Tailing pond displacement dynamic prediction method and system based on multi-algorithm coupling model

Abstract

The invention discloses a tailing pond displacement dynamic prediction method and system based on a multi-algorithm coupling model. The prediction method comprises the following steps: S1, data acquisition and preprocessing; s2, decomposing a displacement signal; s3, trend term displacement prediction; s4, performing feature derivation; s5, periodic term displacement prediction; and S6, predicting result fusion. The prediction system comprises the following modules: a data acquisition module, a data preprocessing module, a displacement signal decomposition module, a trend term displacement prediction module, a feature derivation and optimization module, a periodic term displacement prediction module and a prediction result fusion module. Through the advanced multi-algorithm coupling model, the high-precision dynamic prediction of the displacement of the tailing pond is realized, the precision of the displacement prediction of the tailing dam and the long-term dependence modeling capability are improved, the generalization capability of the model to a complex displacement mode is enhanced, and powerful technical support is provided for the safety monitoring and disaster early warning of the tailing dam.

Description

Technical Field

[0001] The invention relates to a prediction method, in particular to a tailings pond displacement dynamic prediction method and system based on a multi-algorithm coupling model, belonging to the technical field of tailings pond dynamic displacement prediction. Background Art

[0002] Tailings ponds are the main control projects and major sources of danger in mines. They are large in number, small in scale and widely distributed. Once a dam breaks, it will pose a serious threat to the safety of life and property of downstream residents and the surrounding environment.

[0003] Displacement is one of the important parameters that characterizes the development trend and stability of a tailings dam during its service life, and is an important part of the safety monitoring process of a tailings dam. However, due to the loose characteristics of the medium that constitutes the tailings dam and the special geological structure in which it is located, the displacement prediction model currently used for traditional slopes is very limited in applicability to tailings dams. Therefore, research on displacement prediction of tailings dams is of great significance for disaster prediction and reduction of disaster losses in tailings dams.

[0004] Existing studies usually decompose displacement into trend term displacement and periodic term displacement. The periodic term displacement has always been the focus of displacement prediction due to its own nonlinearity, dynamics and the complexity of its influencing factors. At present, the prediction methods for periodic term displacement mainly use traditional machine learning and deep learning algorithms, including support vector machine (SVR) model, extreme learning machine (ELM), BP neural network, convolutional neural network (CNN), etc. Although these prediction models take into account the nonlinearity of periodic term displacement, these studies regard displacement prediction as a static regression problem. In order to integrate the dynamic characteristics of displacement in the prediction process, time series algorithms are used for displacement prediction, including gray prediction model, LSTM, GRU, TCN algorithm, etc. Although these prediction methods take into account the dynamic characteristics of landslide displacement, they usually cannot take into account the long-term and high-precision requirements of displacement prediction. For example, LSTM and GRU models cannot consider long-term series. Although TCN uses dilated convolution to improve the receptive field of the model and has good time series feature extraction and long-term dependency modeling capabilities, its fitting ability is poor when facing nonlinear series. Summary of the invention

[0005] In order to solve the shortcomings of the above technologies, the present invention provides a method and system for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model, comprising the following steps:

[0007] S1. Obtain the monitoring data of the cumulative displacement, reservoir water level and precipitation of the tailings dam, and pre-process the monitoring data;

[0008] S2. The displacement data of the tailings dam is regarded as time series data, and the improved variational mode decomposition algorithm is used to decompose the displacement signal to distinguish the trend term and the periodic term;

[0009] S3. Using the deep belief network DBN and combining the time series characteristics of the trend item displacement, the trend item displacement is modeled and trained to obtain the trend item displacement prediction result;

[0010] S4. Before predicting the periodic term, a linear weighted method is used to derive the characteristics of the influencing factors, namely the cumulative displacement, reservoir water level and precipitation;

[0011] S5. Construct a multi-algorithm coupled periodic term displacement prediction model to obtain the periodic term displacement prediction result;

[0012] S6. The displacement prediction results of the trend term and the periodic term are integrated by means of component superposition to obtain the final dynamic prediction results of the tailings pond displacement.

[0013] As a further implementation of this plan:

[0014] In step S1, the displacement signal decomposition strategy is specifically as follows: based on the variational mode decomposition algorithm, the tailings dam displacement data is decomposed into a number of intrinsic mode functions IMFs to distinguish trend terms and periodic terms;

[0015] The specific process is:

[0016] Each IMF is defined as an amplitude-frequency modulation function, expressed as:

[0017]

[0018] In the formula, u k , A k (t) is the instantaneous amplitude, is the phase, t represents the time variable;

[0019] The variational mode decomposition algorithm is divided into two parts: the construction and solution of the variational problem based on the framework of the variational problem;

[0020] The constructed variational problem is as follows:

[0021]

[0022] In the formula, u k represents the instantaneous amplitude, is the K components obtained after decomposition, is the center frequency of each component; * is the convolution symbol; ω kis the center frequency of the kth IMF, k represents the index of the IMF, and represents the kth IMF; K represents the total number of IMFs, that is, the number of IMFs obtained after decomposition; is the partial derivative with respect to time t, indicating the differential with respect to time; σ(t) is the standard deviation of the Gaussian function, which is used to control the smoothness of the convolution; j is the imaginary unit, which is used to represent complex numbers; st is subject to, indicating the constraints that need to be met in the optimization problem; e is the base of the natural logarithm, which is approximately equal to 2.71828; f(t) represents the original time series signal;

[0023] The variational problem to be solved is to transform the above constraints into an unconstrained variational problem and expand the above formula into a Lagrangian expression:

[0024]

[0025] Where L() represents the Lagrangian function; α is the quadratic penalty factor, λ(t) is the Lagrangian multiplication operator, and the optimal solution of the variational problem can be obtained by using the alternating direction method of the multiplication operator.

