Water conservancy project progress prediction system based on big data

Abstract

The invention discloses a hydraulic engineering progress prediction system based on big data. According to the system, standard engineering multi-source data is used; extracting feature data in the standard engineering multi-source data to obtain feature engineering multi-source data; an improved LSTM long-short term memory network is adopted, a space-time attention mechanism is integrated to establish an LSTM prediction model, and the LSTM prediction model is utilized to learn time attention at different historical moments and differentiated influence weights of space attention of different engineering parts on a prediction target; and inputting the differentiated influence weight and the feature engineering multi-source data into an XGBoost integrated learning model to quantify the comprehensive influence of the complex interaction among multiple factors on the progress, and outputting node completion prediction time and risk level signals. And the accuracy of progress prediction is greatly improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of big data processing, and particularly relates to a water conservancy project progress prediction system based on big data. BACKGROUND

[0002] Traditional water conservancy project progress prediction relies on discrete data collected manually, and the data sampling frequency is low, and it is difficult to integrate geological survey, meteorological environment, resource scheduling and other multi-source heterogeneous data. For example, if the water conservancy hub project does not timely associate the rainstorm early warning data in the construction stage, the foundation pit excavation operation will be delayed, and economic losses will be caused. Meanwhile, a large amount of noise and redundant information is contained in the original data, and the traditional statistical method is difficult to effectively clean, so that the prediction accuracy is low, and the accuracy is not high. SUMMARY

[0003] The application aims to solve the above problems, and designs a water conservancy project progress prediction system based on big data.

[0004] To achieve the above purpose, the technical scheme of the application is that, further, in the water conservancy project progress prediction system based on big data, the water conservancy project progress prediction system comprises the following modules:

[0005] A multi-source data acquisition module is used for acquiring engineering multi-source data, and after similarity analysis and Markov chain are used to eliminate redundancy and noise, normalization processing is performed to obtain standard engineering multi-source data;

[0006] A multi-source feature extraction module is used for extracting feature data in the standard engineering multi-source data to obtain feature engineering multi-source data;

[0007] An influence weight prediction module is used for adopting an improved LSTM long short-term memory network, integrating a space-time attention mechanism to establish an LSTM prediction model, and using the LSTM prediction model to learn the differentiating influence weight of the time attention of different historical moments and the space attention of different engineering parts on the prediction target;

[0008] A node prediction and early warning module is used for inputting the differentiating influence weight and the feature engineering multi-source data into an XGBoost integrated learning model to quantify the comprehensive influence of the complex interaction between multiple factors on the progress, and outputting a node completion prediction time and a risk level signal.

[0009] Further, in the water conservancy project progress prediction system based on big data, the multi-source data acquisition module comprises the following units:

[0010] A data acquisition unit is used for acquiring engineering ontology data, real-time progress data, resource environment data and management coordination data in the water conservancy project construction process to obtain engineering multi-source data;

[0011] A redundant data elimination unit is configured to calculate the similarity between different data, and data with high similarity is regarded as redundant data and eliminated by using a cosine similarity formula, so as to obtain first engineering multi-source data.

[0012] A noise data correction unit is configured to establish a state transition matrix, correct abnormal fluctuation noise data by analyzing state transition probability of the first engineering multi-source data, and obtain second engineering multi-source data.

[0013] A data normalization unit is configured to map the second engineering multi-source data to an interval [0, 1] by using a min-max normalization method, so as to obtain standard engineering multi-source data.

[0014] Further, in the above water conservancy engineering progress prediction system based on big data, the multi-source feature extraction module comprises the following units:

[0015] A feature setting unit is configured to set that the feature engineering multi-source data at least comprises a geological complexity index, a daily effective construction time length, a resource matching degree, an environmental interference factor, a logical constraint strength between processes, and a coordination efficiency index.

[0016] Further, in the above water conservancy engineering progress prediction system based on big data, the multi-source feature extraction module further comprises the following units:

[0017] A complexity calculation unit is configured to calculate the geological complexity index by using a weighted average method based on rock mass classification, joint development degree, and permeability coefficient parameters in a geological survey report.

[0018] A construction time length acquisition unit is configured to obtain the daily effective construction time length by dynamically calculating meteorological environment data and site conditions through a rule base.

[0019] A resource matching degree calculation unit is configured to extract a critical path logical relationship from a construction network plan based on graph theory calculation of process criticality weight, so as to obtain the resource matching degree.

[0020] An interference factor calculation subunit is configured to set interference weights of different factors, a storm weather interference weight is 0.3, a flood is 0.5, and an environmental interference factor is obtained by weighted calculation.

[0021] A constraint strength acquisition unit is configured to assign a constraint strength value to each pair of related processes according to the sequence and influence degree of the processes, so as to obtain the logical constraint strength between processes.

