Dam safety state memory maintenance method and system for digital twin hydraulic engineering large model

Abstract

The invention provides a dam safety state memory maintenance method and system for a digital twin hydraulic engineering large model, and the method comprises the following steps: (1), obtaining and preprocessing the historical monitoring data of a dam, (2), carrying out the self-organizing mapping dimension reduction of an integrated multi-dimensional historical old data set, carrying out the matching of the data set with a new data set monitored in real time, and carrying out the memory maintenance of the dam safety state. (3) calculating a matching combination weight between the old data set and the new data set, and then carrying out model training again; and (4) carrying out long-term safety state prediction and early warning by using the updated hydraulic engineering large model. According to the method, a mechanism capable of dynamically maintaining and updating model memory is constructed, long-term accumulated monitoring data can be effectively managed, new information can be continuously introduced into the model, the long-term prediction and analysis capability of the safety state of the dam is improved, and the problem that memory maintenance is limited in the prior art is effectively solved.

Description

Technical Field

[0001] The present invention relates to the fields of water conservancy engineering and artificial intelligence, and in particular to a dam safety dynamic state memory maintenance method and system for a digital twin water conservancy large model. Background Art

[0002] Digital twins build virtual models of physical systems to achieve real-time monitoring, prediction and analysis of systems. The water conservancy big model is an artificial intelligence model based on a large neural network, designed specifically for the water conservancy industry, and is one of the key technologies of digital twins. By processing massive amounts of industry data (such as monitoring data, meteorological information, historical records, etc.), the water conservancy big model can generate insights, predict trends, and provide intelligent support for real-time monitoring and management of water conservancy projects. For complex engineering structures such as dams, safety state perception requires analysis of a large amount of historical data, including water level changes, stress distribution, temperature effects, etc. Long-term memory maintenance is particularly critical in this context because it can help the model make full use of historical data to understand the long-term safety state and potential risks of the dam. By maintaining and updating the model's memory, the digital twin water conservancy big model can more accurately predict future safety trends, promptly detect potential dam structural risks, and provide scientific early warning and decision support. Therefore, long-term memory maintenance is the key to ensuring that the dam digital twin system can achieve maximum efficiency.

[0003] Model memory plays a vital role in the long-term safety monitoring of dams. The safety status of dam structures is affected by many factors, including seasonal changes, extreme weather events, geological activities, etc., which often have long-term accumulation effects. Model memory can effectively store and manage these historical data to form a global understanding of the safety status of dams. By continuously introducing new monitoring data, the water conservancy large model memory can be dynamically updated to maintain sensitivity to the current status of the dam. At the same time, model memory can also be used to identify and analyze the long-term change trends of dam structures and assist decision makers in formulating maintenance and management strategies. Especially when predicting the long-term safety risks of dams, model memory provides a rich information basis, making the prediction more accurate and reliable.

[0004] Traditional model memory maintenance faces many challenges in dam safety risk prediction. On the one hand, dam safety monitoring involves a large amount of multi-dimensional historical data. Traditional methods often have difficulty in efficiently storing and managing this data, resulting in model memory redundancy or information loss. On the other hand, traditional memory maintenance methods usually lack a dynamic update mechanism, making it difficult to introduce new monitoring data in a timely manner, resulting in the model being unable to accurately reflect the real-time status of the dam. In addition, the aging problem of model memory is also prominent. Over time, the model may gradually lose its sensitivity to early data and cannot effectively use historical information for long-term predictions. These problems are particularly serious in dam safety risk prediction, because the structural state of the dam is complex and long-term, and any inaccurate prediction may lead to serious consequences. Therefore, there is an urgent need for a method that can effectively maintain and update model memory to improve the long-term prediction capability and reliability of the dam digital twin system. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a dam safety status memory maintenance method and system for a digital twin water conservancy large model, and constructs a mechanism that can dynamically maintain and update the model memory, which can not only effectively manage the long-term accumulated monitoring data, but also continuously introduce new information into the model, thereby improving the long-term prediction and analysis capabilities of the dam safety status, and effectively solving the problem of limited memory maintenance in the prior art.

