Water conservancy construction progress prediction system based on multi-source data fusion

Abstract

The invention discloses a water conservancy construction progress prediction system based on multi-source data fusion, and relates to the technical field of artificial intelligence and engineering management, and the system comprises a multi-source sensing collection module which carries out the hierarchical adaptive sampling, quality evaluation and space-time alignment of multiple types of data of a construction site; the data comprises project progress, meteorological environment, geology and hydrology, energy consumption, water consumption and personnel state, and the sampling frequency is dynamically adjusted through an adaptive sampling function according to the prediction error. According to the invention, through hierarchical adaptive sampling and data quality evaluation of the multi-source sensing acquisition module, comprehensive, real-time and high-quality acquisition of construction site data is realized, the problems of data isomerism and uneven quality are effectively solved, the reliability of data input is improved, and the data input efficiency is improved. Through working condition semantic driving and physical consistency constraint fusion of the feature fusion modeling module, internal association among multi-source data can be deeply mined, and fusion features with better interpretation and robustness are generated.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence and engineering management, and in particular to a water conservancy construction progress prediction system based on multi-source data fusion. BACKGROUND

[0002] Water conservancy construction has the characteristics of long cycle, large scale, complex environment, and many uncertain factors. The traditional construction progress management method mainly relies on manual experience, plan charts and regular reports, and has problems such as data acquisition lag, serious information island, low prediction accuracy, and weak ability to deal with emergencies. Meteorological changes, geological condition changes, equipment failures, personnel state fluctuations and other factors on the construction site can have a significant impact on the construction progress. In the construction process, how to dynamically adjust the construction plan according to real-time data to realize fine management and risk avoidance is also a great challenge faced by current water conservancy construction management The prior art has deficiencies in the fusion of multi-source heterogeneous data, intelligent identification of complex working conditions, construction of dynamic time series prediction model, and closed-loop feedback of prediction and regulation, which makes it difficult to meet the demand of modern water conservancy engineering fine management in terms of accuracy and real-time of construction progress prediction, especially in terms of responding to emergencies and optimizing resource allocation, there is a lack of effective intelligent decision support means. SUMMARY

[0003] The application aims to solve the problems in the prior art and provides a water conservancy construction progress prediction system based on multi-source data fusion.

[0004] In order to achieve the above object, the present application adopts the following technical scheme: A water conservancy construction progress prediction system based on multi-source data fusion, comprising: a multi-source perception acquisition module for hierarchical adaptive sampling, quality evaluation and space-time alignment of multiple types of data in the construction site, including engineering progress, meteorological environment, geological hydrology, energy consumption and water consumption and personnel status, dynamically adjusting the sampling frequency according to the prediction error through an adaptive sampling function, and performing signal noise reduction and abnormal correction through a data quality scoring function; a feature fusion modeling module based on a multi-view self-evolution fusion structure driven by working condition semantics, performing semantic alignment, impact event perception and physical consistency constraint fusion on multi-source data, and generating a fusion feature representation conforming to semantic and physical rules; a dynamic time series prediction module based on a working condition hierarchical expert hybrid and intervention sensitive evolution structure, establishing a multi-task nonlinear time series prediction model, outputting a construction progress curve, a completion time and a key path risk index; a self-adaptive weight evaluation module performing double-layer weight optimization and scene switching management according to prediction residuals and control feedback, and adjusting the dynamic weight of each feature factor through a prediction sensitivity optimization function and a control sensitivity optimization function; a closed-loop feedback control module based on multi-objective hierarchical reinforcement learning to generate a construction control strategy, and after verifying the effectiveness of the strategy in a digital twin simulation environment, issuing an execution instruction to realize self-evolution closed-loop control of prediction-control-feedback; a result visual analysis module for visualizing the prediction results and control strategies, including factor contribution explanation, hypothesis deduction, key path dynamic identification and risk warning functions.

[0005] As a further description of the above technical scheme: The multi-source perception acquisition module comprises: an edge acquisition unit accessing meteorological, hydrological, geological and personnel sensors through an Internet of Things communication network, and realizing data time alignment through a PTP time synchronization protocol; an adaptive sampling unit dynamically adjusting the sampling frequency of different data sources according to the prediction error amplitude through a sampling scheduling function; a data quality evaluation unit detecting noise and missing values through a sliding window statistical function and a local anomaly factor algorithm, and repairing abnormal data based on Kalman filtering and time neighborhood interpolation; a physical consistency correction unit performing physical logic correction on the collected data through a water level and flow conservation function and an energy consumption load constraint function.

[0006] As a further description of the above technical scheme: The feature fusion modeling module comprises: a working condition semantic modeling unit extracts historical typical working condition prototypes through a working condition prototype extraction function, calculates the semantic similarity of the current working condition and the historical prototypes by using a working condition matching evaluation function, and maps the current features to a similar working condition semantic space through a working condition alignment mapping function; an impact event perception unit detects sudden events such as rainstorms, landslides or equipment failures through an event candidate detection function, enhances the weight of the event-related modalities by using an event-sensitive feature amplification function, and adjusts the graph structure node relationship by using an event-driven edge weight adjustment function; a physical rule constraint unit detects whether the fused features violate the water balance, energy consumption load or progress operation constraints by using a physical rule consistency evaluation function, and adjusts the features by using a feature correction feedback function in abnormal cases; and a dynamic graph modeling unit establishes a dynamic graph structure with construction units, equipment nodes and monitoring points as vertices, the edges between the nodes include spatial proximity edges, process dependence edges and statistical correlation edges, and the topology adaptive evolution is performed by using a periodic edge weight update function.

[0007] As a further description of the above technical solutions: The dynamic time series prediction module comprises: a working condition hierarchical identification unit determines the current construction stage and selects the corresponding prediction sub-expert model by using a working condition hierarchical identification function; an expert prediction sub-model learns the time-dependent relationship under a specific working condition by using an expert prediction function; a global expert fusion unit dynamically weights the outputs of multiple sub-experts and generates a unified prediction result by using a global expert fusion function; an intervention-sensitive decoding unit converts the future scheduling plan output by the closed-loop feedback control module into an intervention sequence by using a regulation plan encoding function, and generates counterfactual prediction results under different control schemes by using an intervention-sensitive prediction function; and an evolution parameter adaptation unit extracts engineering scale, geological type and main process parameter information by using a project feature encoding function, and dynamically adjusts the model parameters by using an evolution parameter mapping function to complete cross-project transfer learning.

[0008] As a further description of the above technical solutions: The dynamic time series prediction module adopts a multi-task learning mechanism, takes the progress error, risk prediction error and energy consumption prediction error as joint optimization objectives, and improves the prediction accuracy and stability of the model by using a weight sharing and early stopping mechanism.

[0009] As a further description of the above technical solutions: The self-adaptive weight evaluation module comprises: a prediction sensitivity optimization unit calculates the influence degree of each feature factor on the prediction error by using a prediction sensitivity optimization function; a regulation sensitivity optimization unit calculates the influence degree of the change of each feature factor on the regulation result by using a regulation sensitivity optimization function; and a weight collaborative adjustment unit comprehensively updates the weight vector by using a weight collaborative adjustment function according to the prediction and regulation double-layer sensitivity results, so as to balance the prediction accuracy and regulation performance objectives.

[0010] As a further description of the above technical solutions: The self-adaptive weight evaluation module further comprises: a scene mutation detection unit detects scene switching of climate, process or supply mode by monitoring input data distribution and residual change through a scene mutation detection function; a weight reconstruction unit quickly resets weight distribution at scene mutation through a weight reconstruction function, and records and reuses weight configuration through a scene-weight mapping management function; a fairness evaluation and correction unit analyzes balance of multi-objective optimization through a fairness evaluation function, and when detecting that the system is long-term biased towards a single target, calls a fairness correction function to adjust the weight, maintaining balance among safety, livelihood and progress.

[0011] As a further description of the above technical solutions: The closed-loop feedback regulation module comprises: a long-term strategy optimization unit determines total number of personnel, mechanical configuration and construction rhythm on a week or ten-day time scale through a long-term strategy optimization function; a short-term regulation execution unit adjusts construction intensity, operation sequence and equipment running state within a shift time scale through a short-term regulation execution function; a twin simulation evaluation unit simulates and calculates candidate regulation strategies in a virtual construction environment through a twin simulation evaluation function, and selects an optimal scheme according to the simulation result and executes it.

[0012] As a further description of the above technical solutions: The result visual analysis module comprises: an interpretable analysis unit calculates the contribution of each factor to the prediction result and generates an influence visual diagram through an interpretable analysis function; a hypothesis deduction unit simulates prediction curves under different construction or scheduling schemes through a hypothesis deduction function; a key path dynamic identification unit automatically identifies the task node with the greatest impact on the construction period and labels it on the visual interface through a key path dynamic identification function; a comparative display unit displays the progress curve, risk curve and energy consumption curve changes of different schemes through a comparative display function; a risk warning unit sends early warning information to the management terminal through a risk warning function when the predicted risk exceeds the preset threshold.

[0013] As a further description of the above technical solutions: The overall operation process of the system comprises: the multi-source perception acquisition module acquires and cleanses data in real time; the feature fusion modeling module performs semantic alignment, impact event enhancement and physical constraint fusion; the dynamic time series prediction module performs hierarchical expert prediction and intervention-sensitive counterfactual reasoning; the self-adaptive weight evaluation module performs double-layer weight optimization and fairness constraint; the closed-loop feedback regulation module generates multi-objective strategies based on prediction bias and verifies execution in twin simulation; and the result visual analysis module displays prediction curves, risk trends and key paths to realize intelligent decision support.

[0014] The present application has the following beneficial effects: 1、In the present application, firstly, through the hierarchical adaptive sampling and data quality evaluation of the multi-source perception acquisition module, the comprehensive, real-time and high-quality acquisition of the construction site data is realized, the problems of data heterogeneity and uneven quality are effectively solved, the reliability of data input is improved, through the working condition semantic driving and physical consistency constraint fusion of the feature fusion modeling module, the internal correlation between multi-source data can be deeply mined, and impact events can be effectively perceived and responded to, more interpretable and robust fusion features are generated, providing high-quality input for subsequent prediction, through the working condition hierarchical expert mixing and intervention sensitive evolution structure of the dynamic time series prediction module, accurate prediction can be carried out for different construction stages and external intervention scenarios, and the prediction accuracy and the prediction ability of future control scheme are significantly improved.

[0015] 2、In the present application, through the double-layer weight optimization and scene switching management of the self-adaptive weight evaluation module, the system can dynamically adjust the importance of each feature factor according to the prediction error and control feedback, the adaptability and robustness of the system are enhanced, through the multi-objective hierarchical reinforcement learning and digital twin simulation verification of the closed-loop feedback control module, the intelligent closed loop of prediction and control is realized, the optimal control strategy can be generated and risks can be effectively avoided, the intelligent level and decision efficiency of construction management are greatly improved, through the factor contribution explanation, hypothesis deduction and key path dynamic identification function of the result visual analysis module, intuitive and comprehensive decision support information is provided for managers, and the explainability and user friendliness of the system are enhanced. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The figure is a schematic diagram of the system architecture of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] Referring to Figure 1 An embodiment provided by the present application: a water conservancy construction progress prediction system based on multi-source data fusion, comprising a multi-source perception acquisition module, a feature fusion modeling module, a dynamic time series prediction module, a self-adaptive weight evaluation module, a closed-loop feedback control module and a result visual analysis module.