[0022] Further, in the above water conservancy engineering progress prediction system based on big data, the influence weight prediction module further comprises the following units:

[0023] A judgment unit is configured to establish a construction phase identification model, calculate daily average operation intensity through a sliding window, determine a construction period when the intensity is greater than a threshold value, determine a shutdown period otherwise, and determine a dynamic adjustment mechanism of a decay coefficient;

[0024] A processing unit is configured to perform stability processing on numerical values of an input gate combination function, set a lower limit of sigmoid output to 0.01, and obtain an improved LSTM long short-term memory network;

[0025] A establishing unit is configured to establish an LSTM prediction model by using the improved LSTM long short-term memory network and integrating a space-time attention mechanism.

[0026] Further, in the water conservancy project progress prediction system based on big data, the influence weight prediction module further includes the following units:

[0027] A time attention unit is configured to improve parallelism by using multi-head attention and relieve gradient explosion by using a scaling factor, establish a key event dictionary, and identify a corresponding timestamp through text matching;

[0028] A space attention unit is configured to divide according to WBS three-level coding, associate a corresponding BIM component ID for each unit, ensure that a physical space and a data space are mapped consistently, and determine an adjustment coefficient based on a geological complexity index GCI and a segmented function.

[0029] Further, in the water conservancy project progress prediction system based on big data, the node prediction and early warning module further includes the following units:

[0030] A feature processing unit is configured to splice a weighted feature vector and original feature engineering data, and reduce dimensions through PCA principal component analysis;

[0031] A model fusion unit is configured to use a stacking integration strategy, use a prediction result of the LSTM as an input feature of XGBoost, use a linear regression as a meta-model to integrate outputs of each base model, determine a weight through grid search, and output a node to complete prediction time and risk level signals.

[0032] Further, in the method for realizing the water conservancy project progress prediction system based on big data, the method includes the following steps:

[0033] Multi-source engineering data is collected, normalized after similarity analysis and Markov chain are used to eliminate redundancy and noise, and standard engineering multi-source data is obtained;

[0034] Feature data in the standard engineering multi-source data is extracted, and feature engineering multi-source data is obtained;

[0035] An improved LSTM long short-term memory network is used to establish an LSTM prediction model integrated with space-time attention mechanism, and the LSTM prediction model is used to learn the differentiating influence weight of time attention at different historical time and space attention of different engineering parts on the prediction target.

[0036] The differentiating influence weight and the feature engineering multi-source data are input into an XGBoost integrated learning model to quantify the comprehensive influence of complex interactions among multiple factors on progress, and output node completion time and risk level signals.

[0037] Further, in the method for implementing a water conservancy project progress prediction system based on big data, the method comprises the following steps:

[0038] A construction phase identification model is established, the daily average work intensity is calculated by a sliding window, and when the intensity is greater than a threshold, it is determined as a construction period, otherwise it is a shutdown period, and a dynamic adjustment mechanism of the decay coefficient is determined.

[0039] The numerical value of the input gate combination function is processed for stability, and the lower limit of the sigmoid output is set to 0.01 to obtain an improved LSTM long short-term memory network.

[0040] An improved LSTM long short-term memory network is used to establish an LSTM prediction model integrated with space-time attention mechanism.

[0041] Further, in the method for implementing a water conservancy project progress prediction system based on big data, the method comprises the following steps:

[0042] Multi-head attention is used to improve parallelism, and scaling factor is used to alleviate gradient explosion; a key event dictionary is established to identify the corresponding timestamp through text matching;

[0043] According to WBS three-level coding, each unit is associated with a corresponding BIM component ID to ensure that the physical space and data space are mapped consistently, and the determination of the adjustment coefficient is based on the geological complexity index GCI to set a segmented function.

[0044] The beneficial effects are that 1, the prediction accuracy is significantly improved: through deep fusion of multi-source heterogeneous big data and construction of dynamic coupling characteristics reflecting the characteristics of water conservancy projects, combined with the accurate modeling of the mixed intelligent model for the spatio-temporal dependence and complex interaction, the accuracy of progress prediction is greatly improved. 2, precise risk early warning is realized: the model can not only predict the progress, but also quantify the influence degree of each risk factor and output the contribution ranking, provide risk cause tracing, realize the active early warning of "early discovery, accurate positioning", and provide direct basis for management decision. 3, strong environmental adaptability: for the complex environmental factors such as geology, hydrology and meteorology specific to water conservancy projects, through special feature construction and spatio-temporal attention mechanism, the model can effectively perceive and respond to these dynamic changes. 4, high decision support value: not only time point prediction is provided, but also risk distribution, key factor analysis and early warning signal, which greatly improves the foresight and scientificity of project management, and helps resource optimization and risk active prevention and control. BRIEF DESCRIPTION OF DRAWINGS

[0045] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are intended to further aid the full and complete understanding of the preferred embodiments. The drawings are merely illustrative of the preferred embodiments and are not to be considered limiting of the application.