[0026] As a further implementation of this plan:

[0027] The improved variational mode decomposition algorithm is implemented by optimizing the parameters of the variational mode decomposition algorithm through the dung beetle optimization algorithm DBO, that is: the penalty factor and the mode decomposition number are adjusted through the dung beetle optimization algorithm DBO, and the growth entropy is used as the fitness function. Starting from different stages of the dung beetle optimization algorithm, chaotic mapping, spiral search, levy flight and t-distribution mutation strategies are introduced to improve the optimization ability. The optimization goal is to minimize the growth entropy.

[0028] As a further implementation of this plan:

[0029] The specific steps of the dung beetle optimization algorithm are as follows:

[0030] 1) Chaotic mapping is introduced to improve the diversity of the randomly generated initial population, and its expression is:

[0031]

[0032] Where r is a random number between 0 and 1; x represents the individual position in the dung beetle optimization algorithm, i usually represents the index of the individual in the population, μ is a parameter in the chaotic map, which is used to control the intensity of the chaotic behavior; η is another parameter in the chaotic map, which is used to control the scaling and displacement of the map; mod is the modulus operator; when μ∈(0,1), η∈(0,1) and μ∈(0,1), the function is in a chaotic state;

[0033] 2) The spiral search strategy is introduced into the process of egg ball reproduction and dung beetle foraging. The expression after introduction is:

[0034] B i (t+1)=X * +e zl ×cos(2πl)×b1×(B i (t)-Lb * )+e zl ×cos(2πl)×b2×(B i (t)-Ub * )

[0035] x i (t+1)=e zl ×cos(2πl)×x i (t)+C1×(x i (t)-Lb b )+C2×(x i (t)-Ub b )

[0036]

[0037] Where l is a uniformly distributed random number between -1 and 1, and z is the spiral exploration factor. i represents the position of the i-th dung beetle in the search space, X * Indicates the location of the currently found optimal solution, b * represents a benchmark or reference value related to the optimal solution, b1, b2, and b3 are parameters that control the step size or affect the search direction during the search process, U represents the upper bound of the search space, C1, C2, and C3 are coefficients used to adjust the weights of different parts of the search strategy, and b b Represents a benchmark or reference value related to the current solution, used to compare with x i (t) for comparison, i max Indicates the maximum number of iterations or the current number of iterations during the iteration process;

[0038] 3) Introducing the Levy flight strategy in stealing cockroaches:

[0039] x i (t+1)=Levy(d)×X b +S×g×(|x i (t)-X * ∣+∣x i (t)-X b ∣)

[0040]

[0041] Where d is the vector dimension, Γ is the gamma function, β is a constant, r1 and r2 are random numbers between 0 and 1; S is the scaling factor used to control the size of the levy flight step; g is a parameter related to the search strategy, which is used to adjust the intensity or direction of the search;

[0042] 4) Introducing t-distribution mutation strategy in the iteration process:

[0043]

[0044] is the new location of the i-th cuckoo nest in the population after mutation; x i is the individual position before mutation; t(iter) is the T distribution value, with the number of iterations as the degree of freedom.

[0045] As a further implementation of this plan:

[0046] The calculation steps of growth entropy are as follows:

[0047] (1) Perform the first-order difference on the time series with the original sequence length of N to obtain the difference sequence;

[0048] (2) Reconstruct the difference sequence and divide it into Nm subsequences, each of which contains m elements;

[0049] (3) Solve the corresponding pattern vector w(i) for each subsequence, calculate the number of occurrences Q(w(i)) of the pattern vector w(i) in each subsequence, and then calculate the frequency P(w(i)) of each pattern vector. The frequency calculation formula is as follows:

[0050]

[0051] (4) Calculate the growth entropy H(m) of the time series using the following formula:

[0052]

[0053] Wherein, i represents the index of the pattern vector, which is used to traverse all possible pattern vectors; k represents the number of types of pattern vectors or a specific index.

[0054] As a further implementation of this plan:

[0055] In step S4, the feature derivation process is: by calculating the value of the feature at different time points and its correlation coefficient, and using the dung beetle optimization algorithm DBO to optimize the model parameters, wherein the linear weighted weight coefficient is optimized and determined by the dung beetle optimization algorithm, including the following steps:

[0056] (1) Calculate the value of feature time t and feature time t-1;

[0057] (2) Correlation coefficient between feature s time, feature s-1 time and the predicted target at time t;

[0058] (3) Linear weighting is performed according to the correlation coefficient to obtain the derived features.

[0059] As a further implementation of this plan:

[0060] In step S5, the process of constructing the periodic term displacement prediction model is as follows:

[0061] Combine the temporal convolutional network TCN and the bidirectional gated recurrent unit BiGRU to build a TCN-BiGRU model, and then introduce the self-attention mechanism to weight the output of the TCN-BiGRU model;

[0062] The dung beetle optimization algorithm DBO is used to optimize the model parameters of TCN, BriGRU and self-attention mechanism. The loss function is set as the fitness function. The parameters are adjusted until the loss function value is minimized to obtain the optimal periodic item displacement prediction model.

