Dam effect quantity prediction method, system, medium and equipment

Abstract

The invention discloses a dam effect quantity prediction method and system, a medium and equipment, and relates to the technical field of safety engineering, dam monitoring data is acquired, influence factors of a dam effect quantity are determined, and effect quantity influence factor samples are constructed; performing dimension reduction processing on the influence factor sample through a principal component analysis method to obtain an influence factor matrix after decoupling and dimension reduction; coupling the long and short-term memory neural network with an artificial fish swarm algorithm, constructing a dam effect quantity prediction model, performing parameter setting on the dam effect quantity prediction model, and optimizing the initial weight and the number of neurons of the long and short-term memory neural network by using the artificial fish swarm algorithm; training the optimized long and short term memory neural network until convergence, and obtaining a trained dam effect quantity prediction model; and performing prediction and anti-normalization processing to obtain an effect quantity prediction result. And the accuracy, the stability and the efficiency of dam effect quantity prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of safety engineering, and in particular to a dam effect quantity prediction method, system, medium and equipment. Background Art

[0002] In recent years, various machine learning algorithms have been used for data analysis and mining, opening up new avenues for the research of dam safety monitoring modeling. Based on the combination of support vector machine (SVM), artificial neural network (ANN), random forest (RF) and other algorithms with dam monitoring data, scholars apply the powerful computing power of machine learning to the nonlinear relationship between mapping data and construct dam safety monitoring models. Among them, the performance of SVM is greatly affected by parameters. Domestic and foreign research mainly uses intelligent algorithms such as genetic algorithm and particle swarm algorithm to optimize it. When ANN is learning the network, it is necessary to fine-tune the hyperparameters, while the initial value and weight of BP neural network are randomly selected. Therefore, using intelligent optimization algorithm to optimize it can effectively improve the performance of neural network. Although neural network has good performance for large sample data, especially the depth of network structure can improve its learning performance, but with the increase of sample size, traditional shallow neural network has problems such as gradient loss and gradient explosion. In addition, due to its own structural limitations, traditional neural network algorithm cannot effectively process sequence feature data, resulting in poor model adaptability and low precision. To address this problem, domestic and foreign scholars introduced RNN technology into the research and application of dam safety monitoring model. However, RNN may not be able to capture long-term dependencies when processing long sequence data; when the network is deep, the gradient in the back-propagation process may gradually tend to 0 due to repeated multiplication, resulting in the inability to effectively update the model parameters, thus affecting the model performance.

[0003] In summary, the existing dam monitoring models still have problems such as weak adaptability and insufficient forecasting accuracy. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned existing dam monitoring models, such as weak adaptive ability and insufficient prediction accuracy, the main purpose of the present invention is to provide a dam effect quantity prediction method, system, medium and equipment.

[0005] To achieve the above object, the present invention adopts the following technical scheme, a dam effect quantity prediction method, comprising the following steps:

[0006] Collect dam monitoring data, determine the influencing factors of the dam effect quantity, and construct a sample of influencing factors of the effect quantity;

[0007] The impact factor samples are processed by principal component analysis to reduce the dimension, and the impact factor matrix after decoupling and dimension reduction is obtained;

[0008] A dam effect quantity prediction model is constructed by coupling a long short-term memory neural network and an artificial fish school algorithm, and the influencing factor matrix after decoupling and dimensionality reduction is used as the input of the dam effect quantity prediction model to train the dam effect quantity prediction model; wherein the artificial fish school algorithm is used to optimize the initial weights and the number of neurons of the long short-term memory neural network; the long short-term memory neural network after the initial weights and the number of neurons are optimized is trained until the loss function converges, thereby obtaining a trained dam effect quantity prediction model;

[0009] Obtain the influencing factor sample data after decoupling and dimensionality reduction of the effect quantity to be predicted, input it into the trained dam effect quantity prediction model after normalization, and obtain the normalized dam effect quantity prediction value;

[0010] The normalized predicted value of the dam effect is denormalized to obtain the true prediction result of the dam effect.