[0019] The multi-source perception acquisition module, as the bottom layer of data input and quality assurance of the system, realizes high-precision acquisition of multiple heterogeneous data of construction sites and external environment, unified time and space alignment, quality evaluation and physical consistency correction, to ensure that the input data into the upper layer of fusion modeling and prediction has timeliness, integrity, accuracy and physical rationality. It is composed of edge acquisition unit, adaptive sampling unit, data quality evaluation unit and physical consistency correction unit, which cooperatively operate through edge computing nodes to form a hierarchical closed-loop data acquisition system. The edge acquisition unit is responsible for the access, data acquisition and preliminary synchronization of multi-source sensors. Multiple types of sensing nodes are deployed on the construction site, including meteorological, hydrological, geological and personnel status sensors. Meteorological sensors collect air temperature, humidity, rainfall, wind speed and evaporation; hydrological sensors collect groundwater level, seepage rate and water flow; geological sensors collect pore water pressure, soil moisture content and ground settlement displacement; personnel status sensors collect worker heart rate, body temperature, blood oxygen, working time and fatigue index. All sensors are connected to edge node gateways through Internet of Things communication network, including LoRaWAN, NB-IoT and 5G modules to ensure data transmission stability in different distances and environments. The edge acquisition unit uses PTP time synchronization protocol to realize multi-node time alignment. All collected data are attached with high-precision timestamps to ensure unified time matching of multi-source data in cross-device and cross-space conditions. The edge node gateway has real-time data buffering and buffer overflow protection mechanism. When the network is temporarily interrupted, the node can store the data in local cache and automatically report after the network is restored, ensuring the continuity of data without loss. The edge acquisition unit also undertakes basic data format conversion and communication verification tasks. All input data are converted into system standard data structure after collection, described in a unified field encoding format, and CRC verification is performed before uploading to avoid numerical abnormalities caused by transmission errors. The adaptive sampling unit dynamically adjusts the sampling frequency according to the system prediction feedback to balance sampling efficiency and data accuracy. The unit has a sampling scheduling function that immediately increases the sampling frequency of related data sources when the system detects that the progress error output by the dynamic time series prediction module exceeds the preset threshold. For weather, hydrological and equipment load data that have a greater impact on progress fluctuations, the sampling frequency is increased by two times the original frequency; when the prediction error tends to be stable, the system automatically reduces the sampling frequency to the normal value to reduce communication and storage load. The adaptive sampling unit sets independent sampling periods for different types of data. Meteorological data are usually collected with a period of 30 minutes, water level data with a period of 10 minutes, equipment load data with a period of 1 minute, and personnel physiological data with a period of 5 minutes. The sampling period of each type of sensor will be automatically reset according to the progress error change. The system dynamically allocates sampling resources based on residual sensitivity to focus on key factors that affect construction progress fluctuations.The adaptive sampling unit has both asynchronous sampling and time unification calibration mechanisms. The asynchronous sampling mechanism ensures that data with different sampling frequencies can still be fused and analyzed in a unified time coordinate system. The time unification calibration mechanism periodically corrects the time errors of the sampling nodes using PTP clocks, thereby maintaining time consistency. The data quality assessment unit performs noise detection, missing value repair, and outlier processing on the collected raw data to ensure data integrity and reliability. This unit internally runs a sliding window statistical function to calculate the mean, variance, missing rate, and abnormal rate of each data stream within a given time window. When the abnormal rate or missing rate exceeds a predetermined threshold, the system automatically starts the data repair process. Missing value repair is completed by time neighborhood interpolation and spatial neighborhood interpolation. Time neighborhood interpolation performs smooth estimation based on the trend of adjacent time values, while spatial neighborhood interpolation makes corrections based on the spatial relationship of adjacent measurement points or similar sensor data. Outlier detection is performed by the local outlier factor algorithm, which judges the degree of abnormality by comparing the data point with its neighborhood distribution; for data judged to be abnormal, the system uses Kalman filtering for dynamic correction, thereby achieving real-time denoising and repair of multidimensional data. Each processed data stream is assigned a data quality score, which takes into account the completeness, stability, and volatility of the data, ranging from 0 to 1. When the quality score is less than 0.85, the system marks the data stream as low confidence and feeds it back to the feature fusion modeling module, so as to reduce its weight during model training, thereby ensuring the overall data reliability of the system. The physical consistency correction unit's task is to perform physical logic correction on the data assessed by quality to make it comply with the laws of hydraulics and energy balance. This unit contains two correction mechanisms: the water level and flow conservation function and the energy consumption load constraint function. When the system detects that the collected water level and flow data do not satisfy the water conservation relationship, the water level and flow conservation function adjusts the abnormal points proportionally based on the upstream and downstream measurement data to ensure that the water balance error is less than 5%. The energy consumption load constraint function is used to correct the energy consumption and load data of construction equipment. When the system detects that the energy consumption change is inconsistent with the equipment operating state, it automatically corrects the data to make the energy consumption and equipment load logically consistent. The physical consistency correction unit records the values before and after correction and the correction factor when performing physical correction, and generates a physical consistency report to be stored in the data log for subsequent tracing and verification. Physical consistency correction not only improves the reliability of the data, but also provides a physical constraint reference for the feature fusion modeling module, making the entire prediction system consistent in physical logic. The multi-source perception acquisition module forms a closed-loop process of "collection - adaptation - evaluation - correction" during operation. The edge collection unit continuously collects data and performs time alignment; the adaptive sampling unit dynamically adjusts the sampling strategy based on prediction feedback; the data quality assessment unit monitors the data state in real time and performs repair; the physical consistency correction unit ensures that the data complies with physical constraints.All the processed data is attached with data source label, timestamp, quality score and correction identifier before entering the feature fusion modeling module, ensuring that the system can achieve high-precision and traceable data management under multi-source input conditions.

