A dam deformation prediction method and system based on physical information and interpretable artificial intelligence

Abstract

The present application belongs to the field of civil engineering safety monitoring and artificial intelligence technology, more specifically, relates to a dam deformation prediction method and system based on physical information and interpretable artificial intelligence. The method constructs a neural network prediction model trained by a hybrid loss function, the hybrid loss function couples a data-driven item measuring prediction accuracy and a physical information constraint item based on engineering mechanics law; and further introduces an interpretability analysis module to perform attribution analysis on the prediction results of the model, realizing the "physical-data" decomposition of prediction contribution. The present application solves the technical contradiction that in the existing dam safety monitoring technology, high-precision prediction models such as deep learning are difficult to obtain engineering trust due to their "black box" characteristics, while the accuracy of traditional interpretable models is insufficient.

Description

Technical Field

[0001] This invention belongs to the field of civil engineering safety monitoring and artificial intelligence technology, and more specifically, relates to a method and system for predicting dam deformation based on physical information and interpretable artificial intelligence. Background Technology

[0002] As critical infrastructure, the long-term safe and stable operation of dams is of paramount importance. Dam deformation is a key physical quantity that comprehensively reflects its structural health status and response to external loads. Therefore, accurately predicting dam deformation is a core task for implementing effective safety monitoring and early warning.

[0003] Chinese patent document CN111614913A discloses a method, terminal equipment, and storage medium for predicting dam deformation at multiple measurement points. The method acquires the spatial coordinates of the dam's measurement points and historical monitoring data for a certain time period, including the upstream and downstream water levels, upstream and downstream boundary temperatures at various times, and the deformation monitoring values ​​of the measurement points to be predicted. The spatial coordinates of the measurement points, the upstream and downstream water levels, and the boundary temperatures at various times within a certain time period are input into a Bayes-LSTM network model. The deformation monitoring values ​​of the measurement points to be predicted are used as the expected output of the Bayes-LSTM network model to train the model, thus obtaining the prediction model.

[0004] Current dam deformation prediction technologies mainly fall into two categories: First, traditional statistical models (such as HST and MLR), which are easy to interpret but have limited accuracy and struggle to capture complex nonlinear responses. Second, modern data-driven models (such as ANN and LSTM), while offering high prediction accuracy, suffer from opaque decision-making mechanisms due to their "black box" nature, making them difficult to trust in safety-critical fields. Although there have been attempts to apply Physical Information Neural Networks (PINN) to other fields, existing technologies have failed to effectively address the unique "time lag effect" in dam deformation, and interpretability methods have not provided a quantitative assessment of "physical compliance."

[0005] In summary, existing technologies present a technical contradiction between achieving high prediction accuracy and high interpretability simultaneously, necessitating the development of an innovative method that combines both high accuracy and high reliability. Summary of the Invention

[0006] The present invention aims to overcome at least one of the defects of the prior art and provide a dam deformation prediction method based on physical information and interpretable artificial intelligence. This solves the problem that technical solutions that pursue high prediction accuracy sacrifice the interpretability and credibility of the model, while technical solutions that maintain high interpretability cannot meet engineering requirements in terms of prediction accuracy.

[0007] The detailed technical solution of this invention is as follows: A method for predicting dam deformation based on physical information and interpretable artificial intelligence, the method comprising: S1. Obtain the historical deformation data sequence and related environmental quantity data sequence of the dam monitoring points, and perform preprocessing to obtain the preprocessed data; S2. Construct a hybrid prediction model with physical information enhancement. Its network architecture is an optimized physical information-gated long short-term memory network, and it is optimized using a hybrid loss function. S3. Use the preprocessed data and train the constructed PIG-LSTM model using the hybrid loss function; S4. Use the SHAP tool to perform in-depth analysis of the model's prediction results, quantify the contribution of each input variable, and decompose the causes of the predicted values ​​into physical-data contributions. Finally, integrate the prediction results with the explanatory analysis report and present them to the user.

