A soft soil area chemical equipment cluster settlement prediction method based on meta learning and space-time correlation

Abstract

The present application relates to a kind of soft soil area chemical equipment cluster settlement prediction method based on meta-learning and space-time association, comprising: the adaptive space-time graph is built with equipment as node, with the coupling relationship of foundation stress conduction and settlement between equipment as edge, time-varying adjacency matrix is generated by learnable edge weight updater, and space-time association characteristics are obtained by bidirectional space-time interaction fusion;Using meta-learning "meta-training-meta-fine-tuning" strategy, first use the settlement data set of multiple historical factory area to carry out meta-training, obtain the basic model parameter with cross-scene generalization ability;Then, for the cold start scene of new device, the basic model parameter is fine-tuned using a small amount of monitoring data of new device, to realize the rapid adaptation of few samples;In model prediction output, the settlement prediction value and risk level of target equipment cluster in the future specified period are obtained by fusing the soft soil creep and consolidation characteristics.The present application can realize the high-precision, rapid prediction and early warning of soft soil area chemical equipment cluster settlement.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of soft soil engineering and artificial intelligence, specifically involving a method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation. Background Technology

[0002] In soft soil areas (such as coastal and estuarine silt areas), due to the high water content, low compression modulus, and long consolidation period, chemical equipment (with individual units weighing up to thousands of tons) built on these areas is prone to cumulative settlement during long-term operation. Especially when the equipment is clustered, there is a spatiotemporal correlation effect between the diffusion of foundation stress and settlement between the equipment (for example, settlement of tank A can lead to stress concentration in the foundation of tank B, potentially triggering a chain reaction of uneven settlement). This poses a serious threat to the safe operation of pipelines, foundations, and structures throughout the entire plant area.

[0003] Existing settlement prediction technologies for chemical equipment in soft soil areas have the following main shortcomings, making it difficult to meet the needs of actual engineering projects:

[0004] 1) Traditional machine learning models (such as LSTM and XGBoost) usually treat each piece of equipment as an independent entity and make predictions based only on the historical settlement data of a single piece of equipment. They do not consider the spatiotemporal correlation between equipment loads, resulting in excessive prediction bias for equipment at the edge of the cluster and newly commissioned equipment, and failing to reflect the actual settlement coupling effect.

[0005] 2) For newly built chemical plant areas or new equipment, there is often a lack of long-term monitoring data (usually only 1-3 months of data). Traditional models require a large amount of historical data (usually 1-2 years) for training, which makes it impossible for the model to adapt to new scenarios quickly. The implementation cycle can be as long as 6 months or more, thus missing the critical window for early settlement prevention and control.

[0006] 3) Existing models fail to effectively incorporate the creep, consolidation and other time-dependent characteristics of soft soil, and cannot quantify the impact of long-term deformation of soft soil on cluster settlement, further reducing the accuracy of long-term predictions.

[0007] Meta-learning technology enables models to quickly adapt to new scenarios through cross-task learning, while graph neural networks (GNNs) can effectively capture spatial correlation features in network structures. These two technologies offer potential technical paths to solve the aforementioned problems. However, currently, no technical solution deeply integrates meta-learning with spatiotemporal correlation modeling and optimizes it for the time-dependent characteristics of soft soil, thus applying it to the specific scenario of settlement prediction for chemical equipment clusters in soft soil areas. Therefore, there is an urgent need to construct a targeted prediction method to fill this technological gap. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation. This method solves the problems of traditional models ignoring the spatiotemporal coupling between equipment, slow cold start adaptation, and failure to consider the long-term aging characteristics of soft soil. It enables high-precision and rapid prediction of the settlement of new and old chemical equipment clusters, especially the long-term settlement trend.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation, comprising the following steps:

[0010] Step S1. Data Acquisition and Preprocessing: Collect parameters, operating conditions, and historical settlement monitoring data of each piece of equipment in the target chemical equipment cluster, and perform preprocessing.

[0011] Step S2. Spatiotemporal correlation feature construction: Based on the preprocessed data, an adaptive spatiotemporal graph is constructed with equipment as nodes and the relationship between foundation stress transmission and settlement coupling between equipment as edges. A time-varying adjacency matrix is ​​generated through a learnable edge weight updater, and the fused spatiotemporal correlation features are obtained through bidirectional spatiotemporal interaction fusion.

[0012] Step S3. Meta-learning model training and adaptation: Based on the meta-learning method, meta-training is performed using settlement datasets from multiple historical factory areas to obtain basic model parameters with cross-scenario generalization ability; for the cold start scenario of newly built equipment, meta-fine-tuning is performed based on the basic model parameters using a small amount of monitoring data from the target newly built equipment to complete the rapid adaptation of the model.

