A method for predicting the unit grouting volume in curtain grouting
By using singular spectrum analysis and gated cyclic unit model, combined with composite risk weights and two-stage training, the underestimation error problem of high unit grout volume and sudden increase in samples in curtain grouting was solved, achieving more accurate prediction and construction guidance.
Patent Information
- Application Number
- CN202611391885.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively utilize short-sequence information within boreholes during curtain grouting tests, and the underestimation errors caused by high unit grout injection volume and sudden increases in sample size have not been adequately addressed, impacting construction organization and material preparation.
We employ singular spectral analysis to reconstruct component and gated cyclic unit models, and combine composite risk weights and a two-stage training method to construct multi-scale short sequence prediction samples, thereby reducing the underestimation error of high unit gray injection volume and sudden increase in number of segments.
It improves the ability to predict high unit ash injection volume and sudden increase in the number of stages, reduces the underestimation of extreme high value samples, and provides a more reliable construction reference.
Smart Images

Figure CN123115688A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing and computer-aided prediction technology for curtain grouting construction in water conservancy and hydropower projects, and in particular relates to a method for predicting the unit amount of grout injected into curtain grouting. Background Technology
[0002] Curtain grouting is a common treatment method in reservoir dam foundations, abutments, and underground seepage prevention projects. However, because the grout diffusion and filling process occurs within the underground medium, its diffusion path, filling range, and final treatment effect are difficult to observe directly. On-site, it is usually necessary to combine information such as water pressure tests, segmented grouting records, grouting pressure, and injection volume to indirectly determine the groutability of the formation and construction anomalies. During segmented grouting construction, if the grouting level of subsequent segments can be estimated based on the unit grouting volume of completed segments within the same grouting hole before the current main segment is constructed, construction personnel can arrange grout material reserves in advance, verify grout preparation and delivery capabilities, and strengthen monitoring of potentially high grouting or sudden increases in segments. This type of prediction cannot replace water pressure tests, geological assessment, and on-site construction experience, but it can provide quantitative references for material organization, construction scheduling, and risk warning.
[0003] However, current technologies still face several challenges in predicting engineering records from curtain grouting test projects. First, the number of usable main segments within a single grouting hole is limited, making it difficult to form the long sequence samples required for conventional depth time-series models. Second, karst formations are highly heterogeneous, and the unit grout volume sequence often exhibits non-stationary changes and local spikes. Third, high unit grout volume segments account for a relatively low proportion in the samples, but often correspond to higher material consumption, construction organization pressure, and abnormal handling requirements. If the training process primarily pursues the overall average error, the model is easily dominated by the majority of ordinary samples, thus weakening the focus on high grout volume samples and their underestimation errors.
[0004] In recent years, machine learning and deep learning methods have been increasingly used for grouting parameter prediction and construction data analysis. Existing research has shown that methods such as support vector machines, tree models, artificial neural networks, and recurrent neural networks can characterize the nonlinear relationship between grouting parameters and construction records, and improve the automation level of grouting parameter prediction. For construction data with time-series characteristics, recurrent neural networks such as GRU (Gated Recurrent Unit) can utilize prior state information, making them suitable for characterizing local dependencies in short sequences. However, when directly fitting the original unit grouting volume sequence, the overall trend, local fluctuations, and extreme spikes are often mixed on the same scale, increasing the difficulty of model learning. Signal decomposition methods such as SSA (Singular Spectrum Analysis) and ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) can break down non-stationary sequences into several variable components, and then the prediction model learns the characteristics of each component separately, providing an effective approach to improving the prediction accuracy of complex engineering sequences. On the other hand, Huber loss, asymmetric loss, and risk-oriented sample weighting strategies can improve the model's attention to high-risk samples and their underestimation errors, providing new possibilities for reducing the underestimation errors of extremely high-value samples.
[0005] For predicting unit grout volume and related indicators, existing research has also made beneficial explorations. For example, Rahimi et al., based on experimental grouting data from the Bakhtiari and Karun 4 dam sites, used statistical analysis and artificial neural networks to estimate grout volume, showing that the neural network method can better characterize the nonlinear relationship between grout volume and geological and construction parameters than traditional multivariate statistical methods (Rahimi et al., 2019); Niu et al., for karst curtain grouting, established a method for predicting unit grout volume based on rock permeability, pointing out that there is a nonlinear positive correlation between unit grout volume and rock permeability, and verified the engineering applicability of the method through field grouting tests (Niu et al., 2019); Zhu et al. further used the GOA-ELM model to predict grout volume for dam foundation grouting, providing an intelligent prediction approach for grouting material preparation, quality control, and construction parameter adjustment (Zhu et al., 2025). The above studies demonstrate that the unit ash injection volume and its related indicators have a certain degree of predictability, but the modeling approach still mainly focuses on mapping static engineering parameters. However, existing research still has the following shortcomings: 1) Existing studies on the prediction of unit grouting volume and related indicators mostly use geological conditions, permeability and construction parameters as static input variables, focusing on establishing the nonlinear mapping relationship between "engineering parameters and unit grouting volume", and do not make sufficient use of the local continuity and short sequence dependence reflected by the previous main segments in the same grouting hole.
[0006] 2) Existing grouting prediction studies mainly use overall indicators such as R², RMSE and MAE as the main evaluation basis, while there is relatively insufficient special evaluation of samples with high unit grouting volume, samples with sudden increases in adjacent sections and extreme underestimation of error.
[0007] 3) Although the decomposition prediction model helps to alleviate the non-stationarity of the original sequence, it usually still takes the minimization of the overall error as the main goal and does not further consider the engineering cost when the high-level ash injection segment is underestimated.
[0008] 4) If a strong risk loss is applied directly from the early stages of training, a small number of high-value samples may prematurely dominate parameter updates, affecting the model's learning of the overall change pattern.