[0063] As a further implementation of this plan:

[0064] The periodic item displacement prediction model includes a temporal convolutional network (TCN) layer, a bidirectional gated recurrent unit (BiGRU) layer, and a self-attention mechanism layer;

[0065] The temporal convolutional network (TCN) layer extracts local features and long-term dependencies of the time series by using dilated convolution kernels of different scales, wherein the sizes of the dilated convolution kernels include 3, 5, 7, and 9;

[0066] The bidirectional gated recurrent unit BiGRU layer completes information differentiation and result output by introducing a reset gate and an update gate, wherein the number of neurons in the GRU layer is 50;

[0067] The self-attention mechanism layer weights the output of the TCN-BiGRU model, calculates the importance weights of different time steps in the input sequence, and dynamically adjusts the attention paid to key information in the time series, thereby completing the prediction of the periodic item displacement.

[0068] As a further implementation of this plan:

[0069] In step S6, the predicted displacement results of the trend item and the periodic item are superimposed according to the time series principle, that is, the predicted displacement value of the trend item is added to the predicted displacement value of the periodic item to obtain the final predicted displacement value of the tailings dam.

[0070] A prediction system for a tailings pond displacement dynamic prediction method based on a multi-algorithm coupling model, the prediction system comprising:

[0071] Data acquisition module, used to collect the cumulative displacement, reservoir water level and rainfall monitoring data of the tailings dam;

[0072] Data preprocessing module, used for data cleaning and preprocessing of monitoring data, including noise removal, missing value filling and data standardization;

[0073] A displacement signal decomposition module, used for decomposing the displacement signal by using an improved variational mode decomposition algorithm;

[0074] The trend item displacement prediction module is used to establish a trend item displacement prediction model using a deep belief network DBN;

[0075] Feature derivation and optimization module, used for feature derivation and model parameter optimization;

[0076] The periodic item displacement prediction module is used to build a periodic item displacement prediction model by combining the temporal convolutional network (TCN) and the bidirectional gated recurrent unit (BiGRU), and introduces a self-attention mechanism;

[0077] The prediction result fusion module is used to superimpose the predicted displacement results of the trend item and the periodic item to obtain the final predicted displacement value of the tailings dam.

[0078] The present invention discloses a method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model. Through the advanced multi-algorithm coupling model, high-precision dynamic prediction of tailings pond displacement is achieved, which not only improves the accuracy of tailings dam displacement prediction and long-term dependent modeling capabilities, but also enhances the model's generalization ability for complex displacement patterns, providing strong technical support for safety monitoring and disaster warning of tailings dams. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of the technical route of the present invention.

[0080] Figure 2 This is the flow chart of the IDBO optimization algorithm of the present invention.

[0081] Figure 3 This is a schematic diagram of the DBN structure of the present invention.

[0082] Figure 4 Schematic diagram of the coupling model structure of the present invention.

[0083] Figure 5 This is the internal structure diagram of TCN of the present invention.

[0084] Figure 6 This is the GRU structure diagram of the present invention.

[0085] Figure 7 This is a diagram showing the weight calculation principle of the attention mechanism of the present invention.

[0086] Figure 8 This is a diagram of the cumulative displacement decomposition results of the present invention.

[0087] Fig. 9 It is the prediction result and error diagram of the trend item displacement result of the present invention.

[0088] Fig.10 It is the prediction result and error diagram of the periodic term displacement result of the present invention. DETAILED DESCRIPTION

[0089] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0090] The present invention discloses a method for dynamic prediction of tailings dam displacement based on a multi-algorithm coupling model. Combining the advantages of multiple models, starting from different stages of deep learning prediction, a tailings dam displacement prediction idea based on "feature derivation-decomposition prediction-model optimization" is proposed to achieve dynamic prediction of tailings dam displacement. The present invention improves the accuracy and reliability of tailings dam displacement prediction through an advanced algorithm coupling model, as well as the long-term dependent modeling and fitting capabilities of the model, and has a higher accuracy in tailings dam displacement prediction.

[0091] The overall technical process of the solution setting is as follows Figure 1 As shown, the following steps are included:

[0092] Step 100: Data collection and preprocessing.

[0093] The cumulative displacement, reservoir water level and precipitation monitoring data of the tailings dam location are collected and then preprocessed to clean noise, handle missing values ​​and standardize the data to ensure data quality and provide a reliable basis for multi-algorithm coupling model analysis.

[0094] Step 200: Extract displacement characteristics of tailings dam.

[0095] From the preprocessed data, features related to tailings dam displacement are extracted, including three influencing factors: cumulative displacement, reservoir water level and precipitation.

[0096] Step 300: Decomposition of tailings dam displacement time series.

[0097] The displacement data of the tailings dam is regarded as time series data and decomposed. Time series decomposition usually includes the identification and separation of trend terms and periodic terms.