[0013] Step S4. Fusion Prediction and Output: Input the spatiotemporal correlation features obtained in Step S2 into the prediction model obtained after training and adaptation in Step S3, and output the settlement prediction value; then, based on the creep and consolidation characteristics of soft soil, dynamically correct the settlement prediction value output by the model, and finally obtain the settlement prediction value and risk level of the target equipment cluster in the future for a specified period.

[0014] Furthermore, in step S2, the construction of the adaptive spatiotemporal graph specifically involves:

[0015] Taking a single piece of equipment in a chemical equipment cluster as node V, the node characteristics include equipment load, foundation size, soft soil parameters, and historical settlement sequence statistics. Taking the foundation stress transmission and settlement coupling path between equipment as edge E, based on the distance between equipment, soft soil parameters, load changes, and settlement rate, the edge weights are dynamically calculated through a learnable edge weight updater to generate a time-varying adjacency matrix A(t) to characterize the adaptive change of the spatiotemporal coupling strength between equipment with time and operating conditions.

[0016] Furthermore, in step S2, the bidirectional spatiotemporal interactive fusion specifically includes:

[0017] Based on the generated time-varying adjacency matrix A(t), a graph attention network is used to extract the spatial features of the equipment, and a long short-term memory network is used to extract the settlement time-series features of individual equipment, i.e., the temporal features of the equipment. Through cross-attention mechanism and gating fusion, the spatial and temporal features of the equipment are mutually corrected and fused within the same prediction window to obtain the spatiotemporal correlation features h of each node. i (t):

[0018]

[0019] Among them, h i (t) represents the spatiotemporal correlation feature of equipment i at time t; σ(⋅) represents the Sigmoid activation function, whose output value serves as the gating weight; W g This represents the weight matrix of the gated network; z i S (t) represents the spatial characteristics of equipment i at time t; h i T (t) represents the temporal characteristics of equipment i at time t; [;] represents the concatenation operation; ⊙ represents element-wise multiplication; W z Transformation matrix representing spatial characteristics;

[0020] The obtained fusion feature is represented as h i (t) is fed back to the learnable edge weight updater for updating the time-varying adjacency matrix A(t+1) in the next prediction period, thus achieving iterative optimization.

[0021] Furthermore, in step S3, the meta-learning method employs a model-independent meta-learning algorithm, which includes a meta-training phase and a meta-fine-tuning phase;

[0022] In the meta-training phase, the subsidence dataset of each historical factory area is regarded as a meta-task, and the training objective is to minimize the average prediction error of the basic model parameters on all meta-tasks.

[0023] In the meta-fine-tuning stage, the basic model parameters obtained from meta-training are used as initialization, and a small number of iterative trainings are performed using the monitoring data of the newly built target device to complete the adaptation to the target scene.

[0024] Furthermore, the meta-fine-tuning stage also includes similar task retrieval: based on a pre-built prototype memory library, the features of the target newly built equipment are matched with the typical soft soil-load-settlement prototype vectors of historical plant areas, and the basic model parameters are weighted and initialized based on similar prototypes.

[0025] Furthermore, the meta-fine-tuning stage also includes virtual sample enhancement: based on the existing mapping relationship between soft soil parameters in the plant area and equipment loads, virtual settlement data of the target newly built equipment is generated to supplement the training samples.

[0026] Further, in step S3, the loss function of the model is a combination of mean squared error and spatial correlation regularization term, and its expression is:

[0027]

[0028] Where Loss represents the model's loss function, MSE() represents the mean squared error function, and y pred y represents the model's predicted settlement value. true Represents the actual settlement value, where σ is the regularization coefficient, and A... i,j (t) represents an element in the time-varying adjacency matrix, indicating the edge weight between equipment i and equipment j at time t. pred,i y represents the model's predicted settlement value for equipment i. pred,i This represents the model's predicted settlement value for equipment j.

[0029] Furthermore, in step S4, the dynamic correction of the settlement prediction value output by the model based on the creep and consolidation characteristics of soft soil is specifically as follows:

[0030] The Merchant creep model is introduced, and the instantaneous elastic settlement, viscoelastic settlement amplitude, viscoelastic time constant, and viscoplastic settlement rate of the Merchant creep model are transformed into quantifiable features. The mechanistic baseline settlement is calculated and superimposed with the residual prediction output by the model. The calculation formula for the mechanistic baseline settlement is as follows:

[0031]

[0032] Where s(t) represents the mechanistic baseline settlement at time t, s0 represents the instantaneous elastic settlement, s1 represents the viscoelastic settlement amplitude, τ1 represents the viscoelastic time constant, and s2 represents the viscoplastic settlement rate.

