structured feature vector into the evaluation network to obtain the prediction result. Compare the prediction result with the measured value, calculate the mean square error and the crack risk probability deviation, and simultaneously evaluate the energy consumption reduction rate and temperature gradient suppression rate of the adaptive cooling strategy to form a quantitative feedback index. S603. Update the evaluation network parameters based on the quantized feedback index using mini-batch gradient descent. S604. Monitoring environmental mutation information triggers transfer learning, and the evaluation network parameters are updated again.
6. The method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam as described in claim 5, characterized in that, In step S604, monitoring environmental mutation information to trigger transfer learning and updating the evaluation network parameters again includes: S6041. When environmental mutation information is detected, freeze some of the underlying parameters of the evaluation network; S6042. Extract environmental feature data from the environmental mutation information, construct a fine-tuning dataset, and define a transfer learning loss function, wherein the loss function combines the differences between the historical distribution before the mutation and the real-time data distribution after the mutation. S6043. Using the fine-tuning dataset and the transfer learning loss function, quickly fine-tune the unfrozen parameters in the evaluation network; S6044. Synchronize the fine-tuned evaluation network parameters to the control network and perform transfer learning update of the dual-network structure model.
7. The method for cooling the heat of hydration of large-volume concrete in a high-altitude, cold-climate gravity dam as described in claim 6, characterized in that, The expression for the transfer learning loss function is: in, This represents the total loss from transfer learning. This represents the historical-to-new data weighting coefficient. Indicates the Kullback-Leibler divergence. This represents the predicted distribution of the evaluation network on data prior to the environmental abrupt change. Represents the predicted result vector, This represents the input feature vector. This indicates the evaluation network's real-time predicted distribution on post-mutation data. This represents the mean square error in the new environment. Represents the regularization coefficient. This indicates the evaluation of the top-level parameters of the network. This represents the top-level parameters before the mutation.
8. A hydration heat cooling system for large-volume concrete in high-altitude, cold-climate gravity dams, characterized in that, The system includes: processor; Memory used to store processor-executable instructions; The processor is configured to implement the hydration heat cooling method for large-volume concrete of high-altitude and cold-climate gravity dams according to any one of claims 1 to 7 when executing the executable instructions.