[0006] The technical solution provided by the present invention is a dam safety state memory maintenance method for a digital twin water conservancy large model, comprising the following steps:

[0007] (1) Acquire and preprocess the historical monitoring data of the dam to form an integrated multi-dimensional historical monitoring data set;

[0008] (2) Perform dimensionality reduction on the integrated multi-dimensional historical old data set using self-organizing mapping and match it with the new data set monitored in real time;

[0009] (3) Manage and optimize the new and old memory combination nodes of the model after the self-organizing map in step (2), and then retrain the hydropower dam safety state prediction model, and measure the model error and loss by calculating the Euclidean distance between the input vector and the best matching unit (Best Matching Unit, BMU). Specifically, the training error includes quantization error (Quantization Error, QE) and topological error (Topographic Error, TE),

[0010] According to the model's loss and error evaluation of training data, the learning rate is adjusted in real time. Through neighborhood functions and weight updates, the model remains effective in long-term monitoring and accurately captures the distribution patterns of new and historical data.

[0011] (4) Using the model optimized by the self-organizing map in step (3) to make predictions, the model output is a safety prediction signal, and the dam safety risk is predicted based on the optimized model memory, and early warning information is generated.

[0012] Furthermore, the historical monitoring data in step (1) includes water level changes, seepage pressure, stress distribution and temperature. The collected data are comprehensively cleaned and standardized. The standardization formula is:

[0013]

[0014] In the formula, x j is the normalized data value, μ j is the average value of feature j, which represents the arithmetic mean of the feature in the entire sample, σ j is the standard deviation of feature j, which indicates the dispersion of the feature value relative to its mean.

[0015] Furthermore, in step (1), various types of sensors are installed at key structural parts of the dam, including but not limited to the dam body, dam foundation, reservoir and spillway. The sensors include water level sensors, seepage pressure sensors, strain gauges, temperature sensors, rainfall sensors and flow meters. All sensors are connected through a data acquisition system to automatically collect data regularly and attach timestamps.

[0016] Furthermore, the step (2) specifically includes the following steps:

[0017] (2-1) Input data standardization: The standardized historical monitoring data obtained in step (1) is converted into an input vector. The standardization formula is:

[0018] x=[x1,x2,...,x n ]

[0019] Where x represents the standardized data vector, and n represents the dimension of the vector, that is, the number of features;

[0020] (2-2) Construct and initialize the SOM grid, select the grid size m×m, and initialize the weight vector of each node.

[0021] w i =[w i1 , w i2 , ..., w in ]

[0022] Among them, w i is the weight vector of the i-th node, n represents the dimension of the vector, and m represents the size parameter of the grid;

[0023] (2-3) Perform data mapping and find the best matching unit. For each node, calculate the Euclidean distance between the input vector and the node weight vector, map the input data, and use the Euclidean distance to find the best matching unit BMU in the competition stage.

[0024]

[0025] BMU represents the node closest to the input data on the map, helping to map data features. Updates not only act on BMUs, but also affect their neighboring nodes. For the best matching unit BMU determined:

[0026] i * = arg min i d i

[0027] Among them, d i is the distance between the input vector and node i, i * Represents the node index closest to the input;

[0028] (2-4) Combine new and old data to update memory. Each time new data is input, the corresponding BMU is found, and the weights of the BMU and its surrounding nodes are adjusted according to the learning rate and neighborhood function. The influence of the old data is retained, the new data is integrated, and the model is continuously adjusted to adapt to changes. In this step, the new data refers to the data obtained by real-time monitoring, and the old data refers to the historical monitoring data from step (1). After each new data is input, the model will review the historical data to ensure the retention and reference of valid data in order to learn new patterns and trends.

[0029] Furthermore, in step (3), the new and old memory combination nodes of the model in step (2) are managed and optimized, specifically including: in the cooperation stage of the self-organizing map, the neighborhood function is calculated and the node weight is updated, wherein the formula of the neighborhood function is:

[0030]

[0031] Among them, h ci (t) is the neighborhood function, which indicates the degree to which node i is affected by BMU, r c and r i are the coordinates of BMU and node i, σ(t) is the neighborhood radius at time t;

[0032] In the adjustment phase of the self-organizing map, the model optimizes the weight vector by changing the learning rate and neighborhood function, and adjusts the weight vector according to the learning rate:

[0033] w i (t+1)=w i (t)+α(t)·hci (t)·(xw i (t))

[0034] The learning rate and neighborhood decay function are:

[0035]

[0036] Where α(t) is the learning rate at time t, α is the initial learning rate, T is the total training time, σ(t) is the current neighborhood radius, which controls the influence range of BMU on neighboring nodes, σ0 is the initial neighborhood radius, and λ is the time constant, which controls the neighborhood shrinkage speed;

[0037] For the learned and decayed models, the online learning and dynamic adjustment mechanism of adaptive learning is used to ensure the effectiveness in long-term applications. The learning rate is dynamically adjusted over time and model performance to prevent the model from falling into local optimality or oscillation:

[0038]

[0039] Where: α(t) is the current learning rate, β is a constant used to control the attenuation of the learning rate, ensuring that the learning rate can be reduced at an appropriate rate.

