A method and system for optimizing the design of urban drainage networks based on BIM models

Abstract

The present invention relates to the technical field of urban drainage network optimization design, and discloses an urban drainage network optimization design method and system based on a BIM model. Specifically, an urban drainage network optimization design method based on a BIM model includes: collecting multi-source hydrological data and preprocessing the multi-source hydrological data; modeling the drainage system as a spatiotemporal graph structure to predict the bearing capacity of the drainage system; constructing an overflow risk assessment model to identify key influencing factors, quantify node risks and identify high-risk areas; constructing a fault model based on a causal graph to identify the causal chain of overflow events and achieve a transition from correlation analysis to causal understanding; and optimizing drainage system parameters using a BIM model. The present invention realizes dynamic bearing capacity prediction, accurate risk assessment and abnormal overflow source diagnosis of the drainage system by integrating multi-source data analysis, spatiotemporal graph neural network and causal inference technology.

Description

Technical Field

[0001] The present invention relates to the technical field of urban drainage network optimization design, and more specifically, to an urban drainage network optimization design method and system based on a BIM model. Background Art

[0002] Urban drainage networks, as an essential component of urban infrastructure, undertake the critical functions of rainwater collection, transportation, and treatment. With the acceleration of urbanization and the frequent occurrence of extreme climate events, traditional drainage network design and management methods face numerous challenges.

[0003] Traditional urban drainage network design is primarily based on historical statistical data and static design standards, using return period rainfall as the design basis. This design approach typically assumes uniform rainfall distribution and independent operation of each system component, lacking in-depth consideration of complex hydrological processes and the system's dynamic characteristics. In practical applications, this static design approach has the following shortcomings:

[0004] Lack of dynamic adaptability. Traditional design methods mainly rely on historical average data and empirical formulas, which are difficult to adapt to changes in extreme rainfall patterns caused by climate change. When encountering rainfall events that exceed the design standards, the system is prone to local or overall overflow, causing urban waterlogging; insufficient analysis of the system's overall integrity. Existing methods mostly adopt a segmented or unitized design approach, simplifying complex drainage networks into independent subsystems for analysis, ignoring the hydraulic transmission relationship and mutual influence between different areas. In actual operation, flow changes in upstream areas will be transmitted to downstream through the pipe network, forming a cascade effect, and traditional methods are difficult to accurately predict this system response; limited fault diagnosis capabilities, drainage system overflow events are often the result of a combination of multiple factors. The results of the cooperation include pipe blockage, equipment failure, abnormal rainfall, etc. Traditional analysis methods are mainly based on statistical correlation analysis, which makes it difficult to identify the real root causes, resulting in a lack of targeted problem handling; data utilization efficiency is not high. Although modern urban drainage systems are equipped with a large number of monitoring equipment, which can collect rich data including rainfall, flow, water level, etc., these data are often stored in a scattered manner and in an inconsistent format. There is a lack of effective fusion analysis mechanism, and the data value is not fully utilized; design and operation and maintenance are disconnected. Traditional BIM modeling technology is mainly used for static design and three-dimensional display, and lacks an organic combination with dynamic analysis tools such as hydrological analysis and risk assessment, making it difficult to form an integrated management system from design to operation and maintenance.

[0005] In recent years, with the development of artificial intelligence and big data technologies, some research has begun exploring the application of machine learning methods to drainage system analysis. However, existing research has largely focused on improving a single aspect of the technology, such as considering only time series prediction or focusing solely on spatial correlation analysis, lacking comprehensive modeling of the spatiotemporal coupling characteristics of drainage systems. Furthermore, existing methods still have shortcomings in handling multi-source heterogeneous data fusion, modeling complex network structures, and mining causal relationships.

[0006] Therefore, it is urgent to develop an urban drainage network optimization design method that can fully utilize BIM model information, effectively integrate multi-source data, accurately model spatiotemporal correlation relationships, and have causal inference capabilities, so as to improve the design accuracy, operation efficiency and risk response capabilities of the drainage system. Summary of the Invention

[0007] The present invention provides a method and system for optimizing the design of an urban drainage network based on a BIM model, which solves the technical problems in related technologies such as the lack of dynamic adaptability of traditional drainage system design, insufficient system integrity analysis, limited fault diagnosis capability, and low data utilization efficiency.

[0008] The present invention provides an urban drainage network optimization design method based on a BIM model, comprising:

[0009] Collect multi-source hydrological data, pre-process them through spatiotemporal data heterogeneous fusion technology, and generate a comprehensive digital representation of the drainage system;

[0010] Based on comprehensive digital representation, the drainage system is modeled as a spatiotemporal graph structure. A deep learning model combining a temporal graph convolutional network and a gated recurrent unit is applied to analyze the dynamic response characteristics of the system under different time periods and rainfall conditions, and to predict the carrying capacity of the drainage system.