[0009] Therefore, how to fully utilize the short sequence information within the hole while maintaining the overall predictive ability, and how to improve the model's ability to control errors for high unit ash injection volume samples, sudden increase samples, and extreme underestimation errors, are key issues that need to be addressed in unit ash injection volume prediction. Summary of the Invention
[0010] The purpose of this invention is to provide a method for predicting the unit grout volume in curtain grouting, so as to construct a multi-scale short sequence prediction sample with the continuity of the in-hole engineering under the conditions that the number of available main segments in a single hole is limited, there is no real continuity between the records of multiple holes, and the unit grout volume is non-stationary and has local sudden increase characteristics, and reduce the underestimation error of high unit grout volume and sudden increase segments.
[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for predicting the unit grouting volume of curtain grouting, the steps of which include: Step S1: Obtain the grouting hole number, main section sequence number, cement injection volume, and grouting section length of the target curtain grouting test project, and organize them into standardized sequence data; Step S2: Construct a complete and ordered unit ash injection quantity sequence, and obtain the reconstructed components through singular spectrum analysis; Step S3: Map each reconstructed component back to the corresponding grouting hole and main segment. Then, group and sort the mapped data. Based on the review length, form a component-level input vector from the component values in each reconstructed component sequence. All component-level input vectors form a short sequence sample within the component-level hole. Divide the samples into a validation set and a training set through cross-validation. Step S4: Construct a composite risk weight based on the absolute level of the unit ash injection volume and the positive increment in the hole; Step S5: Normalize each reconstructed component, establish gated recurrent unit sub-models respectively, and construct risk-adaptive asymmetric weighted Huber composite loss. Step S6: Train each sub-model in a two-stage manner, and obtain the predicted value of unit gray injection amount through component prediction, inverse normalization and summation reconstruction.
[0012] Furthermore, the specific steps of S1 are as follows: S11, Read the grouting hole number, segment number, segment length and cement injection volume from the grouting results table of the target curtain grouting test project, use the hole number field of the results table as the grouting hole number, extract the numerical part from the beginning of the segment number string as the main segment number, and merge the initial grouting record and re-grouting record with the same main segment number in the same grouting hole. S12, for the first The first grouting hole For each main grouting segment, the cement injection volume of each initial grouting record or re-grouting record contained therein is added together to obtain the total cement injection volume of the main segment. ; S13, use the segment length recorded in the merge group as the segment length of the primary segment. Calculate the unit grouting volume for each main segment of the grouting hole and output a dataset containing the grouting hole number, main segment number, and unit grouting volume.
[0013] Furthermore, the specific steps of S2 are as follows: S21, arrange the unit grouting volume of each effective main segment in all grouting holes into a complete ordered sequence according to the grouting hole number and the order of the main segments in the hole. And save the correspondence between the amount of ash injected per unit and the hole number and segment number. Assume the window length for singular spectrum analysis is... Then the number of columns in the trajectory matrix Then use the length of Sliding window constructs trajectory matrix ; S22, for the trajectory matrix Perform singular value decomposition to obtain Elementary matrices; S23, perform diagonal averaging on each elementary matrix to convert the elementary matrix into a one-dimensional reconstructed component sequence of the same length as the complete ordered unit ash injection quantity sequence, as shown in the following expression: In the formula, Indicates the first A reconstructed component sequence, Indicates the sorted order of the first... The main paragraph is in the first... Component values in a reconstructed component sequence This indicates the total number of primary segment records involved in the decomposition.
[0014] Furthermore, in step S3, the mapped data is numbered according to the main segment number inside each grouting hole. Sort in ascending order, and let the lookback length be... For the first A sequence of reconstructed components, with the same grouting hole as the previous one. The main paragraph is in the first... The component values in each reconstructed component sequence form a component-level input vector. All component-level input vectors form a component-level in-hole short sequence sample. The original unit ash injection amount corresponding to each sample is read. After obtaining all component-level in-hole short sequence samples, the samples are divided into a validation set and a training set through cross-validation.
[0015] Furthermore, the specific steps of S4 are as follows: S41, for the first The first grouting hole For each main grouting section, the positive increment of the unit grouting volume for the same grouting hole relative to the previous main section is calculated using the following formula: In the formula, Indicates the first The first grouting hole The main paragraph is relatively relative to the first The unit ash injection volume of each main section increases positively. Indicates the first The first grouting hole The unit amount of ash injected in each main section, Indicates the first The first grouting hole The unit amount of ash injected in each main section, This indicates taking the maximum value, when hour, ; S42, calculate the absolute level threshold and the positive increment threshold, using the following formula: In the formula, This indicates the absolute horizontal threshold for the unit amount of ash injected. This indicates the positive incremental threshold within the hole. Represents the quantile operator. Indicates the first The first grouting hole The unit amount of ash injected in each main section, Indicates the first The first grouting hole The main paragraph is relatively relative to the first The positive increment of unit ash injection volume in each main section; S43, Determine the smoothing scale for absolute level risk response. and the smoothing scale of positive incremental risk response The formulas for calculating the absolute level risk response and the positive incremental risk response are as follows: In the formula, The standard deviation of the unit amount of ash injected. This represents the standard deviation of the positive increment of the grouting volume per unit volume within the grouting hole. Indicates the first The absolute level of risk response for each main segment, Indicates the first Positive incremental risk response for each main segment; S44, Calculate the composite risk weights The formula is as follows: in, This represents the absolute level risk response coefficient. This represents the positive incremental risk response coefficient.