[0046] Figure 1 The first embodiment schematic diagram of a water conservancy project progress prediction system based on big data in the embodiments of the application;

[0047] Figure 2 The second embodiment schematic diagram of a water conservancy project progress prediction system based on big data in the embodiments of the application;

[0048] Figure 3 The third embodiment schematic diagram of a water conservancy project progress prediction system based on big data in the embodiments of the application. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical scheme and advantages of the present application clearer, the following will further illustrate the present application by combining with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0050] Those skilled in the art can understand that, unless specifically stated, the singular forms "a", "an" and "the" used herein also include the plural forms. It should be further understood that the use of the word "comprise" in the specification of the present application means that the features, integers, steps, operations, elements and / or components exist, but do not exclude the existence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0051] The application will be specifically described below with reference to the drawings, such as Figure 1 As shown in the drawings, a water conservancy project progress prediction system based on big data comprises the following modules:

[0052] A multi-source data acquisition module is configured to acquire project multi-source data, normalize the project multi-source data after eliminating redundancy and noise by similarity analysis and Markov chain, and obtain standard project multi-source data.

[0053] Specifically, the embodiment further comprises a data acquisition unit configured to acquire project ontology data, real-time progress data, resource environment data, and management coordination data in the water conservancy project construction process, and obtain project multi-source data.

[0054] A redundant data elimination unit is configured to calculate the similarity between different data, regard data with high similarity as redundant data by using a cosine similarity formula, and eliminate the redundant data, to obtain first project multi-source data.

[0055] A noise data correction unit is configured to establish a state transition matrix, correct noise data with abnormal fluctuations by analyzing the state transition probability of the first project multi-source data, and obtain second project multi-source data.

[0056] A data normalization unit is configured to map the second project multi-source data to the interval [0, 1] by using a min-max normalization method, and obtain standard project multi-source data.

[0057] Specifically,

[0058] (I) Data acquisition (detailed version)

[0059] 1. Project ontology data acquisition

[0060] Core data items and acquisition content

[0061] Design data: including structural design drawings (dam / channel structure drawings, reinforcement drawings), construction organization design (stage construction plan, key process flowchart), BIM model file (three-dimensional component attributes, process correlation), and design change records (change time, change content and impact range) need to be covered.

[0062] Geological survey data: soil physical parameters (moisture content, void ratio, compression coefficient), rock layer distribution data (lithology, burial depth, integrity coefficient), hydrogeological data (confined water head, permeability coefficient), and adverse geological body distribution (landslide body, cave location and scale).

[0063] Engineering technical parameters: building design indicators (dam height, dam length, impermeable grade), material performance parameters (concrete strength grade, steel yield strength), and construction technology standards (compaction parameters, maintenance period).

[0064] Collection method

[0065] Interface with CAD / BIM system of the design institute, automatic synchronization of drawings and model files (daily incremental update).

[0066] Obtain electronic version of the survey report from the geological survey unit, extract structured data through OCR technology, and establish a geological database.

[0067] The project technical department is responsible for sorting out design change orders and entering the engineering change management system (completed within 24 hours).

[0068] Responsible subject: project technical department + geological survey team

[0069] Data attributes: static data as the main, dynamic update when design changes, data formats include DWG, PDF, CSV, BIM model format (.rvt).

[0070] 2. Real-time progress data collection

[0071] Core data items and collection content

[0072] Process completion data: daily completion quantity of sub-item projects (dam filling volume, concrete pouring strength), process start / end time, acceptance record (pass rate, rectification items).

[0073] On-site working condition data: number and distribution of personnel in construction area (classified by type), mechanical equipment operation status (machine running time, fault record, oil consumption), real-time consumption of materials (cement, steel, sand and stone consumption).

[0074] Abnormal event record: stoppage reason and duration (weather influence, equipment failure, safety accident), rework range and treatment measures.

[0075] Collection method

[0076] Construction teams fill in the "Daily Construction Log" through the mobile APP, and the system automatically checks the data integrity (mandatory items include completion quantity, resource input, abnormal situation).

[0077] Deploy 4K high-definition cameras (1 per 5000㎡ construction area) and drones (take off at 8 am, 12 pm and 6 pm every day), and use computer vision to identify personnel / equipment distribution and construction surface status.

[0078] Install Internet of Things sensors (concrete curing temperature and humidity sensors, dam displacement monitors) at key process points, and transmit data to the cloud platform in real time (sampling frequency 1 / 10 minutes).

[0079] Mechanical equipment is equipped with GPS and working condition sensors, real-time uploading of location and running parameters (once every 30 seconds).

[0080] Subject of responsibility: construction team + monitoring center

[0081] Data attributes: dynamic data, time granularity accurate to hours, storage format is JSON (real-time stream data), JPG / MP4 (image data).

[0082] 3. Resource and environment data collection

[0083] Core data items and collection content

[0084] Resource supply data:

[0085] Human resources: the number of registered workers of each type, attendance records, skill levels, and shift planning.

[0086] Material resources: material inventory (rebar model / quantity, cement grade / inventory), equipment account (model, purchase time, maintenance record), rental equipment access / exit time.

[0087] Financial resources: monthly financial plan, actual disbursement amount, and cost expenditure of each sub-item project.