[0098] The displacement data decomposition adopts the improved variational mode decomposition (IDBO-VMD) algorithm to decompose the tailings dam displacement time series into several intrinsic mode functions (IMFs) and distinguish the trend term and the periodic term. The parameters of the VMD algorithm are optimized using the dung beetle optimization algorithm (DBO) to improve the accuracy and efficiency of the decomposition. Specifically, by adjusting the penalty factor and the number of mode decompositions, using the growth entropy as the fitness function, and improving the diversity and global search ability of parameter optimization through chaotic mapping and spiral search strategies. In the specific implementation, the penalty factor and the number of mode decompositions of the IDBO-VMD algorithm are adjusted by the dung beetle optimization algorithm to minimize the growth entropy, thereby obtaining the best decomposition effect.

[0099] Step 400: DBN model training.

[0100] The deep belief network (DBN) in deep learning is used in combination with the time series characteristics of trend item displacement to model and train the trend item displacement, so as to obtain the trend item displacement prediction value. DBN is a neural network composed of multiple layers of restricted Boltzmann machines (RBM), which can learn the hierarchical representation of data, thereby effectively capturing the complex structure and patterns in the data.

[0101] Step 500: feature derivation and optimization.

[0102] Before the periodic item prediction, the influencing factors (cumulative displacement, reservoir water level and precipitation) are linearly weighted to derive features and enhance the expressiveness of features. The influencing factors here mainly refer to reservoir water level and rainfall. The reservoir water level at the previous time is linearly weighted with the reservoir water level data at this time, and the rainfall at the previous time is linearly weighted with the rainfall at this time.

[0103] Step 600: IDBO-TCN-BiGRU-Self model training.

[0104] Combining the temporal convolutional network (TCN) and the bidirectional gated recurrent unit (BiGRU), a TCN-BiGRU model is constructed to capture the long-term dependency and dynamic change characteristics of displacement data. Then, a self-attention mechanism is introduced to weight the output of the TCN-BiGRU model to highlight the key information in the time series and further improve the accuracy of the prediction.

[0105] The Dung Beetle Optimization (DBO) algorithm is used to optimize the model parameters such as TCN, BiGRU and self-attention mechanism to improve the model performance and obtain the optimal periodic item displacement prediction model, namely the IDBO-TCN-BiGRU-Self model. The preprocessed data is input into the model, and the local and long-term dependency features of the time series are captured by the TCN layer; the TCN layer effectively extracts the local features and long-term dependency features of the time series by using dilated convolution kernels of different scales, where the sizes of the dilated convolution kernels include 3, 5, 7, and 9. The BiGRU layer further processes the bidirectional information of the time series, and completes the information distinction and result output by introducing reset gates and update gates. The number of neurons in the GRU layer is 50; the self-attention mechanism dynamically adjusts the model's attention to the key information in the time series by calculating the importance weights of different time steps in the input sequence. The attention mechanism is used to learn and emphasize the information of important time points, thereby completing the accurate prediction of periodic item displacement.

[0106] Step 700: superimpose components to fuse the trend item prediction result and the period item prediction result.

[0107] The predicted displacement results of the trend term and the periodic term are superimposed according to the time series principle, that is, the predicted displacement value of the trend term is added to the predicted displacement value of the periodic term to obtain the final predicted displacement value of the tailings dam.

[0108] Furthermore, the accuracy and reliability of the prediction model are verified by comparing with the actual monitoring data.

[0109] In the implementation method, the data such as the cumulative displacement, reservoir water level and precipitation monitoring data collected in step 100 must be authentic and reliable, and the pre-processed data must be checked one by one to avoid incorrect data affecting the multi-algorithm coupling model analysis.

[0110] In an embodiment, in step 300, the tailings dam time series decomposition is implemented using a variational mode decomposition (DBO-VMD) algorithm. The DBO-VMD algorithm is a time-frequency analysis method for decomposing a time series signal into multiple intrinsic mode functions (IMFs).

[0111] like Figure 2 As shown, VMD defines each IMF as an amplitude modulation frequency modulation function, which can be expressed as:

[0112]

[0113] In the formula, u k , A k (t) is the instantaneous amplitude, is the phase, and t represents the time variable, which is a continuous parameter used to represent the time position of each data point in the time series.

[0114] The VMD algorithm can be divided into two parts based on the framework of variational problems: the construction and solution of variational problems. The constructed variational problems are as follows:

[0115]

[0116] In the formula, u k That is, the instantaneous amplitude, is the K components obtained after decomposition, is the center frequency of each component; * is the convolution symbol; ω k is the center frequency of the kth IMF, k represents the index of the IMF, and represents the kth IMF; K represents the total number of IMFs, that is, the number of IMFs obtained after decomposition; is the partial derivative with respect to time t, indicating the differential with respect to time; σ(t) is the standard deviation of the Gaussian function, which is used to control the smoothness of the convolution; j is the imaginary unit, which is used to represent complex numbers; st is subject to, indicating the constraints that need to be met in the optimization problem; e is the base of the natural logarithm, which is approximately equal to 2.71828; f(t) represents the original time series signal;

[0117] The solution to the variational problem lies in transforming the above constraints into an unconstrained variational problem and expanding the above formula into a Lagrangian expression:

[0118]

[0119] Where L() represents the Lagrangian function; α is the quadratic penalty factor, λ(t) is the Lagrangian multiplication operator, and the optimal solution of the variational problem can be obtained by using the alternating direction method of the multiplication operator.