[0011] The effect quantity influencing factors include: water pressure component, temperature component, time component and rainfall component.

[0012] The method of performing dimensionality reduction processing on the impact factor samples by the principal component analysis method to obtain the impact factor matrix after decoupling and dimensionality reduction includes the following steps:

[0013] Normalizing the impact factor samples to obtain normalized impact factor samples;

[0014] Obtain the covariance matrix of the influencing factor sample, and obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix of the influencing factor sample;

[0015] Sort the eigenvalues ​​and obtain the cumulative contribution rate of the first m principal components;

[0016] Select the eigenvectors corresponding to the first m principal components, construct a transformation matrix, obtain the first m principal components, obtain the sample matrix of m influencing factors, and then obtain the influencing factor matrix after decoupling and dimensionality reduction.

[0017] The method of optimizing the initial weights and the number of neurons of the long short-term memory neural network by using the artificial fish swarm algorithm comprises the following steps:

[0018] Initializing artificial fish swarm algorithm parameters, the algorithm parameters including population size, field of view, step length, crowding factor, and maximum number of iterations;

[0019] Initializing long short-term memory neural network parameters, the network parameters including network structure, initial weights and number of neurons;

[0020] The positions of the artificial fish swarm are randomly initialized in the solution space, and each position represents a set of initial weights and number of neurons of LSTM;

[0021] Performing behavior selection on the artificial fish to obtain behavior selection results; the behaviors include flocking behavior, tail-chasing behavior, foraging behavior, and random behavior;

[0022] According to the result of the behavior selection, the position of the artificial fish is updated to obtain the updated position, that is, the initial weight and number of neurons of LSTM;

[0023] The LSTM model is trained using the updated positions and its performance is evaluated on the validation set, with the performance indicator being the food concentration until the maximum number of iterations is reached. Among all the artificial fish, the position with the highest food concentration is selected, i.e., the optimal LSTM initial weights and number of neurons;

[0024] The artificial fish swarm algorithm parameters include the number of fish swarms, the maximum number of iterations, the initial weight search range, the hidden layer neuron number search range, and the training rounds.

[0025] The parameter training algorithm of the long short-term memory neural network comprises the following steps:

[0026] Forward calculation of the output value of each neuron;

[0027] Determine the optimization objective function. When the previous output value is inconsistent with the expected value, calculate the error term value of each neuron and construct the loss function.

[0028] Guided by the loss function, the weights and parameters of the neural network are modified until the network error is less than the given value.

[0029] A dam effect quantity prediction system, comprising:

[0030] The module for processing effect quantity influencing factors is used to collect dam monitoring data, determine the influencing factors of the dam effect quantity, and construct effect quantity influencing factor samples; the influencing factor samples are processed by principal component analysis to obtain the influencing factor matrix after decoupling and dimensionality reduction;

[0031] Establish a prediction model module for coupling the long short-term memory neural network and the artificial fish school algorithm, construct a dam effect quantity prediction model, and use the decoupled and dimension-reduced influencing factor matrix as the input of the dam effect quantity prediction model to train the dam effect quantity prediction model; wherein the artificial fish school algorithm is used to optimize the initial weights and the number of neurons of the long short-term memory neural network; the long short-term memory neural network after the initial weights and the number of neurons are optimized is trained until convergence, and the trained dam effect quantity prediction model is obtained;

