Dam seepage prediction method based on improved long-short term memory deep learning model

Abstract

The invention provides a dam seepage prediction method based on an improved long-short term memory deep learning model. The method comprises the following steps: firstly, mining five major factors, namely upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall, which are strong in seepage of the dam from hydrological characteristics, meteorological characteristics and interactive characteristics of the hydrological characteristics, the meteorological characteristics and the interactive characteristics through a principal component analysis (PCA) method; and jointly forming a time sequence data set required by deep learning model training and testing. Secondly, constructing an improved long-short-term memory deep learning model; the innovation points of the model comprise an improved elite genetic algorithm for precise initialization of network model parameters of binary and real number hierarchical hybrid coding, collaborative elite tournament strategy selection and dynamic comprehensive fitness evaluation, and a network back propagation process parameter learning method based on the improved elite genetic algorithm. And thirdly, performing model training based on the constructed model and the data set, and completing seepage prediction. Finally, on the basis of actually monitored data, it is verified that the method provided by the invention has relatively strong advantages in prediction precision and generalization ability, and the method has relatively strong theoretical value for realizing intelligent and accurate prediction of dam seepage.

Description

Technical Field

[0001] The present invention belongs to the field of dam seepage, and specifically is a dam seepage prediction method based on an improved long short-term memory deep learning model. Background Art

[0002] In water conservancy projects, the safety of dams is crucial to flood control, irrigation, hydropower generation and ecological environmental protection. Dam seepage is one of the key factors affecting dam stability and safety. Accurate prediction of dam seepage helps to take preventive and maintenance measures in time to prevent disasters. However, dam seepage is affected by many factors, including upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall, etc. The complex dynamic changes of these factors increase the difficulty of prediction.

[0003] Traditional seepage prediction methods often rely on empirical models or simple statistical methods, which are difficult to fully capture the complex spatiotemporal relationship of seepage changes. With the rapid development of artificial intelligence technology, the long short-term memory deep learning model, as a special recursive neural network, has demonstrated powerful capabilities in processing time series data. The network can remember and capture important information in long time series and is suitable for predicting dynamic changes in seepage.

[0004] However, the performance of LSTM networks is highly dependent on the selection of hyperparameters and the design of network structure. Traditional manual parameter adjustment methods are not only time-consuming and labor-intensive, but also difficult to find the global optimal solution. Genetic algorithms, as a global optimization algorithm, can automatically optimize the hyperparameters of the network by simulating the natural evolution process, but conventional genetic algorithms are prone to fall into local optimality and have a slow convergence speed.

[0005] Therefore, the present invention proposes a dam seepage prediction method based on an improved long short-term memory deep learning model. First, the five main factors with the strongest correlation with dam seepage - upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall - are mined from hydrological characteristics, meteorological characteristics and their interactive characteristics by principal component analysis (PCA) method, and by adding whether seepage annotation information occurs, a time series data set required for deep learning model training and testing is formed together. Secondly, an improved long short-term memory deep learning model is constructed. The innovative points of the model include fast initialization of network model parameters of binary and real hierarchical mixed coding, improved elite genetic algorithm of collaborative elite tournament strategy and dynamic comprehensive fitness evaluation, and network back propagation process parameter learning method based on improved elite genetic algorithm. Again, model training is carried out based on the constructed model and data set, and seepage prediction is completed. Finally, the actual monitored data verifies that the method proposed by the invention has strong advantages in prediction accuracy and generalization ability, and has strong theoretical value for realizing intelligent and accurate prediction of dam seepage. Summary of the invention

[0006] Purpose of the invention: In order to achieve intelligent and accurate prediction of dam seepage, the present invention introduces a dam seepage prediction method based on an improved long short-term memory deep learning model to provide theoretical support for timely prevention of dam seepage.