[0016] Furthermore, the specific steps of S5 are as follows: S51, perform max-min normalization on each reconstructed component; S52, respectively Each reconstructed component sequence establishes a corresponding gated recurrent unit (ROU) sub-model. The gated recurrent unit sub-model is a standard gated recurrent unit (ROU) model, and each sub-model sequentially includes an input layer, a gated recurrent unit layer, a random deactivation layer, and a single-neuron fully connected output layer. Specifically, the first... The prediction process of the gated recurrent unit sub-model corresponding to each reconstructed component sequence is represented as follows: In the formula, Indicates the first The gated cyclic unit sub-model corresponding to each reconstructed component sequence This represents the trainable network parameters of the sub-model. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence, where T represents the transpose; S53, calculate the residuals of the normalized predicted component values and the Huber base loss, using the following formula: In the formula, Indicates the target main paragraph is in the first position. Normalized true component values in a reconstructed component sequence Indicates the target main paragraph is in the first position. The normalized predicted component values are calculated by the gated cyclic unit sub-model of each reconstructed component sequence. Indicates the target main paragraph is in the first position. The residuals between the normalized true component values and the normalized predicted component values of each reconstructed component sequence. This represents Huber's base loss. The threshold for Huber's base loss; S54, Identifying Underestimated Indicator Variables To determine whether an undervaluation has occurred, the asymmetric amplification factor for the undervaluation direction is calculated by combining the composite risk weights and the undervaluation indicator variable, as shown in the following formula: In the formula, This indicates an underestimation of the directional asymmetric magnification factor. Indicates composite risk weights, This indicates an underestimation of the penalty coefficient. This indicates that the indicator variable has been underestimated. A value greater than 0 indicates that the predicted value is lower than the actual value, resulting in underestimation. In this case, the indicator variable is underestimated. ,like ≤0 indicates that no underestimation has occurred. , ; S55 converts the composite risk weights into scaling parameters for the risk gate value and calculates the risk gate value. Then, the sample-level mixing coefficient is calculated based on the risk threshold value, using the following formula: In the formula, As the denominator for risk gating, This means restricting the input value to the range of 0 to 1. Represents the basic mixing coefficient. Indicates the maximum mixing coefficient. The risk term mixing coefficient used for the current training samples. Indicates composite risk weights; Risk-adaptive composite loss for a single training sample for: In the formula, This represents the total training loss for a single component-level training sample. Indicates mean square error. This represents the risk-adaptive composite loss. This indicates an underestimation of the directional asymmetric magnification factor. Indicates the first The residual between the normalized true component value and the normalized predicted component value of each reconstructed component; For all samples in the current training batch The average value is taken, and the batch training loss of the corresponding gated recurrent unit sub-model is obtained by simplification.
[0017] Furthermore, the specific steps of S6 are as follows: S61, the first training phase, for the current cross-validation compromise... For each reconstructed component, the normalized short sequence within the aperture from the training set is input into the gated recurrent unit sub-model corresponding to that reconstructed component, and the normalized true component value corresponding to that reconstructed component is used as the supervision signal. The mean squared error is used as the loss function, as shown in the following formula: In the formula, Indicates the first The mean square error of the predicted component values obtained from the gated recurrent unit sub-model corresponding to each reconstructed component sequence. Indicates the current training batch. This indicates the number of samples included in the current training batch. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... Normalized true component values of a reconstructed component sequence; S62, based on the network weights obtained in the first training phase, a gated recurrent unit sub-model with the same structure as the first training phase is rebuilt, and the network weights obtained in the first training phase are set as the initial network weights for the second training phase. The optimizer is rebuilt, and the second training phase is performed. In this phase, the risk-adaptive composite loss defined in S5 is used to continue training the rebuilt sub-model, and the optimization is performed on the network weights obtained in the first training phase. After the reconstructed component sequences complete two-stage training, we obtain A trained gated recurrent unit sub-model; S63, for the validation samples in the current cross-validation compromise, a normalizer fitted to each reconstructed component sequence in the current training set is used to scale the corresponding historical component input, and the first... The normalized historical short sequences corresponding to the reconstructed component sequences are input into the gated recurrent unit sub-model trained to obtain the normalized predicted component values. Then, using the normalization parameters fitted to the reconstructed components in the current training set, the normalized predicted component values are inversely normalized, as shown in the following formula: In the formula, This indicates the restoration to the original component scale. The first grouting hole The main paragraph is in the first... The predicted component values in the reconstructed component sequence Indicates the first The minimum value of each reconstructed component sequence in the current training set. Indicates the first The maximum value of each reconstructed component sequence in the current training set. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence; S64, summing all the inverse normalized prediction component values, yields the... The first grouting hole The predicted unit ash injection volume for each target main segment is calculated using the following formula: In the formula, Indicates the first The first grouting hole Predicted unit ash injection volume for each main section. Indicates the number of reconstructed component sequences. This indicates the restoration to the original component scale. The first grouting hole The main paragraph is in the first... The predicted component values in a reconstructed component sequence.
[0018] The beneficial effects of this invention are: This invention can more effectively suppress large deviations in the prediction process, improve the prediction ability for high unit grouting volume and sudden increase in stages, and reduce the underestimation of extreme high value samples, thereby making the prediction results more reliable in high-risk stages. It can provide a more effective reference for grout preparation, construction organization and risk warning in curtain grouting construction. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is the overall flowchart of the method for predicting the unit grouting volume of the curtain grouting of the present invention.
[0021] Figure 2 This is the overall architecture diagram of the unit ash injection volume prediction model of the present invention.
[0022] Figure 3 It is a singular spectrum reconstruction component diagram of a complete ordered unit ash injection quantity sequence, where (a) is the original unit ash injection quantity, (b) is the first reconstruction component, (c) is the second reconstruction component, (d) is the third reconstruction component, (e) is the fourth reconstruction component, and (f) is the fifth reconstruction component.
[0023] Figure 4 It is a diagram of the reconstruction of component reflection and the construction of short sequence samples within the well.
[0024] Figure 5 This is a schematic diagram of the risk-adaptive composite loss structure and two-stage training.