[0032] The dam effect quantity prediction module is used to obtain the influencing factor sample data after decoupling and dimensionality reduction of the effect quantity to be predicted, and input it into the trained dam effect quantity prediction model after normalization to obtain the normalized dam effect quantity prediction value; the normalized dam effect quantity prediction value is denormalized to obtain the real dam effect quantity prediction result.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: by collecting various influencing factors such as water pressure component, temperature component, time component and rainfall component, various data in various aspects affecting the dam effect are comprehensively considered, and the comprehensiveness and accuracy of the prediction model are improved. The influencing factor samples are processed by principal component analysis to reduce the dimension, reduce redundant information and noise, simplify the complexity of the data, so that the prediction model can focus more on the main influencing factors, and improve the training efficiency and prediction accuracy of the model. The LSTM neural network can effectively process time series data and capture long-term dependencies, which is particularly suitable for data with time series characteristics such as the dam effect. And the use of the artificial fish swarm algorithm to optimize the initial weight and number of neurons of the LSTM can effectively avoid local optimality and improve the overall optimization performance and prediction ability of the model. And the artificial fish swarm algorithm is used to iteratively train the LSTM to ensure that the parameters of the model are near the optimal solution, and through repeated optimization and adjustment, the overfitting phenomenon of the model is reduced, thereby improving the stability and generalization ability of the model. The system of the present invention includes a processing effect quantity influencing factor module, a prediction model module and a dam effect quantity prediction module. The clear division of labor and modular design make the system easy to maintain and upgrade, and also make the function of each module more focused and optimized. By combining principal component analysis, LSTM neural network and artificial fish school algorithm, the complexity of model training and computing resource consumption are reduced, the prediction efficiency is improved, and the dam effect quantity can be updated and predicted in real time, which meets the demand for real-time data in dam management and monitoring, and helps to take preventive measures in time.

[0034] In summary, this method and system improves the accuracy, stability and efficiency of dam effect quantity prediction by comprehensively utilizing multiple technical means and optimization algorithms, providing strong support for dam management and engineering decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0036] Figure 1 It is a schematic diagram of the flow chart of the AFSA-LSTM seepage prediction model in the present invention;

[0037] Figure 2It is a structural schematic diagram of the factors affecting the seepage of the dam according to the present invention;

[0038] Figure 3 It is a schematic diagram of a comparison diagram of the measured values ​​of the present invention and the prediction results of each model;

[0039] Figure 4 It is a schematic diagram of a comparison diagram of prediction results of the same prediction period and different training periods of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0041] In the operation state of concrete dam, the size of seepage effect is not constant, but changes with various factors such as water level and temperature. In order to effectively monitor the seepage condition of concrete dam, it is necessary to combine the actual engineering data, build a prediction model of the seepage characteristics of concrete dam, understand its changing law with external influencing factors, and conduct dynamic analysis on it, so as to lay the foundation for the safety monitoring of dam. By establishing a concrete dam seepage prediction model under the correlation of multiple influencing factors, it provides a basis for the health diagnosis and performance evaluation of concrete dam.

[0042] Example 1

[0043] The first embodiment of the present invention provides a dam seepage prediction method, please refer to Figure 1 , is a flow chart of the first embodiment of the present invention, comprising the following steps:

[0044] Collect dam monitoring data, determine the seepage influencing factors of the dam, and construct a sample of seepage influencing factors;

[0045] The principal component analysis method is used to reduce the dimension of the seepage influencing factor samples and obtain the influencing factor matrix after decoupling and dimension reduction.

[0046] A dam seepage prediction model is constructed by coupling a long short-term memory neural network and an artificial fish swarm algorithm, and the parameters of the dam seepage prediction model are set. The influencing factor matrix after decoupling and dimensionality reduction is used as the input of the dam effect quantity prediction model, and the dam effect quantity prediction model is trained; wherein, the initial weights and the number of neurons of the long short-term memory neural network are optimized by the artificial fish swarm algorithm; the long short-term memory neural network after the initial weights and the number of neurons are optimized is trained until the loss function converges, and the trained dam seepage prediction model is obtained;

[0047] Obtain the sample data of the seepage influencing factors to be predicted after decoupling and dimensionality reduction, input them into the trained dam seepage prediction model after normalization, and obtain the normalized dam seepage prediction value;

[0048] The normalized dam seepage prediction value is denormalized to obtain the real seepage prediction result.

[0049] The method provided by the first embodiment of the present invention is described in detail.

[0050] For the determination of impact factors, please refer to Figure 2 , Figure 2 The diagram is a structural diagram of factors affecting dam seepage. The seepage of concrete dams is mainly affected by the dam structure and load factors. The dam structure mainly includes the seepage coefficient of the dam material, cracks, foundation, anti-seepage measures and other factors; the load factors mainly include reservoir water level, temperature, aging, precipitation and other factors. The main factors affecting the seepage of concrete dams are as follows: Figure 2 shown.