[0007] Technical solution: To achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a dam seepage prediction method based on an improved long short-term memory deep learning model, comprising the following steps:

[0008] 1. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the method comprises the following steps:

[0009] (1) Construct a deep learning dataset of dam seepage, including the collection of hydrological characteristics, meteorological characteristics and their interactive characteristics associated with dam seepage; based on the principal component analysis (PCA) method, the five main factors with the strongest correlation with dam seepage - upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall - are mined; semantic annotation information on whether seepage occurs is added, and the five main factors and seepage semantic information are combined to form a time series dataset required for deep learning model training and testing;

[0010] (2) Construct an improved long short-term memory deep learning model, including fast initialization of network model parameters with binary and real hierarchical mixed coding, an improved elite genetic algorithm with collaborative elite tournament strategy and dynamic comprehensive fitness evaluation, and a network back-propagation process parameter learning method based on the improved elite genetic algorithm;

[0011] (3) Perform model training based on the constructed model and dataset, and complete seepage prediction.

[0012] 2. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the data set construction process in step (1) includes:

[0013] (1) Construct a set of 13 observation elements related to seepage, including hydrological characteristics, meteorological characteristics and their interactive characteristics, including element observation date, upstream water level, downstream water level, temperature, rainfall, water level difference, seasonal water level, cumulative rainfall, rainfall intensity, temperature change rate, extreme temperature, upstream water level and rainfall accumulation, and temperature and rainfall accumulation;

[0014] (2) For the 13 constructed observation elements, the PCA method was used to mine the main elements and determine the five main factors that have the greatest impact on dam seepage:

[0015] Upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall. In order to improve the accuracy of the model for seepage warning, the seepage semantic annotation information is added on this basis. The sample model D is shown in formula (1):

[0016]

[0017] Among them, D up represents the upstream water level, D down represents the downstream water level, D diff Indicates the upstream and downstream water level difference, D tem Indicates temperature, D rain Indicates rainfall, D tag Represents the percolation semantic annotation information. n represents the number of samples in the dataset D. The final constructed dataset includes 5000 samples, including 4000 samples in the training set, 500 samples in the test set, and 500 samples in the validation set.

[0018] 3. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the rapid initialization of the network model parameters of binary and real hierarchical mixed coding in step (2) includes:

[0019] (1) Initialize the value ranges of five parameters in the network model based on the pre-trained model: time window value, number of neurons, number of network layers, learning rate, and dropout value;

[0020] (2) Discretize the range of learning rate and Dropout value;

[0021] (3) Binary code the boundary values ​​of the five parameter value intervals, and encode the real numbers in the intervals. The binary coding formula is shown in formula (2).

[0022]

[0023] Where x represents the real number to be converted. and Respectively represent the conversion of the integer part and the decimal part, r k Indicates the result of each modulo operation, indicating a binary digit, m is the number of binary digits, and r0 is the lowest digit. j is the rounding result when the decimal part is converted to binary, and p is the number of digits in the decimal part. The real number encoding formula is shown in formula (3):

[0024]

[0025] Among them, actual_value is the actual value of the parameter, and scale_factor is the normalization coefficient;

[0026] (4) Based on the binary and real number layered hybrid coding mode, optimization is performed within the initial range of time window value, number of neurons, number of network layers, learning rate and dropout value;

[0027] (5) Based on the optimization results, the learning rate and Dropout value are fine-tuned based on real number coding. The adjustment formula is as follows.

[0028] p=p0+(1+α×δ) (4)

[0029] Among them, p represents the adjusted parameter value, p0 represents the parameter value obtained by optimization, α is the adjustment coefficient, and its value range is [0.9, 1.1], and δ is the random disturbance term, which obeys the uniform distribution in the interval [-0.1, 0.1].