[0088] Environmental monitoring data:

[0089] Meteorological data: temperature, precipitation, wind force level, and visibility (updated hourly). Hydrological data: river water level, flow rate, flow, and water quality indicators (pH value, sediment content).

[0090] Geological environment: slope stability (displacement rate), groundwater level change, and earthquake monitoring (magnitude, focal depth).

[0091] Collection method

[0092] Integrate with human resource management system (HRM), material management system (ERP), and financial system to automatically synchronize resource data (batch update every morning).

[0093] Access the API of national weather stations and self-built weather station data (set up temperature and humidity, rainfall sensors), and hydrological data interface with the monitoring system of the basin hydrological bureau.

[0094] Install total station and inclinometer on the slope, and transmit data to the monitoring platform through 4G module (once every hour).

[0095] Subject of responsibility: material department + financial department + safety and environmental protection department

[0096] Data attributes: quasi-real-time data, resource data updated daily, environmental data updated according to monitoring frequency, storage format: relational database table, CSV (environmental monitoring data).

[0097] 4. Management of collaborative data collection

[0098] Core data items and collection content

[0099] Collaborative process data: task assignment records (person in charge, deadline, priority), approval process progress (design change approval, payment approval), cross-department coordination records.

[0100] Communication record data: supervision instructions and replies (rectification requirements, acceptance opinions), owner's instructions, and construction unit meeting minutes (resolutions, to-do list).

[0101] Document transfer data: construction plan review records, detection report transmission track, and data archiving status.

[0102] Collection method

[0103] Based on the project management system (ProjectWise, Primavera), automatically capture task progress and approval node data.

[0104] Data attributes: semi-dynamic data, updated according to events, storage format: XML (process data), DOCX (documents), MySQL database (structured records).

[0105] 5. Data collection quality control

[0106] Establish a data verification rule library: set thresholds for key indicators (concrete strength value range, maximum number of personnel attendance), automatically alert when thresholds are exceeded, and review by data officers.

[0107] Implement a three-level audit mechanism: team fills in data → construction team leader preliminary review → technical department final review, ensure that the accuracy of daily data is ≥98%.

[0108] Calibrate monitoring equipment regularly: sensors are calibrated once a month, drones are checked for flight accuracy every quarter, and ensure that data collection error is ≤5%.

[0109] Data backup strategy: real-time data uses off-site dual backup (local server + cloud storage), automatic backup every morning, and retains 30 days of historical versions.

[0110] (2) Data processing

[0111] Redundancy and noise elimination

[0112] Similarity analysis: Calculate the similarity between different data, use cosine similarity formula, data with high similarity (more than 0.9) are considered as redundant data and removed.

[0113] Markov chain: Establish state transition matrix, identify abnormal fluctuation noise data and correct or eliminate by analyzing state transition probability of data sequence.

[0114] Normalization processing: Using min-max normalization method, map data to [0, 1] interval.

[0115] Multi-source feature extraction module, for extracting feature data in standard engineering multi-source data, obtaining feature engineering multi-source data;

[0116] Specifically, the embodiment also includes a feature setting unit for setting the feature engineering multi-source data to include at least geological complexity index, daily effective construction time, resource matching degree, environmental interference factor, inter-process logical constraint strength and coordination efficiency index.

[0117] Complexity calculation unit, for calculating geological complexity index based on rock mass classification, joint development degree and permeability coefficient parameters in geological survey report through weighted average method;

[0118] Construction time acquisition unit, for dynamically calculating meteorological environment data and site conditions through rule base to obtain daily effective construction time;

[0119] Resource matching degree calculation unit, for extracting key path logical relationship from construction network plan based on graph theory to obtain resource matching degree;

[0120] Interference factor calculation subunit, for setting interference weight of different factors, rainstorm weather interference weight is 0.3, flood is 0.5, and environmental interference factor is obtained by weighted calculation;

[0121] Constraint strength acquisition unit, for assigning constraint strength value to each pair of related processes according to the sequence and influence degree of processes to obtain inter-process logical constraint strength.

[0122] Specifically,

[0123] 2.1 Geological complexity index (GCI)

[0124] Based on geological survey data calculation:

[0125] GCI = 0.3 × rock integrity coefficient + 0.2 × fault density + 0.3 × groundwater depth + 0.2 × soil bearing capacity

[0126] Where the rock integrity coefficient is quantified by RQD value (Rock Quality Designation), and the fault density is calculated by the number of faults per square kilometer. The index value ranges from [0, 1], and the higher the value, the more complex the geological conditions.

[0127] 2.2 Daily effective construction duration (DEWD)

[0128] DEWD = planned operation duration x ∏ (influencing factors)

[0129] The influencing factors include: weather influence coefficient (1.0 for sunny days, 0.8 for light rain, and 0 for heavy rain), equipment availability (real-time monitoring value), and personnel attendance rate (face recognition statistics value), accurate to 0.5 hours.