[0020] The feature fusion modeling module is located in the data analysis and modeling core layer of the system, and is the key bridge between the multi-source perception collection module and the dynamic time series prediction module. Its task is to align and conform the multi-source heterogeneous data from meteorology, hydrology, geology, energy consumption, personnel status and engineering progress in the semantic space, and complete the impact event perception, physical rule constraint and topology evolution under the graph structure framework, to realize the high-consistency and high-robustness fusion feature representation. It is composed of working condition semantic modeling unit, impact event perception unit, physical rule constraint unit and dynamic graph modeling unit, and each unit forms a tightly coupled fusion mechanism through internal function and parameter mapping. The core goal of the working condition semantic modeling unit is to realize the semantic alignment and unified representation of multi-source heterogeneous data. First, the working condition prototype extraction function is used to cluster the historical construction data and extract multiple typical working condition prototypes. The working condition prototype extraction function is defined as: The first working condition prototype vector is defined as: , : the semantic feature vector of the first working condition prototype, which is spliced by the environmental feature distribution, the operation load mode, the personnel feature distribution and the progress response curve in order, : the working condition prototype extraction function, which summarizes the historical samples in the same cluster into a typical working condition prototype, : the historical sample set divided into the first working condition cluster : the sample index set contained in the first working condition cluster, : the number of samples in the first working condition cluster, : the multi-source original feature vector corresponding to the first historical sample, : the environmental feature component extracted from the first sample, such as the statistical representation of temperature, humidity, rainfall, etc., : the operation load mode feature component extracted from the first sample, including the statistical representation of mechanical operation frequency, energy consumption change, etc., : the personnel feature component extracted from the first sample, such as the combination of worker heart rate, working time, fatigue index, etc., : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, : the progress response curve extracted from the first sample, The progress response curve feature components extracted from the sample data are used to characterize the time dependency between working conditions and progress speed. Each working condition prototype consists of four types of semantic features: environmental feature distribution, workload pattern, personnel feature distribution, and progress response curve. The environmental feature distribution covers external conditions such as temperature, humidity, and rainfall; the workload pattern describes the frequency of machine operation and energy consumption changes; the personnel feature distribution characterizes worker status and work duration; and the progress response curve reflects the time dependency between working conditions and task progress speed. After real-time data enters the system, the working condition matching evaluation function calculates the semantic similarity between the current working condition and each historical prototype, using a similarity threshold of 0.8 as the judgment criterion. Working condition matching evaluation function: current time... Working condition semantic vector With the Prototype for each working condition Matching similarity is defined as: , Current moment The working conditions and the first The semantic similarity between individual working condition prototypes, and its range of values. The higher the value, the better the match. : The working condition matching evaluation function is used to calculate the similarity between the current working condition and the historical working condition prototypes in the semantic space. Current moment The semantic feature vector of work conditions, constructed from multi-source data, includes, in order, environmental feature distribution, work load pattern, personnel feature distribution, and progress response features. : No. The semantic feature vector of each working condition prototype is calculated by the working condition prototype extraction function. : Current working condition semantic vector and the first The inner product of prototype semantic vectors, The L2 norm of the current working condition semantic vector is used to normalize the similarity. : No. The L2 norm of the semantic vector of each working condition prototype, in actual judgment, when At that time, the current working condition can be considered to be the same as the first... The prototype of the current working condition is semantically matched. When the similarity reaches or exceeds a threshold, the system determines that the current working condition matches the target prototype. The working condition alignment mapping function maps the current multi-source features to the semantic space corresponding to the prototype, thereby making features from different time periods and different engineering projects comparable and aggregated on a unified semantic scale. Working condition alignment mapping function: in relation to the prototype... After matching the prototype of each working condition, the current multi-source feature vector is... Perform semantic alignment mapping to obtain aligned semantic feature vectors. : , : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the multi-source input feature vector of the current time t, including weather, hydrology, geology, energy consumption, personnel status, and engineering progress, etc. original or pre-processed features, : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the semantic feature vector of the i-th working condition prototype, providing semantic reference for determining the mapping parameters, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the mapping parameter matrix corresponding to the i-th working condition prototype, describing the linear combination relationship of different source features in the semantic space of this working condition, : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is : the event candidate confidence score of the current time t, the value range is nonlinear compression function of interval, e.g. logistic function , when exceeds a preset threshold, the current time is recorded as a candidate time of impact event, and subsequent feature amplification and edge weight adjustment are triggered. When abnormal combinations of meteorological, hydrological and equipment operation signals are detected, such as the simultaneous occurrence of rapid increase in rainfall, increase in seepage rate and rise in water level in the construction area, the system marks this combination as a rainstorm event. Similarly, when the output of the geological sensor shows an abnormal increase in pore water pressure accompanied by equipment shutdown signals, the system identifies it as a landslide risk event. After the event is confirmed, the event-sensitive feature amplification function immediately performs weighting operation on the relevant modal features, and the weight of the features closely related to the event is increased by 1.2 to 1.5 times to highlight the impact of the event on subsequent modeling. Event-sensitive feature amplification function: weighting and amplifying the modal features related to the event to obtain the amplified features : , : the amplified feature vector of the first class of modal at time , used to enhance the expression strength of the impact event on this modal, : event-sensitive feature amplification function, weighting and amplifying specific modal features according to event confidence, : the original fusion feature vector of the first class of modal at time , such as meteorological modal features, hydrological modal features, geological modal features or equipment load modal features, : the event confidence score at the current time output by the event candidate detection function, : the amplification coefficient of the first class of modal, when takes a value in the interval and is close to 1, the overall amplification factor is about 1.2 to 1.5, and for the modal unrelated to the event, it can be made close to 0, so that its feature weight remains basically unchanged. The event-driven edge weight adjustment function reweights the nodes affected by the event and their adjacent relationships in the graph structure, so that the edge weight of the event propagation path is adjusted in real time according to the intensity of the event. Event-driven edge weight adjustment function: in the graph structure, the edge weight between node and node is adjusted by the event, and the adjusted edge weight is obtained: , : the edge weight value adjusted by the event at time , used to strengthen or weaken the relationship between node and node Information dissemination capabilities between them Event-driven edge weight adjustment function: Adjusts edge weights based on event confidence and whether the edge is on the event's impact path. Before event mediation, at any time node With nodes The original edge weights between them can be determined by spatial proximity, process dependency, or statistical correlation. The event confidence score at the current moment is output by the event candidate detection function. Indicator node With nodes The binary label indicates whether the edge between them belongs to the path affected by the event. The value is 1 if the edge is directly affected by the current event, and 0 otherwise. : Event edge weight amplification factor, used to control the amplification strength of the event on the relevant edge weights. It is usually taken as a positive value to enhance the propagation path of key information when the event occurs, and when the event subsides, Gradually decreasing, The weights then return to their original levels to prevent the long-term amplification of the event's impact. While the event persists, the system maintains high-weight connections to enhance feature flow; after the event ends, the weights gradually return to normal to prevent excessive information amplification. Furthermore, the impact event perception unit has an event recovery mechanism that performs regression analysis on the data window after the event subsides, calculates the stability index of the system features before and after the event, and feeds the results back to the physical rule constraint unit, providing a reference for subsequent physical consistency correction. The physical rule constraint unit ensures that the fused features conform to the basic engineering laws of hydraulic construction at the physical logic level. Through the physical rule consistency evaluation function, the fused features are verified by multi-dimensional constraints, mainly including water balance constraints, energy consumption load constraints, and schedule constraints. The physical rule consistency evaluation function defines a physical rule consistency index for the consistency of the fused features under the three types of physical constraints: water balance, energy consumption load, and schedule. , :time The physical rule consistency index indicates that the smaller the value, the more the integration characteristics conform to engineering principles such as water conservation, energy consumption-load matching, and schedule-operation correspondence. The physical rule consistency evaluation function weighted aggregates the residuals of various physical constraints into a single scalar evaluation value. The hydrological residuals calculated from the water balance constraints are used to measure the degree of deviation of the combined characteristics such as rainfall, seepage rate, and groundwater level from the water conservation level. The energy consumption-related residuals calculated from energy load constraints are used to measure the consistency deviation between construction equipment start-up and shutdown records, power load, and energy consumption curves. : Progress-related residual calculated from progress operation constraints, used to measure whether there is unreasonable jump between task completion rate and operation duration, equipment start-stop record, : Weight coefficient of water balance constraint residual, used to reflect the importance of hydrological constraints in overall physical consistency, : Weight coefficient of energy consumption load constraint residual, used to reflect the importance of device energy consumption and load matching constraints, : Weight coefficient of progress operation constraint residual, used to reflect the importance of progress and operation correspondence constraints, when exceeds the set threshold, the physical rule constraint unit triggers the feature correction feedback function to back off and reweight the related features. In the water balance constraint, the system checks the rainfall, seepage rate and groundwater level three types of features in linkage, when detecting that the rainfall surges while the water level does not rise synchronously, the system determines that this feature group violates the physical law, and immediately calls the feature correction feedback function to perform correction. Feature correction feedback function: After detecting that the physical rules are obviously violated, the current fusion feature vector is corrected to obtain the modified fusion feature : , : The fusion feature vector after physical constraint correction at time will be used as the input of subsequent dynamic graph modeling and time series prediction, : Feature correction feedback function, which performs directional rollback on the current fusion feature according to the physical consistency index, : The original fusion feature vector at time is obtained after processing by the aforementioned working condition semantic modeling and event perception units, : The physical consistency index at time output by the physical rule consistency evaluation function, used to control the correction step size, : Physical constraint correction step length coefficient, used to adjust the feature rollback degree, when the value is larger, the correction is more aggressive, : The reference fusion feature vector corresponding to time , usually taking the trusted feature state from the recent period that meets the physical rules, : The difference direction between the current fusion feature and the reference trusted feature, : the two-norm of the difference vector, used to unitize the difference direction, facilitating the rollback by scalar step, the modified form embodies the idea of "rollback to the nearest trusted state under physical constraints", and records the values before and after the modification and the call time in the modification log. The modification process ensures that the modified data meets the water conservation requirements by reweighting the characteristics of the abnormal nodes and rolling back to the trusted values at the previous time. The energy load constraint is related to the running state of the construction machinery and the power load data. When the start-stop record of the mechanical equipment does not match the energy consumption trend, the system determines that the data is inconsistent and performs modification, so that the energy consumption curve and the equipment operation signal remain synchronized. The progress constraint ensures the correspondence between the engineering task quantity and the operation time. When the task completion rate is found to rise while the equipment start-stop record does not change, the system automatically modifies the abnormal part of the progress data. All modification operations are recorded in the physical consistency log, which contains the values before and after the modification, the basis for the modification and the function call time, for subsequent auditing and model tracing. Through the above mechanism, the physical rule constraint unit ensures the credibility and engineering rationality of the fused features, making the model consistent at the calculation and physical levels. The dynamic graph modeling unit is responsible for establishing and maintaining the dynamic topology structure of the construction scene to support efficient propagation of multi-source features in the time and space dimensions. The construction unit, equipment node and monitoring point are taken as the vertices of the graph, and each node carries three types of attributes: fused feature vector, data confidence and working condition label. Nodes are connected by three types of edges: spatial proximity edge, process dependence edge and statistical correlation edge. The spatial proximity edge describes the relationship between nodes that are close to each other in geographical space, such as the connection between water level sensors and temperature and humidity sensors in the same construction area; the process dependence edge reflects the dependence relationship in the construction sequence, such as the dependence of the concrete pouring node on the completion of the formwork installation node; the statistical correlation edge is established according to the correlation index in the long-term historical data to describe the coordinated change trend between different parameters. The dynamic graph modeling unit performs edge weight update operation once every prediction period, adjusting the connection strength between nodes through the periodic edge weight update function. Periodic edge weight update function: at the end of each prediction period, the edge weight between nodes and is periodically updated to obtain the edge weight of the next period: , : the updated edge weight between node and node in prediction period , used to describe the connection strength between the two nodes in the graph, : the periodic edge weight update function, which weights and fuses the historical edge weight and the current period feature similarity, : the updated edge weight between node and node in prediction period The current edge weights between them can include the results of event-driven adjustments. During the forecast period internal nodes The carried fusion feature vector contains multi-source information corresponding to data confidence and operating condition labels. During the forecast period internal nodes The carried fusion feature vector :node With nodes The inner product of features is used to measure the semantic similarity between the two. , :node and nodes The L2 norm of the fused feature vectors is used to normalize the similarity. : Smoothing coefficient for edge weight update, range of values , A higher similarity value indicates a stronger influence of the current cycle's feature similarity on the new edge weights. This update method allows the graph structure to adaptively evolve in response to new operational changes and statistical correlation changes while recording historical topology, thus connecting with the calculation of topology stability indices and subsequent re-evaluation mechanisms. When the system detects changes in operational conditions, such as sudden climate changes or shifts in operational phases, the topology structure is automatically reconstructed based on the changes, and the new edge weight distribution is written into the topology weight table for subsequent prediction modules to use. A topology evolution tracking mechanism records the evolution of the graph structure over time. After each structural update, the system calculates a topology stability index to measure the impact of changes in node connections on overall information propagation efficiency. When the stability index drops below a threshold, the system automatically triggers a re-evaluation of the fused features to ensure that the information propagation path remains optimal. Data undergoes four stages within the module: semantic alignment, event enhancement, physical correction, and topology propagation, resulting in a fused feature representation with high confidence and high consistency. The fusion result includes semantic labels, event status identifiers, physical consistency scores, and node weight matrices in the output, serving as input to the dynamic time-series prediction module and providing a stable and reliable feature foundation for progress curve prediction and risk assessment. Through the above structure and operating mechanism, the feature fusion modeling module realizes a fusion system that combines semantic-driven, adaptive, and physical constraints, enabling the water conservancy construction progress prediction system to maintain the stability and reliability of its prediction performance even in the face of complex environments and unexpected events.