[0030] 4. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the collaborative elite tournament strategy and the improved elite genetic algorithm of dynamic comprehensive fitness evaluation in step (2) include:

[0031] (1) Construct a dynamic comprehensive fitness evaluation method to accurately evaluate the advantages and disadvantages of the time window value, number of neurons, number of network layers, learning rate and dropout value in the network model;

[0032] (2) Construct an elite dynamic selection strategy based on tournaments to accelerate parameter optimization while retaining excellent parameters through intra-group competition;

[0033] (3) Construct a non-uniform mutation strategy that integrates multi-point crossover to integrate mutations during crossover evolution and ensure the safety of mutations;

[0034] 5. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the dynamic comprehensive fitness evaluation method constructed in step (4) includes:

[0035] (1) From the perspective of genetic algorithms, the performance indicators for evaluating the quality of parameters in network models are divided into fitness indicators and diversity indicators;

[0036] (2) In the parameter optimization process, the weights of the two types of indicators in the parameter evaluation process are adaptively and dynamically adjusted based on the diversity and fitness targets of each genetic iteration. The corresponding dynamic comprehensive fitness evaluation formula is:

[0037] F=w a ·S+w d ·D+ΔF(t) (5)

[0038] Among them, S is the comprehensive fitness index, D is the diversity index, and w a is the fitness weight, wd is the diversity weight, and ΔF(t) is the fitness increment based on the current iteration number t.

[0039] (3) In the initial iteration, 20% of the hyperparameters are randomly selected as diversity indicators for evaluation to reduce the risk of the population falling into a local optimal solution and improve the overall search capability;

[0040] (4) The fitness score is updated according to the genetic speed and fitness evaluation of each iteration to promote the continuous improvement of the elite characteristics of the population. The fitness update formula is:

[0041] F′(t+1)=F(t)+ΔF(t) (6)

[0042] Where ΔF(t) is the fitness increment based on the current iteration number t:

[0043] 6. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the elite dynamic selection strategy based on the tournament constructed in step (6) includes:

[0044] (1) At the beginning of the iteration, the population is randomly divided into groups, and the tournament selection of genetic individuals is carried out based on the individuals with the strongest fitness in the tournament group. During the iterative evolution process, the number of tournament groups and the number of individuals in each group are adaptively adjusted according to the genetic speed of each iteration. By expanding the number of group members, better winners in the tournament are selected to speed up the genetic speed. The formula for dynamically adjusting the group size is:

[0045]

[0046] Among them, k(t) is the grouping size of the Tth generation, and T is the total number of evolutionary generations;

[0047] (2) Elite individual selection based on the fitness of the group tournament winner:

[0048]

[0049] (3)W is the set of all winners in the tournament group, E is the set of elite individuals, and F(x) represents the fitness score of individual x;

[0050] The elites and tournament winners form the population for the next generation of crossover and mutation operations;

[0051] 7. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the non-uniform mutation strategy of fusion multi-point crossover constructed in step (4) includes:

[0052] (1) In the mutation stage, a calculation formula for the probability of individual mutation based on the evolution process is constructed:

[0053]

[0054] Where P m (i) is the mutation probability of the i-th individual, P0 is the initial mutation probability, G is the current iteration number, G max is the maximum number of iterations, α is the adjustment factor used to control the amplitude of variation, and rand(0,1) is a random number in the interval [0,1]. When the individual variation probability value is greater than the set threshold, the gene position will mutate.

[0055] (1) The winner of the tournament randomly selects multiple crossover points for gene exchange operations. The genes of the parent individuals at the crossover points will be divided into c+1 segments and alternately combined to generate new offspring individuals. The formula is:

[0056] O=(P1[0:p1]+P2[p1:p2]+L+(-1) c P c+1 [p c :L]) (10)

[0057] Among them, O is the newly generated offspring individual, P i represents the i-th parent individual, and the symbol “+” represents the splicing of gene fragments;

[0058] (2) In the initial stage, a mutation threshold is set for each gene position, and a multi-point crossover operation is integrated before each mutation. The safety of the mutation population is guaranteed during the continuous evolution of the mutation, and the mutation probability is continuously updated. When the obtained individual mutation probability value is greater than the set threshold, the gene position mutates;

[0059] 8. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that the network back propagation process parameter learning method based on an improved elite genetic algorithm in step (2) is:

[0060] (1) Construct a network parameter optimization model based on an improved elite genetic algorithm, where the objective function is the loss function of the improved long short-term memory neural network model and the decision variable is the dynamic comprehensive fitness of the population individuals generated in the continuous iteration process;

[0061] (2) Constructing an improved long short-term memory neural network model loss function:

[0062]

[0063] Where y is the actual output, is the model prediction output, N is the number of samples;

[0064] (3) Based on the improved elite genetic algorithm, the optimal solution of network parameters is continuously learned and optimized, and the error of the parameters obtained by solving the optimization target constraint meets the set threshold.

