Transformer substation deep foundation pit deformation time sequence attention prediction method and system

Abstract

This invention discloses a temporal attention prediction method and system for deep foundation pit deformation in substations, relating to the field of attention prediction in artificial intelligence models. The method acquires spatial data, support data, zoning data, and displacement data from multiple monitoring points in the target substation's deep foundation pit, generating a support shielding topology characterizing the actual mechanical propagation relationship at the monitoring points. Based on the support propagation path, it determines the shielding coefficient, enhancement coefficient, and mechanical adjacency weight, and generates spatial fusion features and deformation prediction results in a spatiotemporal attention model. Then, it uses measured deformation data from the current prediction period to generate prediction residual data to determine the mechanical adjacency weight and deformation prediction results for the next prediction period. This system is used to improve the matching between the predicted deformation of deep foundation pits in substations and the actual stress relationship of the support structure.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence model attention prediction technology, and more specifically, this application relates to a temporal attention prediction method and system for deep foundation pit deformation in substations. Background Technology

[0002] During substation construction, deep foundation pit excavation alters the stress state of the surrounding soil, causing displacement or settlement changes in the retaining structure, the pit surface, and adjacent foundations. To ensure construction safety, multiple monitoring points are typically set up on-site to continuously collect deformation data such as displacement and settlement, and these data are combined with historical data to predict subsequent deformation trends.

[0003] In existing technologies, a common approach is to input the historical deformation sequences of each monitoring point into a time-series prediction model, or to further combine information such as spatial distance and historical correlation between monitoring points to establish a multi-point joint prediction relationship. Since these methods typically assume that monitoring points that are closer together have a stronger correlation, or assume that similar historical deformation trends can represent subsequent changes, they can achieve a certain predictive effect in general continuous areas.

[0004] However, in the scenario of deep foundation pits in substations, the deformation relationship between monitoring points is not solely determined by geometric distance. The retaining structures inside and outside the pit, supporting components, excavation zones, corner locations, and stiffness differences in different areas all alter the way deformation influences are transmitted between different monitoring points. In other words, two monitoring points, even if spatially close, may exhibit a weak correlation due to differences in construction zones or structural constraints; conversely, monitoring points that are far apart may exhibit consistent changes due to the influence of the same support system. If existing prediction models still primarily rely on proximity or data similarity to determine the correlation between monitoring points, they are prone to incorrectly transmitting deformation information blocked by structures to the target monitoring point, or omitting long-distance correlation information influenced by the same engineering structure, ultimately leading to inconsistencies between the predicted results and the actual deformation development.

[0005] Therefore, the problem that existing technologies need to solve is: in the deformation prediction of deep foundation pits in substations, how to avoid the misjudgment of deformation transmission caused by establishing the correlation between monitoring points based solely on geometric distance or similarity of historical data, so as to improve the adaptability of the prediction results to the actual stress relationship of the project. Summary of the Invention

[0006] To address the aforementioned technical problems, this technical solution provides a time-series attention prediction method and system for deep foundation pit deformation in substations, resolving the issues raised in the background section.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, this application provides a temporal attention prediction method for deep foundation pit deformation in substations, implemented through a model of spatiotemporal attention constrained by support topology constraints, including: Acquire spatial data, support data, zoning data, and displacement data from multiple monitoring points in the deep foundation pit of the target substation; A support shielding topology that determines the actual mechanical propagation relationship between monitoring points based on spatial data, support data, and zoning data; Based on the support shielding topology, the support propagation path of deformation influence between any two monitoring points is determined. The shielding coefficient and enhancement coefficient are determined along the support propagation path, and the mechanical adjacency weight is generated by combining the geometric distance between the monitoring points. Based on displacement data, the temporal features of monitoring points are generated. In the spatial attention layer of the model, the feature similarity between the temporal features of monitoring points is constrained by mechanical adjacency weights to obtain spatial fusion features. In the temporal attention layer of the model, the deformation prediction result for the current prediction period is generated based on the spatial fusion features of continuous time steps; Obtain the measured deformation data for the current prediction period, and generate prediction residual data based on the difference between the measured data and the corresponding deformation prediction results; The mechanical adjacency weights for the next prediction period are determined based on whether the predicted residual data meets the preset residual adjustment conditions. Based on the determined mechanical adjacency weights and the updated displacement data, the deformation prediction results for the next prediction period are generated and output through the model.

[0008] Secondly, this application provides a temporal attention prediction system for deep substation foundation pit deformation, used to implement the aforementioned temporal attention prediction method for deep substation foundation pit deformation, including: The data acquisition module is used to acquire spatial data, support data, zoning data, and displacement data from multiple monitoring points in the deep foundation pit of the target substation. The topology generation module is used to determine the support shielding topology based on spatial data, support data, and zoning data to identify the actual mechanical propagation relationship between monitoring points; The weight generation module is used to determine the support propagation path of deformation influence propagation between any two monitoring points based on the support shielding topology, determine the shielding coefficient and reinforcement coefficient along the support propagation path, and generate mechanical adjacency weights in combination with the geometric distance between monitoring points. The feature fusion module is used to generate time-series features of monitoring points based on displacement data. In the spatial attention layer of the model, the feature similarity between the time-series features of monitoring points is constrained based on mechanical adjacency weights to obtain spatial fusion features. The prediction output module is used to generate the deformation prediction result for the current prediction period based on the spatial fusion features of continuous time steps in the temporal attention layer of the model. The feedback adjustment module is used to acquire the measured deformation data for the current prediction period and generate prediction residual data based on the difference between the measured data and the corresponding deformation prediction results. It is also used to determine the mechanical adjacency weights for the next prediction period based on whether the predicted residual data meets the preset residual adjustment conditions, and to generate and output the deformation prediction results for the next prediction period based on the determined mechanical adjacency weights and the updated displacement data.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This application determines the shielding coefficient and enhancement coefficient along the support propagation path and generates mechanical adjacency weights by combining geometric distance. This can weaken the non-realistic associations across the excavation zone boundary, corner boundary, or stiffness boundary, while enhancing the effective associations constrained by the same support member, waist beam, or anchor cable. This makes the spatial attention layer subject to engineering mechanical constraints when fusing monitoring point information, and reduces spatial association misjudgments. This application uses the feature similarity between the temporal features of monitoring points based on mechanical adjacency weight constraints and generates spatial fusion features. This allows the shielding effect and force transmission effect of the support structure on deformation propagation to be introduced into the model calculation process while preserving the displacement change trend of each monitoring point, thereby improving the matching degree between the deformation prediction results and the actual stress state of the support system. This application generates prediction residual data by combining the measured deformation data of the current prediction period with the deformation prediction results, and determines the mechanical adjacency weight of the next prediction period based on the prediction residual data. This avoids the need to use measured data for back-calculation of the same prediction period. Instead, the residual feedback is used for spatial correlation correction in subsequent prediction stages, enabling the model to continuously adjust the prediction basis as the construction process and monitoring data change, thereby improving the stability and engineering applicability of subsequent predictions. Attached Figure Description

[0010] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein: Figure 1 This is a flowchart of the time-series attention prediction method for deep foundation pit deformation in substations according to the present invention. Figure 2 This is a schematic diagram of the temporal attention prediction system for deep foundation pit deformation in substations according to the present invention. Detailed Implementation

[0011] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0012] In the monitoring of deep foundation pit construction in substations, the prediction of pit deformation is not simply an extrapolation of displacement trends, but rather an engineering mechanics propagation problem influenced by the retaining structure, supporting components, excavation zones, corner constraints, and differences in support stiffness. Deep foundation pit excavation alters the stress state of the surrounding soil, and the support system also changes the deformation transmission relationship between different monitoring points. Therefore, high-reliability prediction requires not only identifying the displacement trend of individual monitoring points but also determining whether there is a genuine mechanical correlation between multiple monitoring points.

[0013] Existing deep foundation pit deformation prediction technologies typically rely on joint predictions based on geometric distance, historical correlation, or general spatial adjacency between monitoring points, assuming stronger correlation between points that are closer or have similar historical changes. However, in substation deep foundation pits, even if two monitoring points are spatially close, deformation propagation may be weakened due to the boundaries of the excavation zone, corner points, or stiffness boundaries. Conversely, even if two monitoring points are far apart, they may exhibit consistent deformation responses due to constraints from the same supporting structure, girders, or anchor cables. Therefore, relying solely on geometric proximity or historical similarity can easily lead to misjudgments of spatial correlation.

[0014] For example, adjacent monitoring points belonging to different excavation zones should not be simply considered to exhibit the same-direction deformation if there is a significant difference in the construction timeline between the zones. Furthermore, monitoring points located within the influence range of the same support component, even if not the closest, may collectively reflect the stress state of the same support system. If existing models cannot identify these differences in mechanical propagation, they will incorrectly assign contribution weights to monitoring points during spatial information fusion, causing the predicted results to deviate from the actual engineering stress relationships.

[0015] Furthermore, existing residual feedback methods often involve simple adjustments to model parameters or prediction results after discrepancies arise between the predicted and measured data. Directly regenerating results for the same prediction period using already acquired measured data easily leads to post-hoc calibration, failing to reflect the technical value of future-oriented prediction. A more reasonable approach is to utilize the prediction residuals of the current prediction period to identify the spatial consistency of residuals between monitoring points, and accordingly adjust the spatial correlation weights for the next prediction period, making subsequent predictions closer to the actual deformation propagation relationship.