[0033] The degree of consolidation is calculated based on Terzaghi's consolidation theory, and the predicted settlement after superposition is dynamically corrected to make it consistent with the long-term consolidation process of soft soil.

[0034] Furthermore, the method further includes step S5:

[0035] Model updates and optimizations: After the model is deployed, the edge weights of the time-varying adjacency matrix are periodically updated based on the newly received monitoring data in real time, and the model is quickly fine-tuned when a sudden change in operating conditions or an increase in prediction uncertainty is detected.

[0036] The present invention also provides a settling prediction system for chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement the above-mentioned method.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. Precise modeling of spatiotemporal coupling: This invention dynamically depicts the mutual influence of foundation stress and settlement among chemical equipment clusters by constructing an adaptive spatiotemporal graph and a two-way interactive fusion module, which significantly improves the prediction accuracy of equipment inside and at the edge of the cluster.

[0039] 2. Efficiently solves the cold start problem: This invention adopts a two-stage strategy of "meta-training-meta-fine-tuning", combined with prototype memory and virtual sample enhancement, so that the model can quickly complete high-precision adaptation in a very short time (such as 10 days) even in new scenarios with only 1-3 months of monitoring data, shortening the traditional deployment cycle of several months by more than 95%.

[0040] 3. Integration of soft soil mechanical mechanisms: This invention embeds the long-term time-dependent characteristics of soft soil, such as creep and consolidation, into the model in the form of calculable features. This makes the prediction results not only applicable to short-term early warning, but also provide medium- and long-term (1-5 years) settlement trend assessments that conform to the deformation law of soil. The prediction deviation can be controlled within 10%.

[0041] 4. Strong engineering applicability: The model implemented by this invention can be deployed on the edge computing nodes of the park to process sensor data in real time, output predictions and risk levels, and can seamlessly connect with existing monitoring and data acquisition (SCADA) systems to realize real-time early warning and proactive prevention and control of settlement risks. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation provided in the embodiments of the present invention.

[0043] Figure 2 This is a schematic diagram of the "node-edge" graph structure of the chemical equipment cluster in this embodiment of the invention;

[0044] Figure 3 This is a schematic diagram of the meta-learning "meta-training-meta-fine-tuning" and memory enhancement strategies in an embodiment of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0046] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0047] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0048] like Figure 1 As shown, this embodiment provides a method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation, including the following steps:

[0049] Step S1. Data Acquisition and Preprocessing: Collect parameters, operating conditions, and historical settlement monitoring data of each piece of equipment in the target chemical equipment cluster, and perform preprocessing.

[0050] Step S2. Spatiotemporal correlation feature construction: Based on the preprocessed data, an adaptive spatiotemporal graph is constructed with equipment as nodes and the relationship between foundation stress transmission and settlement coupling between equipment as edges. A time-varying adjacency matrix is ​​generated through a learnable edge weight updater, and the fused spatiotemporal correlation features are obtained through bidirectional spatiotemporal interaction fusion.

[0051] Step S3. Meta-learning model training and adaptation: Based on the meta-learning method, meta-training is performed using settlement datasets from multiple historical factory areas to obtain basic model parameters with cross-scenario generalization capabilities; for the cold start scenario of newly built equipment, meta-fine-tuning is performed based on the basic model parameters using a small amount of monitoring data from the target newly built equipment to complete the rapid adaptation of the model.

[0052] Step S4. Fusion Prediction and Output: Input the spatiotemporal correlation features obtained in Step S2 into the prediction model obtained after training and adaptation in Step S3, and output the settlement prediction value; then, based on the creep and consolidation characteristics of soft soil, dynamically correct the settlement prediction value output by the model, and finally obtain the settlement prediction value and risk level of the target equipment cluster in the future for a specified period.

[0053] Step S5. Model Update and Optimization: After the model is deployed, the edge weights of the time-varying adjacency matrix are periodically updated based on the newly received monitoring data in real time. When a sudden change in operating conditions or an increase in prediction uncertainty is detected, the model is quickly fine-tuned.