[0011] Based on the prediction results, an overflow risk assessment model is constructed to identify key influencing factors through the attention mechanism, quantify the node risk and identify high-risk areas;

[0012] Based on the output of the risk assessment model, a fault model based on a causal graph is constructed. Through intervention learning and counterfactual reasoning, the causal chain of the overflow event is identified, achieving a transition from correlation analysis to causal understanding.

[0013] Based on the comprehensive results of risk assessment and causal analysis, the BIM model was used to optimize the drainage system parameters, and the optimization effect was verified through numerical simulation.

[0014] Furthermore, the preprocessing of multi-source hydrological data includes:

[0015] Clean the raw data through outlier detection and processing, missing value filling and data format unification;

[0016] The data with different temporal resolutions are unified to the same time step through resampling technology, and the data with different coordinate systems and spatial resolutions are mapped to a unified spatial reference system through geographic coordinate transformation and spatial interpolation methods;

[0017] The key features reflecting the characteristics of the drainage system are extracted, and the features from different sources are concatenated to form a unified feature vector.

[0018] Furthermore, in the process of constructing the overflow risk assessment model, a node risk scoring function is defined. The node risk score consists of three parts, including:

[0019] Node centrality reflects the topological importance of a node in the network;

[0020] The ratio of predicted traffic to historical average traffic reflects the degree of current traffic anomaly;

[0021] Capacity margin refers to the ratio of the difference between design capacity and predicted flow to design capacity.

[0022] Furthermore, the fault model based on the causal graph includes the following steps:

[0023] Based on the physical characteristics of the drainage system and expert knowledge, a structural causal model of the system state variables is constructed;

[0024] Estimating causal effects between variables through intervention learning, distinguishing between correlation and causation;

[0025] Automatically discover potential hidden causal relationships in data using graph attention networks;

[0026] Conduct counterfactual analysis of overflow events that have occurred to identify key intervention points.

[0027] Furthermore, the optimization of drainage system parameters using the BIM model includes the following steps:

[0028] Establish a mapping relationship between BIM model parameters and risk factors to identify key parameters that affect system performance;

[0029] Construct a multi-objective optimization model that considers flood control capacity, construction cost, and operation and maintenance efficiency;

[0030] The non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and obtain the Pareto optimal solution set;

[0031] The performance of the optimization scheme under extreme working conditions is verified through numerical simulation, and the best scheme is selected for implementation.

[0032] Furthermore, the objective function of the multi-objective optimization model includes:

[0033] The flood control capability objective function represents maximizing the negative value of the maximum risk score of all nodes in the system at all time steps;

[0034] The construction cost objective function represents the sum of the construction costs of all facilities;

[0035] The operation and maintenance efficiency objective function represents the sum of the maintenance cost of the facility and the complexity of the interaction between facilities.

[0036] Furthermore, the deep learning model is trained when predicting the carrying capacity of the drainage system, using the following steps:

[0037] The historical data is divided into training set and validation set in a ratio of 7:3;

[0038] Use training data to train models, including graph convolutional networks, gated recurrent units, and multilayer perceptrons;

[0039] Evaluate model performance on the validation set and use early stopping strategy to avoid overfitting;

[0040] The best performing model is saved for prediction.

[0041] Furthermore, the calculation process of the counterfactual reasoning includes:

[0042] Infer the value of exogenous variables based on observed data and structural causal models;

[0043] A hypothetical intervention is imposed on the target variable, changing its value from the actual observed value to the hypothesized value;

[0044] Calculate the impact of intervention spread on other variables based on the causal graph structure;

[0045] Calculate the predicted value of the target outcome variable under the intervention condition;

[0046] Compare the differences between actual observations and counterfactual predicted values ​​to evaluate the intervention effect.

[0047] Furthermore, the process of modeling the drainage system as a spatiotemporal graph structure includes:

[0048] Extract the spatial coordinates and connection relationships of drainage facilities from the BIM model;

[0049] Attach time series characteristic data to each node, including historical flow, water level and rainfall;

[0050] A time-varying adjacency matrix is ​​constructed, and the edge weights are determined based on the physical parameters of the pipe diameter and slope.

[0051] The present invention provides a BIM-based urban drainage network optimization design system for executing the above-mentioned BIM-based urban drainage network optimization design method, comprising:

[0052] Multi-source data fusion module, used to collect and fuse multi-source hydrological data to generate a comprehensive digital representation of the drainage system;

[0053] The spatiotemporal graph modeling module is used to construct a spatiotemporal graph structure based on comprehensive digital representations and use deep learning models to predict the dynamic carrying capacity of the system;

[0054] Risk quantification module, used to assess overflow risks based on prediction results and identify key influencing factors and high-risk areas;

[0055] Causal analysis module, used to build a fault causal model and identify the causal relationship of overflow events through intervention learning and counterfactual reasoning;

[0056] The optimization and verification module is used to optimize BIM parameters based on risk and causal analysis results and verify the effectiveness of the optimization scheme.