[0016] If the above problems are not solved, the prediction system will continue to suffer from spatial correlation misjudgment and feedback correction lag: the deformation weakening effect caused by the support boundary cannot be entered into the model, and the force transmission coordination relationship formed by the same support component is also difficult to be utilized. Ultimately, the deformation prediction results cannot accurately reflect the real deformation propagation process under the support system, affecting the prediction of foundation pit construction risks and the safety assessment of substation foundations.

[0017] Example 1: like Figure 1 As shown, this application introduces a temporal attention prediction method for deep foundation pit deformation in substations, implemented through a model of spatiotemporal attention constrained by support topology constraints, including: S1. Acquire spatial data, support data, zoning data, and displacement data from multiple monitoring points in the deep foundation pit of the target substation; Specifically, spatial data includes monitoring point numbers, types, and coordinates, which are derived from monitoring point layout maps, measurement result tables, or BIM models. When the coordinate benchmarks of different data sources are inconsistent, the coordinates of each monitoring point are converted to a unified foundation pit coordinate system based on engineering control points, and a monitoring point coordinate table is output.

[0018] The support data includes retaining structure layout data, support component layout data, and support component stiffness data. The retaining structure layout data includes the retaining component number, endpoint coordinates, and connection relationship. The support component layout data includes the component number, endpoint coordinates, and range of action of supports, joists, or anchors. The support component stiffness data includes the stiffness value of the corresponding component. When the stiffness value is not directly given, the equivalent stiffness is determined based on the component material, cross-sectional dimensions, and length, and a support component data table is output.

[0019] The equivalent stiffness satisfies: ;in, This represents the equivalent stiffness of the m-th supporting member. Indicates the elastic modulus of a material. Represents the cross-sectional area. Indicates the length of the component.

[0020] The zoning data includes excavation zone number, excavation zone boundary, construction scope, and construction sequence. The construction scope corresponding to the same excavation zone number is defined as the excavation zone, and the shared or adjacent boundary lines between adjacent excavation zones are defined as the excavation zone boundaries. An excavation zone data table is output. Displacement data includes the horizontal displacement and / or settlement values ​​of each monitoring point within a continuous time step. These values ​​are arranged according to a unified time step, and monitoring values ​​whose timestamps meet the preset time tolerance are assigned to the corresponding time step. A displacement data matrix is ​​output.

[0021] The preset time tolerance can be set according to the sampling frequency of the monitoring equipment and the on-site data upload delay. For example, when displacement data is collected in one hour time step, the preset time tolerance can be 5 to 10 minutes; when displacement data is collected in one 10-minute time step, the preset time tolerance can be 30 seconds to 2 minutes; when manual verification or semi-automatic collection methods are used and there is a large delay in data upload, the preset time tolerance can be 15 to 30 minutes.

[0022] For example, 32 monitoring points were set up in the deep foundation pit of a target substation, with a monitoring frequency of once per hour, continuously collecting horizontal displacement and settlement values ​​for 72 time steps; support data included 36 retaining components, 12 supports or girders, 8 anchor cables and their corresponding stiffness values; zoning data included 4 excavation zones and their construction sequence. The final results included a coordinate table of the 32 monitoring points, a support component data table, an excavation zone data table, and a 32×72×2 displacement data matrix.

[0023] S2. A support shielding topology that determines the actual mechanical propagation relationship between monitoring points based on spatial data, support data, and zoning data; The process of generating the support shielding topology includes: Obtain the coordinates of each monitoring point in the spatial data, and obtain the layout data of the retaining structure, the layout data of the supporting components, and the stiffness data of the supporting components in the support data; Based on the layout data of the enclosure structure, the enclosure structure is divided into continuous layout and mechanical propagation segments, and the turning points of the enclosure structure are determined as corner boundaries. The support zone is divided based on the data of the arrangement of the support components, and is subject to the action of the same support component, lintel or anchor cable. Based on the zoning data, the excavation zones are divided according to the construction organization, and the excavation zone boundaries are determined as the construction boundaries between adjacent excavation zones. Based on the stiffness data of the support components, determine the stiffness boundary when the stiffness difference between adjacent retaining sections or support zones exceeds the preset stiffness difference condition. Based on the coordinates of each monitoring point, each monitoring point is associated with the corresponding retaining section, support area, excavation area, corner boundary, and stiffness boundary to obtain the support shielding topology.

[0024] Specifically, the retaining structure layout data in the support data is obtained. Based on the retaining component number, endpoint coordinates and connection relationship, retaining components whose endpoints are connected and whose extension direction does not meet the preset turning condition are merged into the same retaining segment. When the change in the extension direction of adjacent retaining components meets the preset turning condition, the corresponding connection position is determined as the corner boundary.

[0025] Whether the endpoints are connected can be determined by comparing the minimum distance between the endpoints of two retaining components with a preset endpoint connection threshold. When the minimum distance is not greater than the preset endpoint connection threshold, the endpoints of adjacent retaining components are determined to be connected. The preset endpoint connection threshold can be set according to the coordinate accuracy of the support design drawings or BIM model, for example, 0.1m to 0.5m; 0.1m is used when the design drawings have high accuracy, and 0.3m to 0.5m is used when the endpoints of components are extracted from the field measurement drawings.

[0026] Preset turning conditions are used to determine corner boundaries. The angle between the extension directions of adjacent retaining members is calculated based on the endpoint coordinates of the adjacent retaining members; when the angle between the extension directions is not less than the preset turning angle threshold, the contact position of the adjacent retaining members is determined as the corner boundary. The preset turning angle threshold can be set to 30°, 45°, or 60°; in rectangular or approximately rectangular foundation pits, 45° is preferred to distinguish between continuous connections of ordinary retaining sections and corner turning points.

[0027] Obtain the layout data of the support components, and determine the area subjected to the same support, beam, or anchor cable as the support zone based on the component number and range of action of the support, beam, or anchor cable.

[0028] Obtain the partition data, determine the excavation area based on the excavation area number and construction scope, and determine the shared or adjacent edge lines between adjacent excavation areas as the excavation area boundaries.

[0029] Acquire the stiffness data of the support components and associate the stiffness values ​​with the corresponding retaining sections or support zones. When the stiffness difference between adjacent retaining sections or adjacent support zones meets the preset stiffness difference condition, the corresponding boundary position is determined as the stiffness boundary. Then, based on the coordinates of each monitoring point in the spatial data, associate each monitoring point with the corresponding retaining section, support zone, excavation zone, corner boundary, and stiffness boundary, and output the support shielding topology.

[0030] A preset stiffness difference condition is used to determine stiffness boundaries. The stiffness values ​​of the support components corresponding to adjacent retaining sections or adjacent support zones are obtained, and the stiffness ratio is calculated. When the stiffness ratio is not less than a preset stiffness ratio threshold, a stiffness boundary is determined to exist between adjacent retaining sections or adjacent support zones. The preset stiffness ratio threshold can be between 1.5 and 2.5, preferably 2.0. If stiffness difference is used for judgment, the stiffness boundary is determined when the stiffness difference between adjacent support components is not less than a preset stiffness difference threshold. The preset stiffness difference threshold can be set based on 30% to 50% of the average stiffness of support components in the same project.

[0031] For example, the support data for a deep foundation pit of a target substation includes 36 retaining components. Components with an endpoint distance of less than 0.2m and an included angle of less than 30° between adjacent components are grouped into the same retaining segment, and positions with an included angle of not less than 30° are defined as corner boundaries. The effective range of 12 supports or wainscoting beams is used to delineate the branch protection zones. The adjacent edges of the four excavation zones are defined as excavation zone boundaries. Positions with a stiffness ratio of not less than 2.0 between adjacent retaining segments are defined as stiffness boundaries. Based on the coordinates of 32 monitoring points, each monitoring point is associated with its corresponding retaining segment, support zone, excavation zone, and related boundaries, ultimately yielding the support shielding topology.

[0032] Through the above technical solution, this application solves the problem that existing prediction methods rely solely on geometric distance or historical similarity to establish monitoring point associations, making it difficult to reflect the true force transmission relationship of the support. By using spatial data, support data, and zoning data to determine the retaining section, support zone, excavation zone, corner boundary, and stiffness boundary, each monitoring point is no longer just an isolated spatial point, but is associated with a specific support structure and construction zone. At the same time, by constructing a support shielding topology, the blocking effect of the excavation zone boundary, corner boundary, and stiffness boundary on deformation propagation, as well as the association effect of the same retaining section and the same support zone on deformation propagation, are uniformly expressed. This provides a structured basis for subsequently determining the support propagation path, shielding coefficient, enhancement coefficient, and mechanical adjacency weight, and avoids the model from incorrectly associating monitoring points that are spatially close but mechanically isolated.

[0033] S3. Based on the support shielding topology, determine the support propagation path of deformation influence propagation between any two monitoring points, determine the shielding coefficient and enhancement coefficient along the support propagation path, and generate mechanical adjacency weights by combining the geometric distance between monitoring points. The process of determining the transmission route of support includes: Obtain the target monitoring point and the candidate associated monitoring points that have a spatial correlation effect on the deformation prediction of the target monitoring point, as well as the corresponding retaining section, support area, excavation area, corner boundary and stiffness boundary in the support shielding topology; The retaining path between the two is determined based on the retaining section, and the force transmission path between the two is determined based on the support zone through the support, anchor cable or waist beam. The path that meets the preset path connectivity condition among the retaining path and force transmission path is determined as the support propagation path.