[0008] Furthermore, the historical deformation data sequence and related environmental quantity data sequence of the dam monitoring points include three data sequences: daily horizontal displacement monitoring data sequence for each monitoring point on the dam crest during the required time period, upstream reservoir water level data sequence for the same time period, and temperature monitoring data sequence at the monitoring point location. The preprocessing includes: filling in missing values ​​in the acquired data sequence using Lagrange interpolation and performing Z-score standardization on the data sequence.

[0009] Furthermore, the network architecture of the physically-enhanced hybrid prediction model is an optimized physically-gated long short-term memory network, namely the PIG-LSTM model. The PIG-LSTM model introduces a "physical modulation factor" into the forget gate calculation of the standard LSTM. To dynamically adjust the model's memory level based on real-time load changes, as follows: (1) Define the load variation degree At each time step t, a dimensionless parameter is first calculated to measure the drastic changes in water pressure and temperature stress. The result is obtained by weighted summation of the absolute values ​​of the changes in water pressure and temperature stress between the current moment and the previous moment, as calculated below: (1); In formula (1), and These are the normalized changes in water pressure and temperature stress, respectively. and Preset weights.

[0010] (2) Calculate the physical modulation factor According to the degree of load variation Calculations show that the two are inversely proportional: (2); (3) Modulate the forget gate and output the original forget gate calculated by the standard LSTM. With physical modulation factor Multiplying these together yields the final forget gate output. : (3); The input to the PIG-LSTM model is a time series window. The input vector at each time step contains three features: [current day's reservoir water level, current day's temperature, previous day's displacement]. This input vector is obtained by constructing a time window through sliding sampling based on the upstream reservoir water level data sequence, temperature monitoring data sequence, and daily horizontal displacement monitoring data sequence obtained and preprocessed in S1. The output of the PIG-LSTM model is a single neuron that predicts the displacement value for the next day.

[0011] Furthermore, a hybrid loss function combining data-driven and physical information constraints was employed during training, specifically including: Data-driven loss term The mean squared error (MSE) is used for calculation. (4); In formula (4), For the sample size, It is the predicted displacement of the i-th sample by the PIG-LSTM model. This is the actual monitored displacement of the sample; Physical information constraint loss term Based on the residuals constructed from the elastic, hydraulic, or thermodynamic equations describing the load-response relationship of a dam, under water pressure... and temperature stress Deformation under action Resistance to dam body It should satisfy equilibrium, and its physical residuals Represented as: (5); In formula (5), k is the equivalent stiffness coefficient characterizing the structural response of the dam; components The water pressure load has a magnitude that is related to h. 2 Proportional; Temperature stress Due to temperature change And derived from the coefficient of thermal expansion of the material, The values ​​represent the dam deformation predicted by the PIG-LSTM model. That is, the residual. The L2 norm; This invention employs a dynamic weighting method to optimize the hybrid loss function: (6); In formula (6), the weight coefficients α(t) and β(t) are functions that change with the training round t. To achieve the "data-first, physical correction" ordered training strategy proposed in this invention, the function is specially designed as a linear annealing strategy, specifically implemented as follows: (7); In formula (7), t is the current training round; The total number of rounds for the weight growth phase is set to half the total number of training rounds; The final target weight for β.

[0012] The hybrid loss function is composed of a weighted average of a data-driven loss term and a physical information constraint loss term. The former ensures prediction accuracy, while the latter ensures physical consistency. This setup ensures that the model is primarily driven by the data-driven loss during the early stages of training. Dominant, as training progresses, physical constraints The influence of the smooth enhancement is thus guided to the model to converge stably to the physically consistent optimal solution.

[0013] Furthermore, S3 specifically includes: using the Adam optimizer to train the PIG-LSTM model by minimizing the dynamically changing hybrid loss function L_total(t) through backpropagation algorithm. This strategy can adaptively balance data fitting accuracy and physical constraints at different stages of training.