[0025] Figure 6 This is a comparison chart of the actual unit ash injection volume and the predicted value. Detailed Implementation
[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0027] An embodiment of the present invention provides a method for predicting the unit grouting volume in curtain grouting, the flowchart of which is shown below. Figure 1 As shown, the steps are as follows: Step S1: Obtain the grouting hole number, main segment sequence number, cement injection volume, and grouting segment length of the target curtain grouting test project, and organize them into standardized sequence data that can be used for model learning. The specific steps are as follows: S11: The grouting hole number, segment number, segment length, and cement injection volume in the grouting results table of the target curtain grouting test project are read by the computing device. The hole number field in the results table is used as the grouting hole number. The numerical part of the segment number string is extracted from the beginning position as the main segment number. The initial grouting records and re-grouting records with the same main segment number in the same grouting hole are merged. For example, segments 2, 2F1, and 2F2 are merged into main segment 2.
[0028] S12, for the first The first grouting hole For each main grouting segment, the cement injection volume of each initial grouting record or re-grouting record contained therein is added together to obtain the total cement injection volume of the main segment, i.e.: .
[0029] In the formula, Indicates the first The first grouting hole The total number of grouting records contained in each main section; Indicates the first The first grouting hole The main paragraphs are arranged in order. The amount of cement injected in the initial grouting record, in kg; Indicates the first The first grouting hole Total cement injection volume for each main section, unit: kg.
[0030] S13, use the segment length recorded in the merge group as the segment length of the primary segment. Calculate the unit grouting volume for each main segment of the grouting hole, and output a dataset containing the grouting hole number, main segment number, and unit grouting volume. The formula for calculating the unit grouting volume is as follows: In the formula, Indicates the first The first grouting hole The unit ash injection volume for each main section, in kg / m³; Indicates the length of the corresponding main paragraph, in meters (m). This indicates the amount of cement injected.
[0031] The unit grouting volume described in this invention is calculated based on the grouting volume in the results table, without using the grouting volume, waste volume, or a total volume including waste volume; main sections whose main section numbers cannot be identified, whose cement injection volume after merging is not greater than 0, or whose section length is missing are not included in subsequent processing.
[0032] Step S2: Construct a complete ordered unit gray amount sequence based on the dataset output from S1, and obtain the reconstructed components through singular spectrum analysis. The specific steps are as follows: S21, arrange the unit grouting volume of each effective main segment in all grouting holes into a complete ordered sequence according to the grouting hole number and the order of the main segments in the hole. The system also saves the correspondence between the amount of ash injected per unit and the hole number and segment number (to be permanently saved during subsequent processing). The window length for singular spectrum analysis is set to... (Preferred in this embodiment) Then the number of columns in the trajectory matrix Then use the length of Sliding window constructs trajectory matrix The trajectory matrix is as follows: The trajectory matrix in the first line, number Column elements satisfy: ; In the formula, This represents an ordered sequence of unit ash injection quantities. Indicates the sorted order of the first... The unit amount of ash injected in each main section, This indicates the total number of primary segment records involved in the decomposition. The first element of the trajectory matrix line, number Column elements, , .
[0033] Since adjacent sliding windows share some elements in the original sequence, the same unit of gray ink can appear at different positions in the trajectory matrix.
[0034] S22, for the trajectory matrix Perform singular value decomposition to obtain The elementary matrix is given by the following formula: In the formula, Describes a left singular vector matrix. Represents a singular value matrix. Describes a right singular vector matrix. Representing the trajectory matrix rank, Indicates the first A singular value, Indicates the first The left singular vector corresponding to each singular value. Indicates the first The right singular vector corresponding to each singular value. Indicates the first The elementary matrix corresponding to each singular value, where T represents the transpose.
[0035] S23, perform diagonal averaging on each elementary matrix to transform the elementary matrix into a one-dimensional reconstructed component sequence of the same length as the complete ordered unit ash injection quantity sequence, resulting in... The reconstructed component sequence, the th reconstructed component sequence, The expression for the reconstructed component sequence is as follows: In the formula, Indicates the first A reconstructed component sequence, Indicates the sorted order of the first... The main paragraph is in the first... The component values in a reconstructed component sequence are the supervision targets of this main segment. Indicates the number of reconstructed component sequences. This indicates the total number of primary segment records involved in the decomposition.
[0036] In this embodiment, the window length Preferably 5, used to determine the number of rows in the trajectory matrix; the number of reconstructed component sequences. The preferred value is 5, which is used to determine the number of subsequent prediction sub-models. The resulting singular spectral reconstruction component map is shown below. Figure 3 As shown, (a) is the original unit ash injection amount, (b) is the first reconstruction component, (c) is the second reconstruction component, (d) is the third reconstruction component, (e) is the fourth reconstruction component, and (f) is the fifth reconstruction component.
[0037] Step S3: Based on the reconstructed components from S2 and their correspondence with hole numbers and segment numbers, map each reconstructed component back to the corresponding grouting hole and main segment number, and then process them according to the grouting hole number. The mapped data is grouped (no sliding window needs to be constructed between different grouting holes), and numbered according to the main segment within each grouting hole. Sort in ascending order, and let the lookback length be... For the first A sequence of reconstructed components, with the same grouting hole as the previous one. The main paragraph is in the first... The component values in each reconstructed component sequence form a component-level input vector. All component-level input vectors form component-level in-hole short sequence samples. The original unit injection amount for each sample is read. After obtaining all component-level in-hole short sequence samples, cross-validation is used to divide the samples into a validation set and a training set. The expression for the in-hole short sequence sample is: .
[0038] In the formula, Indicates the first The first grouting hole The main paragraph is in the first... The short sequence sample within the hole is composed of all component-level input vectors consisting of component values in the reconstructed component sequence; The lookback length of the short sequence within the well represents the number of times the same historical main segment of the well is used for each short sequence sample within the well. Indicates the first The first grouting hole The main paragraph is in the first... The component-level input vector is composed of the component values in the reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The component-level input vector consists of the component values in the reconstructed component sequence, where T represents the transpose.