[0034] The process of obtaining candidate correlation monitoring points includes: Acquire the target monitoring point and calculate the geometric distance between the target monitoring point and other monitoring points based on spatial data; The monitoring points whose geometric distances meet the preset candidate distance conditions are determined as the first candidate monitoring points; Based on the support shielding topology, monitoring points belonging to the same retaining section or the same support area as the target monitoring point are identified as the second candidate monitoring points. Based on the support shielding topology, monitoring points that have a support propagation path that meets the preset path connectivity conditions with the target monitoring point are identified as the third candidate monitoring points. The first, second, and third candidate monitoring points are deduplicated and merged to obtain a set of candidate associated monitoring points; When the number of monitoring points in the candidate associated monitoring point set exceeds the preset upper limit of the number of candidates, the monitoring points that have a support propagation path with the target monitoring point are retained first, and the remaining monitoring points are retained in order of increasing geometric distance to obtain the updated candidate associated monitoring point set.

[0035] For example, the preset candidate distance condition is that the geometric distance is no more than 15m, and the preset maximum number of candidates is 20. When 28 candidate associated monitoring points are initially screened for a target monitoring point, the monitoring points with support propagation paths are retained first, and then the number is supplemented to 20 according to the geometric distance from small to large.

[0036] When the number of monitoring points in the deep foundation pit of the target substation is not less than a preset threshold for the number of monitoring points, the set of candidate associated monitoring points is truncated based on a preset upper limit for the number of candidates; the preset upper limit for the number of candidates can be between 8 and 20. Feature similarity and spatial attention weights are calculated only for monitoring points in the set of candidate associated monitoring points.

[0037] The process of determining the shielding coefficient includes: Determine the set of shielding boundaries, which includes the excavation zone boundaries, corner boundaries, and stiffness boundaries through which the support propagation path passes. The shielding multiplier for the excavation zone boundary is determined based on the construction time difference between adjacent excavation zones; the shielding multiplier for the corner point boundary is determined based on the angle between the support propagation path and the turning direction of the retaining structure; and the shielding multiplier for the stiffness boundary is determined based on the stiffness ratio of adjacent support components. Among them, the shielding multiplier of the excavation zone boundary decreases as the construction time difference increases, the shielding multiplier of the corner boundary decreases as the included angle increases, and the shielding multiplier of the stiffness boundary decreases as the ratio increases. The shielding multiplier is used to characterize the degree to which a single shielding boundary weakens deformation propagation. The shielding multiplier of the excavation zone boundary is determined by the construction time difference between adjacent excavation zones; when the construction time difference is 0, the shielding multiplier can be 1; when the construction time difference is one construction stage, the shielding multiplier can be 0.8; and when the construction time difference is not less than two construction stages, the shielding multiplier can be 0.5 to 0.7.

[0038] The shielding multiplier at the corner boundary is determined by the angle between the support propagation path and the turning direction of the enclosure structure; when the angle is less than 30°, the shielding multiplier can be 0.9; when the angle is between 30° and 60°, the shielding multiplier can be 0.7 to 0.8; and when the angle is greater than 60°, the shielding multiplier can be 0.5 to 0.7.

[0039] The shielding multiplier of the stiffness boundary is determined by the stiffness ratio of adjacent support members; when the stiffness ratio is less than 1.5, the shielding multiplier can be 1; when the stiffness ratio is between 1.5 and 2.5, the shielding multiplier can be between 0.7 and 0.8; and when the stiffness ratio is greater than 2.5, the shielding multiplier can be between 0.5 and 0.7.

[0040] Based on the boundary type, shielding multiplier, and adjacency relationship of each shielding boundary in the shielding boundary set, a path coupling correction coefficient is generated. Path coupling correction factor satisfy: ; in, This represents the r-th shielding boundary along the propagation path. Indicates the number of shielding boundaries along the propagation path. Denotes the masking multiplier of the r-th masking boundary. This indicates the boundary type of the r-th shielded boundary. This indicates the preset type coupling weight between two adjacent shielding boundary types. This indicates the preset coupling strength coefficient; The shielding multipliers along the support propagation path are multiplied sequentially and then multiplied with the path coupling correction coefficient to obtain the shielding coefficient.

[0041] The preset type coupling weights are used to characterize the coupling effect of different types of shielding boundaries on the deformation propagation weakening effect when they appear consecutively on the same support propagation path. Shielding boundaries include excavation zone boundaries, corner boundaries, and stiffness boundaries. A type coupling weight table is pre-established, using the boundary type of two adjacent shielding boundaries as an index and the corresponding type coupling weight as the index value.

[0042] The preset type coupling weight typically ranges from 0 to 1. A larger value indicates a stronger coupling weakening effect when two adjacent shielding boundaries appear consecutively. For example, when two adjacent shielding boundaries belong to the same boundary type, the preset type coupling weight can be 0.1 to 0.3; when two adjacent shielding boundaries are an excavation zone boundary and a corner boundary, the preset type coupling weight can be 0.3 to 0.5; when two adjacent shielding boundaries are an excavation zone boundary and a stiffness boundary, the preset type coupling weight can be 0.4 to 0.7; and when two adjacent shielding boundaries are a corner boundary and a stiffness boundary, the preset type coupling weight can be 0.3 to 0.6. Preferably, the preset type coupling weight when the excavation zone boundary and stiffness boundary appear consecutively is greater than the preset type coupling weight when the excavation zone boundary and corner boundary appear consecutively, which characterizes the stronger combined weakening effect on deformation propagation when differences in construction stages and abrupt changes in support stiffness coexist.

[0043] The preset type coupling weight can also be calibrated based on historical displacement data and measured deformation data. Displacement data, measured deformation data, support propagation paths, and shielding boundary sets from historical projects or early construction phases of the current project are acquired. For support propagation paths with the same adjacent boundary type combinations, the degree of measured deformation correlation between the target monitoring point and candidate associated monitoring points is statistically analyzed and compared with the baseline shielding result obtained solely through the cumulative multiplication of each shielding multiplier. When the measured deformation correlation degree is lower than the baseline shielding result, it indicates that the corresponding boundary type combination has an additional coupling weakening effect, and the preset type coupling weight corresponding to that boundary type combination is increased accordingly. When the measured deformation correlation degree matches the baseline shielding result, the preset type coupling weight corresponding to that boundary type combination is maintained or decreased.

[0044] For example, if a certain support propagation path passes through the excavation zone boundary and the stiffness boundary in sequence, and historical monitoring data shows that the deformation correlation degree of the monitoring points on this type of path is consistently lower than the correlation degree obtained by multiplying the shielding multiplier alone, then the preset type coupling weight corresponding to the excavation zone boundary-stiffness boundary is set to 0.6; if another support propagation path passes through two excavation zone boundaries consecutively, and the measured deformation correlation degree is close to the result of multiplying the shielding multiplier, then the preset type coupling weight corresponding to the excavation zone boundary-excavation zone boundary is set to 0.2.

[0045] In one specific implementation, the preset coupling strength coefficient can be selected from 0.1 to 1.0 and determined by the prediction error of the validation set. For example, 0.1, 0.3, 0.5, 0.7 and 1.0 are selected as candidate coupling strength coefficients, and deformation prediction is verified based on historical displacement data; when the coupling strength coefficient is 0.5, the average prediction error in the validation set is the smallest, so 0.5 is determined as the preset coupling strength coefficient.

[0046] The process of determining the enhancement coefficient includes: Obtain the support zones corresponding to the target monitoring point and the candidate associated monitoring points, and determine whether they are constrained by the same supporting component, the same waist beam, or the same anchor cable; When both are constrained by the same supporting member, the same lintel, or the same anchor cable, obtain the force transmission direction of the corresponding supporting member, lintel, or anchor cable; Generate a monitoring point spacing vector based on the coordinates of the target monitoring point and the coordinates of the candidate associated monitoring points; Project the monitoring point spacing vector onto the force transmission direction to obtain the projected distance in the force transmission direction; The projection distance satisfies: ; in, This represents the projected distance between the target monitoring point i and the candidate associated monitoring point j in the force transmission direction. Indicates the coordinates of the target monitoring point. Indicates the coordinates of the candidate associated monitoring points. This indicates the direction of unit force transmission in the supporting member, girders, or anchor cables that are jointly constrained by both. Indicates inner product operation; When the projection distance meets the preset force transmission distance condition, an enhancement coefficient is generated: ; in, This represents the enhancement coefficient between the target monitoring point i and the candidate associated monitoring point j. Indicates the preset enhancement coefficient. This indicates whether the two components are constrained by the same supporting member, the same lintel, or the same anchor cable; a value of 1 indicates yes, and a value of 0 indicates no. Indicates the projection distance. This represents the distance mapping value that increases as the projected distance decreases.

[0047] For example, ; in, This indicates the preset force transmission distance threshold. This applies when neither component is constrained by the same supporting member, the same joist, or the same anchor cable, or... Greater than At that time, Set it to 0, so that the enhancement coefficient is 0. A value of 1 indicates that no enhanced support effect is produced.

[0048] When the projected distance between the target monitoring point and the candidate associated monitoring point in the force transmission direction is not greater than a preset force transmission distance threshold, it is determined that the two meet the preset force transmission distance condition. The preset force transmission distance threshold can be 30% to 60% of the length of the same support action area, or 2 to 3 times the average distance between monitoring points within the adjacent support action area. For example, when the length of the same support action area is 20m, the preset force transmission distance threshold can be 10m.