[0130] 2.3 Resource matching degree (RMD)

[0131] RMD = cosine similarity between actual input resource vector and planned resource vector:

[0132] RMD = (∑(actual resource i x planned resource i)) / (∑actual resource i 2 x ∑planned resource i 2 )

[0133] Resource dimensions include manpower (workdays), machinery (shifts), and materials (tons), with a value range of [0, 1], and <0.6 triggering resource allocation warning.

[0134] 2.4 Environmental interference factor (EIF)

[0135] EIF = 0.4 x rainfall intensity index + 0.3 x wind speed level + 0.2 x temperature deviation value + 0.1 x earthquake warning level

[0136] After standardizing each sub-index, weighting, and determining that the value > 0.7 is a strong interference state, automatically associate the historical similar working condition response scheme.

[0137] 2.5 Process logic constraint strength (PLCS)

[0138] Based on double code network diagram calculation:

[0139] PLCS = 1 - ∑(virtual work duration) / ∑(critical path work duration)

[0140] Reflects the rigidity of process connection, the higher the value, the greater the impact of previous work delay on the subsequent, used to identify key process nodes.

[0141] 2.6 Coordination efficiency index (CEI)

[0142] CEI = 1 - (information transmission delay duration + approval process time consumption) / standard coordination cycle

[0143] The interaction data of each participant (owner, construction, supervision) is recorded through the workflow system, and when the value is less than 0.5, the optimization of the collaborative process is prompted.

[0144] The influence weight prediction module is used to adopt an improved LSTM long short-term memory network, integrate a space-time attention mechanism to establish an LSTM prediction model, and use the LSTM prediction model to learn the differentiating influence weight of the time attention of different historical moments and the space attention of different engineering parts on the prediction target;

[0145] Specifically, the embodiment also includes a judgment unit for establishing a construction phase identification model, calculating the daily average work intensity through a sliding window, determining the construction period when the intensity is greater than the threshold, otherwise determining the shutdown period, and determining the dynamic adjustment mechanism of the decay coefficient;

[0146] The processing unit is used to perform stability processing on the numerical value of the input gate combination function, and set the lower limit of the sigmoid output to 0.01 to obtain an improved LSTM long short-term memory network;

[0147] The establishment unit is used to use the improved LSTM long short-term memory network to integrate a space-time attention mechanism to establish an LSTM prediction model.

[0148] The time attention unit is used to adopt multi-head attention to improve parallelism and alleviate gradient explosion through a scaling factor; a key event dictionary is established to identify the corresponding timestamp through text matching;

[0149] The space attention unit is used to divide according to the WBS three-level coding, each unit is associated with the corresponding BIM component ID to ensure that the physical space and the data space are mapped consistently, and the determination of the adjustment coefficient is based on the geological complexity index GCI to set a segmented function.

[0150] Specifically,

[0151] 3.1 Improved LSTM network architecture

[0152] 3.1.1 Basic unit parameter optimization

[0153] The dynamic adjustment mechanism of the decay coefficient λ: establish a construction phase identification model, calculate the daily average work intensity through a sliding window (window size = 7 days), determine the construction period when the intensity is greater than the threshold (preset as 70% of the planned intensity), otherwise determine the shutdown period.

[0154] Parallel computing of memory cell state: use GPU to accelerate the matrix operation of Ct, divide the time series into blocks (block size = 1000 steps) for large-scale water conservancy projects, and solve the memory limitation problem through gradient accumulation.

[0155] 3.1.2 Gating function engineering implementation

[0156] Numerical stability processing of input gate combination function: set lower limit of sigmoid output to 0.01 (avoid gradient disappearance), and gradient clipping is performed for tanh output (threshold = 5.0).

[0157] Gating parameter initialization strategy: Xavier initialization is used for weight matrices such as Wxi and Whi, and bias terms bi and bg are initialized to 0.1 (to accelerate early convergence).

[0158] 3.2 Spatio-temporal attention mechanism engineering scheme

[0159] 3.2.1 Time attention module deployment

[0160] Self-attention calculation optimization: multi-head attention (number of heads = 8) is used to improve parallelism, and scaling factor √dk (dk = hidden layer dimension / number of heads) is used to alleviate gradient explosion, dk is usually set to 64.

[0161] Key node weight enhancement: establish a key event dictionary (flow interception, flood passage), identify the corresponding timestamp through text matching, automatically multiply the time αt by γ, and determine the optimal value of γ through grid search (step size 0.1 in the range of 1.5-2.0).

[0162] 3.2.2 Spatial attention module implementation

[0163] Spatial unit division standard: divided according to WBS (Work Breakdown Structure) three-level coding (dam body → left bank dam section → 1 dam section), each unit is associated with the corresponding BIM component ID, ensuring that the physical space and data space are mapped consistently.

[0164] Determination of adjustment coefficient δ: based on the geological complexity index GCI, set a piecewise function:

[0165] δ = 1.2 + 0.6 × (GCI - 0.5) when GCI ≥ 0.5

[0166] δ = 1.2 when GCI < 0.5

[0167] 3.3 Model training full-process management

[0168] 3.3.1 Training environment configuration

[0169] Hardware environment: NVIDIA A100 GPU (80GB video memory), CPU is Intel Xeon Platinum 8360Y, memory is 1TB, ensuring no memory overflow when processing 100,000-level samples in batches.