[0040] Furthermore, in step (4), for a given input data x and node weight w i , the similarity is calculated as:

[0041] d(i)=||xw i ||

[0042] Where d(i) is the Euclidean distance between the input data and node i. By calculating the similarity, the node most similar to the current data is selected as the best matching unit BMU.

[0043] BMU=argmin i ||xw i ||

[0044] BMU represents the node position that best matches the current input data. The node matching result is converted into a specific risk assessment. Combined with historical data, the current risk value is generated through the risk calculation function. According to the preset safety threshold θ, the risk value is evaluated to generate an early warning signal:

[0045]

[0046] Among them, f represents the model function, based on the weight vector w i Generates predictions corresponding to the current input data.

[0047] Another technical solution provided by the present invention is a dam safety state memory maintenance system for a digital twin water conservancy large model, comprising:

[0048] The data acquisition and preprocessing module is responsible for collecting and preprocessing the dam's historical monitoring data;

[0049] Model memory initialization module, which uses preprocessed data to build initial model memory;

[0050] Dynamic memory update module, designs and executes the update strategy of model memory;

[0051] Memory management and optimization module, which selectively retains and optimizes model memory;

[0052] The long-term prediction and early warning module uses the updated model to predict the long-term safety status of the dam.

[0053] Beneficial effects of the present invention:

[0054] (1) A mechanism for dynamically maintaining and updating model memory is proposed. By introducing new monitoring data and optimizing model memory, the model's ability to predict the long-term safety state of the dam is improved. The SOM technology is used to map complex, multi-dimensional monitoring data into a low-dimensional topological structure, thereby achieving effective data dimensionality reduction and visualization, especially in the state perception and dynamic simulation of the water conservancy large language model of complex systems. By mapping real-time sensor data into a two-dimensional plane, SOM can help decision makers quickly identify potential patterns and anomalies in the data. The memory of the digital twin system is enhanced, enabling it to continuously track and learn long-term change trends, thereby optimizing system management and control strategies. This capability enables the digital twin to better reflect the operating status of the real world and improve the understanding and prediction capabilities of complex systems.

[0055] (2) The memory maintenance method of the large water conservancy model for long-term dam safety monitoring achieves effective management of the long-term monitoring data of the dam by dynamically maintaining and updating the model memory. The system effectively avoids the redundancy and complexity of memory by selectively retaining and optimizing historical data, and realizes real-time dynamic analysis of the dam's safety status by introducing new monitoring information. This method ensures that the model accurately predicts the long-term safety status of the dam and significantly improves the monitoring and early warning performance of the dam digital twin system.

[0056] (3) Through the memory optimization strategy, the long-term prediction ability and accuracy of the model are effectively improved. The dynamic memory update mechanism can timely reflect the current status of the dam and accurately predict future safety trends. Compared with the prior art, the present invention can not only effectively manage and utilize a large amount of historical monitoring data, but also reduce the redundancy of model memory, improve the efficiency of dam safety monitoring, and provide a more scientific and efficient solution for the long-term safety management of dams. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a characteristic association network diagram of dam monitoring data established by the self-organizing map (SOM) technology in the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with specific embodiments.

[0059] The present invention provides a dam safety state memory maintenance method for a digital twin water conservancy large model, wherein the monitoring period is 3 years, the detection frequency is once per hour, the dam type is a concrete gravity dam, and the specific steps include:

[0060] S1 obtains and preprocesses the historical monitoring data of the dam;

[0061] S2 performs dimensionality reduction on the integrated multi-dimensional historical old data set through self-organizing mapping and matches it with the new data set monitored in real time;

[0062] S3 calculates the matching merging weights between the old and new datasets, and then retrains the model;

[0063] S4 uses the updated large-scale water conservancy model to carry out long-term safety status prediction and early warning.