[0057] The beneficial effects of the present invention are as follows: through multi-source data fusion and spatiotemporal graph neural network, the accuracy of system dynamic bearing capacity prediction is higher than that of traditional methods based on statistical regression or physical models. The high-precision prediction results provide reliable data support for drainage system design, reducing the risk of over-design or under-design;

[0058] The optimized drainage system's ability to cope with extreme weather conditions has been improved, and the risk of overflow at key nodes has been reduced. The dynamic model based on the spatiotemporal graph neural network can accurately capture the system's response under extreme rainfall conditions. The design solution generated by combining multi-objective optimization has greater adaptability to extreme working conditions, reducing the occurrence of waterlogging disasters.

[0059] While ensuring safety, the redundancy of the drainage system is optimized, while reducing construction costs and avoiding overdesign and resource waste. The multi-objective optimization method can balance flood control capacity, economy and operation and maintenance efficiency, providing decision makers with multiple alternatives to achieve optimal resource allocation.

[0060] Using causal inference technology, the system can accurately identify the root cause of overflow incidents, improving diagnostic accuracy compared to traditional methods and making remediation measures more targeted and effective. Compared to traditional correlation analysis, causal inference technology can distinguish between direct causal relationships and indirect associations, providing more precise direction for problem solving.

[0061] Expand BIM from a static design tool to a dynamic decision-making platform, realize the transformation from data-driven to knowledge-driven, and improve the intelligence level of the drainage system throughout its life cycle. Through the integration with multi-source data analysis, spatiotemporal graph neural network and causal inference technology, the BIM model is no longer just a three-dimensional visualization tool, but a comprehensive platform integrating information integration, performance analysis, risk assessment and decision support. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for optimizing the design of an urban drainage network based on a BIM model in the present invention;

[0063] Figure 2 It is a line graph comparing the prediction accuracy of different prediction methods under different rainfall intensity conditions;

[0064] Figure 3 It is a radar chart comparing the method of the present invention with the traditional method in terms of multiple key performance indicators;

[0065] Figure 4 It is an area diagram of the overflow risk change of the system before and after optimization under different rainfall return period conditions;

[0066] Figure 5 It is a scatter plot of the calculation time of the method of the present invention in drainage networks of different scales;

[0067] Figure 6 It is a bar chart showing different solutions generated through multi-objective optimization in three dimensions: flood control capacity, construction cost, and operation and maintenance efficiency. DETAILED DESCRIPTION

[0068] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0069] At least one embodiment of the present invention discloses a method for optimizing the design of an urban drainage network based on a BIM model. Figure 1 As shown, including:

[0070] Step 1: Collect multi-source hydrological data and pre-process them using spatiotemporal data heterogeneous fusion technology to generate a comprehensive digital representation of the drainage system;

[0071] Use spatiotemporal data heterogeneous fusion technology to process multi-dimensional drainage system related data and generate a comprehensive digital representation of the drainage system. Specifically including:

[0072] Step 1.1, data collection;

[0073] The collection includes multi-source heterogeneous data such as historical rainfall, topography, soil permeability, urbanization level, and drainage facility operation status. Data sources include meteorological stations, hydrological monitoring equipment, geographic information systems, urban planning archives, and facility parameter information in BIM models.

[0074] In some implementations, real-time monitoring data, such as data from online rain gauges, water level gauges, and flow meters, can be integrated to enhance the system's responsiveness to emergencies. Alternatively, social media data and citizen reporting can be integrated to supplement traditional monitoring data, particularly in areas with insufficient monitoring coverage.

[0075] Step 1.2, data cleaning and standardization;

[0076] The collected raw data is cleaned, including outlier detection and processing, missing value imputation, and data format standardization. Outlier detection uses statistically based outlier identification algorithms, such as the Z-score or modified interquartile range (IQR) method. Missing value imputation uses appropriate interpolation methods based on the data type, such as time series interpolation for time series data and Kriging interpolation for spatial data. Data standardization uses methods such as minimum-maximum scaling or Z-score standardization to make data of different dimensions comparable.

[0077] Optionally, for data sources with high noise, signal processing methods such as wavelet decomposition or empirical mode decomposition (EMD) can be used to reduce noise and improve data quality. In cases where the data volume is particularly large, a distributed computing framework can be used for parallel processing to improve processing efficiency.