[0014] Furthermore, the physical-data contribution decomposition is achieved through the following steps: The first step is to use the SHAP tool to analyze the complete mixed prediction model and obtain the total contribution of each input variable to the final predicted value. ; The second step is to construct a framework that contains only elements used to define physical residuals. The simplified mechanical equilibrium equations of the pure physical model are also analyzed using SHAP to obtain the contribution values ​​of each input variable in the pure physical case. ; The third step is to calculate the difference between the total contribution value and the pure physical contribution value. This is defined as "data pattern correction contribution"; Finally, and The relative proportions are presented as the final decomposition result.

[0015] In another aspect of the invention, a system is provided for implementing a dam deformation prediction method based on physical information and interpretable artificial intelligence, the system comprising: a data acquisition module, a central server, and a user terminal module; The data acquisition module uses on-site sensors at the dam to monitor the data stream and transmit it to the central server in real time. The central server is responsible for storing, processing, and analyzing the data, and transmitting the analysis results to the user terminal module; The user terminal module is displayed using a web dashboard.

[0016] Specifically, the central server includes: a processing module and a storage module; The processing module includes a PIML model and an XAI engine. The PIML model is a physical information-enhanced hybrid prediction model, PIG-LSTM, which loads the model file and receives data from the monitoring database as input. It then calculates and generates predicted values ​​for future dam deformation. The XAI engine is an interpretability analysis module, implemented using the SHAP tool. It performs attribution analysis on the predicted values ​​generated by the PIML model, quantifies the contribution of each input variable, and achieves "physical-data contribution decomposition," ultimately generating an interpretation report. The storage module includes a monitoring database and model files; the monitoring database is used to store historical and real-time monitoring data input by the data acquisition module, and provides datasets for model training and input data for real-time prediction; the model files are used to store the trained and optimized hybrid prediction model so that the processing module can call it at any time.

[0017] In another aspect of the invention, an electronic device is also provided, comprising: At least one processor; and A memory that stores instructions that, when executed by the at least one processor, cause the at least one processor to perform a dam deformation prediction method based on physical information and interpretable artificial intelligence as described above.

[0018] In another aspect of the invention, a computer-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform a dam deformation prediction method based on physical information and interpretable artificial intelligence as described above.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention provides a method and system for predicting dam deformation based on physical information and interpretable artificial intelligence. By introducing physical constraints into the loss function and innovatively embedding physical laws into the basic unit structure of the neural network, a deep integration of data-driven and physical mechanisms is achieved. This not only ensures that the prediction results conform to the laws of engineering mechanics on a macroscopic level and avoids predictions that violate common sense in physics, but also improves the accuracy and robustness under normal working conditions. More importantly, the deep structural integration enables the model to make a faster dynamic response that is more in line with physical logic when facing sudden load changes such as floods and temperature changes, solving the technical problem of insufficient dynamic response capability of traditional data-driven models.

[0020] (2) The present invention provides a dam deformation prediction method and system based on physical information and interpretable artificial intelligence. It proposes an interpretation method of "physical-data contribution decomposition" for the prediction results. This method not only quantifies the key driving factors affecting dam behavior, but also reveals their dynamic changes, providing engineering technicians with in-depth engineering insights. It can directly assist in understanding the working state of dams, diagnosing faults and tracing causes. This technology of making the decision-making logic of the "black box" model explicit enables engineers not only to know the prediction results of the model, but also to understand its generation logic, fundamentally enhancing the credibility of the model and solving the core technical obstacles of artificial intelligence in safety-critical fields. Attached Figure Description

[0021] Figure 1 This is a flowchart of a dam deformation prediction method based on physical information and interpretable artificial intelligence as described in this invention.

[0022] Figure 2 This is a schematic diagram of the visualization interface for the interpretability analysis and result generation steps in Embodiment 1 of the present invention.