[0120] The parameters of the VMD algorithm are optimized using the dung beetle optimization algorithm (DBO). Starting from different stages of the dung beetle optimization algorithm, chaotic mapping, spiral search, levy flight and t-distribution mutation strategies are introduced to improve the optimization ability of the dung beetle optimization algorithm and find more suitable parameters for the algorithm.

[0121] The specific steps are as follows:

[0122] 1) Chaotic mapping is introduced to improve the diversity of the randomly generated initial population, and its expression is:

[0123]

[0124] Where r is a random number between 0 and 1; x represents the individual position in the dung beetle optimization algorithm, i usually represents the index of the individual in the population, μ is a parameter in the chaotic mapping, which is used to control the intensity of chaotic behavior. When μ is in the range of (0,1), the chaotic mapping can produce a chaotic state and increase the diversity of the population; η is another parameter in the chaotic mapping, which is used to control the scaling and displacement of the mapping. η and μ together affect the behavior of the chaotic mapping; mod is a modulus operator, which is used to ensure that the position x(i+1)x(i+1) of the individual after the chaotic mapping is still within the defined search space, usually in the interval [0,1];

[0125] When μ∈(0,1), η∈(0,1) and μ∈(0,1), the function is in a chaotic state.

[0126] 2) The spiral search strategy is introduced into the process of egg ball reproduction and dung beetle foraging. The expression after introduction is:

[0127] B i (t+1)=X * +e zl ×cos(2πl)×b1×(B i (t)-Lb * )+e zl ×cos(2πl)×b2×(B i (t)-Ub * )

[0128] x i (t+1)=e zl ×cos(2πl)×x i (t)+C1×(x i (t)-Lb b )+C2×(x i (t)-Ub b )

[0129]

[0130] Where l is a uniformly distributed random number between -1 and 1, and z is the spiral exploration factor. i represents the position of the i-th dung beetle in the search space, X * Indicates the location of the currently found optimal solution, b * represents a benchmark or reference value related to the optimal solution, b1, b2, and b3 are parameters that control the step size or affect the search direction during the search process, U represents the upper bound of the search space, C1, C2, and C3 are coefficients used to adjust the weights of different parts of the search strategy, and b b Represents a benchmark or reference value related to the current solution, used to compare with x i (t) for comparison, imax Indicates the maximum number of iterations or the current number of iterations during the iteration process;

[0131] 3) Introducing the Levy flight strategy in stealing cockroaches:

[0132] x i (t+1)=Levy(d)×X b +S×g×(|x i (t)-X * ∣+∣x i (t)-X b ∣)

[0133]

[0134] Where d is the vector dimension, Γ is the gamma function, β is a constant, r1 and r2 are random numbers between 0 and 1; S is a scaling factor used to control the size of the levy flight step; g is a parameter related to the search strategy, which is used to adjust the intensity or direction of the search.

[0135] 4) Introduce the t-distribution mutation strategy in the iterative process to enhance the algorithm's search diversity and global search capabilities:

[0136]

[0137] is the new location of the i-th cuckoo nest in the population after mutation; x i is the individual position before mutation; t(iter) is the T distribution value, with the number of iterations as the degree of freedom.

[0138] Specifically, the calculation steps of the growth entropy used by the fitness function are as follows:

[0139] (1) Perform first-order differences on the original time series u(1), u(2), u(3), ..., u(N) with a length of N to obtain the difference sequence v(i);

[0140] (2) Reconstruct the difference sequence v(i) and divide it into Nm subsequences, each of which contains m elements. Each subsequence is represented by V(i) = v(i), v(i+1), v(i+m-1), 1≤i≤Nm;

[0141] (3) Next, for each subsequence V(i) above, find its corresponding pattern vector w(i). Each w(i) is represented as a combination of the sign and magnitude of each element in V(i), that is:

[0142] w(i)=[(s(1),q(1)),(s(2),q(2)),...,(s(m),q(m))]

[0143] s(k)=sgn(v(k))

[0144]

[0145] Where q represents entropy, k is a constant, sgn represents the sign function, and std represents the standard deviation; (1≤j≤N-14) Calculate the number of occurrences of the pattern vector type in each subsequence Q(w(i)), and then calculate its frequency P:

[0146]

[0147] (4) Finally, calculate the growth entropy H(m) of the time series:

[0148]

[0149] In an embodiment, the DBN model training strategy of step 400 is as follows: Figure 3 As shown in the figure, DBN is initialized using a layer-by-layer greedy algorithm and fine-tuned using a back propagation algorithm to build a multi-layer neural network for prediction and model learning. DBN is combined with the time characteristics of trend item displacement to perform time series modeling and prediction of trend item displacement.

[0150] In an implementation manner, in step 500, a linear model is used in combination with a correlation coefficient to perform feature construction and coupling, and the principle of constructing features is as follows:

[0151]

[0152] In the formula, x t , x t-1 Respectively represent the values ​​of feature time t and feature time t-1; r t , r t-1 Represents the correlation coefficient between the features at time t and time t-1 and the predicted target at time t; f t Represents the characteristics of a structure.

[0153] Specifically, feature derivation includes calculating the values ​​of features and their correlation coefficients at different time points, and optimizing the model parameters using the Dung Beetle Optimization Algorithm (DBO), wherein the weight coefficient of linear weighting is determined by optimizing the DBO algorithm, including the following steps:

[0154] (1) Calculate the value of feature time t and feature time t-1;

[0155] (2) Correlation coefficient between feature s time, feature s-1 time and the predicted target at time t;

[0156] (3) Linear weighting is performed according to the correlation coefficient to obtain the derived features.