[0064] Through the memory maintenance and update mechanism, the effective management of the long-term monitoring data of the dam is realized. This improves the long-term prediction ability and analysis accuracy of the model, and through dynamic memory update, the real-time response ability of the model to the dam safety status is improved. It can achieve accurate prediction of the long-term safety trend of the dam and improve the scientific nature of early warning and maintenance work.

[0065] In the technical solution of the present invention:

[0066] Step S1 involves obtaining the historical monitoring data set of the dam, including multi-dimensional information such as water level changes, seepage pressure, stress distribution, temperature, rainfall and flow. Various types of sensors are deployed at key locations of the dam: Water level sensor: installed in the upstream and downstream areas to monitor water level changes. Seepage pressure sensor: arranged at different depths inside the dam body to monitor the seepage pressure of the dam body. Strain gauge: installed at the dam body and dam foundation to measure the stress distribution of the dam body. Temperature sensor: installed on the surface and inside of the dam to monitor temperature changes. Rainfall sensor: arranged in the area surrounding the dam to monitor rainfall. Flow meter: installed at the spillway to measure flow changes.

[0067] Data is collected automatically on a regular basis. The data acquisition system can be set to collect data within a certain period of time every hour or every day to ensure that the data can be obtained in a timely and complete manner, and a timestamp is attached so that historical changes can be accurately traced. Historical data within 3 years records hourly monitoring information, with a total of approximately 500,000 data records. The system collects data from sensors every hour and uploads it to the central database. The integration of all monitoring data facilitates subsequent analysis and ensures data integrity and accuracy. These sensors are designed to monitor various environmental and structural parameters related to dam safety in real time. The accuracy and consistency of input data are ensured by constructing an initial model memory. In order to address the impact of different data dimensions, the collected data needs to be comprehensively cleaned and standardized. The standardization formula is:

[0068]

[0069] In the formula, x j is the normalized data value. j is the average value of feature j, which represents the arithmetic mean of the feature in the entire sample. j is the standard deviation of feature j, indicating the degree of dispersion of the feature value relative to its mean. Step S1 finally forms an integrated multidimensional historical monitoring data set. The standardized multidimensional historical data set contains the following features: water level (H): 35 m = 50 m, seepage pressure (P): 50 kPa = 200 kPa, stress (S): 10 με-100 με, temperature (T): 10 °C = 35 °C, rainfall (R): 0 mm / h = 100 mm / h, flow (F): 0 m 3 / s=5m 3 / s.

[0070] Table 1 Data preprocessing reference values

[0071]

[0072] Step S2 involves performing a self-organizing map on the memory and contains four sub-steps.

[0073] (1) Input data standardization: convert the standardized historical monitoring data obtained in step S1 into an input vector. The standardization formula is:

[0074] x=[x1,x2,...,x n ]

[0075] Where x represents the normalized data vector and n represents the dimension (number of features) of the vector. The raw data obtained from various sensors and data sources is converted into input vectors in a standard format. This step ensures that the scope and impact of each feature are equal, thereby improving the consistency and accuracy of subsequent model operations.

[0076] (2) Construct and initialize the SOM grid. The initialization of the self-organizing map (SOM) determines the grid size of the SOM, that is, how many nodes will be used to represent the topological structure of the input data. This involves choosing the number and arrangement of nodes, which is usually determined based on the complexity and dimensionality of the data set. The grid size is 20×20 (400 nodes in total), and the weight vector of each node is initialized. Each node has a weight vector w i , whose dimensions are the same as the input data. When the weight vector is initialized, it can be randomly selected or distributed according to a specific rule. This process ensures that each node has different weights at the beginning, thus covering a wide area of ​​the input data space.

[0077] w i =[w i1 , w i2 , ..., w in ]

[0078] Among them, w i is the weight vector of the ith node, and m represents the size parameter of the grid. In the initialization, by organizing the nodes into a predefined grid structure, the initial weights are distributed in a predefined grid, such as a rectangle or hexagon, which provides a basis for the learning of the large water conservancy model in the neighborhood relationship.