[0023] Figure 3 This is a schematic diagram of the hardware and software deployment architecture of the system described in Embodiment 1 of the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terms used here are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0028] Embodiment 1 Refer Figure 1 , this embodiment provides a dam deformation prediction method based on physical information and interpretable artificial intelligence. Taking the deformation prediction of the horizontal direction of the dam crest of a concrete gravity dam downstream as the application scenario, the method specifically includes: S1. Data acquisition and preprocessing: Obtain the historical deformation data sequence of the dam monitoring point and the relevant environmental quantity data sequence, and perform preprocessing to obtain the preprocessed data; For data acquisition and preprocessing, from the automated monitoring system of the dam, three data sequences are obtained, namely, the daily horizontal displacement monitoring data sequence of a certain measuring point at the dam crest from January 1, 2018 to December 31, 2022, the corresponding upstream reservoir water level data sequence within the same time period, and the temperature monitoring data sequence at the measuring point location.

[0029] The preprocessing includes: using the Lagrange interpolation method to fill the missing values in the obtained data sequence, and performing Z-score standardization processing on the data sequence.

[0030] S2: Construct a physically informed enhanced hybrid prediction model; Construct a hybrid prediction model, which is trained through a hybrid loss function that includes a data-driven loss term and a physically informed constraint loss term. The physically informed constraint loss term is constructed based on the physical laws related to the dam. The basic architecture of the hybrid prediction model is one of the long short-term memory network LSTM, the temporal convolutional network TCN, or the gated recurrent unit GRU.

[0031] The network architecture of the physically informed enhanced hybrid prediction model described in the present invention is an optimized physically informed gated long short-term memory network, namely the PIG-LSTM model, hereinafter referred to as the "hybrid prediction model" for short. It should be noted that the "hybrid" feature here is mainly reflected in the use of a hybrid loss function that combines data-driven and physically informed constraints in its training process, rather than a structural combination of multiple different models. The PIG-LSTM model can also be constructed based on other time series networks, such as the temporal convolutional network TCN or the gated recurrent unit GRU.

[0032] The PIG-LSTM model introduces a "physical modulation factor" into the forget gate computation of standard LSTM. To dynamically adjust the model's memory level based on real-time load changes, as follows: (1) Define “load variation degree” At each time step t, a dimensionless parameter is first calculated to measure the drastic changes in water pressure and temperature stress. This parameter is obtained by weighted summation of the absolute values ​​of the changes in water pressure and temperature stress between the current moment and the previous moment, as calculated below: (1); In formula (1), and These are the normalized changes in water pressure and temperature stress, respectively. and Preset weights.

[0033] (2) Calculate the "physical modulation factor" This factor is based on the degree of load variation. Calculations show that the two are inversely proportional: (2); (3) Modulate the forget gate: output the original forget gate calculated by the standard LSTM. With physical modulation factor Multiplying these together yields the final forget gate output. : (3); The above method achieves a clear physical logic: When the load changes drastically A large positive number leads to a modulation factor It will be close to 0. This will make the final forget gate output... It is also close to 0, forcing the model to "forget" historical information in order to quickly adapt to new load conditions.

[0034] When the physical state is stable Approaching 0 results in a modulation factor It is close to 1. At this point... The forgetting gate is largely unaffected by physical information and is primarily driven by data, thus maintaining the stability of memory.

[0035] The input to the PIG-LSTM model is a time series window. The input vector at each time step contains three features: [current day's reservoir water level, current day's temperature, previous day's displacement]. This input vector is obtained by constructing a time window through sliding sampling based on the upstream reservoir water level data sequence, temperature monitoring data sequence, and daily horizontal displacement monitoring data sequence obtained and preprocessed in S1. The output of the PIG-LSTM model is a single neuron that predicts the displacement value for the next day.

[0036] Definition of hybrid loss function: Data-driven loss term The mean squared error (MSE) is used for calculation, and the formula is as follows: (4); in, For the sample size, It is the predicted displacement of the i-th sample by the PIG-LSTM model. This is the actual monitored displacement of the sample.