[0157] In an implementation manner, the IDBO-TCN-BiGRU-Self model training in step 600 is a multi-algorithm fusion.

[0158] like Figure 4 As shown, the specific steps for establishing the IDBO-TCN-BiGRU-Self model are as follows:

[0159] The data are normalized and divided into training set and test set, and six influencing factors (weighted value of reservoir water level at the previous moment and this moment, reservoir water level value at this moment, weighted value of rainfall at the previous moment and this moment, rainfall value at this moment, displacement increment at this moment, displacement increment at the previous moment) and periodic displacement are taken as output;

[0160] 1) Use MATLAB2023 or other deep learning frameworks to build a deep learning multi-algorithm coupling model, that is, build the DBO-TCN-BiGRU-Self model of the present invention. Set the model structure to input layer, TCN layer, forward GRU layer, reversal layer, reverse GRU layer, self-attention mechanism layer, fully connected layer, and finally reach the output layer.

[0161] 2) Use the self-attention mechanism to learn the important information of different time points in the time series and complete the prediction of the displacement of periodic items.

[0162] like Figure 5 As shown in the figure, the TCN model provides sufficient receptive field through dilated convolution, and the dilated convolution calculation formula used is:

[0163]

[0164] Among them, F(s) represents the output value of the dilated convolution, f(i) represents the convolution kernel weight, and X s-d·i represents the input element sequence, d is the dilation rate, and i is the convolution kernel index.

[0165] like Figure 6 As shown in the figure, the Bi-GRU model introduces the reset gate and update gate of GRU to complete the information differentiation and result output. The calculation formula used is:

[0166] r t =σ(W r x t +U r h t-1 )

[0167] z t =σ(W z x t +U zh t-1 )

[0168]

[0169] In the formula, r t 、z t Represent the reset gate and update gate respectively; σ is the Sigmoid activation function; W and U both represent weights; Intermediate memory state is the inner product operation.

[0170] In the implementation method, the dung beetle optimization algorithm is used to optimize the parameters of the periodic term displacement prediction deep model, specifically: assuming that the population size is 20 and the maximum number of iterations is 50, the lower boundaries of the five optimization parameters of TCN convolution size, TCN model dropout ratio, BIGRU neuron number, learning rate, and regularization parameter are set as [2, 0.001, 10, 0.0001, 0.00001]; the upper boundaries are [6, 0.1, 100, 0.1, 0.005]. The loss function is set as the fitness function, and the deep learning model is trained by adjusting the parameters, and the loss function value is minimized.

[0171] like Figure 7 As shown in the figure, the self-attention mechanism is combined with TCN-BiGRU to increase the influence of important information by assigning different weight coefficients to the hidden outputs of TCN-BiGRU at different time steps, automatically learn the important information at different time points in the time series, adjust the representation of different time points based on this information, and capture the long-term dependencies in the time series data. The calculation formula and calculation process of the self-attention mechanism are as follows:

[0172] 1) Calculate the Q, K, and V values ​​of each input sequence element respectively. The calculation formula is as follows:

[0173] Q=XW Q

[0174] K=XW k

[0175] V=XW v

[0176] Among them, Q, K, V represent query vector, key vector and value vector respectively; X represents input value, and W represents weight value;

[0177] 2) Use the obtained Q, K, and V values ​​to calculate the weight configuration of different elements, where the internal calculation process of element a2 is as follows:

[0178]

[0179] Among them, v represents the value vector and b represents the output value.

[0180] 3) In actual situations, the calculation process of different elements is parallel calculation, and the calculation formula is as follows:

[0181]

[0182] Among them, d k is the dimension of Q and K, softmax represents the softmax function, and T represents the transpose of the matrix.

[0183] In the implementation method, the prediction results of step 700 are integrated. According to the principle of time series decomposition, the predicted values ​​of each sub-item displacement are superimposed to the cumulative displacement prediction result of the tailings dam. Assuming that the correlation coefficient R is 0.995 and the root mean square error MAE is 0.092 mm, the overall prediction accuracy is high and the step-type change of the tailings dam displacement can be better predicted.

[0184] The present invention also discloses a tailings pond displacement dynamic prediction system, comprising:

[0185] (a) Data acquisition module, used to collect the cumulative displacement, reservoir water level and rainfall monitoring data of the tailings dam;

[0186] (b) Data preprocessing module, used for data cleaning and preprocessing, including noise removal, missing value filling and data standardization;

[0187] (c) a displacement signal decomposition module, used to decompose the displacement signal using the IDBO-VMD algorithm;

[0188] (d) a trend item displacement prediction module, used to establish a trend item displacement prediction model using DBN;

[0189] (e) a feature derivation and optimization module, used for feature derivation and model parameter optimization;

[0190] (f) Periodic item displacement prediction module, which is used to combine TCN and BiGRU to build a periodic item displacement prediction model and introduce a self-attention mechanism;

[0191] (g) The prediction result fusion module is used to superimpose the predicted displacement results of the trend term and the periodic term to obtain the final predicted displacement value of the tailings dam.