[0051] According to the influencing factors of dam seepage under the condition of dam body stability, the seepage effect of the monitoring point of the concrete dam can be decomposed into influencing components such as water pressure component, temperature component, time component and rainfall component, that is, the seepage effect of the dam can be expressed as:

[0052] Y=Y H +Y T +Y θ +Y R (1)

[0053] Among them, Y represents the seepage effect value of a monitoring point of the dam; Y θ represents the time-dependent component; T represents the temperature component; Y H Represents the water pressure component; Y R Represents the amount of rainfall.

[0054] The seepage influencing factors are decoupled and reduced in dimension. The specific principle of the method is that there is a phenomenon of mutual influence between the influencing factors of the seepage effect of concrete dam. Therefore, in order to reduce the multicollinearity and number of seepage influencing factors, the principal component analysis (PCA) method is used to decouple and reduce the dimension of the influencing factors to obtain new seepage influencing factors.

[0055] (1) Because the number of most influencing factors varies greatly, the numerical differences are large, which is easy to produce distortion when measuring the variability of each factor, resulting in deviations for the larger scale influencing factors. Formula (2) is used to normalize the data of each factor to eliminate the influence of the differences and variations between factors on the seepage effect.

[0056]

[0057] Among them, x i is the i-th value of a certain impact factor sequence.

[0058] (2) Calculate the covariance matrix of the influence factor matrix using Equation (3).

[0059]

[0060] where: Cov(x i * ,x j * ) = E{[x i * - E(x i * )][x j * - E(x j * )]}, i = 1, 2, …, n; j = 1, 2, …, p.

[0061] (3) Calculate the eigenvalues λ x and the corresponding eigenvectors μ i of the covariance matrix C i of the influence factor matrix, where i = 1, 2, …, p.

[0062] (4) Arrange the eigenvalues in descending order. The cumulative contribution rate formula for the first m (m < p) principal components is:

[0063]

[0064] Here, the first m new components with a cumulative contribution rate greater than 95% are taken as the principal components.

[0065] Select the eigenvectors corresponding to the first m principal components, and use Equation (5) to construct the transformation matrix T,

[0066] and use Equation (6) to calculate the first m principal components.

[0067] T = (μ1, μ2, …, μ m ) T (5)

[0068] Z = Tx * (6)

[0069] Finally, obtain the sample matrix of m influence factors.

[0070] The basic principle of the artificial fish swarm algorithm is as follows: In a certain water body, fish can find areas rich in nutrients by themselves or following other fish. Therefore, the area with the largest number of fish is usually the area with the highest nutrient content in the water body. By simulating the foraging, aggregation, and following behaviors of the fish swarm, the purpose of optimizing the search is achieved.

[0071] During the optimization process, each artificial fish individual X(X1, X2, …) can be regarded as a solution, and the corresponding objective optimization function is Y = f(X), which represents the food concentration of the current state in the algorithm. Taking the maximization of the objective function as an example, the optimization process of the artificial fish mainly includes the following behaviors:

[0072] (1) Foraging behavior: During the optimization process, assume the current state of the artificial fish is X. A state X1 can be randomly selected within the perception range (visual) of the artificial fish, and the objective functions of the two are compared. If f(X) < f(X1), it continues to move forward; otherwise, another state X2 is selected within the sensing range. If the forward condition cannot be met after the maximum number of trial moves (trynumber), it moves randomly one step. The artificial fish X i moves one step forward to the X j state, which can be calculated by the following formula: next where step represents the maximum movement step length of the artificial fish, and rand() represents a random number in the range [0, 1).

[0073]

[0074]

[0075] (2) Random behavior: The artificial fish swims freely and randomly in the water body. This process is also a way for the artificial fish to search for new food resources, which helps to jump out of local extrema and assist the algorithm in finding the optimal solution.