[0021] The dynamic time-series prediction module is the core time reasoning and trend analysis layer of the system. It is responsible for establishing a nonlinear time dependency model based on multi-source fusion features, thereby predicting progress changes, completion time, critical path risks, and energy consumption trends during water conservancy construction. It includes a working condition hierarchical identification unit, an expert prediction sub-model, a global expert fusion unit, an intervention-sensitive decoding unit, and an evolution parameter adaptation unit. A multi-task learning mechanism is used to jointly optimize progress errors, risk prediction errors, and energy consumption prediction errors. The working condition hierarchical identification unit determines the current construction stage and its corresponding environmental type, providing a basis for selecting an appropriate expert model in the subsequent prediction process. The system predefines several working condition levels, each corresponding to a specific construction stage and external conditions, such as the foundation excavation stage, high-temperature construction stage, or seepage prevention operation stage. The working condition hierarchical identification function analyzes the input multi-source fusion features to identify the working condition category of the current sample. The working condition hierarchical identification function is the current fused semantic feature vector output by the feature fusion modeling module. , define the first The confidence level for each working condition level is: , In time At that time, the first The recognition confidence level for each working condition level takes a value of [value range missing]. The sum of the confidence levels for all levels is 1. The working condition stratification identification function is used to determine the probability distribution of the current sample belonging to each working condition level based on input features. :time The multi-source fusion semantic feature vector is composed of a comprehensive temporal representation of environmental parameters, mechanical load, personnel status, and geological and hydrological features. : with the The weight vector corresponding to each working condition level is used to characterize the discrimination direction of that level in the semantic space. : with the The offset scalar corresponding to each working condition level is used to adjust the intercept of the hyperplane. The total number of predefined working condition levels in the system, such as different construction stages and environmental combinations like foundation excavation, high-temperature construction, and seepage prevention operations. : No. Each working condition level in time The unnormalized score on The exponential function is used to map scores to positive numbers for normalization. The system uses... The highest level is designated as the current working condition category, and is written to the model state cache when a category change is detected, triggering the expert model switching logic. The recognition process is based on parsing the semantic feature vector output by the feature fusion modeling module. The system calculates the temporal patterns of environmental parameters, mechanical load, personnel status, and geological and hydrological features, and judges working condition transitions through threshold classification rules or clustering indicators. When a working condition switching signal is detected, the new working condition label is written to the model state cache, triggering the expert model switching logic, enabling the system to maintain the continuity and stability of the predicted structure at different construction stages. The working condition layering recognition unit is dynamically updated during operation. After each prediction cycle, the layering boundary is re-evaluated based on the statistical distribution of the input data to ensure that the recognition results match the actual site conditions. The expert prediction sub-model is the core computational unit of the dynamic temporal prediction module, with each sub-model independently corresponding to a specific working condition level. The expert prediction function learns the temporal dependency between the fused features and construction progress changes under that working condition during the training phase. Expert prediction function: For the identified working condition... For samples at each working condition level, the expert prediction sub-model within the time window The input sequence is Predicting future moments The multi-task output vector is defined as: , : By the The time given by the expert model for each working condition The predicted output vector typically contains three components: the progress curve value, the critical path risk indicator, and the energy consumption trend value. : No. An expert prediction function for each operating condition is used to learn the temporal dependency between fused features and future indicators under that operating condition. From time Time Continuous The fused semantic feature sequence at each time point reflects the recent changes in working conditions and operational status. The length of the time window used by the expert model, i.e., the number of historical time steps involved in the prediction. Prediction step size, used to specify the time relative to the current time. Predicting future moments, for example, if For several forecast periods, progress curves and energy consumption curves can be generated. The feature vectors within the time window are concatenated into a long vector in chronological order, which is then used as input for temporal feature extraction. : No. The time series feature extraction matrix of the expert model for each working condition is used to perform linear transformation on the vectorized historical sequence. : No. The intermediate layer bias vector of the working condition expert model is used to add translational degrees of freedom after feature extraction. Nonlinear activation functions are used to enhance the model's ability to fit complex nonlinear time dependencies; for example, ReLU or hyperbolic tangent functions can be used. : No. The output layer weight matrix of the expert model for each working condition is used to map intermediate features to the multi-task output space. : No. The output layer bias vector of each work condition expert model is used to adjust the overall level of the prediction results. It receives feature sequence input from the feature fusion modeling module, extracts time-series features, and performs state recursion to generate prediction outputs for future progress curves and energy consumption trends. In actual operation, the system maintains an independent expert model library for each work condition. Expert models for different work conditions maintain a consistent parameter structure, but their internal weights differ based on the feature distribution differences of historical training data. In this way, the system can quickly switch to the most suitable expert model when climate conditions, geological features, and work intensity change, improving prediction accuracy. The expert prediction sub-model not only outputs the progress curve but also simultaneously generates a project completion time estimate and a critical path risk indicator. The progress curve represents the cumulative completion percentage of the task over time; the completion time estimate is obtained based on time-series extension prediction; and the critical path risk indicator is calculated based on inter-process dependencies and node delay probabilities. All output results are subsequently integrated by the global expert fusion unit. The global expert fusion unit performs dynamic weighted aggregation among multiple expert prediction results to generate a unified prediction output. The global expert fusion function performs a weighted average of the expert outputs based on the real-time work condition weights. Global expert fusion function: expert prediction results for all operating conditions. The global prediction output is defined as: , The global expert fusion unit in time The unified forecast output vector includes comprehensive results such as schedule curves, completion time estimates, energy consumption trends, and critical path risks. The global expert fusion function is used to dynamically weight and average the prediction results of multiple working conditions from various experts. : No. A working condition expert at time The predicted output vector, :time Time and the The fusion weight corresponding to each working condition expert is non-negative and satisfies the following conditions: The confidence level of the operating condition identification and the historical prediction residuals are jointly determined. The total number of working condition levels and corresponding expert models. : the result of weighting and superimposing each expert's prediction output by weight, used to obtain the most reasonable comprehensive prediction for the current working condition, the system can adjust the weight in real time according to the residual error of each expert , the expert with smaller residual error will get higher weight in the subsequent period. The weight distribution is based on the working condition matching confidence provided by the working condition hierarchical identification unit, when the confidence of a certain working condition state is high, the corresponding expert output weight is correspondingly increased. During the fusion process, the system monitors the prediction residual error of each expert model in real time, and adjusts the weight proportion accordingly. The expert with smaller residual error is given higher weight, thereby enhancing the stability and accuracy of the overall prediction result. The fusion unit also maintains a record of historical weight changes for subsequent model evaluation and optimization. In the output stage, the global expert fusion unit generates a complete set of prediction results, including construction progress curve, completion time prediction, energy consumption trend and key path risk assessment results. These results serve as input for the closed-loop feedback regulation module, providing basic data for subsequent strategy generation. The intervention-sensitive decoding unit is used to realize the counterfactual prediction capability, enabling the system to infer future construction progress under different regulation schemes. Receive the future scheduling plan from the closed-loop feedback regulation module, convert the regulation strategy into an intervention sequence through the regulation plan encoding function, and input it together with the multi-source features into the prediction network. Regulation plan encoding function: encode the future time step regulation plan sequence into an intervention semantic vector: , : the intervention semantic vector corresponding to time , used to summarize the overall characteristics of the regulation scheme for a period of time in the future, : regulation plan encoding function, encode the scheduling plan in the form of discrete time sequence into a fixed length vector, : regulation plan sequence from time to time , each vector can contain scheduling instructions such as equipment addition or reduction, shift adjustment, resource reallocation, : single-step regulation action vector at future time , elements can represent the scheduling intensity or regulation level of different equipment and personnel, : the number of future time steps covered by the regulation plan encoding, corresponding to the influence window length of the regulation scheme in the prediction, : regulation plan encoding matrix, used to project the single-step regulation action into the intervention semantic space, : average the encoded regulation action in the time dimension to obtain the comprehensive representation of the overall regulation scheme. In the decoding process, the intervention-sensitive prediction function generates prediction results under different schemes according to the input intervention sequence. Intervention-sensitive prediction function: given the current time implicit state vector with a certain scenario corresponding intervention semantic vector under the condition of, the scenario predicted output is defined as: , : under the scenario , the intervention-sensitive prediction output vector at time , which can include construction progress curve values, risk indicators, and energy consumption levels, : the scenario corresponding intervention-sensitive prediction function, used for counterfactual prediction under the condition of given current state and intervention scheme, : the implicit state vector at time , which is obtained by dynamically encoding the historical multi-source feature sequence and working condition information with a dynamic time sequence network, : the intervention semantic vector output by the regulation plan encoding function under the scenario , different scenarios correspond to different regulation schemes, : the weight matrix for mapping the current implicit state under the scenario , which describes the influence of historical working conditions on future prediction, : the weight matrix for mapping the intervention semantic vector under the scenario , which describes the influence of the regulation scheme on future prediction, : the bias vector under the scenario , used to adjust the overall level of the scenario prediction output, : the output layer nonlinear function, used for joint modeling of multi-task output, for example, piecewise linear or saturated nonlinear, different scenarios may correspond to "maintain existing configuration", "increase equipment investment", "adjust shift schedule" and other schemes, the system provides decision basis for managers by comparing different The system can output progress prediction under multiple scenarios, such as progress curve under the three schemes of maintaining existing resource allocation, increasing equipment investment, or adjusting shift schedule. By comparing different prediction results, managers can evaluate the impact of each regulation scheme on construction efficiency and risk, thereby realizing data-driven decision support. The intervention-sensitive decoding unit also has a regulation response analysis mechanism. When the model detects that the intervention scheme has a significant impact on the prediction result, the system records the mapping relationship between the intervention variable and the progress change, providing a basis for the closed-loop feedback regulation module to optimize strategies. The function of the evolution parameter adaptation unit is to realize the adaptive migration of the model between different projects. Through the project feature encoding function, the project size, geological type, and main process parameter information are extracted to form a project feature vector. Project feature encoding function: encoding the project feature vector describing the project size, geological type, and main process parameters to obtain the project semantic vector: , : Project semantic vector, used to depict the comprehensive characteristics of the current project in terms of scale, hydrogeological conditions and main construction technology, : Project feature encoding function, mapping original project description parameters to the semantic feature space available for the model, : Project feature vector, elements can include total construction scale, main structure form, typical geological type, main hydrological characteristics and key construction technology parameters, etc. static information, : Project feature encoding matrix, used to linearly combine different project features to extract directions that are discriminative for the prediction task, : Project feature encoding bias vector, used to adjust the center position of the feature distribution after encoding. The evolutionary parameter mapping function dynamically adjusts the model parameters according to this feature vector, so that the model output in different engineering environments remains consistent and generalizable. Evolutionary parameter mapping function: Given the base parameter vector and the project semantic vector , the evolved parameter vector under the current engineering environment is defined as: , : Evolved model parameter vector at time , used for dynamic time series prediction of the current project, : Evolutionary parameter mapping function, which adaptively adjusts the base parameters within a limited range according to project information, : Base parameter vector trained in the source project or benchmark scenario, serving as the starting point for cross-project migration, : Project semantic vector, obtained by project feature encoding function, reflecting the static characteristics of the current project, : Parameter evolution weight matrix, used to map the project semantic vector to an intermediate representation of parameter scaling factors, : Hyperbolic tangent function, whose output range is , : Relative adjustment ratio of each parameter dimension, limited to , corresponding to a parameter variation amplitude of no more than ±20%, : All-1 vector with the same dimension as the parameter vector, used to construct a scaling coefficient of "1 + relative adjustment ratio", The element-wise vector multiplication operator is used to scale each dimension of the basic parameter vector according to a scaling factor. This naturally limits the variation of all parameters to 80%–120% of the basic value, satisfying the requirement of "parameter variation range constrained to ±20%". The parameter adaptation process is confined within a control range, with the system limiting parameter variation to ±20% to ensure the transferred model does not deviate from the original stable solution. At the end of each prediction cycle, the adaptation unit compares the current project features with historical records. When a feature change exceeds a threshold, the parameter update process is automatically triggered. After the parameter update, the model saves the new weight configuration as the basis for subsequent project condition predictions. Through the dynamic adjustment of the evolutionary parameter adaptation unit, the system can quickly achieve model transfer and reuse when facing engineering projects with different hydrological conditions, construction scales, or geological structures, thereby reducing retraining time and improving application efficiency. The dynamic time series prediction module uses a multi-task learning mechanism to train and optimize the model. The system uses schedule error, risk prediction error, and energy consumption prediction error as joint optimization objectives, achieving complementary learning among multiple tasks through a shared parameter structure. During the training phase, an adaptive optimization algorithm is used to control the learning rate and gradient convergence speed to ensure that the model remains numerically stable under multi-objective conditions. Adaptive optimization algorithm: In the training iteration... During the step, for the current parameter vector The update rule is defined as follows: , : No. The model parameter vector after the next iteration is used for the next training or validation step. The update function corresponding to the adaptive optimization algorithm is used to adjust the parameters based on the current gradient and the adaptive learning rate. : No. The current model parameter vector at the next iteration contains the trainable parameters of all expert sub-models, fusion units, and decoding units. : No. The total loss function value of this iteration is a weighted combination of the losses from multiple tasks. : No. The progress prediction error loss in each iteration can be calculated, for example, by using the mean square error of the progress curve. : No. The risk prediction error loss for each iteration can be constructed, for example, based on the prediction error of the critical path risk indicator. : No. The energy consumption prediction error loss in each iteration can be addressed, for example, by using a weighted error metric based on energy consumption time series data. The weighting coefficient for schedule forecasting losses is used to control their contribution to the total loss. Weighting coefficients for predicted risk losses. : weight coefficient of energy consumption prediction loss, : gradient operator of parameter vector , used to calculate partial derivative of total loss with respect to each parameter, : adaptive learning rate vector of the th iteration, different learning rates can be taken for different parameter dimensions, and the values are dynamically adjusted according to the historical gradient size and the validation set performance, : vector element-wise multiplication operator, used to scale the gradient by the corresponding learning rate in each dimension, and the early stopping mechanism is set during training by monitoring the change on the validation set combined with the above update rule, to maintain numerical stability and avoid overfitting under multi-task conditions. The early stopping mechanism is set during model training when the validation error does not decrease for several times in a row to stop training and avoid overfitting. The weight parameters of each expert predictor model are saved independently, which can quickly restore to the historical optimal state when the system switches under different working conditions. This mechanism guarantees the long-term running stability of the model and the continuous reliability when the field conditions change. The five units of the dynamic time series prediction module form a complete time prediction closed loop in the system. In the running process, the prediction state, residual index and parameter distribution are updated in real time, and the final output prediction result is transmitted to the adaptive weight evaluation module and the closed-loop feedback regulation module.