[0037] Physical information constraint loss term The physical information constraint loss term is a residual constructed based on the elasticity, hydraulics, or thermodynamics equations describing the load-response relationship of the dam. In this embodiment, the loss term is defined as the residual of the simplified mechanical equilibrium equation of the dam cross-section. The dam body is considered as a cantilever beam under water pressure... and temperature stress Deformation under action Resistance to dam body (simplified to) The physical residuals should satisfy equilibrium. It can be represented as: (5); That is, the residual. The L2 norm. Where k is the equivalent stiffness coefficient characterizing the dam structure response; components The water pressure load is based on the reservoir water level h and the hydrostatic pressure formula ( The total pressure acting per unit width of the dam body, calculated from h, is related to h. 2 Proportional. Temperature stress can be caused by temperature changes. And derived from the coefficient of thermal expansion of the material, The values ​​represent the dam deformation predicted by the PIG-LSTM model.

[0038] This paper uses a dynamic weighting method to optimize the hybrid loss function: (6); Wherein, the weight coefficients α(t) and β(t) are functions that change with the training round t. To implement the "data-first, physical correction" ordered training strategy proposed in this invention, the functions are specifically designed as a linear annealing strategy, as follows: (7); In formula (7), t is the current training round; The total number of rounds for the weight growth phase is set to half the total number of training rounds; The final target weight for β is, for example, 0.5.

[0039] This setup causes the model to be primarily subject to data-driven loss during the early stages of training (when t is relatively small). Dominant, as training progresses, physical constraints The influence of the smooth enhancement is thus guided to the model to converge stably to the physically consistent optimal solution.

[0040] S3. Model Training and Prediction: The PIG-LSTM model is trained using the preprocessed data and the hybrid loss function described above. Model training and prediction used preprocessed data from 2018 to 2021 as the training set and data from 2022 as the test set. The PIG-LSTM model was trained using the Adam optimizer, minimizing the dynamically changing mixed loss function L_total(t) via backpropagation. After training, the model was deployed to predict daily horizontal displacements over the next 7 days.

[0041] S4. Interpretability Analysis and Result Generation: Use the trained model to generate future dam deformation predictions; The SHAP tool is used to perform in-depth analysis of the model's prediction results, quantify the contribution of each input variable, and decompose the causes of the prediction values ​​into physical-data contributions. Finally, the prediction results are integrated with the explanatory analysis report and presented to the user. Specifically, an interpretability analysis is performed on the predicted values ​​to generate an interpretive report that includes the quantification of key influencing factors and the decomposition of physical-data contributions. Interpretability analysis and result generation were performed using the SHAP (SHapley Additive exPlanations) analysis tool to interpret the trained model. When the model predicted a displacement increment of +2.5 mm on the third day, the SHAP analysis module generated the following result: Figure 2 The explanation report shown. The physical-data contribution decomposition is achieved through the following steps: The first step is to use the SHAP tool to analyze the complete mixed prediction model and obtain the total contribution of each input variable (water level, temperature, etc.) to the final predicted value. .

[0042] The second step is to construct a framework that contains only elements used to define physical residuals. The simplified mechanical equilibrium equations of the pure physical model are also analyzed using SHAP to obtain the contribution values ​​of each input variable in the pure physical case. .

[0043] The third step is to calculate the difference between the total contribution value and the pure physical contribution value. This is defined as 'data pattern correction contribution'.

[0044] Finally, and The relative proportions are presented as the final decomposition result.

[0045] The visualization is presented in a web dashboard application. The main interface displays historical displacement and predicted displacement for the next 7 days as a line graph. When the user hovers the mouse pointer over any predicted data point, an information box pops up. The layout and content of this information box strictly follow the attached... Figure 2 The diagram is shown below: the top displays "Predicted Displacement: +2.5mm"; the middle section is a bar chart showing "Reservoir Water Level Contribution: +1.8mm", "Temperature Contribution: -0.5mm", and "Historical State Contribution: +1.2mm"; the bottom is a pie chart showing "Physical Law Dominance: 80%" and "Data Pattern Correction: 20%".

[0046] Example 2 This embodiment provides a system for implementing a dam deformation prediction method based on physical information and interpretable artificial intelligence. The overall system architecture is as follows: Figure 3 As shown, the system includes: a data acquisition module, a central server, and a user terminal module; The data acquisition module uses on-site sensors at the dam to monitor the data stream and transmit it to the central server in real time. The central server is the core of this system, responsible for data storage, processing, and analysis, and includes processing and storage modules.