[0076] (3) Swarming behavior: During the swarming process of artificial fish, the swarming behavior of artificial fish should follow the following two criteria: one is to move towards the center of its companions as much as possible; the other is to prevent the fish swarm from being too dense. When the above two criteria are met, the artificial fish can move towards the center of its companions; otherwise, it executes the foraging behavior.

[0077] (4) Following behavior: After the artificial fish discovers that there is enough food near its companions, it will follow the companions and swim towards the companions in the best state. When there is a companion in the best state within the perception range of the artificial fish, if it is not crowded around, the artificial fish moves one step in the direction of the optimal partner; otherwise, it executes the foraging behavior.

[0078] Based on the above four behaviors, the artificial fish realizes the following process of the algorithm:

[0079] (1) Set initial parameters;

[0080] (2) Calculate the fitness value of each initial artificial fish, find the optimal artificial fish state, and assign its value to the bulletin board;

[0081] (3) Evaluate the state of each artificial fish to select the behavior to be executed next; ​

[0082] (4) Execute the behavior of the artificial fish, update itself, and generate a new school of fish;

[0083] (5) Evaluate all artificial fish; if a better artificial fish status appears, the result on the bulletin board is updated to the value of the artificial fish;

[0084] (6) The algorithm ends when the optimal solution on the bulletin board reaches a satisfactory error or the number of iterations reaches the upper limit, otherwise go to step 3.

[0085] There is also the LSTM parameter training algorithm, which is still the back propagation algorithm. There are mainly three steps:

[0086] (1) Forward calculation of the output value of each neuron.

[0087] (2) Determine the optimization objective function. When the previous output value is inconsistent with the expected value, calculate the error term value of each neuron and construct the loss function.

[0088] (3) Using the loss function as a guide, the weights and parameters of the neural network are modified.

[0089] (4) Then jump to the first step and repeat (1), (2) and (3) until the network error is less than the given value.

[0090] Example 2

[0091] The second embodiment of the present invention provides an application example of the concrete dam seepage prediction method described in the first embodiment.

[0092] The model was verified using the monitoring data of seepage pressure at the foundation of a concrete dam. The 42# dam section piezometer UP42 was selected as the prediction object, and a total of 366 sets of monitoring data from June 21, 2017 to June 21, 2022 were used to construct the mathematical model. There is one water level measurement point in front of the dam and five thermometers in the dam body: T 42-8 , T 42-16 , T 42-20 , T 42-21 , T 42-22 .

[0093] Since the dam lacks downstream water level and precipitation data, the influencing factor set for dam seepage pressure prediction is θ=0.01t, where t is the cumulative number of monitoring days from the monitoring day to the base day.

[0094] After normalizing the 366 sets of monitoring data, the 13-dimensional influencing factor vectors have a certain correlation and high dimension, which will increase the difficulty and time of neural network learning. The PCA method can be used to reconstruct the influencing factor vectors and generate n linearly uncorrelated new influencing factors, thereby improving the speed and accuracy of model prediction. First, the covariance matrix of the influencing factor sample is calculated as shown in Table 1. From the covariance matrix table, it can be found that the correlation between most influencing factors exceeds 85%, indicating that decoupling and dimensionality reduction are very necessary.

[0095] Table 1 Covariance matrix of influencing factor samples

[0096]

[0097] Secondly, the eigenvalues ​​of the impact factor sample matrix and the cumulative contribution rate of the eigenvalues ​​are calculated as shown in Table 2.

[0098] Table 2 Eigenvalues ​​of the sample matrix of influencing factors and their cumulative variance contribution rates

[0099] Impact Factor Eigenvalue Variance percentage (%) Cumulative variance percentage (%) 1 7.16121 55.09 55.09 2 3.58651 27.59 82.67 3 1.25663 9.67 92.34 4 0.63843 4.91 97.25 5 0.12911 0.99 98.25 6 0.10474 0.81 99.05 7 0.05853 0.45 99.50 8 0.03749 0.29 99.79 9 0.01661 0.13 99.92 10 0.00848 0.07 99.98 11 0.00166 0.01 100.00 12 5.81E-04 0.00 100.00 13 1.60E-05 0.00 100.00 total 13 100

[0100] It can be seen that the cumulative variance of the first six principal components reaches 99%, so the original data is reconstructed into 6 principal components, which not only eliminates multicollinearity but also greatly reduces the dimension of the input vector. Thus, the eigenvectors of the first six principal components are obtained, as shown in Table 3.