[0022] The adaptive weight evaluation module is the dynamic optimization core layer of the system, which is used to evaluate the influence intensity of various feature factors on the prediction target under different construction stages and different regulation scenarios in real time, and to realize the weight balance and dynamic self-adaptive adjustment of multiple targets. Through the double-layer structure of prediction layer weight optimization and regulation layer weight optimization, a multi-objective optimization framework is established, and combined with the scene mutation detection mechanism and the fairness constraint mechanism, the system realizes global coordination between progress prediction accuracy, safety risk control, energy consumption minimization and people's livelihood water balance. It contains prediction sensitivity optimization unit, regulation sensitivity optimization unit, weight collaborative adjustment unit, scene mutation detection unit, weight reconstruction unit and fairness evaluation and correction unit, each unit forms a closed-loop update system through the weight vector interaction mechanism. The prediction sensitivity optimization unit is used to calculate the influence degree of each feature factor on the prediction error, and to update the basic weight based on the result. Through the prediction sensitivity optimization function, the error between the input multi-source fusion features and the output of the dynamic time series prediction module is analyzed for sensitivity, and the contribution of each feature variable to the progress prediction result is measured. The prediction sensitivity optimization function is: : feature vector : multi-task prediction residual vector , the prediction sensitivity optimization function updates the feature basic weight vector as: , The prediction layer feature base weight vector, used for the next prediction cycle after the prediction cycle ends, with each dimension corresponding to the prediction layer weight of a feature factor. The prediction sensitivity optimization function adaptively adjusts the feature weights based on the current prediction residuals. : The basic weight vector of the prediction layer features at the start of the current prediction period. The multi-source fusion feature vector used for modeling in the current forecast period includes indicators such as meteorology, hydrology, geology, energy consumption, and personnel status. The multi-task forecast residual vector at the end of the current forecast period, whose components may include schedule error, risk forecast error, and energy consumption forecast error. The step size coefficient for updating the prediction layer weights takes values ​​within... Between these values, the strength of the impact of the new sensitivity on weight adjustments is controlled. Regarding the residual sensitivity vector of the eigenvector, its first... Dimension is defined as: , indicating the first The direction and magnitude of the influence of individual feature changes on the overall predicted residual sum of squares The vector obtained by taking the absolute value of each element of the sensitivity vector is used to characterize the intensity of the influence of each feature, regardless of the positive or negative direction. :and A vector of all 1s with the same dimension is used to calculate the sum of absolute sensitivity in the denominator. The sum of the absolute sensitivities of all features is used for normalization, transforming the sensitivity into a weight distribution. After the update, features with greater prediction error impact receive higher base weights. During operation, the system first receives prediction residual input from the dynamic time-series prediction module. The residual consists of three parts: schedule error, risk prediction error, and energy consumption prediction error. The prediction sensitivity optimization function performs gradient direction calculation and amplitude comparison for each feature factor to quantify the impact of feature changes on the prediction residual. Subsequently, the system updates the base weights of each feature factor according to the magnitude of the impact, increasing the weights of high-impact features and decreasing the weights of low-impact features. The prediction sensitivity optimization unit completes an update at the end of each prediction cycle and writes the updated weight vector into the system weight cache table, providing initial input for the regulation layer optimization. This mechanism ensures that the system maintains the real-time optimality of prediction results under dynamically changing data distribution conditions. The regulation sensitivity optimization unit is used to evaluate the impact of feature factor changes on the closed-loop feedback regulation effect. It receives the regulation result input from the closed-loop feedback regulation module through the regulation sensitivity optimization function and calculates the impact of different feature variable adjustments on the completion probability, risk reduction rate, and energy consumption indicators. Regulation sensitivity optimization function: a multi-index improvement vector formed from the closed-loop feedback regulation results. With feature adjustment vector , the regulation sensitivity optimization function updates the regulation layer weight vector is: , : the feature regulation weight vector for the next round of closed-loop regulation after the current regulation period ends, : the regulation sensitivity optimization function, which updates the weight of the feature in the regulation layer according to the regulation performance, : the feature regulation weight vector at the beginning of the current regulation period, : the multi-index improvement vector under the current regulation strategy, whose components can be the shortening ratio of the project duration, the risk reduction rate, and the energy consumption reduction rate, : the feature change vector of the current regulation strategy relative to the baseline condition, each dimension represents the adjustment amplitude of a certain feature factor, : the regulation layer weight update step size coefficient, taking values between , : the regulation sensitivity vector, the th dimension of which can be represented as: , where represents the th regulation performance index, represents the importance weight of the index in the comprehensive regulation performance, : the result of taking the absolute value of the regulation sensitivity vector element by element, which is used to measure the influence of different features on the regulation effect, : a full 1 vector with the same dimension as : the sum of the absolute values of all feature regulation sensitivities, which is used for normalization to obtain a regulation weight distribution similar to a probability distribution. In this way, the more significant the feature is in improving the shortening of the project duration, the reduction of risk, or the reduction of energy consumption, the greater the regulation weight will be. In the analysis stage, the system first compares the simulation evaluation results under different regulation strategies, including the shortening ratio of the project duration, the energy consumption reduction rate, and the safety risk reduction amplitude. The regulation sensitivity optimization function establishes a mapping relationship between these results and the change amount of the input features, thereby identifying the feature factors that have a key contribution to the regulation effect. For the factors that have a significant impact, the system increases their regulation weight, so that they are given priority in the next round of prediction-regulation closed loop. Weight adjustment is performed once after each regulation period, and the weight update log is shared with the prediction sensitivity optimization unit. Through this two-way information transmission mechanism, the system can achieve a dynamic balance between prediction accuracy and regulation performance. The weight coordination adjustment unit is responsible for integrating the double-layer optimization results of the prediction layer and the regulation layer to form a unified weight vector. Through the comprehensive calculation of the weight coordination adjustment function on the two-layer weight results, the overall control effect is improved without sacrificing the prediction accuracy. The weight coordination adjustment function: after obtaining the prediction layer weight vector and the regulation layer weight vector , the weight coordination adjustment function generates a unified master control weight vector : , The comprehensive weight vector, after collaborative adjustment and written into the system's master control weight table, will be used in both the prediction and control modules in the next cycle. The weighted adjustment function is used to balance the weights between prediction accuracy and regulatory performance. The prediction layer weight vector output by the prediction sensitivity optimization unit. : The control layer weight vector output by the control sensitivity optimization unit. Inter-layer synergy coefficient, with values ​​ranging from Between these two levels, the weights are used to control the proportion of the control layer's weights in the overall weights. The value can be dynamically determined based on the weight deviation between the two layers and historical performance indicators. When the value is small, the bias is to ensure prediction accuracy; when... When the deviation is large, the system tends to improve the control effect. The system first calculates the deviation between the prediction layer weights and the control layer weights. When the deviation exceeds a threshold, the weight coordination adjustment function performs a weighted balancing operation, fusing the prediction sensitivity and control sensitivity results according to a preset ratio. The fusion ratio is automatically determined by the system based on historical performance indicators to ensure consistent optimization direction across different scenarios. The weight coordination adjustment unit performs a stability check after each weight update, using weight variance and entropy regularization to constrain the smoothness of the weight distribution and prevent extreme bias. The final generated comprehensive weight vector is written into the system's main control weight table, providing input parameters for the prediction and control modules in the next cycle. The scenario mutation detection unit continuously monitors the distribution of input data and changes in prediction residuals to determine scenario switching in climate, processes, or supply modes. The scenario mutation detection function analyzes the statistical characteristics of the input data in real time to identify changes in scenario status. Scenario mutation detection function: [Further details about time and context are needed for accurate translation.] Feature statistics and residual statistics within the sliding window; scene shift intensity index output by the scene abrupt change detection function. : , :time The intensity of scene abrupt change is a metric; a higher value indicates a more significant change in the distribution of the input data and the prediction residual. : Scene abrupt change detection function, used to determine whether a scene change has occurred by combining feature distribution and residual changes. By time Within the sliding window at the end, it is the dimensionally averaged vector of the multi-source feature vectors. The multi-source feature mean vector, which takes the previous window as the interval, is used for comparison with the mean of the current window. By time Within the sliding window at the end, the multi-source feature's standard deviation vector is used to characterize the fluctuation intensity. : the standard deviation vector of the multi-source feature in the previous window, : the average amplitude of the prediction residual in the sliding window with time as the end point, which can be defined as the average two-norm of the residual vector, : the average amplitude of the residual in the previous window, : the two-norm of the vector, used to measure the comprehensive amplitude of the mean and standard deviation changes, : the weight coefficient of the standard deviation change term, used to adjust the importance of fluctuation changes in scene mutation detection, : the weight coefficient of the residual offset term, used to adjust the importance of prediction error mutations in scene detection, when exceeds the preset threshold, the system marks the current time as a scene mutation point and sends a scene switching signal to the weight reconstruction unit. The detection process is carried out in a sliding window manner, and the system calculates the feature mean, variance and residual offset in the time dimension. When the feature distribution in the continuous window appears significant deviation, or the prediction residual increases significantly beyond the set threshold, the system determines that a scene mutation has occurred. Typical scene mutations include meteorological data mutations caused by the arrival of flood season, changes in work intensity caused by process conversion, and abnormal energy consumption distribution caused by material supply mode adjustment. After detecting the scene mutation, a scene switching signal is generated and transmitted to the weight reconstruction unit, triggering the latter to perform the weight redistribution process. Through the real-time scene detection mechanism, the system can maintain the continuity and effectiveness of weight evaluation under complex construction environment changes. The weight reconstruction unit, upon receiving the scene mutation signal, performs the recalculation and update of the weight distribution. The old weights are redistributed according to the contribution degree through the weight reconstruction function, and the new configuration is recorded using the scene-weight mapping management function. The weight reconstruction function: upon receiving the scene mutation signal, the weight reconstruction function generates a new scene weight vector based on the feature contribution vector in the new scene and the last scene terminal weight vector : , : the redistributed feature weight vector obtained in the new scene, which is written into the system as the weight configuration of the current scene, : the weight reconstruction function, used to determine the weight distribution after scene switching by combining the historical state and new contribution degree, : the last weight vector at the end of the previous scene, which is frozen and saved before scene switching, : the feature contribution vector calculated based on the difference in prediction error before and after the mutation in the new scene, each dimension representing the relative contribution of a certain feature to the error change, : the weight reconstruction coefficient, taking a value between , used to control the proportion of the influence of new contribution degree information on weight reconstruction, : and all-ones vector of the same dimension, : the sum of all feature contributions, used to normalize the contribution vector to obtain a new contribution proportion distribution, when is larger, the weight distribution is closer to the contribution structure of the new scene; when is smaller, the continuity of the last scene is emphasized. Scene-weight mapping management function: for the current scene semantic vector , the scene-weight mapping management function generates the corresponding weight vector from the stored scene library: , : the current scene recommended weight vector output by the scene-weight mapping management function, which can be used as a candidate result for weight reconstruction or initialization, : scene-weight mapping management function, used to establish a mapping relationship between the scene space and the weight configuration space, : time scene semantic vector, which can be encoded by information such as working condition label, climate feature, operation stage, and supply mode, : the number of historical scene entries stored in the scene-weight mapping table, : the weight configuration vector corresponding to the th historical scene in the mapping table, : the semantic vector prototype of the th historical scene in the mapping table, used to represent the center position of the scene feature, : the similarity weight of the current scene and the th historical scene, satisfying , : the Euclidean distance between the current scene semantic vector and the th scene prototype, : temperature parameter, used to control the sharpness of the similarity distribution, : when is smaller, it is more biased towards selecting individual highly similar scenes,When the scale is large, multiple similar scenarios are integrated. When a state similar to a historical scenario appears in the future, the system can quickly retrieve the corresponding weight configuration through this function without needing to completely reconstruct it. The weight reconstruction process includes three steps: First, the system freezes the final weight state of the previous scenario and records the corresponding working condition label; second, based on the difference in prediction error before and after the change, the contribution ratio of each feature factor is recalculated; finally, the weight vector is reset according to the new contribution ratio to generate a new weight distribution. The generated weight configuration is marked as a weight mapping entry for the current scenario and stored in the scenario-weight mapping table. When a similar scenario reappears in the future, the system can directly call this configuration without recalculation, thereby improving the response speed and stability of weight adjustment. The weight reconstruction unit ensures the system's adaptive evolution capability in complex construction environments, enabling the weight system to automatically adjust according to changes in the external environment, working conditions, and tasks, achieving continuous optimization of prediction and control. The fairness assessment and correction unit is used to monitor the multi-objective balance of the system to prevent safety or social impacts caused by over-optimization of a single objective. The multi-objective optimization process is analyzed using a fairness evaluation function. When a long-term bias towards a single objective is detected, a fairness correction function is invoked to adjust the weight allocation. The fairness evaluation function evaluates the multi-objective performance index vector at the current time step. The fairness assessment function outputs the objective balance index. : , :time The fairness index indicates that the smaller the value, the more balanced the multiple objectives are; the larger the value, the more significant the bias or neglect of certain objectives. : Fairness evaluation function, used to measure whether a multi-objective optimization process is biased towards a certain type of objective in the long run. : A multi-objective performance index vector, whose components can be normalized scores for safety risk indicators, domestic water use indicators, construction efficiency indicators, and energy consumption indicators, etc. : No. The goal in time The performance score can be standardized or normalized. The number of targets considered simultaneously, such as the number of possible targets. Corresponding to the four goals of safety, people's livelihood, efficiency and energy consumption, In time The arithmetic mean of the performance scores for different targets is used as a balancing reference. : No. The weight coefficient of each objective in the fairness assessment is used to reflect the importance of that objective in the balance considerations. A consistently large value indicates that the system is biased towards certain objectives in the long term, necessitating a fairness correction process. The fairness correction function is defined given the current objective weight vector. With performance index vector Under the condition that the fairness correction function is triggered, the target weight vector is updated as: , : target weight vector after fairness correction, each component corresponds to the relative importance of different objectives in the current period, : fairness correction function, used to automatically adjust the target weight when long-term bias towards a single objective is detected, : target weight vector of the current period, the components can correspond to the weights of progress, risk, safety, livelihood and energy consumption, etc. : multi-objective performance index vector of the current period, each component is the actual performance value of a different objective, : expected target performance reference vector, used to represent the ideal level when each objective is in the "balance interval", which can be obtained from planning indicators or historical balance period statistics, : fairness correction step parameter, used to control the influence strength of target performance deviation on weight adjustment, : deviation vector of the current performance of each objective relative to the reference level, the corresponding component is positive when the performance of a certain objective is low, thereby promoting the weight of the objective to rise, : all-1 vector with the same dimension as : the denominator : sum of the modified weight vector, used for normalization so that the sum of the components of the new weight vector is 1. Through this correction form, the target that is sacrificed more will get a higher weight, so it will be given priority in subsequent prediction and control. In the running process, the system continuously calculates the safety risk index, livelihood water index and construction efficiency index. When any index deviates from the balance interval for a long time, for example, to speed up the progress, the energy consumption or the fatigue of workers increases, the fairness evaluation function will trigger the weight correction process. The correction strategy increases the weight of the features related to the damaged target and correspondingly reduces the weight of other targets, ensuring the balance between objectives. After correction, the system generates a fairness correction report, recording the adjustment amplitude, correction time and corrected target balance index, providing optimization decision basis for managers. Through this mechanism, the system can realize the collaborative optimization of construction progress and resource utilization under the premise of ensuring safety and livelihood. The adaptive weight evaluation module runs in the system in a periodic closed loop. The entire weight update process is synchronized with the prediction period of the dynamic time series prediction module, and the weight evaluation and update are performed once an hour. The updated weight vector is fed back to the dynamic time series prediction module and the closed-loop feedback control module, realizing the complete closed-loop process from prediction to control to optimization.