[0079] (3) Perform data mapping and find the best matching unit. For each node, calculate the Euclidean distance between the input vector and the node weight vector. Map the input six dimensions (water level, seepage pressure, stress, temperature, rainfall, flow). In the competition stage, use the Euclidean distance to find the best matching unit (BMU). The formula is:

[0080]

[0081] For a determined Best Matching Unit (BMU):

[0082] i * = arg min i d i

[0083] Among them, d i is the distance between the input vector and node i, i * Indicates the node index closest to the input (BMU)

[0084] (4) Combine new and old data to update memory. In this step, new data refers to data obtained from real-time monitoring, and old data refers to historical monitoring data from step S1. After each new data is input, the model will review the historical data to ensure the retention and reference of valid data in order to learn new patterns and trends. The historical monitoring data obtained in S1 provides the basis for the model memory design in S2. By combining new and old data, more flexible model memory maintenance can be achieved, the adaptability of the model can be enhanced, and the foundation for subsequent neighborhood adjustments and weight updates can be laid. It is also the key to SOM's ability to effectively identify and classify data patterns. According to experience, the typical range of memory update parameter values ​​is shown in Table 2.

[0085] Table 2 Memory update strategy parameters

[0086]

[0087] Step S3 involves managing and optimizing the new and old memory combination nodes of the model in S2. In the cooperative stage of the self-organizing map (SOM), calculating the neighborhood function and updating the node weights are key steps. The formula of the neighborhood function is:

[0088]

[0089] Among them, h ci (t) is the neighborhood function, which indicates the degree to which node i is affected by BMU, r c and r i are the coordinates of the BMU and node i, and σ(t) is the neighborhood radius at time t. The node weights are adjusted by applying the neighborhood function to ensure that nodes closer to the BMU receive larger weight adjustments. This forces the network to gradually adapt to the distribution pattern of the input data. In the initial training phase, the network makes adjustments over a larger range, and then gradually narrows the scope of influence to focus on finer adjustments. This strategy enables the network to effectively capture the global structure of the data and refine local features in the later stage. This forms an accurate reflection of the topological structure of the data and improves the visualization and analysis capabilities of the data. This process not only ensures the gradual stability of the model, but also enhances the expressiveness and accuracy of the model when processing complex data sets, making the data easier to understand and use.

[0090] In the adjustment phase of the self-organizing map, the model optimizes the weight vector by changing the learning rate and the neighborhood function. The weight vector is adjusted according to the learning rate:

[0091] w i (t+1)=w i (t)+α(t)·h ci (t)·(xw i (t))

[0092] Table 3 Self-organizing map cooperation weights corresponding to several typical dam safety data parameters

[0093]

[0094]

[0095] This means that the weight vector is obtained by the learning rate α(t) and the neighborhood function h ci (t) is adjusted to make it closer to the input data x.

[0096] The learning rate and neighborhood decay function are:

[0097]

[0098] Among them, α(t) is the learning rate at time t, α0 is the initial learning rate, T is the total training time, and σ(t) is the current neighborhood radius, which controls the influence range of BMU on neighboring nodes. σ0 is the initial neighborhood radius, and λ is the time constant, which controls the neighborhood shrinkage speed. The learning rate and neighborhood decay function ensure that the learning rate and neighborhood radius gradually decrease during the training process, so that the network shifts from global adaptation to local fine-tuning. This adjustment strategy helps SOM grasp the data structure from the macro to the micro optimization details, making the model's understanding of the data more comprehensive and accurate, and improving the effectiveness of data visualization and clustering.

[0099] For the learned and decayed models, the online learning and dynamic adjustment mechanism of adaptive learning is used to ensure the effectiveness in long-term applications. The learning rate is dynamically adjusted over time and model performance to prevent the model from falling into local optimality or oscillation:

[0100]

[0101] In this formula: α(t) is the current learning rate, and β is a constant used to control the decay of the learning rate, ensuring that the learning rate can be reduced at an appropriate rate. The learning rate is adjusted in real time according to the loss and error evaluation of the model on the training data. This adjustment enables the model to quickly respond to the global characteristics of the data in the early stage, and focus on refining local characteristics in the later stage to achieve a balance between sensitivity and stability. The model parameters are updated through regular incremental learning, so that it can continuously optimize itself through training with new data. This incremental update effectively integrates historical data mapping and current data features to ensure the memory utility of the model. The model memory mechanism designed in S2 is implemented and optimized in this step. Through neighborhood functions and weight updates, the model remains effective in long-term monitoring and accurately captures the distribution patterns of new and historical data. Through these steps, the large water conservancy model can effectively and accurately update and maintain its model memory, so as to better serve and support decision-making in the management of complex water conservancy projects.