[0101] Table 3. Eigenvector statistics

[0102]

[0103] The reconstructed impact factor matrix is ​​obtained by multiplying the first six principal component eigenvector matrices with the impact factor sample matrix.

[0104] Matlab code was used to compile the BP model, AFSA-BP model, LSTM model and AFSA-LSTM model, and the prediction was compared. The parameter settings of the four models are as follows: the hidden layer activation function of the BP neural network model is "logsig", and the output layer activation function is "purelin". The LSTM neural network model uses the Adam optimization algorithm. The batch size batch_size is set to 64, that is, 64 groups of data are extracted each time for network training. The specific parameters of the artificial fish school algorithm in the AFSA-BP and AFSA-LSTM models are: the number of fish schools is 15, the maximum number of iterations is 50, the search range of the initial weight is [-1,1], the search range of the number of hidden layer neurons is [6,15], and the training round is 50.

[0105] In order to verify the accuracy of the seepage prediction model, the prediction results of the BP, AFSA-BP, LSTM and the prediction model of the present invention and the AFSA-LSTM model without PCA method were compared during the same training and prediction period. The training set and prediction set of the model were divided before the seepage prediction. The monitoring data from June 21, 2017 to December 20, 2021 were the training set, and the monitoring data from December 21, 2021 to June 20, 2022 were the prediction set. The comparison results of the measured values ​​and the predicted values ​​are as follows:

[0106] Depend on Figure 1 The results show that the measured value of seepage pressure decreased by about 0.5 kPa from December 2021 to January 2022, and then continued to rise by more than 1 kPa in March 2022. After reaching the peak, it continued to decline and began to gradually rise in June 2022. According to the comparison and analysis between the measured values ​​and the prediction results of each model, it is found that:

[0107] (1) The prediction results of BP and the BP model optimized by AFSA have large errors. BP shows a trend of first rising and then falling during the prediction period, and the turning point is in May 2022. AFSA-BP has been on an upward trend and cannot reflect the overall change trend of the measured values.

[0108] (2) Although the prediction results of the LSTM model in the figure show a trend of first rising, then falling, and then rising again, which is the same as the change trend of the measured values ​​and can clearly reflect the trend of the measured values, the decline range exceeds 5 kPa and the rise range exceeds 9 kPa, which is much larger than the measured values. Overfitting may have occurred in the learning prediction process.

[0109] (3) The prediction results of the AFSA-LSTM model without using the PCA method to process the data set first increased, then decreased, and then increased again, and the amplitude of change did not exceed 0.5 kPa, and it could not reflect the changing trend of the measured values. The data set without PCA method increased the difficulty of model prediction due to the multicollinearity and high dimensionality of the data, resulting in unsatisfactory prediction results.

[0110] (4) The error between the AFSA-LSTM model and the measured value is not large. Although the first rising segment is missing, the rest of the change trends are the same.

[0111] According to the prediction results, the performance test indicators of each model are calculated as shown in Table 4.

[0112] Table 4 Statistics of performance test indicators of each model

[0113]

[0114]

[0115] As can be seen from Table 4, the AFSA-LSTM model has higher prediction accuracy than other models.

[0116] Compared with the BP model, AFSA-BP model, and LSTM model, the MAE of the AFSA-LSTM model was reduced by 73.91%, 71.25%, and 59.56%, respectively; the MAPE was reduced by 73.91%, 71.12%, and 60.26%, respectively; and the RMSE was reduced by 68.89%, 67.04%, and 58.86%, respectively. The MAE, MAPE, and RMSE of the AFSA-LSTM model without PCA processing are 81.06%, 79.19%, and 68.49% higher than those of the AFSA-LSTM model after PCA processing, respectively. This shows that the quality of the influencing factor samples is also one of the important factors that determine the quality of the prediction results. After PCA processing, the difficulty of analyzing and calculating the influencing factor samples is reduced, thereby improving the accuracy of model prediction. Compared with the LSTM model after PCA processing, the MAE, MAPE, and RMSE are 26.77%, 28.79%, and 30.68% lower, respectively. This shows that the artificial fish swarm algorithm optimization has a huge improvement in the prediction accuracy of the LSTM model, and its ability to improve the prediction accuracy is greater than that of the PCA processing of the influencing factor samples.