[0078] Step 1.3, spatiotemporal alignment;

[0079] Temporal and spatial alignment is performed on data from different sources, sampling frequencies, and spatial scales. Temporal alignment uses resampling techniques to unify data of different temporal resolutions to the same time step; spatial alignment uses geographic coordinate transformation and spatial interpolation methods to map data of different coordinate systems and spatial resolutions to a unified spatial reference system.

[0080] Step 1.4, feature extraction and fusion;

[0081] Based on the aligned multi-source data, key features reflecting the characteristics of the drainage system are extracted and then fused. Feature extraction utilizes methods such as time-frequency domain analysis, statistical feature calculation, and spatial feature extraction. Feature fusion employs an early fusion strategy, concatenating features from different sources into a unified feature vector that serves as input for subsequent analysis.

[0082] Step 2: Based on comprehensive digital representation, the drainage system is modeled as a spatiotemporal graph structure. A deep learning model combining a temporal graph convolutional network and a gated recurrent unit is applied to analyze the system's dynamic response characteristics under different time periods and rainfall conditions, and to predict the drainage system's carrying capacity.

[0083] The drainage system is modeled as a spatiotemporal graph structure, and a deep learning model combining a temporal graph convolutional network and a gated recurrent unit is applied to analyze the dynamic response characteristics of the system under different times and rainfall conditions, and to predict the carrying capacity of the drainage system. Specifically, it includes:

[0084] Step 2.1, modeling the spatiotemporal diagram of the drainage system;

[0085] Based on the drainage network topology in the BIM model, a spatiotemporal graph network representation is constructed:

[0086] ;

[0087] in Representing spatiotemporal graph networks; Represents a set of nodes, corresponding to key facilities in the drainage system, such as inspection wells, pumping stations, and overflow outlets; Represents the edge set, corresponding to the pipeline connection relationship; represents the time-varying adjacency matrix; represents the time variable; Represents the node feature matrix.

[0088] The construction process of the spatiotemporal graph network includes the following sub-steps: first, the spatial coordinates and connection relationships of the drainage facilities are extracted from the BIM model; then, time series feature data are attached to each node, including historical flow, water level, rainfall, etc.; finally, a time-varying adjacency matrix is ​​constructed, in which the edge weights can be determined based on physical parameters such as pipe diameter and slope, or data-driven weight learning can be performed in combination with historical data.

[0089] Step 2.2, spatial dependency modeling;

[0090] Apply graph convolutional networks to extract the spatial dependency characteristics of drainage systems. Graph convolution is used to extract spatial connections between nodes using input features. This process can be understood as the propagation and aggregation of information on the graph structure: each node updates its feature representation by aggregating information from its neighboring nodes. Graph convolution, through the convolution function defined on the graph, effectively captures the spatial propagation effects in the drainage network.

[0091] In the embodiments of this application, the graph convolutional network architecture comprises three layers of graph convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The first layer of graph convolution maps the original node features to a 64-dimensional hidden representation, the second layer maintains the dimensionality, and the third layer outputs a 32-dimensional node representation. Furthermore, a residual connection mechanism is used to add the input features to the graph convolution output, which helps alleviate the training difficulties of deep graph networks and improves model convergence.

[0092] Step 2.3, temporal evolution modeling;

[0093] Based on the extracted spatial features, a Gated Recurrent Unit (GRU) is applied to capture the temporal evolution of the system state. By updating and resetting gates, the GRU selectively retains historical information and integrates current input, effectively modeling long-term temporal dependencies.

[0094] The GRU calculation process includes four steps: update gate calculation, reset gate calculation, candidate hidden state generation, and final hidden state update. The update gate controls the proportion of historical information retained, the reset gate controls the degree of influence of historical information on candidate states, and the final hidden state is a weighted combination of historical and candidate states via the update gate. In this example, the hidden state dimension is set to 128, the time window length is 24 hours, and predictions are performed in hourly units.

[0095] In some implementations, a Long Short-Term Memory (LSTM) network can be used instead of a GRU to achieve stronger long-term dependency modeling capabilities, particularly when dealing with long-term problems such as seasonal variations. Alternatively, an attention-enhanced recurrent neural network, such as the Transformer architecture, can be used to better capture long-range dependencies in time series.

[0096] Step 2.4, flow prediction and carrying capacity assessment;

[0097] Based on the GRU's hidden state, a multilayer perceptron is used to predict node flows in future time steps. The multilayer perceptron maps the GRU's hidden state to a prediction target space and outputs the predicted flow value for the future time step. By comparing the predicted flow with the design capacity, the drainage system's carrying capacity under different conditions is assessed, and potential overflow risks are identified.