[0054] 1. Adaptive Spatiotemporal Graph Construction

[0055] The chemical equipment cluster is abstracted as a weighted directed graph G=(V,E,W). The design of nodes and edges expresses the physical coupling and spatiotemporal changes between the equipment, thus providing unified data for subsequent spatiotemporal feature extraction and interactive fusion. Node V represents a single piece of chemical equipment (storage tank, reactor, etc.), and its features include: equipment load (static self-weight + dynamic operating load), foundation dimensions, soft soil parameters (moisture content, compression modulus, consolidation coefficient, etc.), and historical settlement sequence statistics. These node features describe the bearing capacity and foundation conditions of the individual equipment and also serve as input for subsequent edge weight learning and temporal encoding, enabling the model to simultaneously consider differences in equipment attributes and soft soil environment. Edge E represents the stress transmission and settlement coupling path between equipment. The construction of edge E considers not only geometric proximity but also the long-range effects caused by stress diffusion under soft soil conditions; therefore, it can be constructed in two stages: candidate neighborhood and weight learning. First, candidate adjacency relationships are generated based on distance thresholds, influence radii, or soil layer distribution. Then, a learnable edge weight updater is introduced to learn and correct the candidate edge weights online, thus avoiding insufficient coupling characterization caused by relying solely on static distance formulas. This two-stage approach allows the candidate neighborhood to first identify potentially coupled objects, and then edge weight learning to determine the strength and direction of coupling, thereby providing a reliable spatial relationship foundation for subsequent interaction and fusion.

[0056] The learnable edge weight updater dynamically adjusts the edge weights based on recent settlement rate and load changes in each prediction period, thereby generating and updating the time-varying adjacency matrix A(t) online.

[0057] (1)

[0058] Where D is the spacing matrix, E s Let Δq(t) be the compression modulus of soft soil, and Δq(t) be the change in load under the working condition. For the settling rate, It is a multilayer perceptron and subject to physical prior constraints (edge ​​weights decrease with increasing distance). This design allows the spatial coupling strength to adapt to changes in time and operating conditions. When the load on a piece of equipment suddenly changes or the settlement rate increases abnormally, its influence on surrounding equipment can be amplified or redistributed in a timely manner. When the equipment spacing increases or the soil stiffness is high, the edge weights will naturally decrease, thereby avoiding unreasonable long-distance strong coupling.

[0059] In this embodiment, the "node-edge" graph structure of the chemical equipment cluster is as follows: Figure 2 As shown.

[0060] After obtaining the time-varying adjacency matrix A(t), this invention uses an LSTM network to extract the time-series features of single equipment settlement (such as daily settlement amount and settlement rate), i.e., the time features of the equipment; at the same time, it uses a graph attention network (GAT) to capture the spatial features of the equipment (such as the impact of adjacent equipment settlement on the target equipment), and finally integrates them into a spatiotemporal correlation feature with a dimension of N×T×F (N is the number of equipment, T is the time series length, and F is the feature dimension).

[0061] 2. Two-way spatiotemporal interactive fusion

[0062] This method utilizes a time-varying adjacency matrix constructed from an adaptive spatiotemporal graph and dynamic edge weight information to perform bidirectional spatiotemporal interactive fusion. On one hand, it leverages the spatial features calculated from the time-varying adjacency matrix. On the other hand, the time features obtained by LSTM are utilized. Furthermore, by employing bidirectional information flow, spatial influences and temporal evolution can mutually correct each other within the same prediction window, thus avoiding the shortcomings of traditional serial structures where space influences only once and time accumulates only unidirectionally. The core objective of bidirectional spatiotemporal interactive fusion is to enable deep interaction between spatial and temporal features through cross-attention mechanisms and gating fusion. Among these, spatial features... Used as a gating signal, the time encoder is modulated to update its memory of the settlement sequence, causing the spatial coupling relationship of the equipment to affect the temporal feature extraction process of each node. When the overall settlement of neighboring equipment accelerates or the operating disturbance increases, the gating signal guides the time encoder to increase its sensitivity to recent changes. When the influence of the neighboring area is weak, the time encoder relies more on the historical evolution of the equipment itself.

[0063] Cross-attention and gating fusion determines the respective representations of the final node based on the interaction of spatial and temporal features at each time step, i.e., spatiotemporal correlation features. Contributions:

[0064] (2)

[0065] Among them, h i (t) represents the spatiotemporal correlation feature of equipment i at time t; σ(⋅) represents the Sigmoid activation function, whose output value serves as the gating weight; W g This represents the weight matrix of the gated network; z i S (t) represents the spatial characteristics of equipment i at time t; h i T (t) represents the temporal characteristics of equipment i at time t; [;] represents the concatenation operation; ⊙ represents element-wise multiplication; W z Transformation matrix representing spatial features; splicing terms This enables the gating network to simultaneously sense the strength of spatial coupling and changes in temporal state. Weights are generated to enable the same device to adaptively select spatial and temporal information across different dimensions.