[0049] The process of generating mechanical adjacency weights includes: Obtain the target monitoring point and candidate associated monitoring points that have a spatial correlation effect on the deformation prediction of the target monitoring point, and construct a set of candidate associated monitoring points based on the candidate associated monitoring points; Obtain the geometric distance, shielding coefficient, and enhancement coefficient between the target monitoring point and the candidate associated monitoring points; Distance attenuation data is generated based on geometric distance, and the distance attenuation data, shielding coefficient, and enhancement coefficient are multiplied to obtain the initial association strength; The initial association strength in the set of candidate associated monitoring points corresponding to the target monitoring point is normalized to obtain the mechanical adjacency weight: ; in, This represents the mechanical adjacency weight between the target monitoring point i and the candidate associated monitoring point j. Indicates the geometric distance between the two. This indicates the preset spatial distance normalization scale. This represents the set of candidate associated monitoring points corresponding to the target monitoring point i. This represents the candidate associated monitoring points in the candidate associated monitoring point set.

[0050] The preset spatial distance normalization scale can be set to the average spacing between monitoring points, the average spacing between adjacent monitoring points within the same retaining section, or the average distance within the candidate associated monitoring point set. If the average spacing between monitoring points around the target foundation pit is 5m, then the preset spatial distance normalization scale can be 5m; if it is necessary to retain a longer distance for force transmission association within the same support, it can be 10m.

[0051] Specifically, taking the target monitoring point as the calculation object, the corresponding retaining section, support zone, excavation zone, corner boundary, and stiffness boundary are obtained, and candidate associated monitoring points are determined from other monitoring points. The retaining path is determined based on the retaining section corresponding to the target monitoring point and the candidate associated monitoring point, and the force transmission path formed by the support, waist beam, or anchor cable is determined based on the support zone corresponding to the two. The retaining path or force transmission path that meets the preset path connectivity conditions is determined as the support propagation path.

[0052] Preset path connectivity conditions are used to determine the support propagation path. Based on the support shielding topology, the retaining path and force transmission path between the target monitoring point and candidate associated monitoring points are obtained. A path is considered to meet the preset path connectivity conditions when it simultaneously satisfies the following conditions: the path ends connect the target monitoring point and the candidate associated monitoring point respectively; the retaining section or support area in the path is continuous in the support shielding topology; and the boundaries traversed by the path have all been identified as excavation zone boundaries, corner boundaries, or stiffness boundaries. If no retaining path or force transmission path satisfying the above conditions exists between two monitoring points, the path is not considered a support propagation path.

[0053] Subsequently, the excavation zone boundaries, corner boundaries, and stiffness boundaries traversed by the support propagation path are identified to form a set of shielded boundaries. Based on the boundary type, shielding multiplier, and adjacency relationship of each shielded boundary in the set, a path coupling correction coefficient is generated. The shielding multipliers on the support propagation path are multiplied sequentially and then multiplied with the path coupling correction coefficient to obtain the shielding coefficient. Furthermore, it is determined whether the target monitoring point and candidate associated monitoring points are constrained by the same support, beam, or anchor cable, and an enhancement coefficient is generated based on their projected distance in the corresponding force transmission direction.

[0054] Finally, distance attenuation data is generated based on the geometric distance between the target monitoring point and the candidate associated monitoring points. The distance attenuation data, the shielding coefficient, and the enhancement coefficient are multiplied to obtain the initial association strength. The initial association strength of each candidate associated monitoring point corresponding to the same target monitoring point is normalized to obtain the mechanical adjacency weight.

[0055] For example, if the geometric distance between a target monitoring point and a candidate associated monitoring point is 8m, and the support propagation path between them passes through a corner boundary with a shielding multiplier of 0.7, and since the support propagation path only passes through one shielding boundary, there is no coupling weakening relationship between adjacent shielding boundaries, the path coupling correction coefficient is set to 1, resulting in a shielding coefficient of 1 × 0.7 = 0.7. If both are subject to the same support constraint and their projected distance in the support force transmission direction is 3m, satisfying the preset force transmission distance condition, an enhancement coefficient greater than 1 is generated. Multiplying the distance attenuation data corresponding to the geometric distance, the shielding coefficient, and the enhancement coefficient yields the initial association strength of the candidate associated monitoring point. This initial association strength is then normalized to the initial association strengths of other candidate associated monitoring points to obtain the mechanical adjacency weight from the target monitoring point to the candidate associated monitoring point.

[0056] Through the above technical solution, this application solves the problem that existing prediction methods are unable to uniformly quantify the shielding effect, force transmission effect, and geometric distance of the support structure. By determining the support propagation path between the target monitoring point and the candidate associated monitoring point through the support shielding topology, the shielding effect corresponding to the excavation zone boundary, corner boundary, and stiffness boundary is identified along the support propagation path, and a shielding coefficient is generated. At the same time, an enhancement coefficient is generated by combining whether the monitoring point is constrained by the same support component, waist beam, or anchor cable and the projection distance between the two in the force transmission direction. Then, the geometric distance, shielding coefficient, and enhancement coefficient are combined into mechanical adjacency weights, so that the model can both weaken the erroneous association across the shielding boundary and enhance the effective association on the same support force transmission path, providing a weight basis that conforms to the engineering mechanics propagation relationship for subsequent spatial attention calculation.

[0057] S4. Generate time-series features of monitoring points based on displacement data. In the spatial attention layer of the model, constrain the feature similarity between the time-series features of monitoring points based on mechanical adjacency weights to obtain spatial fusion features. The process of obtaining spatial fusion features includes: Based on displacement data, determine the displacement value, displacement increment, displacement rate of change, and displacement direction of each monitoring point at the corresponding time step, and combine them into the time series characteristics of the monitoring points; Linear mapping is performed on the time-series characteristics of the monitoring points to obtain query vector, key vector, and value vector; Feature similarity is obtained based on query vectors and key vectors; Multiply the elements at the same dimension positions in the query vector and key vector respectively, and sum all the multiplication results to obtain the vector dot product value; then, perform scale normalization based on the vector dimension to obtain the feature similarity: Feature similarity satisfies: ; in, Indicates time step Feature similarity between target monitoring point i and candidate associated monitoring point j Represents the query vector. Represents the key vector. This indicates the dimensions of the query vector and key vector. Feature similarity is used to characterize the degree of matching between two monitoring points in terms of displacement state, displacement increment, displacement rate of change, and displacement direction at the current time step.

[0058] Furthermore, spatial attention weights are obtained by weighting feature similarity based on mechanical adjacency weights. Spatial attention weights satisfy: ; in, This represents the spatial attention weight of target monitoring point i to candidate associated monitoring point j at time t. Represents the query vector. Represents the key vector. Represents the vector dimension. This indicates a preset zero-prevention constant, which can be taken as... to , Indicates the preset topology constraint strength; The preset topological constraint strength is used to control the degree to which the mechanical adjacency weight constrains the spatial attention weight, and can be set from 0.5 to 2.0, with 1.0 being preferred. When the preset topological constraint strength increases, the spatial attention weight is more affected by the mechanical adjacency weight; when the preset topological constraint strength decreases, the spatial attention weight retains more of the similarity influence between the temporal features of the monitoring points.

[0059] Spatial fusion features are obtained based on spatial attention weights and value vectors; Spatial integration characteristics satisfy: ; in, Represents a value vector. This indicates spatial fusion characteristics.

[0060] Specifically, the displacement data of each monitoring point at the current time step and adjacent historical time steps are obtained. The displacement increment is determined based on the displacement difference between adjacent time steps, the displacement change rate is determined based on the displacement increment and the time interval, and the displacement change direction is determined based on the sign of the displacement increment. Subsequently, the displacement value, displacement increment, displacement change rate and displacement change direction of the same monitoring point at the same time step are combined into the time series characteristics of the monitoring point.

[0061] Linear mapping is performed on the temporal features of the monitoring points to obtain query vector, key vector, and value vector. The inner product of the query vector of the target monitoring point and the key vector of the candidate associated monitoring points is calculated and normalized based on the vector dimension to obtain the feature similarity.

[0062] Then, based on the mechanical adjacency weight, the feature similarity is weighted and modulated, and the modulation results of each candidate associated monitoring point corresponding to the same target monitoring point are normalized to obtain the spatial attention weight; finally, based on the spatial attention weight, the value vectors of each candidate associated monitoring point are weighted and summed to obtain the spatial fusion feature of the target monitoring point at the current time step.

[0063] For example, if a monitoring point has a horizontal displacement of 12.6 mm at the current time step and 12.1 mm at the previous time step, with a sampling interval of 1 hour, then the displacement increment is 0.5 mm, the displacement rate is 0.5 mm / h, and the displacement direction is positive. This displacement value, displacement increment, displacement rate, and displacement direction are combined to form the temporal feature of the monitoring point. If this monitoring point corresponds to three candidate associated monitoring points, and the mechanical adjacency weights obtained in step S3 are 0.60, 0.30, and 0.10 respectively, then after calculating the feature similarity, the spatial attention layer will constrain the contribution of the candidate associated monitoring points according to the aforementioned mechanical adjacency weights, so that candidate associated monitoring points with a strong support-propagation relationship with the target monitoring point occupy a higher weight in the spatial fusion feature.