[0047] The storage module includes: Monitoring Database: This database stores historical and real-time monitoring data received from the data acquisition module. It provides the dataset for model training and input data for real-time prediction.

[0048] Model file: Used to store the physical information-enhanced hybrid prediction model obtained after training and optimization in the S3 step, with fixed network weights and parameters, so that the processing module can call it at any time.

[0049] The processing module includes: PIML (Physics-Informed Machine Learning) model: This refers to the physical information-enhanced hybrid prediction model (such as the PIG-LSTM model) constructed in Example 1. Its function is to load the model file, receive data from the monitoring database as input, and generate predicted values ​​for future dam deformation through calculation.

[0050] The XAI (Explainable AI) engine, also known as the interpretability analysis module in Example 1, can be implemented using tools such as SHAP. Its function is to perform attribution analysis on the predicted values ​​generated by the PIML model, quantify the contribution of each input variable, and achieve an innovative "physical-data contribution decomposition," ultimately generating an explanatory report.

[0051] The central server transmits the analysis results to the user terminal module, which uses a web dashboard.

[0052] Example 3 This embodiment also provides an electronic device, including: At least one processor; and A memory that stores instructions that, when executed by the at least one processor, cause the at least one processor to perform a dam deformation prediction method based on physical information and interpretable artificial intelligence as described above.

[0053] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable computing device, consumer electronic device, etc.

[0054] Example 4 This embodiment also provides a machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform a dam deformation prediction method based on physical information and interpretable artificial intelligence as described above.

[0055] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0056] In this case, the program code read from the readable medium itself can perform the functions of any of the above embodiments, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification.

[0057] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0058] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0059] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for predicting dam deformation based on physical information and interpretable artificial intelligence, characterized in that, The method includes: S1. Obtain the historical deformation data sequence and related environmental quantity data sequence of the dam monitoring points, and perform preprocessing to obtain the preprocessed data; S2. Construct a hybrid prediction model with physical information enhancement. Its network architecture is an optimized physical information-gated long short-term memory network, and it is optimized using a hybrid loss function. S3. Use the preprocessed data and train the constructed PIG-LSTM model using the hybrid loss function; S4. The SHAP tool is used to perform in-depth analysis of the model's prediction results, quantify the contribution of each input variable, and decompose the causes of the predicted values ​​into physical-data contributions. Finally, the prediction results are integrated with the explanatory analysis report and presented to the user.

2. The dam deformation prediction method based on physical information and interpretable artificial intelligence according to claim 1, characterized in that, The historical deformation data sequence and related environmental quantity data sequence of the dam monitoring points include three data sequences: daily horizontal displacement monitoring data sequence of each monitoring point on the dam crest for the required time period, upstream reservoir water level data sequence within the same time period, and temperature monitoring data sequence at the monitoring point location. The preprocessing includes: filling in missing values ​​in the acquired data sequence using Lagrange interpolation and performing Z-score standardization on the data sequence.

3. The dam deformation prediction method based on physical information and interpretable artificial intelligence according to claim 2, characterized in that, The network architecture of the physically-enhanced hybrid prediction model is an optimized physically-gated long short-term memory network, namely the PIG-LSTM model. The PIG-LSTM model introduces a physical modulation factor into the forget gate calculation of the standard LSTM. To dynamically adjust the model's memory level based on real-time load changes, as follows: (1) Define the load variation degree At each time step t, a dimensionless parameter is first calculated to measure the drastic changes in water pressure and temperature stress. The result is obtained by weighted summation of the absolute values ​​of the changes in water pressure and temperature stress between the current moment and the previous moment, as calculated below: （1）； In formula (1), and These are the normalized changes in water pressure and temperature stress, respectively. and Preset weights; (2) Calculate the physical modulation factor According to the degree of load variation Calculations show that the two are inversely proportional: （2）； (3) Modulate the forget gate and output the original forget gate calculated by the standard LSTM. With physical modulation factor Multiplying these together yields the final forget gate output. : （3）； The input to the PIG-LSTM model is a time series window. The input vector at each time step contains three features: [current day's reservoir water level, current day's temperature, and previous day's displacement]. The output of the PIG-LSTM model is a single neuron that predicts the displacement value for the next day.