[0098] The multilayer perceptron consists of two fully connected layers, with a ReLU activation function and a dropout layer (dropout rate 0.3) added between them to prevent overfitting. The first layer maps the hidden state of the GRU to 64 dimensions, and the second layer outputs the dimension of the predicted target, which is the flow value at the next time step. The model uses mean squared error as the loss function and is trained with the Adam optimizer. The learning rate is initially set to 0.001, and a learning rate decay strategy is used to improve model stability.

[0099] Step 3: Based on the prediction results, an overflow risk assessment model is constructed to identify key influencing factors through the attention mechanism, quantify the node risk and identify high-risk areas;

[0100] Based on the dynamic prediction results, an overflow risk assessment model is constructed, which uses the attention mechanism to identify key influencing factors, quantify node risks, and identify high-risk areas. Specifically, it includes:

[0101] Step 3.1, risk scoring function construction;

[0102] A node risk scoring function is defined, which comprehensively considers the topological importance, predicted traffic and capacity margin of the node.

[0103] Specifically, a node's risk score is composed of three components: node centrality (reflecting the node's topological importance in the network), the ratio of predicted traffic to historical average traffic (reflecting the degree of current traffic anomaly), and capacity margin (the ratio of the difference between the designed capacity and predicted traffic to the designed capacity). These three factors are weighted and combined to form a comprehensive risk score.

[0104] Step 3.2, calculation of key factor weights;

[0105] The attention mechanism adaptively calculates the weight coefficients of each factor, enabling the model to dynamically adjust the importance of each factor based on different nodes and scenarios. The attention mechanism uses learnable parameters to assign weights to different factors based on the importance of the current input features, ensuring that the sum of the weights is 1. In this way, the model can automatically identify the most critical factors for risk assessment, improving the accuracy and interpretability of the assessment.

[0106] Step 3.3, risk hotspot identification;

[0107] Based on the calculated risk scores, a threshold clustering method is used to identify high-risk node clusters and form a risk hotspot map. First, a risk threshold is set, marking nodes with risk scores above the threshold as high-risk nodes. Then, a spatial clustering algorithm, such as DBSCAN (density-based spatial clustering), is applied to group spatially adjacent high-risk nodes into the same risk hotspot area. Finally, based on the clustering results, a risk hotspot distribution map is generated, visually displaying the high-risk areas in the system.

[0108] Step 3.4, risk dynamic tracking;

[0109] Establish a time series model for risk scoring to track the dynamic evolution of risk hotspots. For identified high-risk areas, construct a time series forecasting model, such as the Autoregressive Integrated Moving Average (ARIMA) model, to predict future risk trends and provide a time window for preventative regulation.

[0110] Step 4: Based on the output of the risk assessment model, a fault model based on a causal graph is constructed. Through intervention learning and counterfactual reasoning, the causal chain of the overflow event is identified, achieving a transition from correlation analysis to causal understanding.

[0111] For nodes identified as high-risk or overflowing, a fault model based on a causal graph is constructed, and the causal chain of overflow events is identified through intervention learning and counterfactual reasoning, achieving a transition from correlation analysis to causal understanding. Specifically, it includes:

[0112] Step 4.1, structural causal model construction;

[0113] Based on the physical characteristics of the drainage system and expert knowledge, a structural causal model of the system's state variables is constructed. A structural causal model describes the causal relationships between variables using a set of structural equations, where each variable is determined by its direct causal parent node and unobservable exogenous variables. SCM explicitly defines causal relationships between variables, rather than simple statistical correlations.

[0114] The initial causal diagram construction is based on the following rules:

[0115] Causal relationships directly caused by physical connections, such as the water level at an upstream node affecting the water level at a downstream node;

[0116] Causal relationships caused by equipment control relationships, such as the impact of the pump station's open state on the flow rate of connected pipelines;

[0117] The causal relationship between meteorological conditions and system status, such as rainfall affecting surface runoff and infiltration;

[0118] The causal relationship between urban underlying surface characteristics and rainfall runoff coefficient.

[0119] The initial causal graph is represented by a directed acyclic graph (DAG), where nodes are system state variables, edges represent causal relationships, and edge weights represent causal strengths. It can be initially set based on expert experience and subsequently optimized through data learning.

[0120] Step 4.2, intervention learning;

[0121] Intervention learning estimates causal effects between variables and distinguishes between correlation and causation. Based on a "do-calculus" framework, intervention learning evaluates the direct causal impact of a variable on a target variable by calculating an intervention distribution rather than a conditional distribution, eliminating the interference of indirect associations.