[0066] The interactive fusion of spatial and temporal features yields This feedback will be sent to the edge weight updater to update the time-varying adjacency matrix A(t+1) and enhance the interaction, meaning that the spatial and temporal characteristics between nodes will be enhanced through bidirectional spatiotemporal interaction. And dynamically adjusted. And obtained through two-way spatiotemporal interaction. This will serve as the primary input for subsequent soft soil property fusion predictions, used to output future settlement predictions and their calibration. After spatiotemporal interactive fusion, the model further enhances its learning of spatiotemporal relationships through K iterations. In each iteration, the current fused features are used to update the edge weights and node feature representations of the graph, thereby gradually optimizing the ability to capture spatiotemporal coupling and temporal evolution.

[0067] 3. Soft Soil Property Integration Prediction

[0068] To adapt to the time-dependent characteristics of soft soil, this invention proposes a soft soil property fusion prediction method. Based on the original model, the mechanical parameters of soft soil are embedded into the model, and the accuracy of long-term predictions is further improved through dynamic corrections to creep and degree of consolidation. The instantaneous elastic deformation, viscoelastic deformation, and viscoplastic deformation components of the Merchant creep model are transformed into quantifiable features, and the mechanistic baseline settlement is calculated and superimposed with the residual prediction output by the model. These features enable the model to more accurately represent the settlement process of soft soil at different time stages and effectively reflect the trend of settlement rate decay. The calculation formula for mechanistic baseline settlement is as follows:

[0069] (3)

[0070] Where s(t) represents the mechanistic baseline settlement at time t, s0 represents the instantaneous elastic settlement, s1 represents the viscoelastic settlement amplitude, τ1 represents the viscoelastic time constant, and s2 represents the viscoplastic settlement rate.

[0071] Furthermore, dynamic correction of the degree of consolidation can further ensure the long-term predictive performance of the model. This method calculates the degree of consolidation U at different times based on the soft soil consolidation formula (Texas-Gesellschaft consolidation theory). t The superimposed predicted settlement is dynamically corrected to ensure that the long-term prediction (1-5 years) is consistent with the actual consolidation process of soft soil. This correction enables the model's prediction output to be used not only for short-term early warning but also to provide reasonable trend support for medium- and long-term settlement assessment. The final output will then be incorporated into the model training loss calculation and early warning deployment logic.

[0072] 4. Meta-learning cold start adaptation

[0073] To address the cold start problem of newly built equipment, this invention employs a two-stage meta-learning strategy of "meta-training-meta-fine-tuning," such as... Figure 3 As shown, meta-learning provides the subject with training mechanisms for "cross-scene transferable parameter initialization" and "rapid adaptation with few samples," enabling the model to work stably even when data from new equipment is insufficient. This strategy allows the model to quickly adapt to new scenes with only a small amount of monitoring data after learning general settlement patterns across scenes. Simultaneously, a prototype memory is introduced to store typical soft soil-load-settlement prototype vectors from historical plant areas, used for similar task retrieval and weighted initialization in new scenes, further reducing the risk of a cold start. Cross-scene transfer involves first matching similar prototypes in the new scene, then initializing using the parameter or feature distribution of the similar prototypes, so that subsequent fine-tuning does not have to start from a random starting point, further reducing the risk of a cold start.

[0074] Phase 1: Meta-training, i.e., building a cross-scenario base model. Collect a dataset of chemical equipment clusters in multi-source soft soil areas (historical data from multiple plant areas, each containing 1-3 years of settlement data for 5-12 pieces of equipment). Treat each plant area as a "meta-task," and use the MAML (Model-Agnostic Meta-Learning) algorithm to train the base model parameters (such as the weights and thresholds of the GAT-LSTM fusion model) with the goal of minimizing the average prediction error of all meta-tasks. This enables the model to quickly adapt to new scenarios and generalize.

[0075] Phase 2: Meta-fine-tuning. For newly built equipment (with only 1-3 months of monitoring data), it is used as the "target meta-task." Based on the base model obtained from meta-training, only 10%-20% of the data from the target task is used for 10-20 rounds of fine-tuning to complete model adaptation, significantly shortening the training cycle. To further improve accuracy, virtual sample augmentation is introduced. Based on the mapping relationship between existing soft soil parameters in the plant area and equipment loads, virtual settlement data for the newly built equipment is generated. This is particularly suitable for scenarios with extremely scarce data or unstable initial monitoring points, solving the problem of extreme data scarcity.