[0102] S4 involves using the updated water conservancy model for long-term safety status prediction and early warning. After completing the weight update and model optimization in S3, S4 performs data monitoring and safety early warning on the optimized model, using the model optimized by self-organizing mapping in S3 for prediction. The model output is a safety prediction signal: based on the optimized model memory, the dam safety risk prediction is performed and early warning information is generated. The currently monitored data: water level 48 meters, seepage pressure 180kPa, stress 90με, temperature 30℃, rainfall 60mm / h, flow 3m 3 / s. When. By Figure 1 The long-term memory model after the self-organizing map shown in the figure processes and aggregates the real-time data to obtain a smooth warning information flow. For a given input data x and node weight w i , the similarity can be calculated as:

[0103] d(i)=||xw i ||

[0104] Where d(i) is the Euclidean distance between the input data and node i. By calculating the similarity, the node that is most similar to the current data is selected as the best matching unit (BMU). Specifically, the selection of BMU is based on minimizing the Euclidean distance:

[0105] BMU=arg min i ||xw i ||

[0106] BMU represents the node position that best matches the current input data. The node matching result is converted into a specific risk assessment, combined with historical data, and the current risk value is generated through the risk calculation function. According to the preset safety threshold θ, the risk value is evaluated to generate an early warning signal:

[0107]

[0108] Where f represents the model function, based on the weight vector, w i Generate a corresponding prediction for the current input data. If the risk value exceeds the set threshold (0.75), a warning signal is generated. The risk value of the current input data is 0.82, which exceeds the warning threshold, and the system automatically issues an alarm.

[0109] Step S1=S4 Through the effective management and optimization of model memory, the system achieves accurate maintenance of the long-term monitoring data of the dam, avoids the accumulation of redundant information, and ensures the effective retention of important historical data. The optimized model can maintain effectiveness and accuracy in long-term monitoring, accurately predict the long-term safety status of the dam, and provide early warning information in a timely manner. The final effect is to significantly improve the performance of the large water conservancy model in long-term monitoring and risk prediction, and provide more reliable and scientific technical support for the safety management of the dam.

[0110] To achieve the above technical goals, the system framework of the proposed method is as follows:

[0111] Module 1: Data acquisition and preprocessing module, responsible for collecting historical monitoring data of the dam and preprocessing it.

[0112] Module 2: Model memory initialization module, which uses preprocessed data to establish the initial model memory.

[0113] Module 3: Dynamic memory update module, designs and executes the update strategy of model memory.

[0114] Module 4: Memory management and optimization module, which selectively retains and optimizes model memory.

[0115] Module 5: Long-term prediction and early warning module, using the updated model to predict the long-term safety status of the dam.

[0116] Through these modules, the long-term prediction and early warning module finally realizes the accurate prediction of the long-term safety status of the dam. This system framework realizes the memory maintenance and dynamic update of dam monitoring data, providing scientific and reliable technical support for long-term safety prediction.

[0117] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A dam safety dynamic state memory maintenance method for a digital twin water conservancy large model, characterized in that: The steps include: (1) Acquire and preprocess the historical monitoring data of the dam to form an integrated multi-dimensional historical monitoring data set; (2) Perform dimensionality reduction on the integrated multi-dimensional historical old data set using self-organizing mapping and match it with the new data set monitored in real time; (3) managing and optimizing the new and old memory combination nodes of the model after the self-organizing mapping in step (2), and then retraining the water conservancy dam safety state prediction model. According to the loss and error evaluation of the model on the training data, the learning rate is adjusted in real time. Through the neighborhood function and weight update, the model maintains its effectiveness in long-term monitoring and accurately captures the distribution pattern of new data and historical data. (4) Using the model optimized by the self-organizing map in step (3) to make predictions, the model output is a safety prediction signal, and the dam safety risk is predicted based on the optimized model memory, and early warning information is generated.

2. According to claim 1, a dam safety dynamic state memory maintenance method for a digital twin water conservancy large model is characterized in that: The historical monitoring data in step (1) include water level changes, seepage pressure, stress distribution, temperature, rainfall and flow rate. The collected data are comprehensively cleaned and standardized. The standardization formula is: In the formula, x j is the normalized data value, μ j is the average value of feature j, which represents the arithmetic mean of the feature in the entire sample, σ j is the standard deviation of feature j, which indicates the dispersion of the feature value relative to its mean.