[0122] Intervention learning is implemented using an adjustment set approach: first, based on the causal graph, an adjustment set between variables is identified—the minimum set of variables that blocks all non-causal paths. Conditional and marginal probabilities are then estimated based on the observed data. Finally, a weighted average is calculated to obtain the intervention distribution. For continuous variables, nonparametric estimation methods such as kernel density estimation or Gaussian process regression can be used. For high-dimensional adjustment sets, a dual robust estimation method is used, combining propensity scores and regression models to reduce estimation errors.

[0123] Step 4.3, causal discovery;

[0124] Graph Attention Networks (GATs) are used to automatically discover potential hidden causal relationships in data. Graph Attention Networks (GATs) capture the dependency structure in the data by learning attention weights between nodes, where attention weights can be interpreted as the strength of influence between nodes. By analyzing the learned attention weights, potential causal relationships in the data can be identified, supplementing the initial causal graph constructed based on expert knowledge.

[0125] The computational process of the Graph Attention Network (GAT) includes four steps: attention coefficient calculation, attention weight normalization, neighbor information aggregation, and multi-head attention integration. The GAT is trained to minimize traffic prediction error and then analyzes the learned attention weights. Edges with larger weights are likely to indicate potential causal relationships.

[0126] Step 4.4, counterfactual reasoning;

[0127] Conduct counterfactual analysis of past overflow events to identify key intervention points. Counterfactual reasoning answers the question, "What would have happened if conditions had been different?" By comparing actual observed outcomes with predicted outcomes after a hypothetical intervention, the effectiveness of the intervention can be evaluated.

[0128] The computational process of counterfactual reasoning includes:

[0129] Infer the value of exogenous variables based on observed data and structural causal models;

[0130] A hypothetical intervention is imposed on the target variable, changing its value from the actual observed value to the hypothesized value;

[0131] Calculate the impact of intervention spread on other variables based on the causal graph structure;

[0132] Calculate the predicted value of the target outcome variable under the intervention condition;

[0133] Compare the differences between actual observations and counterfactual predicted values ​​to evaluate the intervention effect.

[0134] For the drainage system, we focus on analyzing the impact of the operating strategies of pipeline control facilities (such as pumping stations and gates) on overflow risks, identifying the key facilities with the most intervention value and the optimal operation time, and providing a basis for the formulation of emergency plans.

[0135] In some implementations, reinforcement learning techniques can be combined to construct a Markov Decision Process (MDP) that includes system states, operational actions, and reward functions. Through Deep Q-Learning or Policy Gradient Methods, optimal control strategies can be learned to achieve real-time intelligent control of drainage systems. Optionally, context-aware mechanisms can be introduced to adaptively adjust control strategies based on different rainfall scenarios and system states, improving system adaptability and robustness.

[0136] Step 5: Based on the comprehensive results of risk assessment and causal analysis, the BIM model is used to optimize the drainage system parameters, and the optimization effect is verified through numerical simulation;

[0137] Based on the results of risk assessment and causal analysis, the BIM model is used to optimize drainage system parameters, including key parameters such as pipe network layout, pipe diameter, slope, pump station capacity, etc., and the optimization effect is verified through numerical simulation. Specifically, it includes:

[0138] Step 5.1, mapping BIM parameters and risk factors;

[0139] Establish a mapping relationship between BIM model parameters and risk factors to identify key parameters that affect system performance. First, extract the geometric parameters, material properties, and functional parameters of the drainage facilities from the BIM model. Then, based on the results of the cause-and-effect analysis, determine the corresponding relationship between these parameters and risk factors. Finally, establish a parameter-risk mapping matrix to guide the subsequent optimization process.

[0140] The mapping relationship is established using the following method:

[0141] Parameter sensitivity analysis based on physical models, such as changing a single parameter while keeping other parameters constant to observe changes in system response;

[0142] Parameter importance assessment based on causal inference, using intervention learning to quantify the causal effects of different parameters;

[0143] Based on statistical analysis of historical data, calculate the correlation between parameter changes and system performance;

[0144] Structured integration of expert knowledge and engineering experience.

[0145] Map each element in the matrix Representation parameters Risk factors The influence intensity of is in the range of [-1, 1]. Positive values ​​indicate positive correlation, negative values ​​indicate negative correlation, and the absolute value indicates the correlation strength.

[0146] Step 5.2, multi-objective optimization model construction;

[0147] Based on the mapping relationship, a multi-objective optimization model considering flood control capacity, construction cost and operation and maintenance efficiency is constructed:

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] in, is the multi-objective function vector; is the decision variable vector, which contains the BIM model parameters that need to be optimized; 、 、 They represent the objective function of flood control capacity, construction cost and operation and maintenance efficiency respectively; For the inequality constraints; For the equality constraints; is the inequality constraint index; is the equality constraint index; is the total number of inequality constraints; is the total number of equality constraints; is the decision variable vector; is the lower limit constraint vector of the parameters; is the upper limit constraint vector of the parameters; represents the minimization optimization objective; Indicates constraints.