[0023] The closed-loop feedback control module utilizes a multi-objective hierarchical reinforcement learning structure, a digital twin simulation evaluation mechanism, and a field execution mechanism to achieve the self-evolutionary generation, verification, and implementation of control strategies. It fully leverages the prediction results from the dynamic time-series prediction module and the dynamic weight vector output by the adaptive weight evaluation module to generate executable construction control plans based on the current working conditions, system prediction deviations, and control requirements. It consists of a long-term strategy optimization unit, a short-term control execution unit, and a digital twin simulation evaluation unit. These three units form a control closed loop through four steps: strategy input, state feedback, simulation verification, and field execution. This ensures the system maintains stable construction progress even when working conditions change, schedule deviates from the plan, or equipment and human resource constraints change. The long-term strategy optimization unit is responsible for generating control strategies suitable for medium- to long-term construction coordination on a weekly or ten-day timescale. Through the long-term strategy optimization function, based on the completion time prediction, critical path risk indicators, and energy consumption trend analysis results output by the dynamic time-series prediction module, and considering factors such as personnel numbers, machinery and equipment configuration, energy consumption patterns in the construction area, and historical task completion efficiency, a long-term control strategy is established. The long-term strategy optimization function generates a long-term control strategy vector based on schedule deviations, risk deviations, and energy consumption deviations. : , The long-term control strategy vector generated on the current weekly or ten-day timescale may include long-term executable parameters such as "personnel increase / decrease ratio, equipment working section adjustment parameters, construction rhythm change plan, and resource allocation benchmark". The long-term strategy optimization function is used to generate the overall control direction based on prediction results and deviation information on medium- to long-term time scales. :time The schedule deviation scalar can be defined as the difference between the predicted cumulative schedule and the planned cumulative schedule. A negative value indicates that the task is lagging behind, while a positive value indicates that the schedule is ahead of schedule. :time The risk deviation scalar can be defined as the difference between the critical path risk indicator and the safety control target. A positive value indicates that the risk is higher than the target level. :time The energy consumption deviation scalar can be defined as the difference between the energy consumption trend and the energy-saving benchmark. A positive value indicates that energy consumption is too high, and a negative value indicates that energy consumption is too low. : Long-term baseline strategy vector, corresponding to the default resource allocation and construction pace when "progress is basically normal, risks are controllable, and energy consumption is within a reasonable range". : Schedule deviation gain coefficient vector or matrix, used to map schedule deviations to adjustments for long-term strategies; when When it is negative (schedule is delayed), For positive direction adjustment, it can reflect the actions such as "improve equipment operation rate, increase total personnel, speed up process rhythm", etc. : risk deviation gain coefficient vector or matrix, used to modify the content such as "high-risk process rhythm, operation rearrangement, redundant resource allocation" in the strategy according to risk deviation, : energy consumption deviation gain coefficient vector or matrix, used to modify the content such as "equipment load level, shift density, energy consumption optimization measures" in the strategy according to energy consumption deviation; when energy consumption is high, it can produce "reduce operation intensity, slow down resource consumption" adjustment, and when progress lags behind, : improve construction rhythm, when progress is ahead of schedule or energy consumption is high, : etc. to moderately reduce the negative, achieve the medium and long-term balance of progress, risk and energy consumption. In the strategy generation process, the long-term strategy optimization function first reads the overall progress deviation of the current construction period. If the progress deviation is negative, indicating that the task is lagging behind, the long-term strategy optimization function increases the mechanical equipment operation rate, moderately increases the total number of personnel, optimizes the process rearrangement, etc. to improve the construction rhythm. If the progress deviation is positive, indicating that the progress is ahead of schedule, the long-term strategy optimization function moderately reduces the equipment operation intensity, adjusts the shift density or slows down the resource consumption rate to achieve energy consumption optimization and risk balance. The long-term strategy optimization unit writes the strategy content into the strategy cache table after generating the strategy, including personnel increase and decrease ratio, equipment work section adjustment parameter, construction rhythm change scheme and resource allocation benchmark. The strategy cache table is inherited by the short-term regulation execution unit for real-time control at the shift level. The short-term regulation execution unit is responsible for dynamic regulation in the shift time scale according to the long-term strategy and real-time working conditions. Through the short-term regulation execution function, the strategy generated by the long-term strategy optimization unit is refined into an executable short-period scheduling scheme. The short-term regulation execution function: under the condition of given long-term strategy vector, real-time working condition observation and sudden event indication, the short-term regulation execution function generates a shift-level regulation action vector : , , : time corresponding short-term regulation execution vector, its components can include "equipment start and stop instruction, operation sequence adjustment, personnel distribution change within shift, non-critical operation reduction", etc. specific shift-level scheduling decision, : short-term regulation execution function, used to combine the long-term strategy with real-time working conditions to generate an executable regulation scheme refined to the shift granularity, : the current long-term strategy vector output by the long-term strategy optimization function, providing a high-level guidance framework for short-term regulation, : real-time working condition observation vector at time , containing "personnel status, equipment start and stop record, process completion progress, energy consumption and water consumption change", etc. real-time data on site, : Time burst vector, where components can represent binary or intensity indicators such as "equipment shutdown alarm, water level surge alarm, personnel status anomaly alarm", : Long-term policy mapping matrix, used to project weekly / ten-day scale policy parameters to shift scale regulation space, to achieve policy refinement, : Real-time working condition response matrix, used to fine-tune in-shift operation intensity, equipment operation and process sequence according to current observation, : Time backup solution regulation vector, corresponding to pre-stored emergency regulation solutions in the policy cache table, such as "emergency shutdown sequence, local process rearrangement, personnel rapid evacuation and redistribution", etc. : Time event response coefficient, taking values in the interval , used to balance the weights of "regular regulation solutions" and "backup emergency solutions", : Compression function that maps real numbers to the interval , which can be taken as the logistic function , used to convert event comprehensive intensity into probability type weight, : Event intensity weight vector, used to depict the contribution of different burst event indicator components to the overall event severity, : Weighted sum of various burst event indicators at the current time, used to measure the overall event severity within the shift, when shows that the event is severe, is close to 1, and short-term regulation is mainly based on backup solutions ; when the event is stable, Approaching zero, short-term control primarily relies on long-term strategies and real-time operating conditions for refined scheduling. It receives real-time data input, including personnel status, equipment start-up and shutdown records, process completion status, and changes in energy consumption and water usage. Within each shift, it adjusts construction intensity, work sequence, and equipment operating status based on real-time changes. The short-term control execution function first determines whether an emergency has occurred, such as equipment shutdown, sudden water level rise, or abnormal personnel status, based on the real-time data collected for the current shift. When an emergency affecting construction efficiency is detected, the short-term control execution function immediately calls the backup plan in the strategy cache table, changing the equipment start-up and shutdown sequence, rearranging local processes, and altering the personnel distribution in work sections to ensure that the construction progress does not deviate further. The short-term control execution unit has resource linkage control capabilities. When the dynamic time-series prediction module predicts that the energy consumption increase trend in the future cycle reaches the warning threshold, the short-term control execution function reduces resource burden by lowering the mechanical load of non-critical work sections and prioritizing the execution of low-energy-consumption work sections, keeping energy consumption within a safe range. All short-term control decisions are recorded in the control log and sent to the twin simulation evaluation unit for unified verification after the control is completed. The twin simulation evaluation unit is responsible for virtually simulating the strategies generated by the long-term strategy optimization unit and the short-term control execution unit in a digital twin environment to verify the feasibility, effectiveness, and security of the strategies. By calling the digital twin model of the construction site through the twin simulation evaluation function, it simulates terrain, water level changes, equipment operation behavior, and personnel work behavior, generating virtual progress curves, risk curves, and energy consumption curves for strategy execution. The twin simulation evaluation function, given the current twin scenario state vector and control strategy vector, outputs deviation vectors in three aspects: progress, risk, and energy consumption. : , In time The deviation vector after performing twin simulation on the control strategy includes three simulation evaluation results: schedule deviation, risk deviation, and energy consumption deviation. The twin simulation evaluation function calls the digital twin model to simulate the strategy execution process in a virtual construction environment and provides the deviation from the predicted results. :time The twin scene state vector contains virtual scene states composed of "meteorological parameters, seepage status, water level distribution, equipment start-up and shutdown modes, personnel characteristic distribution, and process status". In time The vector of the control strategy to be executed can be generated by combining the long-term strategy optimization unit and the short-term control execution unit. The cumulative completion rate or virtual progress curve of a task simulated in a digital twin environment over time. The value of , The dynamic timing prediction module, without considering twin simulation feedback, predicts time... progress prediction value of the progress, : the value of the risk indicator generated in the twin simulation environment at time , used to describe the virtual evolution of safety risk or critical path risk during construction, : the predicted value of the risk indicator by the dynamic timing prediction module under the original prediction, : the value of the energy consumption curve simulated in the twin simulation environment at time , : the predicted value of the energy consumption trend by the dynamic timing prediction module based on historical data and feature fusion results, when three components are within the preset tolerance range, it indicates that the strategy is basically consistent with the prediction model in terms of progress, risk, and energy consumption, and can be considered as an "executable solution"; if the deviation is too large, the adjustment signal is returned by the function to prompt the long-term or short-term regulation unit to regenerate the strategy and enter the simulation verification again. The twin simulation evaluation function first loads the twin scene containing the current working condition input during the strategy verification process, including meteorological parameters, seepage state, equipment start-stop mode, and personnel characteristic distribution. Then, according to the strategy content, the virtual construction environment is simulated, and the strategy is applied to the equipment nodes, personnel nodes, and process nodes in the model to generate construction task completion time prediction values, risk fluctuation curves, and energy consumption trend curves. When the simulation results deviate from the future prediction results of the dynamic time series prediction module, the system determines that the strategy cannot be directly executed. The twin simulation evaluation unit automatically returns the strategy adjustment instruction when the strategy is not executable, and the long-term strategy optimization unit or the short-term regulation execution unit regenerates the strategy and performs simulation verification again. Only when the simulation output meets the executable conditions in terms of progress deviation, risk indicators, and energy consumption changes, the strategy will be marked as the optimal solution. After the strategy passes the verification, the twin simulation evaluation unit transfers it to the execution subsystem to form an executable task instruction set, including device operation power adjustment commands, shift arrangement schemes, process reordering lists, and personnel scheduling schemes. The closed-loop feedback regulation module, dynamic time series prediction module, and self-adaptive weight evaluation module form a complete closed-loop control link. At the end of each prediction period, the dynamic time series prediction module outputs the prediction deviation index as the basis for the long-term strategy optimization unit to generate the strategy; the strategy is refined by the short-term regulation execution unit and verified by the twin simulation evaluation unit; the strategy that passes the verification is executed, and the field data feedback enters the multi-source perception collection module, and then the new prediction results are generated by the feature fusion modeling module and the dynamic time series prediction module. The system continuously verifies and adjusts the strategy effect in the closed-loop operation, so that the regulation behavior automatically evolves according to the real-time working conditions, realizing continuous optimization. The strategy verified by the twin simulation evaluation unit is finally issued to the field execution system, and the edge control device converts the strategy content into executable instructions. Shift-level regulation instructions include device start-stop operations, process segment adjustments, and personnel rotation arrangements; week-level regulation instructions include device scheduling adjustments, construction intensity adjustments, and resource allocation changes. These execution instructions are implemented on site by the device control system and the operation scheduling system. The execution log is recorded and returned to the system as the basis for correction for the next round of prediction and regulation.