[0076] 5. Model Training and Deployment

[0077] The model of this invention calculates a time-varying adjacency matrix A(t) based on the distance between equipment, soft soil parameters, and load changes and settlement rates. Graph attention encoding is used to weighted aggregate neighborhood information on A(t) to extract spatial coupling features between equipment. Temporal evolution features are extracted from the settlement sequence, settlement rate, and load change sequence of a single piece of equipment through temporal encoding. Based on this, gated interactive fusion is used to adaptively weight and fuse the obtained spatial and temporal features to obtain a fused representation. And can Feedback is fed to the edge weight updater to iteratively update the graph structure. Then, the soft soil property fusion prediction is applied to the fused representation. Based on this, soft soil mechanical parameters (such as compression modulus, consolidation coefficient, and creep-related characteristics) are embedded, and these parameters are superimposed with the network residuals to output the settlement prediction values ​​for a single piece of equipment for the next 1–30 days. Simultaneously, to improve cross-plant generalization and rapid adaptation to newly built equipment, this invention introduces "meta-training—meta-fine-tuning" meta-learning during the model training phase. Initialization parameters θ0 (covering dynamic edge weight updater, graph attention encoder, temporal encoder, fusion layer, and prediction head parameters) that can be quickly transferred are obtained through MAML training. When only a small amount of monitoring data is available for the target newly built equipment, rapid adaptation is achieved through a few rounds of fine-tuning based on θ0, thus ensuring that the above architecture can still stably output prediction results in data-scarce scenarios.

[0078] Loss function: The loss function of this method adopts a combination of mean squared error (MSE) and spatial correlation regularization term to minimize the deviation between the predicted value and the true value, and to constrain the correlation between the predicted values ​​of adjacent equipment.

[0079] (4)

[0080] Where Loss represents the model's loss function, MSE() represents the mean squared error function, and y pred y represents the model's predicted settlement value. true Represents the actual settlement value, where σ is the regularization coefficient (ranging from 0.1 to 0.3), and A... i,j (t) represents an element in the time-varying adjacency matrix, indicating the edge weight between equipment i and equipment j at time t. pred,i y represents the model's predicted settlement value for equipment i. pred,i This represents the model's predicted settlement value for equipment j.

[0081] Deployment and Application: The model is deployed on the edge computing node of the chemical industrial park to receive equipment monitoring data (such as Beidou positioning and fiber optic sensor data) in real time, output cluster settlement prediction values ​​and risk levels (such as "high risk" for settlement rate > 1mm / d), and connect to the SCADA system to realize early warning.

[0082] In this embodiment, the specific implementation process of the prediction method proposed in this invention is as follows:

[0083] Step 1: Data Acquisition and Preprocessing

[0084] Historical plant cluster data (equipment parameters, soft soil parameters, working load and settlement monitoring data) and data from newly built equipment (1-3 months of monitoring data) were collected. To ensure the comparability of subsequent graph construction and model training, missing data were filled in and outliers were removed (wavelet thresholding / robust filtering), and a unified time reference was constructed (aligning sampling frequency, timestamp, and window boundaries). Subsequently, static and dynamic features were standardized and normalized respectively.

[0085] Step 2: Construction of Spatiotemporal Correlation Features

[0086] Based on the obtained aligned data, an initial 'node-edge' graph structure is constructed. Using equipment as nodes and foundation stress transmission and settlement coupling relationships as edges, candidate adjacency relationships are generated by combining distance thresholds between equipment, influence radii, or soil layer distribution. Initial prior weights are formed using distance and soft soil compression modulus. Based on recent settlement rates and load changes, an edge weight updater is invoked to generate a time-varying adjacency matrix A(t), allowing the spatial coupling strength to adaptively adjust with time and load conditions. Within each prediction window, K spatial message passing and temporal encoding interactions are performed, and bidirectional interactive fusion is achieved through cross-attention mechanisms and gating fusion to obtain a bidirectional fused spatiotemporal feature representation.

[0087] Step 3: Meta-learning model training and adaptation

[0088] The initialization parameter θ0 is obtained through MAML meta-training using the factory area as the meta-task. Multiple historical factory areas are treated as a multi-task set, and a few updates are performed within each meta-task. The meta-parameters are then updated with the overall goal of minimizing the validation error of each meta-task, thus learning an initialization capable of rapid adaptation to different soft soil conditions and equipment layouts. For newly constructed equipment targeting the target, meta-fine-tuning is performed using a small amount of observation data, enabling the model to adapt to the specific settlement patterns of the target scenario in a short time. Simultaneously, similar tasks are retrieved from the prototype memory and weighted initialization is performed. The settlement characteristics of the target scenario are compared with historical prototypes, and the most similar prototypes are selected to weight and guide the model parameters or feature representations, reducing cold-start instability. Virtual sample augmentation is also used to supplement scarce interval data, alleviating the problems of insufficient sample size or incomplete coverage of operating conditions in the early stages of monitoring.