[0065] Beneficial effects: The dam seepage prediction method based on the improved long short-term memory deep learning model constructed by the present invention can effectively solve the problem that traditional seepage prediction methods are difficult to capture complex spatiotemporal relationships. Since traditional prediction methods rely on empirical models and simple statistics, it is difficult to cope with the prediction needs of dynamic changes in multiple factors. Therefore, the research on intelligent prediction methods based on deep learning has important practical value. At the same time, the innovation of this method in the optimization of long short-term memory network hyperparameters also has important theoretical value. By introducing a fast initialization method of binary and real hierarchical mixed coding, a collaborative elite tournament strategy, and an improved elite genetic algorithm with dynamic comprehensive fitness evaluation, it not only significantly improves the prediction accuracy, but also realizes the automatic optimization of model parameters. Compared with the traditional manual parameter adjustment method, it can save a lot of labor costs and provide reliable technical support for intelligent monitoring of dam safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of the present invention for realizing prediction based on the combination of long short-term memory model and optimization algorithm.

[0067] Figure 2 This invention is based on the improved long short-term memory deep learning model structure diagram.

[0068] Figure 3 It is a schematic diagram of model performance data of an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0070] Combination Figure 1 Describing the technical details of the present invention, an embodiment of the present invention discloses a dam seepage prediction method based on an improved long short-term memory deep learning model, which mainly includes the following steps:

[0071] Step 1: Collect the seepage characteristics of the dam, construct an observation element set, use the principal component analysis (PCA) method to extract the main factors, and add the seepage semantic annotation information on this basis to construct a dam seepage deep learning dataset.

[0072] First, the basic observation data related to the seepage of the dam were collected and summarized, including the original monitoring data such as the upstream water level, downstream water level, temperature and rainfall. On this basis, 13 observation elements closely related to seepage were constructed through feature engineering: date, upstream water level, downstream water level, temperature, rainfall, water level difference, seasonal water level, cumulative rainfall, rainfall intensity, temperature change rate, extreme temperature, upstream water level and rainfall accumulation, temperature and rainfall accumulation. Among them, the water level difference is calculated by the real-time difference between the upstream and downstream water levels, the seasonal water level takes into account the periodic change characteristics of the water level, the cumulative rainfall counts the cumulative precipitation 15 days ago, the rainfall intensity reflects the rainfall in a unit time, the temperature change rate represents the dynamic change trend of the temperature, the extreme temperature records the daily maximum and minimum temperatures, and the interactive features such as upstream water level and rainfall accumulation, temperature and rainfall accumulation are used to characterize the coupling effect of multiple factors. For the 13 constructed observation elements, the principal component analysis (PCA) method is used for feature importance analysis and dimensionality reduction. Specifically, all features are first standardized to eliminate the dimension effect. Then the covariance matrix of the standardized feature matrix was calculated, and the eigenvalue decomposition was performed to obtain the eigenvalues ​​and corresponding eigenvectors. The features were ranked according to their importance according to the eigenvalues, and the principal components with a cumulative contribution rate of more than 85% were selected. By analyzing the composition of the principal components, the five main factors that most significantly affected the seepage of the dam were finally determined: upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall.

[0073] In order to improve the accuracy of the model's seepage warning, the semantic annotation information of seepage is added on the basis of the five main factors. Specifically, according to the seepage threshold (set as Q) in the historical monitoring data, the seepage state is binarized: when the actual seepage exceeds the threshold Q, it is marked as seepage (marked as 1); when the actual seepage is lower than the threshold Q, it is marked as no seepage (marked as 0). The threshold Q is determined by combining expert experience with the statistical analysis of historical data to select a critical value that can effectively distinguish normal seepage from abnormal seepage.