[0024] The result visual analysis module is located in the human-computer interaction and interpretation layer of the system, and is an important interface connecting prediction, regulation and decision. It visually displays the construction progress prediction results, risk assessment results, energy consumption trend and regulation strategy execution effect, and through factor contribution explanation, key path identification, scheme comparison display and risk warning, management personnel can complete analysis, judgment and decision in an understandable interface. It includes five functional units: interpretable analysis unit, hypothesis deduction unit, key path dynamic identification unit, comparison display unit and risk warning unit. Each unit forms a top-down display link according to the order of information input. From the dynamic time series prediction module, the self-adaptive weight evaluation module and the closed-loop feedback regulation module, it receives prediction results, weight vectors, future regulation strategies and risk values, processes the results into structured visual content through internal function structures, and synchronously updates with the digital twin scene. The role of the interpretable analysis unit is to explain the source of the output results of the dynamic time series prediction module, and to reveal the influence degree of multi-source feature factors on the prediction value. The unit calculates the feature importance used by the prediction model in the reasoning process according to the influence contribution through the interpretable analysis function, and generates a visual graph. The interpretable analysis function first receives the complete prediction sequence output by the dynamic time series prediction module, including the construction progress curve, the completion time prediction, the key path risk index and the energy consumption trend. Then, the function calculates the contribution degree of each type of feature factor to the prediction result by using the change relationship between the fused features and the prediction residual. For example, when the change of the weather feature in the current period is significantly synchronized with the progress change, the interpretable analysis function will assign a higher contribution degree; while the personnel state feature has a weak influence on the progress in the current period, the contribution degree is lower. The interpretable analysis unit displays the contribution degree results in the form of bar chart, heat map or stacked weight distribution chart, and projects the contribution change trend to the key equipment nodes, construction location nodes or monitoring points in the digital twin scene, so that management personnel can quickly identify the important factor source affecting the construction progress. The unit also generates a contribution explanation report, which records the contribution sequence change of each factor, for long-term perspective strategy making and construction planning evaluation. The hypothesis deduction unit is used to simulate the prediction curve changes under different construction or scheduling schemes, and realizes multi-scheme prediction comparison through the hypothesis deduction function, so that management personnel can evaluate the effect of potential strategies before decision making. The hypothesis deduction function first receives the user-specified hypothesis variables in the interface, which come from adjustable construction elements, including equipment operation capacity change, personnel scheduling adjustment, work order change or shift construction intensity improvement. The function injects these variables into the corresponding intervention-sensitive prediction structure of the dynamic time series prediction module as intervention items, generating counterfactual prediction sequences under different schemes. The hypothesis deduction unit synchronously displays multiple prediction curves in the interface, including the current scheme prediction curve and several hypothetical scheme prediction curves, all curves are superimposed according to the same time axis.The system allows users to replay the dynamic changes of different schemes in the digital twin environment, enabling users to intuitively observe the differences between schemes from the spatial and temporal dimensions. The unit generates a deduction report when outputting, which records the hypothetical variables, counterfactual prediction curves, progress difference values, energy consumption difference values, and risk change values, enabling managers to make reasonable judgments based on quantitative comparisons. The critical path dynamic identification unit is used for automatic identification and visual annotation of the critical path of construction progress, and the critical path dynamic identification function is used to realize real-time updating of the critical path nodes. The critical path dynamic identification function first reads the critical path risk indicators and construction task dependency graphs output by the dynamic time series prediction module to confirm the critical task nodes in the current stage and the completion probability change trend of these nodes. When the delay probability of a node increases, the function immediately marks the node as a high-risk node. The critical path dynamic identification unit dynamically annotates the critical task nodes in the visual interface. The annotation method is to highlight the critical node area in the three-dimensional digital twin scene and use red flashing to prompt the risk nodes. When the node risk increases, the flashing frequency increases, and when the node returns to normal, the highlight intensity decreases. It also supports key path evolution recording function, which continuously records the changes of key path structure over time during construction, enabling management personnel to monitor path evolution for a long time and identify structural bottleneck problems in construction organization, providing basis for long-term construction planning. The comparison display unit is used to display the progress curve, risk curve and energy consumption curve of different strategy schemes in the same interface, so that the management personnel can determine the optimal scheme in visual comparison. The comparison display unit uses the comparison display function to structure the curves before and after regulation, between schemes or in different scenarios. The comparison display function first receives at least two sets of prediction sequences, one of which is the actual prediction sequence of the current scheme, and the other is the optimal strategy prediction sequence generated by the closed-loop feedback regulation module and verified by the twin simulation evaluation unit. The system performs time alignment processing on the two sets of data to ensure that the visualization comparison is performed under the same time reference. The comparison display unit displays the progress curve in polyline form, the energy consumption curve in segmented trend line, and the risk curve in change band. The interface also provides difference visualization based on color and transparency, such as using blue to represent the current scheme and orange to represent the optimized scheme, so that management personnel can intuitively identify the differences. The comparison display unit provides a detailed index comparison window, including the amount of time saved, the amount of risk reduced, and the amount of energy saved, all of which are derived from the structured data of the prediction module and the simulation evaluation module, ensuring comparability and accuracy. The risk warning unit is used to send warning information to the management terminal when the predicted risk exceeds the system preset threshold, and the risk warning function is used to realize automatic risk detection and hierarchical warning. The risk warning function receives the critical path risk indicators, energy consumption anomaly indicators, and progress deviation indicators output by the dynamic time series prediction module, and sets different thresholds according to the risk categories.When any index exceeds its threshold value, the function immediately generates an alarm event, writes the alarm content, risk type, triggering time and affected area to the alarm record table. The risk alarm unit reminds the risk in the form of a prompt window, a highlighted area or a flashing identifier in the visual interface. The system supports a three-level risk identification method, corresponding to a slight risk, a moderate risk and a serious risk. When the risk belongs to the serious level, the unit will synchronize the alarm information to the mobile device of the management terminal and continuously flash the warning in the digital twin environment for the associated device nodes or construction sections. The risk alarm unit is also responsible for recording the system response after the risk occurs, including the short-term regulation execution unit's regulation action, the risk change trend and the final risk removal time, in order to constitute a traceable risk management record and provide support for subsequent risk management strategy optimization. The result visual analysis module maintains a strict data synchronization relationship with the upstream module. The dynamic time series prediction module and the closed-loop feedback regulation module push the prediction results and strategy results to this module after each prediction period, which are interpreted by the interpretable analysis unit, simulated by the hypothesis deduction unit, visualized by the key path dynamic identification unit, evaluated by the comparison display unit, and monitored by the risk alarm unit.