[0064] The process of linearly mapping the time-series characteristics of monitoring points to obtain query vectors, key vectors, and value vectors includes: Based on preset query mapping parameters, a linear transformation is performed on the time-series characteristics of the target monitoring point to obtain the query vector corresponding to the target monitoring point. Based on preset key mapping parameters, linear transformation is performed on the time-series features of candidate associated monitoring points to obtain the key vectors corresponding to the candidate associated monitoring points. Based on preset value mapping parameters, a linear transformation is performed on the time-series features of candidate associated monitoring points to obtain the value vector corresponding to the candidate associated monitoring points. The query vector is used to calculate the feature similarity between the target monitoring point and the candidate associated monitoring point with the key vector, and the value vector is used to generate the spatial fusion feature of the target monitoring point based on the spatial attention weight.

[0065] The preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters are used to convert the time-series features of monitoring points into a vector representation with a unified dimension.

[0066] Specifically, the temporal features of the monitoring points are represented as feature vectors, and linear transformations are performed on these feature vectors using preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters to obtain query vectors, key vectors, and value vectors. The query vector represents the deformation state retrieval features of the target monitoring point at the current time step, the key vector represents the deformation state features of candidate associated monitoring points that can be matched at the current time step, and the value vector represents the deformation information of candidate associated monitoring points that can be fused into the target monitoring point.

[0067] Preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters can be represented as query mapping matrix, key mapping matrix, and value mapping matrix, respectively. If the dimension of the time-series features of the monitoring points is... The query vector, key vector, and value vector all have the same dimension. Then the dimensions of the query mapping matrix, key mapping matrix, and value mapping matrix are all 1. Generate the query vector, key vector, and value vector using the following methods: , , ; in, This represents the temporal characteristics of monitoring point i at time step t. This represents the temporal characteristics of candidate associated monitoring point j at time step t. This indicates the preset query mapping parameters. This indicates the preset key mapping parameters. This indicates the preset value mapping parameter. Represents the query vector. Represents the key vector. Represents a value vector.

[0068] The preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters can be determined based on historical displacement data and corresponding measured deformation data during the model training phase. During training, the temporal features of monitoring points within the historical time window are used as input, and the measured deformation data for the corresponding prediction period are used as supervision labels. The query mapping matrix, key mapping matrix, and value mapping matrix are iteratively updated by minimizing the error between the predicted deformation and the measured deformation data. After training, the updated query mapping matrix, key mapping matrix, and value mapping matrix are stored as the preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters, and are directly invoked during the prediction phase.

[0069] For example, when the temporal features of a monitoring point include five components—horizontal displacement value, settlement value, displacement increment, displacement rate of change, and displacement direction of change—the dimension of the temporal features is 5. If the spatial attention layer sets the vector dimension to 64, then the preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters can all be set to 5×64 matrices. The 5-dimensional temporal features of each monitoring point at the current time step are mapped to 64-dimensional query vectors, 64-dimensional key vectors, and 64-dimensional value vectors, respectively. Feature similarity is calculated based on the query vectors and key vectors, and spatial fusion features are generated based on the value vectors.

[0070] The preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters are model parameters obtained by training historical displacement data and measured deformation data. They are used to convert the temporal features of monitoring points into a vector representation that can be computed by the spatial attention layer.

[0071] Through the above technical solution, this application solves the problem that existing spatial attention prediction methods rely solely on the similarity of temporal features of monitoring points for information fusion, easily neglecting the constraints of the support structure on deformation propagation. By generating temporal features of monitoring points containing displacement values, displacement increments, displacement change rates, and displacement change directions through displacement data, the model can characterize the deformation state of each monitoring point at the current time step. At the same time, the feature similarity between monitoring points is calculated through query vectors, key vectors, and value vectors, and the mechanical adjacency weight is used to constrain the feature similarity, thereby weakening the similar deformation information across the shield boundary and preserving the effective deformation information on the same support propagation path. Finally, spatial fusion features that conform to the topological constraints of the support are generated, providing an input basis that is closer to the actual engineering force relationship for subsequent temporal prediction.

[0072] S5. In the temporal attention layer of the model, the deformation prediction result for the current prediction period is generated based on the spatial fusion features of continuous time steps. The process of generating deformation prediction results includes: Acquire spatial fusion features of the same monitoring point over multiple consecutive preset time steps; In the temporal attention layer of the model, temporal attention weights are generated based on the spatial fusion features of multiple consecutive preset time steps, and weighted fusion is performed based on the temporal attention weights to obtain the predicted representation data; Based on the predicted characterization data, at least one of the following is generated for the corresponding prediction period: predicted horizontal displacement, predicted settlement, and predicted displacement change rate, which serves as the deformation prediction result for the corresponding prediction period.

[0073] Specifically, spatial fusion features of the same monitoring point within a preset historical time window are acquired and input into the temporal attention layer of the model in chronological order. The temporal attention layer generates temporal attention weights based on the correlation between spatial fusion features at different time steps, with historical time steps contributing more to the current prediction period assigned higher weights. The preset historical time window length determines the number of consecutive time steps input to the temporal attention layer. The preset historical time window length is set according to the monitoring frequency and the foundation pit deformation response cycle; when the monitoring frequency is once per hour, the preset historical time window length can be 24, 48, or 72 time steps; when the monitoring frequency is once per day, it can be 7, 14, or 30 time steps. This time window is used to form a sequence of spatial fusion features for consecutive time steps.

[0074] Subsequently, the spatial fusion features of continuous time steps are weighted and fused based on temporal attention weights to obtain prediction characterization data. The prediction characterization data is then input into the prediction output layer to generate at least one of the horizontal displacement prediction value, settlement prediction value, and displacement change rate prediction value within the corresponding prediction period, and is used as the deformation prediction result for the current prediction period.

[0075] Specifically, a temporal attention layer and a prediction output layer are pre-configured. The temporal attention layer is used to identify historical time steps that contribute differently to the prediction period from the continuous spatial fusion features of the same monitoring point; the prediction output layer is used to convert the prediction representation data generated by the temporal attention layer into at least one of the following: horizontal displacement prediction value, settlement prediction value, and displacement change rate prediction value. This model can be trained based on historical engineering monitoring data. The training samples include the spatial fusion features of continuous time steps and the measured deformation data of the corresponding prediction period. The training objective is to reduce the error between the prediction output and the measured deformation data. The reason for this configuration is that deep foundation pit deformation has time accumulation and stage continuity. The deformation of the current prediction period is not determined solely by the most recent time step, but is jointly related to the spatial fusion features of multiple historical time steps.

[0076] Specifically, let the i-th monitoring point be at time step The spatial fusion features of the previous L consecutive time steps are: ; in, This represents the spatial fusion feature sequence of the i-th monitoring point used for the current prediction period. This indicates that the i-th monitoring point is at time step The spatial fusion characteristics are defined by L, which represents the preset historical time window length. The preset historical time window length can be set according to the monitoring frequency and the construction stage of the foundation pit. For example, when the monitoring frequency is once per hour, 24, 48, or 72 time steps can be selected.

[0077] In the temporal attention layer of the model, the spatial fusion features of consecutive time steps are linearly mapped to obtain the temporal query vector, temporal key vector, and temporal value vector: , , ; in, This indicates that the i-th monitoring point is at time step Time-series query vector, Represents the time key vector. Represents a time-series value vector. , , This represents the linear mapping parameters of the temporal attention layer.

[0078] Set the current time step The time-series query vector is multiplied by the time-series key vector of each historical time step, and then scaled according to the vector dimension to obtain the time-series similarity between the current time step and each historical time step: ; in, This indicates that the i-th monitoring point is at the current time step. Step by step with historical time Temporal similarity between them This represents the dimension of the time-series query vector and the time-series key vector. This time-series similarity is used to characterize the degree of correlation between the spatial fusion features of historical time steps and the current prediction requirement.

[0079] Subsequently, the temporal similarity of each historical time step is normalized to obtain the temporal attention weights: Temporal attention weights satisfy: ; in, This indicates that the i-th monitoring point is at the current time step. Historical time steps The temporal attention weight. The larger the temporal attention weight, the greater the contribution of the spatial fusion features of the corresponding historical time step to the current prediction period.

[0080] The temporal value vectors of consecutive time steps are weighted and fused based on temporal attention weights to obtain the predicted representation data: ; in, This represents the predicted representation data for the i-th monitoring point in the current prediction period. This predicted representation data is derived from the spatial fusion features of continuous time steps, including both historical deformation trends and spatial propagation information after support topology constraints.

[0081] Finally, the predicted representation data is input into the prediction output layer to generate the deformation prediction results for the current prediction period: The deformation prediction results satisfy: ; in, This represents the deformation prediction result for the i-th monitoring point during the current prediction period. Indicates the prediction step size. Indicates the predicted output layer weights. This indicates the bias of the predicted output layer.

[0082] For example, a monitoring point uses a 1-hour sampling interval to acquire the spatial fusion features of the most recent 72 time steps and inputs them into a temporal attention layer. If the temporal attention layer determines that the spatial fusion features of the most recent 6 time steps contribute more to the deformation trend in the next hour, it assigns a higher temporal attention weight to these time steps. Then, the spatial fusion features of the 72 time steps are weighted and fused to obtain the predicted representation data, and based on this, the predicted horizontal displacement value for the next hour is 13.4 mm, and the predicted settlement value is 4.8 mm, which are used as the deformation prediction results for the current prediction period.