[0039] In this embodiment Preferably, 2 is used. Taking 305 main segment records as an example, after constructing according to a review length of 2, 253 component-level in-well short sequence samples are obtained. Each sample forms 5 sets of reconstructed component inputs and supervision targets. The process of reconstructed component replay and in-well short sequence sample construction is as follows: Figure 4 As shown.
[0040] In this embodiment, a 5-fold cross-validation method is used. All short sequence samples within each well are divided into 5 folds according to their current order, without random shuffling. Each time, one fold is selected as the validation set, and the remaining four folds are used as the training set. This process is repeated 5 times, ensuring each sample serves as a validation sample once. For each round of cross-validation, the training set is used for subsequent risk weight calculation, component normalization parameter fitting, and gated recurrent unit sub-model training. The validation set does not participate in the calculation or fitting of these parameters; it is only used for prediction and evaluation after training.
[0041] Step S4: Construct a composite risk weight based on the absolute level of the unit ash injection volume and the positive increment within the hole. The specific steps are as follows: S41, for the first The first grouting hole For each main grouting section, the positive increment of the unit grouting volume for the same grouting hole relative to the previous main section is calculated using the following formula: In the formula, Indicates the first The first grouting hole The main paragraph is relatively relative to the first The positive increment of unit ash injection volume in each main segment is used to describe the degree of positive surge between adjacent main segments; Indicates the first The first grouting hole The unit amount of ash injected in each main section, Indicates the first The first grouting hole The unit amount of ash injected in each main section, This indicates taking the maximum value, when hour, .
[0042] S42, based on the unit grouting volume and the positive increment within the grouting hole, calculate the absolute horizontal threshold and the positive increment threshold, using the following formulas: In the formula, This indicates the absolute horizontal threshold for the unit amount of ash injected. This indicates the positive incremental threshold within the hole. This represents the quantile operator, used to calculate the quantile number of the sample set within the parentheses. quantile, in this embodiment, quantile level The preferred value is 0.90; Indicates the first The first grouting hole The unit amount of ash injected in each main section, Indicates the first The first grouting hole The main paragraph is relatively relative to the first The unit ash injection volume of each main section increases positively.
[0043] S43, Determine the smoothing scale for absolute level risk response. and the smoothing scale of positive incremental risk response The formulas for calculating absolute level risk response and positive incremental risk response are as follows: In the formula, The standard deviation of the unit amount of ash injected. This represents the standard deviation of the positive increment of the grouting volume per unit volume within the grouting hole. Indicates the first The absolute level risk response of each main segment, with a value range of 0 to 1; Indicates the first The positive incremental risk response for each main segment ranges from 0 to 1.
[0044] S44, Calculate the composite risk weights The formula is as follows: in, The absolute level risk response coefficient is used to control the additional weight of samples with consistently high unit ash injection volume. In this embodiment, it is preferably 1.5. The positive incremental risk response coefficient is used to control the additional weight of samples with sudden increases within the hole; in this embodiment, it is preferably 2.0. The constant term 1 retains the basic weight for all samples; the level risk term is used to increase the training weight of samples with high unit ash injection volume; the sudden increase risk term strengthens the error cost of samples with abnormal increases in adjacent main segments. Based on the above parameter settings, The theoretical value range is from 1 to 4.5.
[0045] Step S5: Based on the component-level in-hole short sequence samples and composite risk weights in the current training set, normalize each reconstructed component, establish gated recurrent unit sub-models respectively, and construct the risk-adaptive asymmetric weighted Huber composite loss. The specific steps are as follows: S51, normalize each reconstructed component, for the first... The normalization formula for the component values of the reconstructed component sequence is: In the formula, Indicates the first The minimum value of each reconstructed component sequence in the current training set. Indicates the first The maximum value of each reconstructed component sequence in the current training set. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... Component values in a reconstructed component sequence.
[0046] S52, respectively Each reconstructed component sequence establishes a corresponding gated recurrent unit (ROU) sub-model. Each sub-model has the same network structure but does not share trainable parameters. The gated ROU sub-model selects the standard gated ROU model, and each sub-model sequentially includes an input layer, a gated ROU layer, a random deactivation layer, and a single-neuron fully connected output layer. The number of hidden units in the gated ROU layer controls the state representation dimension of the sub-model; the random deactivation rate controls the proportion of neurons randomly deactivated during training; the fully connected output layer has one neuron, used to output the first corresponding to the target main segment. The normalized predicted component values of the reconstructed component sequence. In this embodiment, the preferred value for the number of hidden units in the gated recurrent unit layer is 32, the preferred value for the random inactivation rate is 0.2, and the preferred value for the number of neurons in the fully connected output layer is 1. Wherein, the... The prediction process of the gated recurrent unit sub-model corresponding to each reconstructed component sequence is represented as follows: In the formula, Indicates the first The gated cyclic unit sub-model corresponding to each reconstructed component sequence This represents the trainable network parameters of the sub-model, including the weights and biases in the network. Their values are determined iteratively through an optimization algorithm based on the current training set during model training. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence, where T represents the transpose.
[0047] S53, calculate the residuals of the normalized predicted component values and the Huber base loss, using the following formula: In the formula, Indicates the target main paragraph is in the first position. Normalized true component values in a reconstructed component sequence Indicates the target main paragraph is in the first position. The normalized predicted component value calculated by the gated cyclic unit sub-model of the reconstructed component sequence, i.e., the value at the th reconstructed component sequence... Normalized predicted component values in a reconstructed component sequence; Indicates the target main paragraph is in the first position. The residuals between the normalized true component values and the normalized predicted component values of each reconstructed component sequence. This represents Huber's base loss. In this embodiment, the turning point threshold for Huber's basic loss is... The preferred value is 0.05. 。
[0048] S54, based on the residual between the normalized true component values and the normalized predicted component values. Identify undervalued indicator variables This determines whether an undervaluation has occurred, and combines the composite risk weights and the undervaluation indicator variable to calculate the asymmetric amplification factor for the undervaluation direction, as shown in the following formula: In the formula, This indicates an underestimation of the directional asymmetric magnification factor. Indicates composite risk weights, This indicates an underestimated penalty coefficient, used to control the intensity of additional penalty when high-risk training samples are underestimated. In this embodiment... The preferred value is 2.0; This indicates that the indicator variable has been underestimated. A value greater than 0 indicates that the predicted value is lower than the actual value, meaning that the value has been underestimated. In this case, the indicator variable has been underestimated. ,like ≤0 indicates that no underestimation has occurred. , .