[0025] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for the purpose of limiting the present application, although the foregoing detailed description of the present application is made with reference to the foregoing embodiments, for those skilled in the art, it still can be modified to the technical solutions recorded in the foregoing embodiments, or to replace some of the technical features, within the spirit and principles of the present application, any modification, equivalent replacement, improvement, etc., should be included within the scope of protection of the present application.

Claims

1. A water conservancy construction progress prediction system based on multi-source data fusion, characterized in that: include, The multi-source sensing and acquisition module performs hierarchical adaptive sampling, quality assessment, and spatiotemporal alignment of various types of data at the construction site. The data includes project progress, meteorological environment, geology and hydrology, energy consumption and water use, and personnel status. The adaptive sampling function dynamically adjusts the sampling frequency according to the prediction error, and the data quality scoring function performs signal noise reduction and anomaly correction. The feature fusion modeling module is based on a multi-perspective self-evolving fusion structure driven by working condition semantics. It performs semantic alignment, impact event perception and physical consistency constraint fusion on multi-source data to generate a fusion feature representation that conforms to semantic and physical rules. The dynamic time series prediction module is based on the hierarchical expert hybrid and intervention-sensitive evolution structure of the working condition, establishes a multi-task nonlinear time series prediction model, and outputs the construction progress curve, completion time and critical path risk indicators. The adaptive weight evaluation module performs two-layer weight optimization and scene switching management based on prediction residuals and control feedback, and adjusts the dynamic weights of each feature factor by jointly using the prediction sensitivity optimization function and the control sensitivity optimization function. The closed-loop feedback control module generates construction control strategies based on multi-objective hierarchical reinforcement learning, and issues execution instructions after verifying the effectiveness of the strategies in a digital twin simulation environment, thereby realizing self-evolving closed-loop control of prediction-control-feedback. The results visualization analysis module provides a visual representation of the prediction results and control strategies, including factor contribution explanation, hypothesis deduction, dynamic identification of critical paths, and risk warning functions.

2. The water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The multi-source sensing and acquisition module includes: The edge acquisition unit connects to meteorological, hydrological, geological, and personnel sensors via an IoT communication network and uses the PTP time synchronization protocol to achieve data time alignment. The adaptive sampling unit dynamically adjusts the sampling frequency of different data sources based on the magnitude of the prediction error using a sampling scheduling function. The data quality assessment unit detects noise and missing values ​​using a sliding window statistical function and a local anomaly factor algorithm, and repairs abnormal data based on Kalman filtering and time neighborhood interpolation. The physical consistency correction unit performs physical logic correction on the collected data through the water level and flow conservation function and the energy consumption and load constraint function.

3. The water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The feature fusion modeling module includes: The working condition semantic modeling unit extracts historical typical working condition prototypes through the working condition prototype extraction function, calculates the semantic similarity between the current working condition and the historical prototypes using the working condition matching evaluation function, and then maps the current features to the similar working condition semantic space through the working condition alignment mapping function. The impact event sensing unit detects sudden events such as rainstorms, landslides, or equipment failures through an event candidate detection function, enhances the weights of event-related modalities using an event-sensitive feature amplification function, and adjusts the relationships between nodes in the graph structure through an event-driven edge weight adjustment function. The physical rule constraint unit detects whether the fused features violate water balance, energy load or schedule constraints through the physical rule consistency evaluation function, and calls the feature correction feedback function to make adjustments in abnormal situations. The dynamic graph modeling unit establishes a dynamic graph structure with construction units, equipment nodes, and monitoring points as vertices. The edges between nodes include spatial proximity edges, process dependency edges, and statistical correlation edges, and topological adaptive evolution is performed through a periodic edge weight update function.

4. The water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The dynamic time series prediction module includes: The working condition layering identification unit determines the current construction stage and selects the corresponding prediction sub-expert model through the working condition layering identification function; The expert prediction sub-model learns the time dependencies under specific working conditions through the expert prediction function; The global expert fusion unit dynamically weights the outputs of multiple sub-experts and generates a unified prediction result through the global expert fusion function. The intervention-sensitive decoding unit transforms the future scheduling plan output by the closed-loop feedback control module into an intervention sequence through the control plan encoding function, and generates counterfactual prediction results under different control schemes through the intervention-sensitive prediction function; The evolution parameter adaptation unit extracts information on project scale, geological type, and main process parameters through the project feature encoding function, and dynamically adjusts the model parameters through the evolution parameter mapping function to complete cross-project transfer learning.

5. A water conservancy construction progress prediction system based on multi-source data fusion according to claim 4, characterized in that: The dynamic time series prediction module adopts a multi-task learning mechanism, with schedule error, risk prediction error and energy consumption prediction error as joint optimization objectives, and improves the prediction accuracy and stability of the model through weight sharing and early stopping mechanisms.

6. The water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The adaptive weight evaluation module includes: The prediction sensitivity optimization unit calculates the influence of each feature factor on the prediction error through the prediction sensitivity optimization function; The regulation sensitivity optimization unit calculates the impact of changes in each characteristic factor on the regulation result through the regulation sensitivity optimization function; The weighted collaborative adjustment unit updates the weight vector comprehensively through the weighted collaborative adjustment function based on the dual-layer sensitivity results of prediction and regulation, so as to balance the goals of prediction accuracy and regulation performance.

7. A water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The adaptive weight evaluation module further includes: The scene change detection unit monitors the distribution of input data and residual changes through the scene change detection function to detect scene switching of climate, process or supply mode; The weight reconstruction unit quickly resets the weight distribution when the scene changes abruptly through the weight reconstruction function, and records and reuses the weight configuration through the scene-weight mapping management function. The fairness assessment and correction unit analyzes the balance of multi-objective optimization through the fairness assessment function. When it detects that the system is biased towards a single objective in the long term, it calls the fairness correction function to adjust the weights and maintain a balance between safety, people's livelihood and progress.

8. A water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The closed-loop feedback control module includes: The long-term strategy optimization unit determines the total number of personnel, machinery configuration, and construction rhythm on a weekly or ten-day time scale through the long-term strategy optimization function. The short-term control execution unit adjusts the construction intensity, work sequence, and equipment operating status within the shift time scale through the short-term control execution function; The twin simulation evaluation unit performs simulation calculations on candidate control strategies in a virtual construction environment using the twin simulation evaluation function, and selects the optimal solution for execution based on the simulation results.

9. A water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The results visualization analysis module includes: The interpretable analysis unit calculates the contribution of each factor to the prediction results and generates an influence visualization using interpretable analysis functions; The simulation unit assumes that the simulation function simulates the prediction curves under different construction or scheduling schemes. The critical path dynamic identification unit automatically identifies the task nodes that have the greatest impact on the project schedule through the critical path dynamic identification function and marks them on the visual interface. The comparison and display unit uses a comparison and display function to show the changes in the progress curves, risk curves, and energy consumption curves of different schemes; When the risk alarm unit predicts that the risk exceeds the preset threshold, it sends a warning message to the management terminal through the risk alarm function.

10. A water conservancy construction progress prediction system based on multi-source data fusion according to claim 1, characterized in that: The overall operation flow of the system includes: a multi-source sensing and acquisition module that collects and cleans data in real time; a feature fusion and modeling module that performs semantic alignment, impact event enhancement, and physical constraint fusion; a dynamic time series prediction module that performs hierarchical expert prediction and intervention-sensitive counterfactual reasoning; an adaptive weight evaluation module that performs two-layer weight optimization and fairness constraints; a closed-loop feedback control module that generates multi-objective strategies based on prediction bias and verifies their execution in twin simulation; and a results visualization and analysis module that displays prediction curves, risk trends, and critical paths to achieve intelligent decision support.

Images