[0089] Step 4: Soft Soil Property Fusion Prediction

[0090] The baseline settlement is calculated based on the creep and consolidation characteristics of soft soil, and then combined with the network residual output to obtain the predicted settlement, outputting the predicted settlement value and risk level for the next 1-30 days. That is, the residual prediction driven by spatiotemporal fusion characteristics is superimposed with the mechanism trend, taking into account both short-term sensitivity and long-term rationality.

[0091] Step 5: Model Update and Optimization

[0092] After deployment, the model receives new monitoring data in real time and updates the edge weight updater and calibration head at fixed intervals. The edge weight updater updates A(t) using the latest settlement rate and load changes to maintain the time-varying adaptation of spatial coupling relationships. The calibration head corrects the mapping relationship with the residuals based on the update mechanism of new data, ensuring that the predicted trend is consistent with the actual settlement process. When a sudden change in operating conditions or an increase in prediction uncertainty is detected, rapid fine-tuning (a few steps of updating) is triggered to avoid the lag caused by long-cycle retraining, achieving online adaptation and continuous performance maintenance.

[0093] Experimental Verification: Prediction of Settlement of a Cluster of Chemical Storage Tanks in a Coastal Soft Soil Area

[0094] 1. Experimental parameters

[0095] Cluster size: 8 storage tanks (nodes) with a capacity of 100,000 m³, spaced 15-25 m apart, with a foundation size of φ20m×1.5m;

[0096] Soft soil parameters: moisture content 45%, compression modulus 2.5MPa, internal friction angle 15°;

[0097] Data details: Historical plant area data (3 coastal soft soil areas, 8 storage tanks per plant area, 2 years of settlement data), and data on newly built equipment (1 month of monitoring data).

[0098] Model parameters: 100 meta-training epochs, 15 meta-fine-tuning epochs, 4 attention heads in the GAT layer, 64 neurons in the LSTM layer, and regularization coefficient λ=0.2.

[0099] 2. Experimental Results

[0100] Prediction accuracy: The 30-day settlement prediction deviation for newly built equipment is 8.2%, and the prediction deviation for cluster edge storage tanks is 9.5%, both of which are better than the traditional LSTM model (deviation 21.3%).

[0101] Adaptation efficiency: The adaptation time for newly built equipment models is 10 days, which is 95% shorter than that for traditional models (6 months);

[0102] Long-term verification: The measured settlement data over one year deviated from the model prediction by 7.8%, proving the effectiveness of the integration of soft soil characteristics.

[0103] This embodiment also provides a settlement prediction system for chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation, including a memory, a processor, and computer program instructions stored in the memory and capable of being executed by the processor. When the processor executes the computer program instructions, it can implement the above-mentioned method.

[0104] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0107] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation, characterized in that, Includes the following steps: Step S1. Data Acquisition and Preprocessing: Collect parameters, operating conditions, and historical settlement monitoring data of each piece of equipment in the target chemical equipment cluster, and perform preprocessing. Step S2. Spatiotemporal correlation feature construction: Based on the preprocessed data, an adaptive spatiotemporal graph is constructed with equipment as nodes and the relationship between foundation stress transmission and settlement coupling between equipment as edges. A time-varying adjacency matrix is ​​generated through a learnable edge weight updater, and the fused spatiotemporal correlation features are obtained through bidirectional spatiotemporal interaction fusion. Step S3. Meta-learning model training and adaptation: Based on the meta-learning method, meta-training is performed using subsidence datasets from multiple historical factory areas to obtain basic model parameters with cross-scenario generalization capabilities; For cold start scenarios of newly built equipment, based on the basic model parameters, a small amount of monitoring data of the target newly built equipment is used for meta-fine-tuning to achieve rapid model adaptation; Step S4. Fusion Prediction and Output: Input the spatiotemporal correlation features obtained in step S2 into the prediction model obtained after training and adaptation in step S3, and output the settlement prediction value; Then, based on the creep and consolidation characteristics of soft soil, the settlement prediction values ​​output by the model are dynamically corrected, and finally the settlement prediction values ​​and risk levels of the target equipment cluster for a specified future period are obtained.

2. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 1, characterized in that, In step S2, the construction of the adaptive spatiotemporal graph specifically involves: Taking a single piece of equipment in a chemical equipment cluster as node V, the node characteristics include equipment load, foundation size, soft soil parameters, and historical settlement sequence statistics. Taking the foundation stress transmission and settlement coupling path between equipment as edge E, based on the distance between equipment, soft soil parameters, load changes, and settlement rate, the edge weights are dynamically calculated through a learnable edge weight updater to generate a time-varying adjacency matrix A(t) to characterize the adaptive change of the spatiotemporal coupling strength between equipment with time and operating conditions.

3. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 2, characterized in that, In step S2, the bidirectional spatiotemporal interactive fusion specifically includes: Based on the generated time-varying adjacency matrix A(t), the spatial features of the equipment are extracted using a graph attention network, and the settlement time series features of a single piece of equipment, i.e., the time features of the equipment, are extracted using a long short-term memory network. By employing cross-attention mechanisms and gating fusion, the spatial and temporal characteristics of the equipment are mutually corrected and fused within the same prediction window to obtain the spatiotemporal correlation characteristics h of each node. i (t): Among them, h i (t) represents the spatiotemporal correlation feature of equipment i at time t; σ(⋅) represents the Sigmoid activation function, whose output value serves as the gating weight; W g This represents the weight matrix of the gated network; z i S (t) represents the spatial characteristics of equipment i at time t; h i T (t) represents the temporal characteristics of equipment i at time t; [;] represents the concatenation operation; ⊙ represents element-wise multiplication; W z Transformation matrix representing spatial characteristics; The obtained fusion feature is represented as h i (t) is fed back to the learnable edge weight updater for updating the time-varying adjacency matrix A(t+1) in the next prediction period, thus achieving iterative optimization.

4. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 1, characterized in that, In step S3, the meta-learning method employs a model-independent meta-learning algorithm, which includes a meta-training phase and a meta-fine-tuning phase. In the meta-training phase, the subsidence dataset of each historical factory area is regarded as a meta-task, and the training objective is to minimize the average prediction error of the basic model parameters on all meta-tasks. In the meta-fine-tuning stage, the basic model parameters obtained from meta-training are used as initialization, and a small number of iterative trainings are performed using the monitoring data of the newly built target device to complete the adaptation to the target scene.

5. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 4, characterized in that, The meta-fine-tuning stage also includes similar task retrieval: based on a pre-built prototype memory, the features of the target new equipment are matched with the typical soft soil-load-settlement prototype vectors of historical plant areas, and the basic model parameters are weighted and initialized based on similar prototypes.

6. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 4, characterized in that, The meta-fine-tuning stage also includes virtual sample enhancement: based on the mapping relationship between existing soft soil parameters in the plant area and equipment loads, virtual settlement data of the target newly built equipment is generated to supplement the training samples.

7. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 1, characterized in that, In step S3, the loss function of the model is a combination of mean squared error and spatial correlation regularization term, and its expression is: Where Loss represents the model's loss function, MSE() represents the mean squared error function, and y pred y represents the model's predicted settlement value. true Represents the actual settlement value, where σ is the regularization coefficient, and A... i,j (t) represents an element in the time-varying adjacency matrix, indicating the edge weight between equipment i and equipment j at time t. pred,i y represents the model's predicted settlement value for equipment i. pred,i This represents the model's predicted settlement value for equipment j.

8. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 1, characterized in that, In step S4, the dynamic correction of the settlement prediction value output by the model based on the creep and consolidation characteristics of soft soil is specifically as follows: The Merchant creep model is introduced, and the instantaneous elastic settlement, viscoelastic settlement amplitude, viscoelastic time constant, and viscoplastic settlement rate of the Merchant creep model are transformed into quantifiable features. The mechanistic baseline settlement is calculated and superimposed with the residual prediction output by the model. The calculation formula for the mechanistic baseline settlement is as follows: Where s(t) represents the mechanistic baseline settlement at time t, s0 represents the instantaneous elastic settlement, s1 represents the viscoelastic settlement amplitude, τ1 represents the viscoelastic time constant, and s2 represents the viscoplastic settlement rate. The degree of consolidation is calculated based on Terzaghi's consolidation theory, and the predicted settlement after superposition is dynamically corrected to make it consistent with the long-term consolidation process of soft soil.

9. The method for predicting the settlement of chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation according to claim 1, characterized in that, The method further includes step S5: Model updates and optimizations: After the model is deployed, the edge weights of the time-varying adjacency matrix are periodically updated based on the newly received monitoring data in real time, and the model is quickly fine-tuned when a sudden change in operating conditions or an increase in prediction uncertainty is detected.

10. A settlement prediction system for chemical equipment clusters in soft soil areas based on meta-learning and spatiotemporal correlation, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, wherein when the processor executes the computer program instructions, it can implement the method as described in any one of claims 1-9.

Images