[0117] The AFSA-LSTM model of the present invention is used to perform a prediction scale test: the prediction period is set to 3 months (June 21 to September 2021) and 4 different training periods are proposed: 12 months (June 21, 2020 to June 20, 2021), 24 months (June 21, 2019 to June 20, 2021), 36 months (June 21, 2018 to June 20, 2021), and 48 months (June 21, 2017 to June 20, 2021). The prediction results are as follows: Figure 2 and Table 5.

[0118] Table 5 Statistics of prediction model performance indicators for the same prediction period and different training periods

[0119] Training 12 months Training 24 months Training 36 months Training 48 months MAE(kPa) 1.48 1.24 0.96 0.48 MAPE(%) 5.69 4.76 3.61 1.81 RMSE(kPa) 1.58 1.35 1.20 0.53

[0120] Depend on Figure 2As can be seen from Table 5, as the training period continues to extend, the prediction effect continues to improve. Among them, although the model performance indicators are constantly improving from 12 months to 36 months of training, they cannot reflect the changing trend of the measured values. When the training period reaches 48 months, not only are the prediction results basically the same as the changing trend of the measured values, but the model performance indicators MAE, MAPE, and RMSE are also reduced to 0.48kPa, 1.81%, and 0.53kPa, which are 50%, 50%, and 55.83% higher than those of 36 months of training, respectively, and the accuracy is significantly improved. Therefore, in practical applications, it should be suitable for engineering projects with long-term observation data and at least 48 months of training.

[0121] It should be noted that, in the present invention, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0122] The above embodiments are merely examples of the present invention and do not limit the protection scope of the present invention. All designs that are the same or similar to the present invention fall within the protection scope of the present invention.

Claims

1. A method for predicting dam effect quantity, characterized in that: The following steps are involved: Obtain dam monitoring data, determine the influencing factors of the dam effect quantity, and construct a sample of influencing factors of the effect quantity; The impact factor samples are processed by principal component analysis to reduce the dimension and obtain the impact factor matrix after decoupling and dimension reduction. A dam effect quantity prediction model is constructed by coupling a long short-term memory neural network and an artificial fish school algorithm, and the influencing factor matrix after decoupling and dimensionality reduction is used as the input of the dam effect quantity prediction model to train the dam effect quantity prediction model; wherein the artificial fish school algorithm is used to optimize the initial weights and the number of neurons of the long short-term memory neural network; the long short-term memory neural network after the initial weights and the number of neurons are optimized is trained until the loss function converges, thereby obtaining a trained dam effect quantity prediction model; Obtain the influencing factor sample data after decoupling and dimensionality reduction of the effect quantity to be predicted, input it into the trained dam effect quantity prediction model after normalization, and obtain the normalized dam effect quantity prediction value; The normalized predicted value of the dam effect is denormalized to obtain the true prediction result of the dam effect.

2. The dam effect quantity prediction method according to claim 1, characterized in that: The influencing factors include: water pressure component, temperature component, time component and rainfall component.

3. The dam effect quantity prediction method according to claim 1, characterized in that: The method of performing dimensionality reduction processing on the impact factor samples by the principal component analysis method to obtain the impact factor matrix after decoupling and dimensionality reduction includes the following steps: Normalizing the impact factor samples to obtain normalized impact factor samples; Obtain the covariance matrix of the influencing factor sample, and obtain the eigenvalues ​​and corresponding eigenvectors of the covariance matrix of the influencing factor sample; Sort the eigenvalues ​​and obtain the cumulative contribution rate of the first m principal components; Select the eigenvectors corresponding to the first m principal components, construct a transformation matrix, obtain the first m principal components, obtain the sample matrix of m influencing factors, and then obtain the influencing factor matrix after decoupling and dimensionality reduction.