[0153] The specific objective function is defined as follows:

[0154] Flood control capacity objective function:

[0155] ;

[0156] in, is the objective function of flood control capacity; is the decision variable vector; Indicates the maximum value operation; is the node index; is the time step index; For nodes In time The risk scoring function.

[0157] That is, maximizing the negative value of the maximum risk score of all nodes in the system at all time steps is equivalent to minimizing the maximum risk;

[0158] Construction cost objective function:

[0159] ;

[0160] in, is the construction cost objective function; is the decision variable vector; represents the summation symbol; is the parameter index; is the total number of parameters; For parameters The corresponding cost function; For the decision variables.

[0161] Operation and maintenance efficiency objective function:

[0162] ;

[0163] in, is the operation and maintenance efficiency objective function; is the decision variable vector; Indicates the index From 1 to sum; represents the summation symbol; For facilities Maintenance cost function; Display facilities and The interaction complexity function between them; For the decision variables; For the decision variables; is the total number of parameters; Indexing facilities; Indexing facilities; Index of facilities.

[0164] Constraints include engineering specification requirements, physical feasibility constraints, budget limitations, etc., such as pipe diameter range, slope requirements, minimum cover depth, maximum flow rate limit, etc.

[0165] Step 5.3, non-dominated sorting genetic algorithm solution;

[0166] An improved non-dominated sorting genetic algorithm (NSGA-III) is used to solve multi-objective optimization problems and obtain Pareto-optimal solutions. This algorithm simulates the natural evolutionary process, selecting individuals with better performance for reproduction in each generation, gradually approaching the Pareto frontier. Compared to the traditional NSGA-II algorithm, NSGA-III introduces a reference point mechanism to better maintain the diversity of the solution set, making it particularly suitable for optimization problems with three or more objectives.

[0167] The implementation steps of NSGA-III include:

[0168] Initialization: Randomly generate an initial population that meets the constraints, with a population size of N;

[0169] Reference point generation: Generate uniformly distributed reference points in the regularized target space;

[0170] Evolutionary Cycle:

[0171] Selection: Use binary tournaments to select parent individuals;

[0172] Mutation and crossover: Apply simulated binary crossover (SBX) and polynomial mutation to generate offspring individuals;

[0173] Evaluation: Calculate the objective function value of the offspring individual;

[0174] Merge: merge the parent and offspring to form an intermediate population;

[0175] Non-dominated sorting: Perform non-dominated sorting on the intermediate population to determine the levels;

[0176] Select the next generation: Starting from the front non-dominated layer, select individuals layer by layer until the population size N is reached; for the critical layer, select based on the reference point and the vertical distance of the individual to the reference line to ensure the diversity of the solution set;

[0177] Termination condition: reaching the maximum number of iterations or no significant improvement after several consecutive generations.

[0178] In this embodiment, the population size is set to 200, the maximum number of iterations is 500, and the reference points are generated in the three-dimensional target space using the Das and Dennis method.

[0179] In some implementations, other multi-objective optimization algorithms, such as Multi-Objective Particle Swarm Optimization (MOPSO), Multi-Objective Differential Evolution (MODE), or Multi-Objective Ant Colony Optimization (MOACO), can be used to adapt to different types of optimization problems. Alternatively, adaptive strategies can be introduced to dynamically adjust algorithm parameters, such as crossover probability and mutation rate, to improve algorithm convergence and diversity.

[0180] Step 5.4, solution verification and implementation;

[0181] Numerical simulations were conducted to verify the performance of multiple optimized design options under extreme conditions, and the optimal solution was selected for implementation. These simulations employed either a one-dimensional hydrodynamic model or a two-dimensional shallow water equation model to simulate the drainage system's response under varying rainfall conditions, verifying the effectiveness and stability of the optimized solution. Based on the simulation results and the decision-makers' preferences, the final implementation plan was determined, and the BIM model was updated to guide project implementation.

[0182] The verification process specifically includes the following steps:

[0183] Solution screening: Select representative solutions from the Pareto optimal solution set, which usually includes the optimal solution of each objective function and some compromise solutions;

[0184] Extreme working condition settings: Construct a variety of extreme working condition scenarios including heavy rain with different recurrence periods, continuous rainfall, local facility failures, etc.

[0185] Numerical simulation: Import the optimization scheme parameters into the hydrodynamic model, perform dynamic simulation, and calculate hydraulic parameters such as water level and flow at each node;

[0186] Performance evaluation: Calculate comprehensive performance indicators of the solution based on simulation results, such as overflow risk, system resilience, and recovery time;

[0187] Scheme ranking: comprehensively consider the performance of each scheme under different working conditions and combine the decision maker's preferences to rank the schemes;

[0188] BIM update: Update the final selected scheme parameters to the BIM model and generate construction drawings and bill of quantities.