4. The dam deformation prediction method based on physical information and interpretable artificial intelligence according to claim 1, characterized in that, During training, a hybrid loss function combining data-driven and physical information-constrained approaches was employed, specifically including: Data-driven loss term Calculations are performed using mean square error. （4）； In formula (4), For the sample size, It is the predicted displacement of the i-th sample by the PIG-LSTM model. This is the actual monitored displacement of the sample; Physical information constraint loss term Based on the residuals constructed from the elastic, hydraulic, or thermodynamic equations describing the load-response relationship of a dam, under water pressure... and temperature stress Deformation under action Resistance to dam body It should satisfy equilibrium, and its physical residuals Represented as: （5）； In formula (5), k is the equivalent stiffness coefficient characterizing the structural response of the dam; components The water pressure load has a magnitude that is related to h. 2 Proportional; Temperature stress Due to temperature change And derived from the coefficient of thermal expansion of the material, The values ​​represent the dam deformation predicted by the PIG-LSTM model. That is, the residual. The L2 norm; Therefore, the dynamic weighting method is used to optimize the hybrid loss function as follows: （6）； In formula (6), the weight coefficients α(t) and β(t) are functions that change with the number of training rounds t. These functions are designed as a linear annealing strategy, specifically implemented as follows: （7）； In formula (7), t is the current training round; The total number of rounds for the weight growth phase is set to half the total number of training rounds; The final target weight for β.

5. The dam deformation prediction method based on physical information and interpretable artificial intelligence according to claim 3, characterized in that, Specifically, S3 includes: using the Adam optimizer to train the PIG-LSTM model by minimizing the hybrid loss function L_total(t) that dynamically changes with training through the backpropagation algorithm.

6. The dam deformation prediction method based on physical information and interpretable artificial intelligence according to claim 3, characterized in that, The physical-data contribution decomposition is achieved through the following steps: The first step is to use the SHAP tool to analyze the complete mixed prediction model and obtain the total contribution of each input variable to the final predicted value. ; The second step is to construct a framework that contains only elements used to define physical residuals. The simplified mechanical equilibrium equations of the pure physical model are also analyzed using SHAP to obtain the contribution values ​​of each input variable in the pure physical case. ; The third step is to calculate the difference between the total contribution value and the pure physical contribution value. ) is defined as the data pattern correction contribution; Finally, and The relative proportions are presented as the final decomposition result.

7. A system for implementing a dam deformation prediction method based on physical information and interpretable artificial intelligence, characterized in that, The system includes: a data acquisition module, a central server, and a user terminal module; The data acquisition module uses on-site sensors at the dam to monitor the data stream and transmit it to the central server in real time. The central server is responsible for storing, processing, and analyzing the data, and transmitting the analysis results to the user terminal module; The user terminal module is displayed using a web dashboard; The central server includes: a processing module and a storage module; The processing module includes a PIML model and an XAI engine. The PIML model is a constructed physical information-enhanced hybrid prediction model, PIG-LSTM, which loads the model file and receives data from the monitoring database as input to generate future dam deformation predictions. The XAI engine is an interpretability analysis module, implemented based on the SHAP tool, which performs attribution analysis on the predictions generated by the PIML model, quantifies the contribution of each input variable, decomposes the physical-data contribution, and finally generates an interpretation report. The storage module includes a monitoring database and model files; the monitoring database is used to store historical and real-time monitoring data input by the data acquisition module, and provides datasets for model training and input data for real-time prediction; the model files are used to store trained and optimized hybrid prediction models for use by the processing module.

8. An electronic device, characterized in that, The electronic device includes: processor; A memory on which computer programs that can run on the processor are stored; When the computer program is executed by the processor, it implements a dam deformation prediction method based on physical information and interpretable artificial intelligence as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images