[0074] Finally, the time series data of the five main factors obtained by screening are aligned with the percolation semantic annotation information to construct a data set containing 5,000 samples. The constructed data set is divided into training set, validation set and test set in a ratio of 8:1:1, which are used for model training, validation and testing respectively.

[0075] Step 2: Propose a fast initialization method for network model parameters with binary and real layered mixed coding to achieve fast initialization of long short-term memory deep learning model parameters.

[0076] The proposed method for fast initialization of network model parameters with binary and real hierarchical hybrid coding is as follows: First, the initial value ranges of five key parameters, namely, time window value, number of neurons, number of network layers, learning rate and Dropout value, are determined based on the pre-trained model. Specifically, the time window value is limited to the interval [10,50], the number of neurons is limited to the interval [32,256], the number of network layers is limited to the interval [1,5], the learning rate is limited to the interval [0.0001,0.001], and the Dropout value is limited to the interval [0.1,0.5]. Secondly, the continuous value intervals of the learning rate and Dropout value are discretized. Considering the sensitivity of the learning rate to model training, the discretization step size of the learning rate is set to 0.0001; considering that the Dropout value is mainly used to prevent overfitting, its discretization step size is set to 0.01. Discretization helps to reduce the complexity of the parameter search space and improve the efficiency of parameter optimization. Again, binary coding is used to encode the boundary values ​​of the five parameter value intervals: time window value, number of neurons, number of network layers, learning rate, and Dropout value. The binary coding formula is as follows:

[0077]

[0078] Where x represents the real number to be converted. and Respectively represent the conversion of the integer part and the decimal part, r k Indicates the result of each modulo operation, indicating a binary digit, m is the number of binary digits, and r0 is the lowest digit. j is the rounding result when the decimal part is converted to binary, and p is the number of digits in the decimal part. Real numbers in this interval are encoded with real numbers. Real number encoding can maintain the continuity of parameters and is conducive to more precise parameter adjustment. The real number encoding formula is as follows:

[0079]

[0080] Among them, actual_value is the actual value of the parameter, and scale_factor is the normalization coefficient.

[0081] Based on the above coding, a binary and real number layered mixed coding mode is used to search for optimization in the initial parameter space. Specifically, the encoded parameters are combined to form a parameter vector, and the optimal initial value is searched in the parameter space through the optimization algorithm. Finally, the learning rate and Dropout value are fine-tuned based on the optimization results and real number coding. The adjustment formula is as follows:

[0082] p=p0+(1+α×δ) (14)

[0083] Among them, p represents the adjusted parameter value, p0 represents the parameter value obtained by optimization, α is the adjustment coefficient, and its value range is [0.9, 1.1], and δ is the random disturbance term, which obeys the uniform distribution in the interval [-0.1, 0.1].

[0084] Step 3: Propose an improved elite genetic algorithm with collaborative elite tournament strategy and dynamic comprehensive fitness evaluation, including a dynamic comprehensive fitness evaluation method, a tournament-based elite dynamic selection strategy, and a non-uniform mutation strategy integrating multi-point crossover.

[0085] The proposed dynamic comprehensive fitness evaluation method is as follows: First, from the perspective of genetic algorithms, the performance indicators for evaluating the quality of network model parameters are divided into two categories: fitness indicators and diversity indicators. In the parameter optimization process, based on the diversity and fitness targets of each genetic iteration, the weights of these two types of indicators in the parameter evaluation process are adaptively and dynamically adjusted. The specific dynamic comprehensive fitness evaluation formula is:

[0086] F=w a ·S+w d ·D+ΔF(t) (15)

[0087] Among them, s is the comprehensive fitness index, D is the diversity index, and w a is the fitness weight, w d is the diversity weight, and ΔF(t) is the fitness increment based on the current iteration number t.

[0088] Secondly, in the early stage of iteration, in order to reduce the risk of the population falling into the local optimal solution and improve the overall search ability, 20% of the hyperparameters are randomly selected as diversity indicators for evaluation.