[0049] S55 converts the composite risk weights into scaling parameters for the risk gate value and calculates the risk gate value. Then, the sample-level mixing coefficient is calculated based on the risk threshold value, using the following formula: In the formula, The denominator for risk gating is preferably 3.5 in this embodiment; This means restricting the input value to the range of 0 to 1. This represents the basic mixing coefficient, which is preferably 0.1 in this embodiment; This represents the maximum mixing coefficient, which is preferably 0.4 in this embodiment; The risk term mixing coefficient used for the current training samples. This represents the composite risk weight.
[0050] Risk-adaptive composite loss for a single training sample for: In the formula, This represents the total training loss for a single component-level training sample. Indicates mean square error. This represents the risk-adaptive composite loss. This indicates an underestimation of the directional asymmetric magnification factor. Indicates the first The residuals between the normalized true component value and the normalized predicted component value of each reconstructed component.
[0051] For all samples in the current training batch The average value is taken, and the batch training loss of the corresponding gated recurrent unit sub-model is obtained by simplification.
[0052] Step S6: Based on component-level samples, composite risk weights, gated recurrent unit sub-models, and the total training loss, each sub-model is trained in a two-stage manner. The predicted unit gray volume is obtained through component prediction, inverse normalization, and summation reconstruction. The composite risk weights, risk-adaptive composite loss, and the two-stage training process are as follows: Figure 5 As shown. The specific steps are as follows: S61, the first training phase, for the current cross-validation compromise... For each reconstructed component, the normalized short sequence within the aperture from the training set is input into the gated recurrent unit sub-model corresponding to that reconstructed component, and the normalized true component value corresponding to that reconstructed component is used as the supervision signal. The mean squared error is used as the loss function, as shown in the following formula: In the formula, Indicates the first The mean square error of the predicted component values obtained from the gated recurrent unit sub-model corresponding to each reconstructed component sequence. Indicates the current training batch. This indicates the number of samples included in the current training batch. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence.
[0053] The first training phase enables the gated recurrent unit sub-models of each reconstructed component sequence to learn the overall variation relationship of component values of adjacent main segments of the same grouting hole in the corresponding reconstructed component sequence, and obtains the network weights of each sub-model in the first phase. An adaptive moment estimation optimizer is used. In this phase, the learning rate controls the step size of each parameter update, the number of training epochs controls the number of times the training data is fully used, and the batch size controls the number of samples used for each parameter update. In this embodiment, the preferred learning rate for the first phase is 0.005, the preferred number of training epochs is 30, and the preferred batch size is 8.
[0054] S62, based on the network weights obtained in the first training phase, re-establish a gated recurrent unit sub-model with the same structure as the first training phase, and set the network weights obtained in the first training phase as the initial network weights for the second training phase. Re-establish the optimizer (without inheriting the internal state of the optimizer from the first training phase) and conduct the second training phase. In this phase, the risk-adaptive composite loss defined in S5 is used to continue training the re-established sub-model, enabling the sub-model to maintain its ability to learn overall changes while increasing its attention to samples with high unit ash injection volume, positive sudden increases in samples within holes, and their underestimated errors. This is done separately for... After the reconstructed component sequences complete two-stage training, we obtain A trained gated recurrent unit sub-model.
[0055] The second-stage learning rate is used to limit the parameter adjustment range during risk fine-tuning, and the number of training epochs is used to control the duration of risk fine-tuning. In this embodiment, the second-stage learning rate is preferably 0.001, the number of training epochs is preferably 10, and the batch size follows the preferred value of 8 from the first training stage.
[0056] S63, for the validation samples in the current cross-validation compromise, a normalizer fitted to each reconstructed component sequence in the current training set is used to scale the corresponding historical component input, and the first... The normalized historical short sequences corresponding to the reconstructed component sequences are input into the gated recurrent unit sub-model trained to obtain the normalized predicted component values. Then, using the normalization parameters fitted to the reconstructed components in the current training set, the normalized predicted component values are inversely normalized, as shown in the following formula: In the formula, This indicates the restoration to the original component scale. The first grouting hole The main paragraph is in the first... The predicted component values in the reconstructed component sequence Indicates the first The minimum value of each reconstructed component sequence in the current training set. Indicates the first The maximum value of each reconstructed component sequence in the current training set. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence.
[0057] S64, summing all the inverse normalized prediction component values, yields the... The first grouting hole The predicted unit ash injection volume for each target main segment is calculated using the following formula: In the formula, Indicates the first The first grouting hole The predicted unit ash injection volume for each main section, in kg / m³; Indicates the number of reconstructed component sequences. This indicates the restoration to the original component scale. The first grouting hole The main paragraph is in the first... The predicted component values in a reconstructed component sequence.
[0058] After cross-validation, the prediction results of each cross-validation sample are merged according to the original prediction sample order to obtain the out-of-cross unit grouting volume prediction results for all prediction samples. This can be used to calculate the prediction error of grouting volume in curtain grouting test projects, the prediction error of high unit grouting volume samples, and the underestimation error of extreme samples. The overall architecture diagram of the unit grouting volume prediction model of this invention is as follows: Figure 2 As shown, the comparison between the measured unit ash injection volume and the predicted value is as follows: Figure 6 As shown.