[0192] In summary, the present invention designs a method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model in view of the loose characteristics of the medium constituting the tailings dam body and the characteristics of the special geological structure in which it is located. Through the advanced multi-algorithm coupling model, high-precision dynamic prediction of tailings pond displacement is achieved, thereby effectively preventing tailings dam disasters and protecting environmental safety and people's lives and property. Compared with the prior art, the beneficial effects are:

[0193] (1) The present invention effectively combines the advantages of multiple algorithms through the idea of ​​"feature derivation-decomposition prediction-model optimization", thereby improving the accuracy of tailings dam displacement prediction and long-term dependent modeling capabilities.

[0194] (2) The method adopted in the present invention not only enhances the generalization ability of the model for complex displacement patterns, but also effectively captures the dynamic and nonlinear characteristics of displacement changes by optimizing feature processing and parameter selection.

[0195] (3) Aiming at the core problem of dynamic displacement prediction of tailings ponds, the present invention proposes a dynamic displacement prediction method for tailings ponds based on a multi-algorithm coupling model, which has the advantage of high displacement prediction accuracy and reduces the model's sensitivity to outliers and noise, thereby improving displacement prediction efficiency and model robustness.

[0196] The above implementation modes are not limitations of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the scope of the technical solution of the present invention also belong to the protection scope of the present invention.

Claims

1. A method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model, characterized in that: The following steps are involved: S1. Obtain the monitoring data of the cumulative displacement, reservoir water level and precipitation of the tailings dam, and pre-process the monitoring data; S2. The displacement data of the tailings dam is regarded as time series data, and the improved variational mode decomposition algorithm is used to decompose the displacement signal to distinguish the trend term and the periodic term; S3. Using the deep belief network DBN and combining the time series characteristics of the trend item displacement, the trend item displacement is modeled and trained to obtain the trend item displacement prediction result; S4. Before predicting the periodic term, a linear weighted method is used to derive the characteristics of the influencing factors, namely the cumulative displacement, reservoir water level and precipitation; S5. Construct a multi-algorithm coupled periodic term displacement prediction model to obtain the periodic term displacement prediction result; S6. The displacement prediction results of the trend term and the periodic term are integrated by means of component superposition to obtain the final dynamic prediction results of the tailings pond displacement.

2. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 1, characterized in that: In step S1, the displacement signal decomposition strategy is specifically as follows: based on the variational mode decomposition algorithm, the tailings dam displacement data is decomposed into a number of intrinsic mode functions IMFs to distinguish trend terms and periodic terms; The specific process is: Each IMF is defined as an amplitude-frequency modulation function, expressed as: In the formula, u k , A k (t) is the instantaneous amplitude, is the phase, t represents the time variable; The variational mode decomposition algorithm is divided into two parts: the construction and solution of the variational problem based on the framework of the variational problem; The constructed variational problem is as follows: In the formula, u k represents the instantaneous amplitude, is the K components obtained after decomposition, is the center frequency of each component; * is the convolution symbol; ω k is the center frequency of the kth IMF, k represents the index of the IMF, and represents the kth IMF; K represents the total number of IMFs, that is, the number of IMFs obtained after decomposition; is the partial derivative with respect to time t, indicating the differential with respect to time; σ(t) is the standard deviation of the Gaussian function, which is used to control the smoothness of the convolution; j is the imaginary unit, which is used to represent complex numbers; st is subject to, indicating the constraints that need to be met in the optimization problem; e is the base of the natural logarithm, which is approximately equal to 2.71828; f(t) represents the original time series signal; The variational problem to be solved is to transform the above constraints into an unconstrained variational problem and expand the above formula into a Lagrangian expression: Where L() represents the Lagrangian function; α is the quadratic penalty factor, λ(t) is the Lagrangian multiplication operator, and the optimal solution of the variational problem can be obtained by using the alternating direction method of the multiplication operator.

3. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 2, characterized in that: The improved variational mode decomposition algorithm is implemented by optimizing the parameters of the variational mode decomposition algorithm through the dung beetle optimization algorithm DBO, that is: the penalty factor and the mode decomposition number are adjusted through the dung beetle optimization algorithm DBO, and the growth entropy is used as the fitness function. Starting from different stages of the dung beetle optimization algorithm, chaotic mapping, spiral search, levy flight and t-distribution mutation strategies are introduced to improve the optimization ability. The optimization goal is to minimize the growth entropy.

4. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 3 is characterized in that: The specific steps of the dung beetle optimization algorithm are as follows: 1) Chaotic mapping is introduced to improve the diversity of the randomly generated initial population, and its expression is: Where r is a random number between 0 and 1; x represents the individual position in the dung beetle optimization algorithm, i usually represents the index of the individual in the population, and μ is a parameter in the chaotic mapping, which is used to control the intensity of chaotic behavior; η is another parameter in the chaotic map, which is used to control the scaling and displacement of the map; mod is the modulus operator; when μ∈(0,1), η∈(0,1) and μ∈(0,1), the function is in a chaotic state; 2) The spiral search strategy is introduced into the process of egg ball reproduction and dung beetle foraging. The expression after introduction is: B i (t+1)=X * +e zl ×cos(2πl)×b1×(B i (t)-Lb * )+e zl ×cos(2πl)×b2×(B i (t)-Ub * ) x i (t+1)=e zl ×cos(2πl)×x i (t)+C1×(x i (t)-Lb b )+C2×(x i (t)-Ub b ) Where l is a uniformly distributed random number between -1 and 1, and z is the spiral exploration factor. i represents the position of the i-th dung beetle in the search space, X * Indicates the location of the currently found optimal solution, b * represents a benchmark or reference value related to the optimal solution, b1, b2, and b3 are parameters that control the step size or affect the search direction during the search process, U represents the upper bound of the search space, C1, C2, and C3 are coefficients used to adjust the weights of different parts of the search strategy, and b b Represents a benchmark or reference value related to the current solution, used to compare with x i (t) for comparison, i max Indicates the maximum number of iterations or the current number of iterations during the iteration process; 3) Introducing the Levy flight strategy in stealing cockroaches: x i (t+1)=Levy(d)×X b +S×g×(∣x i (t)-X * ∣+∣x i (t)-X b ∣) Where d is the vector dimension, Γ is the gamma function, β is a constant, r1 and r2 are random numbers between 0 and 1; S is the scaling factor used to control the size of the levy flight step; g is a parameter related to the search strategy, which is used to adjust the intensity or direction of the search; 4) Introducing t-distribution mutation strategy in the iteration process: is the new location of the i-th cuckoo nest in the population after mutation; x i is the individual position before mutation; t(iter) is the T distribution value, with the number of iterations as the degree of freedom.

5. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 3, characterized in that: The calculation steps of growth entropy are as follows: (1) Perform the first-order difference on the time series with the original sequence length of N to obtain the difference sequence; (2) Reconstruct the difference sequence and divide it into Nm subsequences, each of which contains m elements; (3) Solve the corresponding pattern vector w(i) for each subsequence, calculate the number of occurrences Q(w(i)) of the pattern vector w(i) in each subsequence, and then calculate the frequency P(w(i)) of each pattern vector. The frequency calculation formula is as follows: (4) Calculate the growth entropy H(m) of the time series using the following formula: Wherein, i represents the index of the pattern vector, which is used to traverse all possible pattern vectors; k represents the number of types of pattern vectors or a specific index.

6. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 3, characterized in that: In step S4, the feature derivation process is: by calculating the value of the feature at different time points and its correlation coefficient, and using the dung beetle optimization algorithm DBO to optimize the model parameters, wherein the weight coefficient of the linear weighting is determined by the dung beetle optimization algorithm, including the following steps: (1) Calculate the value of feature time t and feature time t-1; (2) Correlation coefficient between feature s time, feature s-1 time and the predicted target at time t; (3) Linear weighting is performed according to the correlation coefficient to obtain the derived features.

7. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 6, characterized in that: In step S5, the construction process of the periodic term displacement prediction model is as follows: Combine the temporal convolutional network TCN and the bidirectional gated recurrent unit BiGRU to build a TCN-BiGRU model, and then introduce the self-attention mechanism to weight the output of the TCN-BiGRU model; The dung beetle optimization algorithm DBO is used to optimize the model parameters of TCN, BriGRU and self-attention mechanism. The loss function is set as the fitness function. The parameters are adjusted until the loss function value reaches the minimum, and the optimal periodic item displacement prediction model is obtained.

8. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 7, characterized in that: The periodic item displacement prediction model includes a temporal convolutional network (TCN) layer, a bidirectional gated recurrent unit (BiGRU) layer, and a self-attention mechanism layer; The temporal convolutional network (TCN) layer extracts local features and long-term dependencies of the time series by using dilated convolution kernels of different scales, wherein the sizes of the dilated convolution kernels include 3, 5, 7, and 9; The bidirectional gated recurrent unit BiGRU layer completes information differentiation and result output by introducing a reset gate and an update gate, wherein the number of neurons in the GRU layer is 50; The self-attention mechanism layer weights the output of the TCN-BiGRU model, calculates the importance weights of different time steps in the input sequence, and dynamically adjusts the attention paid to key information in the time series, thereby completing the prediction of the periodic item displacement.

9. The method for dynamic prediction of tailings pond displacement based on a multi-algorithm coupling model according to claim 8, characterized in that: In step S6, the predicted displacement results of the trend item and the periodic item are superimposed according to the time series principle, that is, the predicted displacement value of the trend item is added to the predicted displacement value of the periodic item to obtain the final predicted displacement value of the tailings dam.

10. A prediction system for the tailings pond displacement dynamic prediction method based on a multi-algorithm coupling model as claimed in claim 1, characterized in that: The forecasting system includes: Data acquisition module, used to collect the cumulative displacement, reservoir water level and rainfall monitoring data of the tailings dam; Data preprocessing module, used for data cleaning and preprocessing of monitoring data, including noise removal, missing value filling and data standardization; A displacement signal decomposition module, used for decomposing the displacement signal by using an improved variational mode decomposition algorithm; The trend item displacement prediction module is used to establish a trend item displacement prediction model using a deep belief network DBN; Feature derivation and optimization module, used for feature derivation and model parameter optimization; The periodic item displacement prediction module is used to build a periodic item displacement prediction model by combining the temporal convolutional network (TCN) and the bidirectional gated recurrent unit (BiGRU), and introduces a self-attention mechanism; The prediction result fusion module is used to superimpose the predicted displacement results of the trend item and the periodic item to obtain the final predicted displacement value of the tailings dam.

Images