3. According to claim 2, a dam safety state memory maintenance method for a digital twin water conservancy large model is characterized in that: In the step (1), various types of sensors are installed at key structural parts of the dam, including but not limited to the dam body, dam foundation, reservoir and spillway. The sensors include water level sensors, seepage pressure sensors, strain gauges, temperature sensors, rainfall sensors and flow meters. All sensors are connected through a data acquisition system to automatically collect data regularly and attach timestamps.

4. According to a method for maintaining dam safety dynamic state memory for a digital twin water conservancy large model according to claim 2, it is characterized in that: The step (2) specifically comprises the following steps: (2-1) Input data standardization: The standardized historical monitoring data obtained in step (1) is converted into an input vector. The standardization formula is: x=[x1,x2,...,x n ] Where x represents the standardized data vector, and n represents the dimension of the vector, that is, the number of features; (2-2) Construct and initialize the SOM grid, select the grid size m×m, and initialize the weight vector of each node. In i =[in i1 ,In i2 ,...,In in ] Among them, w i is the weight vector of the i-th node, n represents the dimension of the vector, and m represents the size parameter of the grid; (2-3) Perform data mapping and find the best matching unit. For each node, calculate the Euclidean distance between the input vector and the node weight vector, map the input data, and use the Euclidean distance to find the best matching unit (BMU) in the competition stage. For a determined Best Matching Unit (BMU): in * =angry me i d i Among them, d i is the distance between the input vector and node i, i * Represents the node index closest to the input; (2-4) Combine new and old data to update memory. In this step, new data refers to data obtained through real-time monitoring, and old data refers to historical monitoring data from step (1). After each new data input, the model reviews historical data to ensure that valid data is retained and referenced in order to learn new patterns and trends.

5. According to claim 4, a dam safety dynamic state memory maintenance method for a digital twin water conservancy large model is characterized in that: In the step (3), the new and old memory combination nodes of the model in step (2) are managed and optimized, specifically including: in the cooperation stage of the self-organizing map, the neighborhood function is calculated and the node weight is updated, wherein the formula of the neighborhood function is: Among them, h ci (t) is the neighborhood function, which indicates the degree to which node i is affected by BMU, r c and r i are the coordinates of BMU and node i, σ(t) is the neighborhood radius at time t; In the adjustment phase of the self-organizing map, the model optimizes the weight vector by changing the learning rate and neighborhood function, and adjusts the weight vector according to the learning rate: w i (t+1)=w i (t)+α(t)·h ci (t)·(x-w i (t)) The learning rate and neighborhood decay function are: Where α(t) is the learning rate at time t, α0 is the initial learning rate, T is the total training time, σ(t) is the current neighborhood radius, which controls the influence range of BMU on neighboring nodes, σ0 is the initial neighborhood radius, and λ is the time constant, which controls the neighborhood shrinkage speed. For the learned and decayed models, the online learning and dynamic adjustment mechanism of adaptive learning is used to ensure the effectiveness in long-term applications. The learning rate is dynamically adjusted over time and model performance to prevent the model from falling into local optimality or oscillation: Where: α(t) is the current learning rate, β is a constant used to control the attenuation of the learning rate, ensuring that the learning rate can be reduced at an appropriate rate.

6. A dam safety state memory maintenance method for a digital twin water conservancy large model according to claim 5, characterized in that: In step (4), for a given input data x and node weight w i , the similarity is calculated as: d(i)=||x-w i || Where d(i) is the Euclidean distance between the input data and node i. By calculating the similarity, the node most similar to the current data is selected as the best matching unit BMU. BMU=argmin i ||xw i || BMU represents the node position that best matches the current input data. The node matching result is converted into a specific risk assessment. Combined with historical data, the current risk value is generated through the risk calculation function. According to the preset safety threshold θ, the risk value is evaluated to generate an early warning signal: Among them, f represents the model function, based on the weight vector w i Generates predictions corresponding to the current input data.

7. A dam safety state memory maintenance system for a digital twin water conservancy large model, implemented based on the method described in any one of claims 1 to 6, characterized in that: include: Data acquisition and preprocessing module, responsible for collecting and preprocessing the dam's historical monitoring data; Model memory initialization module, which uses preprocessed data to build initial model memory; Dynamic memory update module, designs and executes the update strategy of model memory; Memory management and optimization module, which selectively retains and optimizes model memory; The long-term prediction and early warning module uses the updated model to predict the long-term safety status of the dam.

Images