[0083] Through the above technical solution, this application solves the problem that existing prediction methods rely solely on a single time point or simple historical sequence extrapolation, making it difficult to reflect the cumulative and phased continuity of deep foundation pit deformation. By acquiring the spatial fusion features of the same monitoring point at multiple consecutive time steps, the model can simultaneously utilize support constraint deformation information from different historical moments. At the same time, by using a temporal attention layer to assign weights to the spatial fusion features of different time steps, the model highlights historical deformation states that contribute more to the current prediction period and weakens the influence of time steps that contribute less. Ultimately, it generates prediction characterization data and deformation prediction results, ensuring that the prediction process retains the spatial propagation relationship under support topological constraints while reflecting the temporal pattern of deep foundation pit deformation as it continues to develop during the construction process.

[0084] S6. Obtain the measured deformation data for the current prediction period, and generate prediction residual data based on the difference between the measured data and the corresponding deformation prediction results. The mechanical adjacency weights for the next prediction period are determined based on whether the predicted residual data meets the preset residual adjustment conditions. Based on the determined mechanical adjacency weights and the updated displacement data, the deformation prediction results for the next prediction period are generated and output through the model.

[0085] The process of determining the mechanical adjacency weights for the next prediction period includes: Obtain the measured deformation data for the current prediction period, and generate prediction residual data based on the difference between the measured data and the deformation prediction results for the current prediction period. Based on the predicted residual data, the consistency of residual sign and residual change rate between the target monitoring point and the candidate associated monitoring point is determined, and the consistency of support propagation path is determined based on the support propagation path. Among them, the consistency of the support propagation path means that the support propagation path between the target monitoring point and the candidate associated monitoring point does not pass through the shielding boundary where the shielding multiplier is not greater than the preset shielding multiplier threshold. When the predicted residual data meets the preset residual adjustment conditions, a residual space consistency marker is generated: ; in, Indicates the consistency marker of the residual space; The preset residual adjustment condition is used to determine whether to adjust the mechanical adjacency weights for the next prediction period. The preset residual adjustment condition is satisfied when the absolute value of the predicted residual data or the norm of the multi-component residuals is not less than the preset residual adjustment threshold. The preset residual adjustment threshold can be set based on the monitoring alarm threshold or historical prediction error statistics, for example, the 75th percentile of the absolute value of the residuals in the historical validation set, or 10% to 20% of the corresponding monitoring item's early warning threshold.

[0086] Based on the residual space consistency label, the mechanical adjacency weights are adjusted to obtain the mechanical adjacency weights for the next prediction period: ; in, Indicates the preset adjustment step size. This indicates that the mechanical adjacency weights in the set of candidate associated monitoring points corresponding to the target monitoring point are normalized. This represents the mechanical adjacency weight for the next prediction period; The preset adjustment step size controls the adjustment magnitude of the mechanical adjacency weights by the residual space consistency marker. The preset adjustment step size is greater than 0 and less than 1, and can be between 0.05 and 0.2, preferably 0.1. If the residual space consistency marker is 1, the corresponding mechanical adjacency weight is increased before normalization; if the residual space consistency marker is -1, the corresponding mechanical adjacency weight is decreased before normalization; if the residual space consistency marker is 0, the corresponding mechanical adjacency weight is not adjusted directionally. After normalization, the sum of the mechanical adjacency weights of the candidate associated monitoring points corresponding to the same target monitoring point in the next prediction period is 1.

[0087] Based on the mechanical adjacency weights for the next prediction period and the updated displacement data, the deformation prediction results for the next prediction period are generated by the model.

[0088] The residual sign consistency is used to characterize whether the prediction deviation directions of the target monitoring point and the candidate associated monitoring point are consistent. Residual sign labels are generated based on the predicted residual data: when the predicted residual data is greater than the preset residual zero tolerance, the residual sign label is positive; when the predicted residual data is less than the opposite of the preset residual zero tolerance, the residual sign label is negative; when the predicted residual data is within the preset residual zero tolerance range, the residual sign label is zero. If the residual sign labels of the target monitoring point and the candidate associated monitoring point are both positive or both negative, then they are determined to satisfy residual sign consistency; if one is positive and the other is negative, then they are determined not to satisfy residual sign consistency.

[0089] Residual change rate consistency is used to characterize whether the prediction error change trends of the target monitoring point and the candidate associated monitoring point are consistent. Prediction residual data for the target monitoring point and the candidate associated monitoring point are obtained for the current prediction period and the previous prediction period. The residual change rate is calculated based on the difference in prediction residuals and the time interval between the two prediction periods. If the signs of the residual change rates are consistent and the absolute value of the difference in residual change rates is not greater than a preset change rate difference threshold, then the two are determined to satisfy residual change rate consistency; otherwise, they are determined not to satisfy residual change rate consistency. The preset change rate difference threshold can be taken as 0.5 to 1.0 times the standard deviation of the residual change rate in the historical validation set, or as 2 to 3 times the change rate corresponding to the monitoring accuracy.

[0090] Support propagation path consistency is used to characterize whether the consistent change in residuals between the target monitoring point and the candidate associated monitoring point is supported by the support propagation path. The support propagation path between the target monitoring point and the candidate associated monitoring point, as well as the set of shielding boundaries traversed by this path, are obtained. The set of shielding boundaries includes at least one of the excavation zone boundary, corner boundary, and stiffness boundary. Then, the shielding multiplier corresponding to each shielding boundary is obtained. If the support propagation path does not pass through a shielding boundary with a shielding multiplier not greater than a preset shielding multiplier threshold, the two are determined to satisfy support propagation path consistency; otherwise, they are determined not to satisfy support propagation path consistency. The preset shielding multiplier threshold can be 0.6 or 0.7. The smaller the shielding multiplier, the greater the weakening effect of the corresponding boundary on deformation propagation.

[0091] Specifically, the difference between the measured deformation data and the predicted deformation result of the same monitoring point in the current prediction period is calculated to obtain the prediction residual data. When the prediction residual data meets the preset residual adjustment conditions, based on the prediction residual data of the target monitoring point and the candidate associated monitoring points, it is determined whether the two satisfy the conditions of residual change in the same direction, consistent residual change rate, and consistent support propagation path. If the above conditions are met, a residual space consistency marker with a value of 1 is generated; if the two residuals change in opposite directions or the support propagation path does not meet the support propagation path consistency, a residual space consistency marker with a value of -1 is generated; otherwise, a residual space consistency marker with a value of 0 is generated.

[0092] Subsequently, the current mechanical adjacency weights are adjusted and normalized based on the residual space consistency label to obtain the mechanical adjacency weights for the next prediction period. If the predicted residual data does not meet the preset residual adjustment conditions, the current mechanical adjacency weights are used as the mechanical adjacency weights for the next prediction period. The measured deformation data for the current prediction period are also added to the original displacement data to form updated displacement data. Based on the mechanical adjacency weights for the next prediction period and the updated displacement data, the model generates and outputs the deformation prediction results for the next prediction period.

[0093] For example, if the predicted horizontal displacement of a monitoring point in the current prediction period is 13.4 mm and the measured horizontal displacement is 13.9 mm, then the prediction residual is 0.5 mm. If the preset residual adjustment threshold is 0.3 mm, then the prediction residual meets the preset residual adjustment condition. If the residuals of this monitoring point and a candidate associated monitoring point are both positive, the difference in the rate of change of the residuals is not greater than the preset rate of change difference threshold, and the support propagation path between the two meets the support propagation path consistency requirement, then the mechanical adjacency weight corresponding to the candidate associated monitoring point is increased; if the residual direction of the other candidate associated monitoring point is opposite, or its support propagation path does not meet the support propagation path consistency requirement, then its mechanical adjacency weight is decreased. The adjusted and normalized mechanical adjacency weight is used for model calculation in the next prediction period, and the measured horizontal displacement value of 13.9 mm is added to the displacement data to generate the deformation prediction result for the next prediction period.

[0094] Through the above technical solution, this application solves the problem that existing residual feedback is prone to being stuck in the same prediction period for post-calibration and is difficult to serve subsequent predictions. By obtaining the measured deformation data of the current prediction period and calculating the difference with the corresponding deformation prediction results, prediction residual data is generated, enabling the model to identify the deviation between the current prediction results and the actual deformation. At the same time, by judging whether the error between monitoring points conforms to the actual deformation propagation relationship through the consistency of residual sign, the consistency of residual change rate, and the consistency of support propagation path, the mechanical adjacency weight of the next prediction period is adjusted accordingly, so that the residual feedback is no longer used to backcalculate the current results, but is used to correct the spatial correlation in subsequent predictions, ultimately improving the matching degree between the deformation prediction results of the next prediction period and the actual engineering deformation trend.

[0095] The spatiotemporal attention model with topological constraints for support is used to generate deformation prediction results based on spatial data, support data, zoning data, and displacement data. This model includes a data input layer, a topological constraint layer, a spatial attention layer, a temporal attention layer, a prediction output layer, and a residual adjustment layer. The data input layer receives displacement data and support shielding topology from the training samples; the topological constraint layer generates mechanical adjacency weights; the spatial attention layer generates spatial fusion features; the temporal attention layer generates prediction representation data; the prediction output layer generates deformation prediction results; and the residual adjustment layer determines the mechanical adjacency weights for the next prediction period based on the prediction residual data.