[0059] The prediction model in this invention, consisting of singular spectrum analysis of complete ordered sequences, component-level gated cyclic unit sub-models, mean square error pre-training, and risk-adaptive asymmetric weighted Hubell fine-tuning, is denoted as RA-SGOH.
[0060] Experimental verification Using the method of this invention, a non-randomized 50% out-of-range evaluation was performed on 253 short-sequence samples from 26 grouting holes. The coefficient of determination (R²) for RA-SGOH was 0.9651, the root mean square error (RMSE) was 1085.76 kg / m, and the mean absolute error (MAE) was 770.06 kg / m. Compared with SSA-GRU, which uses the same singular spectrum analysis decomposition and parallel gated cyclic unit structure, RA-SGOH's RMSE decreased from 1289.90 kg / m to 1085.76 kg / m; the MAE of the top 10% of samples with high unit grout volume decreased from 2239.16 kg / m to 1186.97 kg / m; the MAE of the three samples with the largest true values decreased from 7488.62 kg / m to 3434.93 kg / m; and the prediction ratio of the sample with the largest true unit grout volume increased from 79.88% to 93.03%. Meanwhile, the MAE of RA-SGOH is slightly higher than that of SSA-GRU (742.29 kg / m), indicating that the technical effect of this invention is mainly reflected in the improvement of overall squared error, high unit ash injection sample error, and underestimation control of extreme high values, rather than all evaluation indicators reaching their optimal levels simultaneously. Comparison of measured unit ash injection volume with predicted values is shown below. Figure 6 As shown.
[0061] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.
Claims
1. A method for predicting the unit grouting volume in curtain grouting, characterized in that the steps include... include: Step S1: Obtain the grouting hole number, main section sequence number, cement injection volume, and grouting section length of the target curtain grouting test project, and organize them into standardized sequence data; Step S2: Construct a complete and ordered unit ash injection quantity sequence, and obtain the reconstructed components through singular spectrum analysis; Step S3: Map each reconstructed component back to the corresponding grouting hole and main segment. After sorting, determine the component-level input vector and the short sequence samples in the component-level hole according to the review length. Then, divide the samples into a validation set and a training set through cross-validation. Step S4: Construct a composite risk weight based on the absolute level of the unit ash injection volume and the positive increment in the hole; Step S5: Normalize each reconstructed component, establish gated recurrent unit sub-models respectively, and construct risk-adaptive asymmetric weighted Huber composite loss. Step S6: Train each sub-model in a two-stage manner, and obtain the predicted value of unit gray amount through component prediction, inverse normalization and summation reconstruction.
2. The method for predicting the unit grouting volume of curtain grouting according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Read the grouting hole number, segment number, segment length and cement injection volume from the grouting results table of the target curtain grouting test project, use the hole number field of the results table as the grouting hole number, extract the numerical part from the beginning of the segment number string as the main segment number, and merge the initial grouting record and re-grouting record with the same main segment number in the same grouting hole. S12, for the first The first grouting hole For each main grouting segment, the cement injection volume of each initial grouting record or re-grouting record contained therein is added together to obtain the total cement injection volume of the main segment. ; S13, use the segment length recorded in the merge group as the segment length of the primary segment. Calculate the unit grouting volume for each main segment of the grouting hole and output a dataset containing the grouting hole number, main segment number, and unit grouting volume.
3. The method for predicting the unit grouting volume of curtain grouting according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, arrange the unit grouting volume of each effective main segment in all grouting holes into a complete ordered sequence according to the grouting hole number and the order of the main segments in the hole. And save the correspondence between the amount of ash injected per unit and the hole number and segment number. Assume the window length for singular spectrum analysis is... Then the number of columns in the trajectory matrix Then use the length of Constructing the trajectory matrix using a sliding window ; S22, for the trajectory matrix Perform singular value decomposition to obtain Elementary matrices; S23, perform diagonal averaging on each elementary matrix to convert the elementary matrix into a one-dimensional reconstructed component sequence of the same length as the complete ordered unit ash injection quantity sequence, as shown in the following expression: In the formula, Indicates the first A reconstructed component sequence, Indicates the sorted order of the first... The main paragraph is in the first... Component values in a reconstructed component sequence This indicates the total number of primary segment records involved in the decomposition.
4. The method for predicting the unit grouting volume of curtain grouting according to claim 1, characterized in that, In step S3, the mapped data is numbered according to the main segment number inside each grouting hole. Sort in ascending order, and let the lookback length be . For the first A sequence of reconstructed components, with the same grouting hole as the previous one. The main paragraph is in the first... The component values in each reconstructed component sequence form a component-level input vector. All component-level input vectors form a component-level in-hole short sequence sample. The original unit ash injection amount corresponding to each sample is read. After obtaining all component-level in-hole short sequence samples, the samples are divided into a validation set and a training set through cross-validation.
5. The method for predicting the unit grouting volume of curtain grouting according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, for the first The first grouting hole For each main grouting section, the positive increment of the unit grouting volume for the same grouting hole relative to the previous main section is calculated using the following formula: In the formula, Indicates the first The first grouting hole The main paragraph is relatively relative to the first The unit ash injection volume of each main section increases positively. Indicates the first The first grouting hole The unit amount of ash injected in each main section, Indicates the first The first grouting hole The unit amount of ash injected in each main section, This indicates taking the maximum value, when hour, ; S42, calculate the absolute level threshold and the positive increment threshold, using the following formula: In the formula, This indicates the absolute horizontal threshold for the unit amount of ash injected. This indicates the positive incremental threshold within the hole. Describes the quantile operator. Indicates the first The first grouting hole The unit amount of ash injected in each main section, Indicates the first The first grouting hole The main paragraph is relatively relative to the first The positive increment of unit ash injection volume in each main section; S43, Determine the smoothing scale for absolute level risk response. and the smoothing scale of positive incremental risk response The formulas for calculating the absolute level risk response and the positive incremental risk response are as follows: In the formula, The standard deviation of the unit amount of ash injected. This represents the standard deviation of the positive increment of the grouting volume per unit volume within the grouting hole. Indicates the first The absolute level of risk response for each main segment. Indicates the first Positive incremental risk response for each main segment; S44, Calculate the composite risk weights The formula is as follows: in, This represents the absolute level risk response coefficient. This represents the positive incremental risk response coefficient.