4. The dam effect quantity prediction method according to claim 1, characterized in that: The method of optimizing the initial weights and the number of neurons of the long short-term memory neural network by using the artificial fish swarm algorithm comprises the following steps: Initializing artificial fish swarm algorithm parameters, the algorithm parameters including population size, field of view, step length, crowding factor, and maximum number of iterations; Initializing long short-term memory neural network parameters, the network parameters including network structure, initial weights and number of neurons; The positions of the artificial fish swarm are randomly initialized in the solution space, and each position represents a set of initial weights and number of neurons of LSTM; Performing behavior selection on the artificial fish to obtain behavior selection results; the behaviors include flocking behavior, tail-chasing behavior, foraging behavior, and random behavior; According to the result of the behavior selection, the position of the artificial fish is updated to obtain the updated position, that is, the initial weight and number of neurons of LSTM; The LSTM model was trained using the updated positions and its performance was evaluated on the validation set, with the performance indicator being the food concentration until the maximum number of iterations was reached. Among all the artificial fish, the position with the highest food concentration was selected, i.e., the optimal LSTM initial weights and number of neurons.

5. The dam effect prediction method according to claim 4, characterized in that: The flocking behavior of the artificial fish is that the artificial fish flocks to its neighboring partners. If the state is better after the flocking, the artificial fish executes the flocking behavior, otherwise it continues to forage; The tail-chasing behavior of the artificial fish is that the artificial fish tries to follow the neighboring partner. If there is food in front of the partner, the artificial fish performs the tail-chasing behavior, otherwise it continues to forage for food; The foraging behavior of the artificial fish is that the artificial fish randomly selects a direction to move forward according to the current state. If the food is found, that is, the objective function value is better, it moves forward in this direction, otherwise it continues to randomly select a direction; The random behavior of the artificial fish is that the artificial fish randomly selects a direction to move in the current state to jump out of the local optimal solution.

6. The dam effect prediction method according to claim 4, characterized in that: The artificial fish swarm algorithm parameters include the number of fish swarms, the maximum number of iterations, the initial weight search range, the hidden layer neuron number search range, and the training rounds.

7. The dam effect prediction method according to claim 1, characterized in that: The training of the long short-term memory neural network after the initial weights and the number of neurons are optimized comprises the following steps: Forward calculation of the output value of each neuron; Determine the optimization objective function. When the previous output value is inconsistent with the expected value, calculate the error term value of each neuron and construct the loss function. Guided by the loss function, the weights and parameters of the neural network are modified until the network error is less than the given value.

8. A dam effect quantity prediction system, characterized in that: include: The module for processing effect quantity influencing factors is used to collect dam monitoring data, determine the influencing factors of dam effect quantity, and construct effect quantity influencing factor samples; The impact factor samples are processed by principal component analysis to reduce the dimension and obtain the impact factor matrix after decoupling and dimension reduction. Establish a prediction model module for coupling the long short-term memory neural network and the artificial fish school algorithm to construct a dam effect quantity prediction model, and use the decoupled and reduced dimension influencing factor matrix as the input of the dam effect quantity prediction model to train the dam effect quantity prediction model, wherein the artificial fish school algorithm is used to optimize the initial weights and number of neurons of the long short-term memory neural network; Training the long short-term memory neural network after the initial weight and the number of neurons are optimized until convergence, thereby obtaining a trained dam effect quantity prediction model; The dam effect quantity prediction module is used to obtain the influencing factor sample data after decoupling and dimensionality reduction of the effect quantity to be predicted, and input it into the trained dam effect quantity prediction model after normalization to obtain the normalized dam effect quantity prediction value; the normalized dam effect quantity prediction value is denormalized to obtain the real dam effect quantity prediction result.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

10. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method described in any one of claims 1 to 7 is implemented.

Images