[0096] Before model training, spatial data, support data, zoning data, displacement data, and measured deformation data collected from historical or current projects are acquired, and training samples are constructed according to a preset historical time window. Each training sample includes displacement data for multiple consecutive time steps, as well as measured deformation data for the corresponding prediction period after that historical time window. For example, when the monitoring frequency is once per hour, the preset historical time window is 72 time steps, and the prediction step size is 1 hour, the displacement data of the most recent 72 hours is used as input, and the measured deformation data of the 73rd hour is used as supervision label.

[0097] After the training samples are constructed, a support shielding topology is generated based on spatial data, support data, and zoning data, and the support propagation path between the target monitoring point and candidate associated monitoring points is determined. For any target monitoring point and candidate associated monitoring point, a shielding coefficient is determined based on the support propagation path, and an enhancement coefficient is determined based on whether they are constrained by the same support, girder, or anchor cable, as well as the projected distance in the force transmission direction. A mechanical adjacency weight is then generated by combining the geometric distance between them. This mechanical adjacency weight serves as the topological constraint input for the spatial attention layer.

[0098] During training, temporal features of monitoring points are generated based on displacement data. These features include displacement value, displacement increment, displacement rate of change, and displacement direction of change. Subsequently, based on preset query mapping parameters, preset key mapping parameters, and preset value mapping parameters, the temporal features of the monitoring points are converted into query vectors, key vectors, and value vectors. Feature similarity is then calculated based on the query vectors and key vectors. Mechanical adjacency weights are used to constrain the feature similarity, generating spatial attention weights. Spatial fusion features are then generated based on these spatial attention weights and the value vectors.

[0099] The spatial fusion features of the same monitoring point at consecutive time steps are then input into the temporal attention layer to generate temporal attention weights. Based on these weights, the spatial fusion features at consecutive time steps are weighted and fused to obtain predictive representation data. The prediction output layer generates deformation prediction results for the corresponding prediction period based on the predictive representation data. These deformation prediction results include at least one of the following: predicted horizontal displacement, predicted settlement, and predicted displacement change rate.

[0100] The model training uses the error between the predicted deformation and the measured deformation data as the prediction loss. For example, the prediction loss can be represented by the mean squared error: ; in, This represents the predicted loss, where N represents the number of monitoring points participating in the training. This indicates that the i-th monitoring point is in the current prediction period. The deformation prediction results This indicates that the i-th monitoring point is in the current prediction period. The measured deformation data, This represents the squared error operation. The query mapping parameters, key mapping parameters, value mapping parameters, temporal attention layer parameters, and prediction output layer parameters are updated by minimizing the prediction loss.

[0101] To avoid backcalculations for the same prediction period, during training, the measured deformation data for the current prediction period is not used to regenerate the deformation prediction result. Instead, prediction residual data is generated based on the difference between the measured deformation data and the deformation prediction result for the current prediction period. The mechanical adjacency weights for the next prediction period are then determined based on these prediction residual data. Specifically, when the prediction residual data meets preset residual adjustment conditions, a residual spatial consistency label is generated based on residual sign consistency, residual rate of change consistency, and support propagation path consistency. The mechanical adjacency weights are then adjusted based on this residual spatial consistency label to obtain the mechanical adjacency weights for the next prediction period. ; in, This indicates the mechanical adjacency weight for the next prediction period. Indicates the current mechanical adjacency weight. Indicates the preset adjustment step size. Indicates the consistency marker of the residual space. This indicates normalization processing.

[0102] Subsequently, the measured deformation data for the current prediction period is added to the displacement data to form updated displacement data. Based on the mechanical adjacency weights for the next prediction period and the updated displacement data, the deformation prediction result for the next prediction period is generated. During training, the prediction error for the next prediction period can also be used as an auxiliary loss, enabling the model to learn the continuous data flow of "current prediction - residual feedback - next period prediction".

[0103] After training, the query mapping parameters, key mapping parameters, value mapping parameters, temporal attention layer parameters, and prediction output layer parameters are stored. The support shielding topology, shielding coefficient, enhancement coefficient, and mechanical adjacency weights generated from spatial data, support data, and partition data are also stored. During the prediction phase, the aforementioned model parameters and topological constraint data are invoked to generate the deformation prediction result for the current prediction period based on the latest displacement data. After obtaining the measured deformation data for the current prediction period, prediction residual data is generated to determine the mechanical adjacency weights and deformation prediction result for the next prediction period.

[0104] For example, a deep foundation pit of a target substation has 32 monitoring points, with a monitoring frequency of once per hour. Displacement data from the most recent 72 time steps is used to predict the deformation result for the next hour. If the displacement data includes both horizontal displacement and settlement components, the displacement data matrix of the training samples is 32×72×2. A mechanical adjacency weight matrix is ​​generated based on the support and shielding topology and used as the topological constraint input for the spatial attention layer. If the predicted horizontal displacement value for a monitoring point in the current prediction period is 13.4 mm, and the measured horizontal displacement value is 13.9 mm, then the prediction residual is 0.5 mm. When this prediction residual meets the preset residual adjustment conditions, the mechanical adjacency weights for the next prediction period are adjusted based on the residual spatial consistency marker, and 13.9 mm is added to the displacement data for deformation prediction in the next hour.

[0105] When the trained model is migrated to a new substation deep foundation pit project, the spatial data, support data and zoning data of the new project are reacquired, and the corresponding support shielding topology, shielding coefficient, enhancement coefficient and mechanical adjacency weight are regenerated. For the model parameters in the spatial attention layer, temporal attention layer and prediction output layer, the parameters obtained from the training of the historical project are called as the initial parameters, and incremental training is carried out using the displacement data and measured deformation data of the initial monitoring stage of the new project.

[0106] During incremental training, the parameters of the prediction output layer and the topological constraint strength parameters are updated first, while the main structure of the query mapping parameters, key mapping parameters, and value mapping parameters is preserved. When the support form of the new project differs from that of the historical project, the parameters of the spatial attention layer and the temporal attention layer are jointly fine-tuned. As a result, the model can reuse existing training results in foundation pit scenarios with different support forms and different excavation depths, and reduce the cost of retraining.

[0107] When it is necessary to generate long-term prediction results, the first prediction step size result is generated based on the deformation prediction results of the current prediction period, and the first prediction step size result is used as one of the inputs of the subsequent prediction step size. The deformation prediction results of multiple prediction periods are generated in a rolling manner according to the preset prediction step size sequence. When the measured deformation data of any prediction period is acquired, the prediction input in the corresponding prediction step size is replaced with the measured deformation data, and the deformation prediction results of the subsequent prediction period are regenerated.

[0108] The preset prediction step sequence can include prediction periods corresponding to the next 1 hour, 6 hours, 24 hours or 7 days; when the monitoring frequency is once per hour, the prediction for the next 7 days can be composed of 168 consecutive prediction steps.

[0109] Example 2: like Figure 2As shown, this application provides a temporal attention prediction system for deep substation foundation pit deformation, used to implement the aforementioned temporal attention prediction method for deep substation foundation pit deformation, including: The data acquisition module 100 is used to acquire spatial data, support data, zoning data and displacement data of multiple monitoring points in the deep foundation pit of the target substation; Topology generation module 200 is used to determine the support shielding topology based on spatial data, support data and zoning data to determine the actual mechanical propagation relationship between monitoring points; The weight generation module 300 is used to determine the support propagation path of deformation influence propagation path between any two monitoring points based on the support shielding topology, determine the shielding coefficient and enhancement coefficient along the support propagation path, and generate mechanical adjacency weights in combination with the geometric distance between monitoring points. The feature fusion module 400 is used to generate time-series features of monitoring points based on displacement data. In the spatial attention layer of the model, the feature similarity between the time-series features of monitoring points is constrained based on the mechanical adjacency weight to obtain spatial fusion features. The prediction output module 500 is used to generate deformation prediction results for the current prediction period based on the spatial fusion features of continuous time steps in the temporal attention layer of the model. The feedback adjustment module 600 is used to acquire the measured deformation data for the current prediction period and generate prediction residual data based on the difference between the measured data and the corresponding deformation prediction results. It is also used to determine the mechanical adjacency weights for the next prediction period based on whether the predicted residual data meets the preset residual adjustment conditions, and to generate and output the deformation prediction results for the next prediction period based on the determined mechanical adjacency weights and the updated displacement data.

[0110] The system proposed in this embodiment and the temporal attention prediction method for deep foundation pit deformation in substations proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0111] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A time-series attention prediction method for deformation of deep foundation pits in substations, characterized in that, The implementation of a model that supports topologically constrained spatiotemporal attention includes: Acquire spatial data, support data, zoning data, and displacement data from multiple monitoring points in the deep foundation pit of the target substation; A support shielding topology that determines the actual mechanical propagation relationship between monitoring points based on spatial data, support data, and zoning data; Based on the support shielding topology, the support propagation path of deformation influence between any two monitoring points is determined. The shielding coefficient and enhancement coefficient are determined along the support propagation path, and the mechanical adjacency weight is generated by combining the geometric distance between the monitoring points. Based on displacement data, the temporal features of monitoring points are generated. In the spatial attention layer of the model, the feature similarity between the temporal features of monitoring points is constrained by mechanical adjacency weights to obtain spatial fusion features. In the temporal attention layer of the model, the deformation prediction result for the current prediction period is generated based on the spatial fusion features of continuous time steps; Obtain the measured deformation data for the current prediction period, and generate prediction residual data based on the difference between the measured data and the corresponding deformation prediction results; The mechanical adjacency weights for the next prediction period are determined based on whether the predicted residual data meets the preset residual adjustment conditions. Based on the determined mechanical adjacency weights and the updated displacement data, the deformation prediction results for the next prediction period are generated and output through the model.

2. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 1, characterized in that, The process of generating a support shielding topology includes: Obtain the coordinates of each monitoring point in the spatial data, and obtain the layout data of the retaining structure, the layout data of the supporting components, and the stiffness data of the supporting components in the support data; Based on the layout data of the enclosure structure, the enclosure structure is divided into continuous layout and mechanical propagation segments, and the turning points of the enclosure structure are determined as corner boundaries. The support zone is divided based on the data of the arrangement of the support components, and is subject to the action of the same support component, lintel or anchor cable. Based on the zoning data, the excavation zones are divided according to the construction organization, and the excavation zone boundaries are determined as the construction boundaries between adjacent excavation zones. Based on the stiffness data of the support components, determine the stiffness boundary when the stiffness difference between adjacent retaining sections or support zones exceeds the preset stiffness difference condition. Based on the coordinates of each monitoring point, each monitoring point is associated with the corresponding retaining section, support area, excavation area, corner boundary, and stiffness boundary to obtain the support shielding topology.

3. The time-series attention prediction method for deep substation foundation pit deformation according to claim 2, characterized in that, The process of determining the transmission route of support includes: Obtain the target monitoring point and the candidate associated monitoring points that have a spatial correlation effect on the deformation prediction of the target monitoring point, as well as the corresponding retaining section, support area, excavation area, corner boundary and stiffness boundary in the support shielding topology; The retaining path between the two is determined based on the retaining section, and the force transmission path between the two is determined based on the support zone through the support, anchor cable or waist beam. The path that meets the preset path connectivity condition among the retaining path and force transmission path is determined as the support propagation path.

4. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 3, characterized in that, The process of determining the shielding factor includes: Determine the set of shielding boundaries, which includes the excavation zone boundaries, corner boundaries, and stiffness boundaries through which the support propagation path passes. The shielding multiplier for the excavation zone boundary is determined based on the construction time difference between adjacent excavation zones; the shielding multiplier for the corner point boundary is determined based on the angle between the support propagation path and the turning direction of the retaining structure; and the shielding multiplier for the stiffness boundary is determined based on the stiffness ratio of adjacent support components. Among them, the shielding multiplier of the excavation zone boundary decreases as the construction time difference increases, the shielding multiplier of the corner boundary decreases as the included angle increases, and the shielding multiplier of the stiffness boundary decreases as the ratio increases. Based on the boundary type, shielding multiplier, and adjacency relationship of each shielding boundary in the shielding boundary set, a path coupling correction coefficient is generated. Path coupling correction factor satisfy: ; in, This represents the r-th shielding boundary along the propagation path. Indicates the number of shielding boundaries along the propagation path. Denotes the masking multiplier of the r-th masking boundary. This indicates the boundary type of the r-th shielded boundary. This indicates the preset type coupling weight between two adjacent shielding boundary types. This indicates the preset coupling strength coefficient; The shielding multipliers along the support propagation path are multiplied sequentially and then multiplied with the path coupling correction coefficient to obtain the shielding coefficient.

5. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 3, characterized in that, The process of determining the enhancement factor includes: Obtain the support zones corresponding to the target monitoring point and the candidate associated monitoring points, and determine whether they are constrained by the same supporting component, the same waist beam, or the same anchor cable; When both are constrained by the same supporting member, the same lintel, or the same anchor cable, obtain the force transmission direction of the corresponding supporting member, lintel, or anchor cable; Generate a monitoring point spacing vector based on the coordinates of the target monitoring point and the coordinates of the candidate associated monitoring points; Project the monitoring point spacing vector onto the force transmission direction to obtain the projected distance in the force transmission direction; When the projection distance meets the preset force transmission distance condition, an enhancement coefficient is generated: ; in, This represents the enhancement coefficient between the target monitoring point i and the candidate associated monitoring point j. Indicates the preset enhancement coefficient. This indicates whether the two are constrained by the same supporting member, the same lintel, or the same anchor cable. Indicates the projection distance. This represents the distance mapping value that increases as the projected distance decreases.

6. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 3, characterized in that, The process of generating mechanical adjacency weights includes: Obtain the target monitoring point and candidate associated monitoring points that have a spatial correlation effect on the deformation prediction of the target monitoring point, and construct a set of candidate associated monitoring points based on the candidate associated monitoring points; Obtain the geometric distance, shielding coefficient, and enhancement coefficient between the target monitoring point and the candidate associated monitoring points; Distance attenuation data is generated based on geometric distance, and the distance attenuation data, shielding coefficient, and enhancement coefficient are multiplied to obtain the initial association strength; The initial association strength in the set of candidate associated monitoring points corresponding to the target monitoring point is normalized to obtain the mechanical adjacency weight: ; in, This represents the mechanical adjacency weight between the target monitoring point i and the candidate associated monitoring point j. Indicates the geometric distance between the two. This indicates the preset spatial distance normalization scale. This represents the set of candidate associated monitoring points corresponding to the target monitoring point i. This represents the candidate associated monitoring points in the candidate associated monitoring point set.

7. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 6, characterized in that, The process of obtaining spatial fusion features includes: Based on displacement data, determine the displacement value, displacement increment, displacement rate of change, and displacement direction of each monitoring point at the corresponding time step, and combine them into the time series characteristics of the monitoring points; Linear mapping is performed on the time-series characteristics of the monitoring points to obtain query vector, key vector, and value vector; Feature similarity is obtained based on query vector and key vector, and spatial attention weight is obtained by weighting the feature similarity based on mechanical adjacency weight. Spatial attention weights satisfy: ; in, This represents the spatial attention weight of target monitoring point i to candidate associated monitoring point j at time t. Represents the query vector. Represents the key vector. Represents the vector dimension. This indicates a preset zero-prevention constant. Indicates the preset topology constraint strength; Spatial fusion features are obtained based on spatial attention weights and value vectors; Spatial integration characteristics satisfy: ; in, Represents a value vector. This indicates spatial fusion characteristics.

8. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 7, characterized in that, The process of generating deformation prediction results includes: Acquire spatial fusion features of the same monitoring point over multiple consecutive preset time steps; In the temporal attention layer of the model, temporal attention weights are generated based on the spatial fusion features of multiple consecutive preset time steps, and weighted fusion is performed based on the temporal attention weights to obtain the predicted representation data; Based on the predicted characterization data, at least one of the following is generated for the corresponding prediction period: predicted horizontal displacement, predicted settlement, and predicted displacement change rate, which serves as the deformation prediction result for the corresponding prediction period.

9. The time-series attention prediction method for deep foundation pit deformation in substations according to claim 8, characterized in that, The process of determining the mechanical adjacency weights for the next prediction period includes: Obtain the measured deformation data for the current prediction period, and generate prediction residual data based on the difference between the measured data and the deformation prediction results for the current prediction period. Based on the predicted residual data, the consistency of residual sign and residual change rate between the target monitoring point and the candidate associated monitoring point is determined, and the consistency of support propagation path is determined based on the support propagation path. Among them, satisfying the consistency of the support propagation path means that the shielding multiplier corresponding to the shielding boundary passed through in the support propagation path between the target monitoring point and the candidate associated monitoring point is not greater than the preset shielding multiplier threshold. When the predicted residual data meets the preset residual adjustment conditions, a residual space consistency marker is generated: ； in, Indicates the consistency marker of the residual space; Based on the residual space consistency label, the mechanical adjacency weights are adjusted to obtain the mechanical adjacency weights for the next prediction period: ; in, Indicates the preset adjustment step size. This indicates that the mechanical adjacency weights in the set of candidate associated monitoring points corresponding to the target monitoring point are normalized. This represents the mechanical adjacency weight for the next prediction period; Based on the mechanical adjacency weights for the next prediction period and the updated displacement data, the deformation prediction results for the next prediction period are generated by the model.

10. A time-series attention prediction system for deformation of deep foundation pits in substations, characterized in that, The time-series attention prediction method for substation deep foundation pit deformation as described in any one of claims 1-9 includes: The data acquisition module is used to acquire spatial data, support data, zoning data, and displacement data from multiple monitoring points in the deep foundation pit of the target substation. The topology generation module is used to determine the support shielding topology based on spatial data, support data, and zoning data to identify the actual mechanical propagation relationship between monitoring points; The weight generation module is used to determine the support propagation path of deformation influence propagation between any two monitoring points based on the support shielding topology, determine the shielding coefficient and reinforcement coefficient along the support propagation path, and generate mechanical adjacency weights in combination with the geometric distance between monitoring points. The feature fusion module is used to generate time-series features of monitoring points based on displacement data. In the spatial attention layer of the model, the feature similarity between the time-series features of monitoring points is constrained based on mechanical adjacency weights to obtain spatial fusion features. The prediction output module is used to generate the deformation prediction result for the current prediction period based on the spatial fusion features of continuous time steps in the temporal attention layer of the model. The feedback adjustment module is used to acquire the measured deformation data for the current prediction period and generate prediction residual data based on the difference between the measured data and the corresponding deformation prediction results. It is also used to determine the mechanical adjacency weights for the next prediction period based on whether the predicted residual data meets the preset residual adjustment conditions, and to generate and output the deformation prediction results for the next prediction period based on the determined mechanical adjacency weights and the updated displacement data.

Images