[0170] Software Stack: PyTorch1.13 is used for deep learning framework, and DDP (Distributed Data Parallel) is used for distributed training. When the sample size is greater than 1 million, mixed precision training (FP16) is enabled to accelerate convergence.

[0171] 3.3.2 Training parameter fine-tuning settings

[0172] Adam optimizer parameters: β1 = 0.9, β2 = 0.999, ε = 1e-8, weight decay = 1e-5 (to suppress overfitting).

[0173] Learning rate scheduling strategy: cosine annealing learning rate is used, initial LR = 0.001, minimum LR = 0.0001, period = 50 epochs, and restart period = 30 epochs.

[0174] 3.3.3 Loss function optimization

[0175] Dynamic adjustment of penalty coefficient μ: when the proportion of overdue samples is greater than 20%, μ is automatically increased by 0.2 (upper limit = 2.0), otherwise it is decreased by 0.1 (lower limit = 0.8), to achieve adaptive attention to overdue risk.

[0176] Early stopping strategy engineering adaptation: set the sliding average of the validation set loss (window = 3), trigger the stop training when the average value of the last 10 epochs does not decrease, and save the model parameters at this time as the optimal model.

[0177] 3.4 Weight learning and visualization system

[0178] 3.4.1 Weight learning process monitoring

[0179] Gradient tracking of backpropagation: record the gradient changes of spatiotemporal weights through TensorBoard, when the average gradient is less than 1e-6, use learning rate multiplication strategy (temporarily amplify by 10 times) for the corresponding layer parameters.

[0180] Cross-weight calculation method: construct spatiotemporal attention tensor (Kronecker product), extract the spatiotemporal coordinates corresponding to the elements with a value greater than 0.7 in the tensor, and mark them as key influence points.

[0181] 3.4.2 Weight output and calibration mechanism

[0182] Heat map generation tool: use Matplotlib to draw time-space weight heat map, x-axis is time (weekly granularity), y-axis is engineering site, color depth represents weight value, and automatically label the top 5 spatiotemporal points.

[0183] Weight calibration process: when CV>0.5, freeze the main parameters of the model, only fine-tune the linear transformation matrix (Wt, Ws) of the attention module, retrain for 5 epochs with small batch size (batch_size=32), and if the calibrated CV still fails to meet the standard, trigger full retraining.

[0184] Node prediction and early warning module, for inputting the differentiated influence weight and feature engineering multi-source data into the XGBoost ensemble learning model to quantify the comprehensive influence of the complex interaction between multiple factors on progress, and output node completion prediction time and risk level signal.

[0185] Specifically, the embodiment also includes a feature processing unit for splicing the weighted feature vector and the original feature engineering data, and reducing the dimension through PCA principal component analysis;

[0186] Model fusion unit, for adopting stacking ensemble strategy, taking the prediction result of LSTM as the input feature of XGBoost, using linear regression to integrate the output of each base model in the meta model, determining the weight through grid search, and outputting node completion prediction time and risk level signal.

[0187] Specifically,

[0188] 4.1 Model integration strategy

[0189] 4.1.1 Feature fusion scheme

[0190] LSTM output feature processing: splice the weighted feature vector h and the original feature engineering data, reduce the dimension through principal component analysis (PCA) (retain 95% variance), and reduce the input dimension to within 200 dimensions.

[0191] XGBoost parameter configuration: learning rate=0.05, number of trees=500, maximum depth=8, minimum_child_weight=3, and cross-validation (k=5) is used to optimize parameters.

[0192] 4.1.2 Multi-model fusion mechanism

[0193] Adopt stacking ensemble strategy: take the prediction result of LSTM as one of the input features of XGBoost, use linear regression to integrate the output of each base model in the meta model (second layer), and determine the weight through grid search.

[0194] 4.2 Prediction result output and risk assessment

[0195] 4.2.1 Node completion time prediction

[0196] Output: Provide interval prediction (90% confidence interval), e.g. "Dam cresting node predicted completion time: 2024-10-15 ± 7 days", interval width dynamically adjusted according to historical prediction errors.

[0197] Accuracy verification: Adopt rolling prediction method (re-predict every 7 days), record prediction errors Trigger warning when ε > warning threshold (5% of planned duration).

[0198] 4.2.2 Risk level signal generation

[0199] Risk quantification model: Risk value R = 0.6 × prediction error rate + 0.4 × weighted coefficient of variation CV, where error rate = ε / yi.

[0200] Classification criteria:

[0201] Low risk: R < 0.3, output green signal, recommend routine monitoring

[0202] Medium risk: 0.3 ≤ R < 0.6, output yellow signal, suggest developing response plans

[0203] High risk: R ≥ 0.6, output red signal, trigger risk review meeting.