[0089] Finally, the fitness score is updated according to the genetic speed and fitness evaluation of each generation to promote the continuous improvement of the elite characteristics of the population. The fitness update formula is:

[0090] F′(t+1)=F(t)+ΔF(t) (16)

[0091] Where ΔF(t) is the fitness increment based on the current iteration number t.

[0092] The proposed tournament-based elite dynamic selection strategy is as follows: First, at the beginning of the iteration, the population is randomly divided into groups, and the individuals with the best fitness in the group are selected based on the tournament method. During the iterative evolution process, the number of groups in the tournament and the number of individuals in each group are adaptively adjusted according to the genetic speed of each iteration. By expanding the number of group members, better winners in the tournament can be selected, thereby speeding up the genetic speed. The formula for dynamically adjusting the group size is:

[0093]

[0094] Among them, k(t) is the grouping size of the tth generation, and T is the total number of evolutionary generations;

[0095] Elite individual selection is done based on the fitness of the group tournament winners:

[0096]

[0097] Where W is the set of all winners in the tournament group, E is the set of elite individuals, and F(x) represents the fitness score of individual x; the elites and tournament winners form the population for the next generation of crossover and mutation operations;

[0098] The proposed non-uniform mutation strategy integrating multi-point crossover is as follows: First, in the mutation stage, an individual mutation formula based on the evolution process is constructed:

[0099]

[0100] Where P m (i) is the mutation probability of the i-th individual, P0 is the initial mutation probability, G is the current iteration number, G max is the maximum number of iterations, α is the adjustment factor used to control the amplitude of variation, and rand(0,1) is a random number in the interval [0,1]. When the individual variation probability value is greater than the set threshold, the gene position will mutate.

[0101] Secondly, multiple crossover points are randomly selected based on the tournament for gene exchange operations. The genes of the parent individuals at the crossover points will be divided into c+1 segments and alternately combined to generate new offspring individuals. The formula is:

[0102] O=(P1[0:p1]+P2[p1:p2]+L+(-1) c P c+1 [p c :L]) (20)

[0103] Among them, O is the newly generated offspring individual, P i represents the i-th parent individual, and the symbol “+” represents the splicing of gene fragments.

[0104] Finally, in the initial stage, a mutation threshold is set for each gene position, and a multi-point crossover operation is integrated before each mutation. The safety of the mutation population is guaranteed during the continuous evolution of the mutation, and the mutation probability is continuously updated. When the obtained individual mutation probability value is greater than the set threshold, the gene position mutates.

[0105] Step 4: Propose a network back-propagation process parameter learning method based on an improved elite genetic algorithm to achieve continuous learning and optimization of the parameters of the long short-term memory deep learning model.

[0106] The proposed network back-propagation process parameter learning method based on the improved elite genetic algorithm is as follows: First, a network parameter optimization model based on the improved elite genetic algorithm is constructed. In this model, the loss function of the improved long short-term memory deep learning model is used as the objective function, and the dynamic comprehensive fitness of the population individuals generated in the continuous iteration process is used as the decision variable. This design can transform the network parameter optimization problem into an optimization problem based on the improved elite genetic algorithm, thereby improving the efficiency and accuracy of parameter learning. Specifically, the loss function of the improved long short-term memory neural network model is:

[0107]

[0108] Where y is the actual output, is the model prediction output, and N is the number of samples. The loss function adopts the mean square error form, which can effectively measure the deviation between the model prediction value and the actual value, and provide an accurate evaluation index for parameter optimization. In the parameter learning process, the present invention uses an improved elite genetic algorithm to continuously learn and optimize the optimal solution of network parameters. When the error of the parameter obtained by solving the optimization objective constraint meets the set threshold, the algorithm stops iterating and outputs the optimal parameter configuration.

[0109] Step 5: Network model training and percolation prediction

[0110] The preprocessed training data is input into the improved LSTM deep learning model, and a certain number of training rounds (1000 iterations) is set. The network weights are updated in each iteration, and the model is continuously optimized until the stop condition is met or the predetermined maximum number of iterations is reached. The trained model is verified using independent test data to test its generalization ability. Finally, the model is used to predict the future seepage of the dam, and the results are recorded for further analysis.