6. The method for predicting the unit grouting volume of curtain grouting according to claim 1, characterized in that, The specific steps of S5 are as follows: S51, perform max-min normalization on each reconstructed component; S52, respectively Each reconstructed component sequence establishes a corresponding gated recurrent unit (ROU) sub-model. The gated recurrent unit sub-model is a standard gated recurrent unit (ROU) model, and each sub-model sequentially includes an input layer, a gated recurrent unit layer, a random deactivation layer, and a single-neuron fully connected output layer. Specifically, the first... The prediction process of the gated recurrent unit sub-model corresponding to each reconstructed component sequence is represented as follows: In the formula, Indicates the first The gated cyclic unit sub-model corresponding to each reconstructed component sequence This represents the trainable network parameters of the sub-model. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... The normalized true component values of a reconstructed component sequence, where T represents the transpose; S53, calculate the residuals of the normalized predicted component values and the Huber base loss, using the following formula: In the formula, Indicates the target main paragraph is in the first position. Normalized true component values in a reconstructed component sequence Indicates the target main paragraph is in the first position. The normalized predicted component values are calculated by the gated cyclic unit sub-model of each reconstructed component sequence. Indicates the target main paragraph is in the first position. The residuals between the normalized true component values and the normalized predicted component values of each reconstructed component sequence. This represents Huber's base loss. The threshold for Huber's base loss; S54, Identifying Underestimated Indicator Variables To determine whether an undervaluation has occurred, the asymmetric amplification factor for the undervaluation direction is calculated by combining the composite risk weights and the undervaluation indicator variable, as shown in the following formula: In the formula, This indicates an underestimation of the directional asymmetric magnification factor. Indicates composite risk weights, This indicates an underestimation of the penalty coefficient. This indicates that the indicator variable has been underestimated. A value greater than 0 indicates that the predicted value is lower than the actual value, resulting in underestimation. In this case, the indicator variable is underestimated. ,like ≤0 indicates that no underestimation has occurred. , ; S55 converts the composite risk weights into scaling parameters for the risk gate value and calculates the risk gate value. Then, the sample-level mixing coefficient is calculated based on the risk threshold value, using the following formula: In the formula, As the denominator for risk gating, This means restricting the input value to the range of 0 to 1. Represents the basic mixing coefficient. Indicates the maximum mixing coefficient. The risk term mixing coefficient used for the current training samples. Indicates composite risk weights; Risk-adaptive composite loss for a single training sample for: In the formula, This represents the total training loss for a single component-level training sample. Indicates mean square error. This represents the risk-adaptive composite loss. This indicates an underestimation of the directional asymmetric magnification factor. Indicates the first The residual between the normalized true component value and the normalized predicted component value of each reconstructed component; For all samples in the current training batch The average value is taken, and the batch training loss of the corresponding gated recurrent unit sub-model is obtained by simplification.
7. The method for predicting the unit grouting volume of curtain grouting according to claim 1, characterized in that, The specific steps of S6 are as follows: S61, the first training phase, for the current cross-validation compromise... For each reconstructed component, the normalized short sequence within the aperture from the training set is input into the gated recurrent unit sub-model corresponding to that reconstructed component, and the normalized true component value corresponding to that reconstructed component is used as the supervision signal. The mean squared error is used as the loss function, as shown in the following formula: In the formula, Indicates the first The mean square error of the predicted component values obtained from the gated recurrent unit sub-model corresponding to each reconstructed component sequence. Indicates the current training batch. This indicates the number of samples included in the current training batch. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence. Indicates the first The first grouting hole The main paragraph is in the first... Normalized true component values of a reconstructed component sequence; S62, based on the network weights obtained in the first training phase, a gated recurrent unit sub-model with the same structure as the first training phase is rebuilt, and the network weights obtained in the first training phase are set as the initial network weights for the second training phase. The optimizer is rebuilt, and the second training phase is performed. In this phase, the risk-adaptive composite loss defined in S5 is used to continue training the rebuilt sub-model, and the optimization is performed on the network weights obtained in the first training phase. After the reconstructed component sequences complete two-stage training, we obtain A trained gated recurrent unit sub-model; S63, for the validation samples in the current cross-validation compromise, a normalizer fitted to each reconstructed component sequence in the current training set is used to scale the corresponding historical component input, and the first... The normalized historical short sequences corresponding to the reconstructed component sequences are input into the gated recurrent unit sub-model trained to obtain the normalized predicted component values. Then, using the normalization parameters fitted to the reconstructed components in the current training set, the normalized predicted component values are inversely normalized, as shown in the following formula: In the formula, This indicates the restoration to the original component scale. The first grouting hole The main paragraph is in the first... The predicted component values in the reconstructed component sequence Indicates the first The minimum value of each reconstructed component sequence in the current training set. Indicates the first The maximum value of each reconstructed component sequence in the current training set. Indicates the first The first grouting hole The main paragraph is in the first... Normalized predicted component values of each reconstructed component sequence; S64, summing all the inverse normalized prediction component values, yields the... The first grouting hole The predicted unit ash injection volume for each target main segment is calculated using the following formula: In the formula, Indicates the first The first grouting hole Predicted unit ash injection volume for each main section. Indicates the number of reconstructed component sequences. This indicates the restoration to the original component scale. The first grouting hole The main paragraph is in the first... The predicted component values in a reconstructed component sequence.