[0204] Its beneficial effects lie in that, by collecting engineering multi-source data, using similarity analysis and Markov chain to eliminate redundancy and noise and then performing normalization processing, standard engineering multi-source data is obtained; feature data in the standard engineering multi-source data is extracted to obtain feature engineering multi-source data; an improved LSTM long short-term memory network is adopted, a spatial-temporal attention mechanism is integrated to establish an LSTM prediction model, and the LSTM prediction model is used to learn the differentiating influence weight of time attention at different historical time and space attention at different engineering parts on the prediction target; the differentiating influence weight and the feature engineering multi-source data are input into an XGBoost integrated learning model to quantify the comprehensive influence of complex interactions among multiple factors on progress, and a node is output to complete the prediction time and risk level signal. 1. The prediction accuracy is significantly improved: by deeply fusing multi-source heterogeneous big data and constructing dynamic coupling features reflecting the characteristics of water conservancy projects, and combining the hybrid intelligent model for accurate modeling of spatial-temporal dependence and complex interaction, the accuracy of progress prediction is greatly improved. 2. Precise risk early warning is realized: the model can not only predict progress, but also quantify the influence degree of each risk factor and output the contribution ranking, provide risk cause tracing, realize "early detection, accurate positioning" active early warning, and provide direct basis for management decision-making. 3. Strong environmental adaptability: for the complex environmental factors such as geology, hydrology and meteorology specific to water conservancy projects, through special feature construction and spatial-temporal attention mechanism, the model can effectively perceive and respond to these dynamic changes. 4. High value of decision support: not only time point prediction is provided, but also risk distribution, key factor analysis and early warning signal, which greatly improves the foresight and scientificity of project management, and helps resource optimization and risk active prevention and control.

[0205] Please refer to Figure 2 In a water conservancy project progress prediction system based on big data, the multi-source data acquisition module comprises the following units:

[0206] A data acquisition unit is configured to acquire engineering ontology data, real-time progress data, resource environment data and management coordination data in the construction process of a water conservancy project, and obtain engineering multi-source data.

[0207] A redundant data elimination unit is configured to calculate the similarity between different data, use a cosine similarity formula to regard data with high similarity as redundant data and eliminate the redundant data, and obtain first engineering multi-source data.

[0208] A noise data correction unit is configured to establish a state transition matrix, correct abnormal fluctuation noise data by analyzing the state transition probability of the first engineering multi-source data, and obtain second engineering multi-source data.

[0209] A data normalization unit is configured to map the second engineering multi-source data to the interval [0, 1] by a min-max normalization method, and obtain standard engineering multi-source data.

[0210] Please refer to Figure 3 In a water conservancy project progress prediction system based on big data, the multi-source feature extraction module comprises the following units:

[0211] The complexity calculation unit is configured to calculate a geological complexity index based on rock mass classification, joint development degree and permeability coefficient parameters in the geological survey report by using a weighted average method.

[0212] The construction duration acquisition unit is configured to dynamically calculate meteorological environment data and site conditions by using a rule base to obtain daily effective construction duration.

[0213] The resource matching degree calculation unit is configured to extract a critical path logical relationship from a construction network plan based on graph theory calculation of process criticality weight to obtain resource matching degree.

[0214] The interference factor calculation sub-unit is configured to set interference weights of different factors, with a storm weather interference weight being 0.3 and a flood interference weight being 0.5, and to obtain an environmental interference factor by weighted calculation.

[0215] The constraint strength acquisition unit is configured to assign constraint strength values to each pair of related processes according to the sequence and influence degree of the processes to obtain logical constraint strength between the processes.

[0216] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application,

[0217] These changes and improvements all fall within the scope of the claimed present application. The scope of the present application is defined by the appended claims and their equivalents.

[0218] The scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A big data based water works progress prediction system, characterized by, The water conservancy project progress prediction system comprises the following modules: A multi-source data acquisition module, configured to acquire project multi-source data, normalize the project multi-source data after eliminating redundancy and noise by similarity analysis and Markov chain, and obtain standard project multi-source data; A multi-source feature extraction module, configured to extract feature data in the standard project multi-source data, and obtain feature project multi-source data; An influence weight prediction module, configured to use an improved LSTM long short-term memory network, integrate a space-time attention mechanism to establish an LSTM prediction model, and use the LSTM prediction model to learn the differentiating influence weight of time attention at different historical time points and space attention of different engineering parts on a prediction target; A node prediction and early warning module, configured to input the differentiating influence weight and the feature project multi-source data into an XGBoost integrated learning model to quantify the comprehensive influence of complex interactions among multiple factors on progress, and output a node completion prediction time and a risk level signal.

2. The big data based water project progress prediction system of claim 1, wherein, The multi-source data acquisition module comprises the following units: A data acquisition unit, configured to acquire project ontology data, real-time progress data, resource environment data, and management coordination data in a water conservancy project construction process, and obtain project multi-source data; A redundant data elimination unit, configured to calculate the similarity between different data, regard data with high similarity as redundant data, and eliminate the redundant data, and obtain first project multi-source data; A noise data correction unit, configured to establish a state transition matrix, analyze the state transition probability of the first project multi-source data, correct abnormal fluctuation noise data, and obtain second project multi-source data; A data normalization unit, configured to map the second project multi-source data to the interval [0, 1] by a min-max normalization method, and obtain standard project multi-source data.