[0111] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The method comprises the following steps: (1) Construct a deep learning dataset of dam seepage, including the collection of hydrological characteristics, meteorological characteristics and their interactive characteristics associated with dam seepage; based on the principal component analysis (PCA) method, the five main factors with the strongest correlation with dam seepage - upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall - are mined; semantic annotation information on whether seepage occurs is added, and the five main factors and seepage semantic information are combined to form a time series dataset required for deep learning model training and testing; (2) Construct an improved long short-term memory deep learning model, including fast initialization of network model parameters with binary and real hierarchical mixed coding, an improved elite genetic algorithm with collaborative elite tournament strategy and dynamic comprehensive fitness evaluation, and a network back-propagation process parameter learning method based on the improved elite genetic algorithm; (3) Perform model training based on the constructed model and dataset, and complete seepage prediction.

2. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The data set construction process in step (1) includes: (1) Construct a set of 13 observation elements related to seepage, including hydrological characteristics, meteorological characteristics and their interactive characteristics, including element observation date, upstream water level, downstream water level, temperature, rainfall, water level difference, seasonal water level, cumulative rainfall, rainfall intensity, temperature change rate, extreme temperature, upstream water level and rainfall accumulation, and temperature and rainfall accumulation; (2) For the 13 constructed observation elements, the PCA method is used to mine the main elements and determine the five main factors that have the greatest impact on the seepage of the dam: upstream water level, downstream water level, upstream and downstream water level difference, temperature and rainfall. In order to improve the accuracy of the model for seepage warning, the semantic annotation information of seepage is added on this basis. The sample model D is shown in formula (1): Among them, D up represents the upstream water level, D down represents the downstream water level, D diff Indicates the upstream and downstream water level difference, D tem Indicates temperature, D rain Indicates rainfall, D tag represents the percolation semantic annotation information, n represents the number of samples in the dataset D, and the final constructed dataset includes 5000 samples, including 4000 samples in the training set, 500 samples in the test set, and 500 samples in the validation set.

3. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The network model parameter rapid initialization of binary and real number layered mixed coding in step (2) includes: (1) Initialize the value ranges of five parameters in the network model based on the pre-trained model: time window value, number of neurons, number of network layers, learning rate, and dropout value; (2) Discretize the range of learning rate and Dropout value; (3) Binary code the boundary values ​​of the five parameter value intervals, and encode the real numbers in the intervals. The binary coding formula is shown in formula (2): Where x represents the real number to be converted. and Respectively represent the conversion of the integer part and the decimal part, r k Indicates the result obtained by each modulus, indicating a binary digit, m is the number of binary digits, and r0 is the lowest digit, b j is the rounding result when the decimal part is converted to binary, p is the number of digits of the decimal part, and the real number encoding formula is shown in formula (3): Among them, actual_value is the actual value of the parameter, and scale_factor is the normalization coefficient; (4) Based on the binary and real number layered hybrid coding mode, optimization is performed within the initial range of time window value, number of neurons, number of network layers, learning rate and dropout value; (5) Based on the optimization results, the learning rate and Dropout value are fine-tuned based on real number coding. The adjustment formula is as follows: p=p0+(1+α×δ) (4) Among them, p represents the adjusted parameter value, p0 represents the parameter value obtained by optimization, α is the adjustment coefficient, and its value range is [0.9, 1.1], and δ is the random disturbance term, which obeys the uniform distribution in the interval [-0.1, 0.1].

4. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The collaborative elite tournament strategy and the improved elite genetic algorithm for dynamic comprehensive fitness evaluation in step (2) include: (1) Construct a dynamic comprehensive fitness evaluation method to accurately evaluate the advantages and disadvantages of the time window value, number of neurons, number of network layers, learning rate and dropout value in the network model; (2) Construct an elite dynamic selection strategy based on tournaments to accelerate parameter optimization while retaining excellent parameters through intra-group competition; (3) Construct a non-uniform mutation strategy that integrates multi-point crossover to integrate mutations during crossover evolution and ensure the safety of mutations.

5. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The dynamic comprehensive fitness evaluation method constructed in step (4) includes: (1) From the perspective of genetic algorithms, the performance indicators for evaluating the quality of parameters in network models are divided into fitness indicators and diversity indicators; (2) In the parameter optimization process, the weights of the two types of indicators in the parameter evaluation process are adaptively and dynamically adjusted based on the diversity and fitness targets of each genetic iteration. The corresponding dynamic comprehensive fitness evaluation formula is: F=w a ·S+w d ·D+AP(t) (5) Among them, S is the comprehensive fitness index, D is the diversity index, and w a is the fitness weight, w d is the diversity weight, ΔF(t) is the fitness increment based on the current iteration number Z; (3) In the initial iteration, 20% of the hyperparameters are randomly selected as diversity indicators for evaluation to reduce the risk of the population falling into a local optimal solution and improve the overall search capability; (4) The fitness score is updated according to the genetic speed and fitness evaluation of each iteration to promote the continuous improvement of the elite characteristics of the population. The fitness update formula is: F′(t+1)=F(t)+ΔF(t) (6) Where ΔF(t) is the fitness increment based on the current iteration number t.

6. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The elite dynamic selection strategy based on the tournament constructed in step (6) includes: (1) At the beginning of the iteration, the population is randomly divided into groups, and the tournament selection of genetic individuals is carried out based on the individuals with the strongest fitness in the tournament group. During the iterative evolution process, the number of tournament groups and the number of individuals in each group are adaptively adjusted according to the genetic speed of each iteration. By expanding the number of group members, better winners in the tournament are selected to speed up the genetic speed. The formula for dynamically adjusting the group size is: Among them, k(t) is the grouping size of the tth generation, and T is the total number of evolutionary generations; (2) Elite individual selection based on the fitness of the group tournament winner: Where W is the set of all winners in the tournament group, E is the set of elite individuals, and F(x) represents the fitness score of individual x; The elites and tournament winners form the population for the next generation of crossover and mutation operations.

7. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The non-uniform mutation strategy of fusion multi-point crossover constructed in step (4) includes: (1) In the mutation stage, a calculation formula for the probability of individual mutation based on the evolution process is constructed: Where P m (i) is the mutation probability of the i-th individual, P0 is the initial mutation probability, G is the current iteration number, G max is the maximum number of iterations, α is the adjustment factor used to control the amplitude of variation, and rand(0,1) is a random number in the interval [0,1]. When the individual variation probability value is greater than the set threshold, the gene position will mutate. (2) The winner of the tournament randomly selects multiple crossover points for gene exchange operations. The genes of the parent individuals at the crossover points will be divided into c+1 segments and alternately combined to generate new offspring individuals. The formula is: O=(P1[0:p1]+P2[p1:p2]+L+(-1) c P c+1 [p c :L]) (10) Among them, O is the newly generated offspring individual, P i represents the i-th parent individual, and the symbol "+" represents the splicing of gene fragments; (3) In the initial stage, a mutation threshold is set for each gene position, and a multi-point crossover operation is integrated before each mutation. The safety of the mutation population is guaranteed during the continuous evolution of the mutation, and the mutation probability is continuously updated. When the obtained individual mutation probability value is greater than the set threshold, the gene position mutates.

8. A dam seepage prediction method based on an improved long short-term memory deep learning model, characterized in that: The network back propagation process parameter learning method based on the improved elite genetic algorithm in step (2) is: (1) Construct a network parameter optimization model based on an improved elite genetic algorithm, where the objective function is the loss function of the improved long short-term memory neural network model and the decision variable is the dynamic comprehensive fitness of the population individuals generated in the continuous iteration process; (2) Constructing an improved long short-term memory neural network model loss function: Where y is the actual output, is the model prediction output, N is the number of samples; (3) Based on the improved elite genetic algorithm, the optimal solution of network parameters is continuously learned and optimized until the error of the parameters solved by the optimization objective constraint meets the set threshold.

Images