3. The big data based water works progress prediction system as claimed in claim 1, wherein, The multi-source feature extraction module comprises the following units: A feature setting unit, configured to set the feature project multi-source data to at least include a geological complexity index, a daily effective construction duration, a resource matching degree, an environmental interference factor, a logical constraint strength between processes, and a coordination efficiency index.

4. The big data based water works progress prediction system of claim 1, wherein, The multi-source feature extraction module further comprises the following units: A complexity calculation unit, configured to calculate the geological complexity index by a weighted average method based on rock mass classification, joint development degree, and permeability coefficient parameters in a geological survey report; A construction duration acquisition unit, configured to obtain the daily effective construction duration by dynamically calculating meteorological environment data and site conditions through a rule base; A resource matching degree calculation unit, configured to extract a key path logical relationship from a construction network plan based on graph theory to calculate a process criticality weight, and obtain the resource matching degree; An interference factor calculation subunit, configured to set interference weights of different factors, set a rainstorm weather interference weight to 0.3 and a flood interference weight to 0.5, and obtain the environmental interference factor by weighted calculation; A constraint strength acquisition unit, configured to assign a constraint strength value to each pair of related processes according to the sequence and influence degree of the processes, and obtain the logical constraint strength between processes.

5. The big data based water works progress prediction system as claimed in claim 1, wherein, The influence weight prediction module further comprises the following units: The judgment unit is used for establishing a construction phase identification model, calculating daily average operation intensity through a sliding window, determining a construction period when the intensity is greater than a threshold value, otherwise determining a shutdown period, and determining a dynamic adjustment mechanism of a decay coefficient; The processing unit is used for performing stability processing on numerical values of input gate combination functions, setting a lower limit of sigmoids output to 0.01, and obtaining an improved LSTM long short-term memory network. The establishing unit is used for establishing an LSTM prediction model by using the improved LSTM long short-term memory network and integrating a space-time attention mechanism.

6. The big data based water works progress prediction system of claim 1, wherein, The influence weight prediction module further includes the following units: The time attention unit is used for improving parallelism by adopting multi-head attention, relieving gradient explosion by a scaling factor, establishing a key event dictionary, and identifying corresponding timestamps through text matching; The space attention unit is used for dividing according to WBS three-level coding, associating corresponding BIM component IDs for each unit, ensuring that physical space and data space are mapped consistently, and determining an adjustment coefficient based on a geological complexity index GCI and setting a segmented function.

7. The big data based water works progress prediction system as claimed in claim 1, wherein, The node prediction and early warning module further includes the following units: The feature processing unit is used for splicing a weighted feature vector and original feature engineering data, and performing dimension reduction through PCA principal component analysis; The model fusion unit is used for adopting a stacking integration strategy, taking prediction results of the LSTM as input features of XGBoost, integrating outputs of each base model by using linear regression for a meta model, determining weights through grid search, and outputting node completion prediction time and risk level signals.

8. A method of implementing a big data based water works progress prediction system as claimed in claim 1, wherein, The method includes the following steps: Collecting engineering multi-source data, performing normalization processing after eliminating redundancy and noise by using similarity analysis and Markov chains, and obtaining standard engineering multi-source data; Extracting feature data in the standard engineering multi-source data, and obtaining feature engineering multi-source data; Using an improved LSTM long short-term memory network, integrating a space-time attention mechanism to establish an LSTM prediction model, and learning time attention at different historical times and spatial attention of different engineering parts to learn the differential influence weight of the prediction target; Inputting the differential influence weight and the feature engineering multi-source data into an XGBoost integrated learning model to quantify the comprehensive influence of complex interactions among multiple factors on progress, and outputting node completion prediction time and risk level signals.

9. A method of implementing a big data based water works progress prediction system as claimed in claim 1, wherein, The method includes the following steps: Establishing a construction phase identification model, calculating daily average operation intensity through a sliding window, determining a construction period when the intensity is greater than a threshold value, otherwise determining a shutdown period, and determining a dynamic adjustment mechanism of a decay coefficient; Performing stability processing on numerical values of input gate combination functions, setting a lower limit of sigmoids output to 0.01, and obtaining an improved LSTM long short-term memory network; Using an improved LSTM long short-term memory network, integrating a space-time attention mechanism to establish an LSTM prediction model.

10. A method of implementing a big data based water works progress prediction system as claimed in claim 1, wherein, The method includes the following steps: Improving parallelism by adopting multi-head attention, relieving gradient explosion by a scaling factor, establishing a key event dictionary, and identifying corresponding timestamps through text matching; According to the WBS three-level coding, each unit is associated with the corresponding BIM component ID, ensuring that the physical space and data space are mapped consistently, and the determination of the adjustment coefficient is based on the geological complexity index GCI, setting a segmented function.

Images