[0189] A BIM-based urban drainage network optimization design system, used to implement the above-mentioned BIM-based urban drainage network optimization design method, includes:

[0190] Multi-source data fusion module, used to collect and fuse multi-source hydrological data to generate a comprehensive digital representation of the drainage system;

[0191] The spatiotemporal graph modeling module is used to construct a spatiotemporal graph structure based on comprehensive digital representations and use deep learning models to predict the dynamic carrying capacity of the system;

[0192] Risk quantification module, used to assess overflow risks based on prediction results and identify key influencing factors and high-risk areas;

[0193] Causal analysis module, used to build a fault causal model and identify the causal relationship of overflow events through intervention learning and counterfactual reasoning;

[0194] The optimization and verification module is used to optimize BIM parameters based on risk and causal analysis results and verify the effectiveness of the optimization scheme.

[0195] Here, the present invention provides an implementation example:

[0196] To verify the effectiveness of the proposed BIM-based urban drainage network optimization design method, a drainage system renovation project in a new eastern district of a certain city was used as an example for application testing. This area covers approximately 12 square kilometers, with an existing drainage network totaling approximately 65 kilometers, including over 800 inspection wells, four pumping stations, and over 1,200 stormwater inlets. In recent years, the area has frequently suffered from flooding, particularly during heavy summer rainfall, with overflows at multiple key nodes, severely impacting residents' lives and traffic safety.

[0197] The new district's drainage system was originally designed 20 years ago. With the acceleration of urbanization, the increase in impervious areas, and the frequent occurrence of extreme rainfall events due to climate change, the original drainage system is no longer able to meet current needs. A complete BIM model has been established for the area, containing 3D information on all drainage facilities. It is also equipped with 18 rainfall monitoring stations, 32 flow monitoring points, and 45 water level monitoring points, collecting relevant data in real time. The goal of the renovation is to optimize the drainage system design within a limited budget, improve its ability to cope with extreme rainfall, and reduce the frequency of urban waterlogging disasters.

[0198] The data collected in the application example includes:

[0199] Historical rainfall data: hourly rainfall records for the past five years;

[0200] Terrain data: 1-meter resolution digital elevation model;

[0201] Soil permeability data: infiltration test results at different locations within the region;

[0202] Urbanization data: proportion of impervious area and land use type;

[0203] Drainage facility operation data: real-time monitoring data such as pipe network flow, water level, and pump station operation status.

[0204] During the data cleaning process, approximately 5% of the flow data contained outliers, which were identified and addressed using the modified interquartile range (IQR) method. Approximately 8% of the monitoring data was missing, and these were interpolated using the ARIMA model for time series data and kriging for spatial data. All data were normalized using the Z-score to ensure comparability across different dimensions.

[0205] Spatiotemporal alignment unifies data at different temporal resolutions (5-minute, 15-minute, and 1-hour resolutions, etc.) to the hourly level, and spatial data to a geographic coordinate system using meters. Feature extraction generates a comprehensive feature vector that includes temporal features (such as weekdays / holidays and seasonality), spatial features (such as elevation and slope), and hydrological features (such as rainfall intensity and duration).

[0206] Based on the BIM model, a spatiotemporal graph network consisting of 800 nodes and 1,050 edges was constructed. Nodes corresponded to manholes and pumping stations, while edges corresponded to pipeline connections. Node features included 10 parameters, such as historical flow rate, water level, pipe diameter, and slope. Edge features included five parameters, such as pipe length, diameter, and material.

[0207] The model training uses the most recent four years of historical data (split into training and validation sets at a ratio of 7:3), with the most recent year of data reserved as the test set. The model configuration includes:

[0208] Graph Convolutional Network: 3 graph convolution layers, hidden dimension 64, residual connections;

[0209] GRU (Gated Recurrent Unit, gated recurrent unit): hidden state dimension 128, time window length 24 hours;

[0210] Multilayer Perceptron: 2 fully connected layers, middle layer dimension 64, dropout rate 0.3;

[0211] Training parameters: batch size 32, learning rate 0.001, Adam optimizer, early stopping strategy (stop if validation set loss does not improve for 10 consecutive rounds).

[0212] The model's flow prediction root mean square error (RMSE, RootMeanSquareError) on the test set is 0.023 cubic meters per second, the mean absolute percentage error (MAPE, MeanAbsolutePercentageError) is 6.8%, and the R² value is 0.92, which is better than the traditional physical model (RMSE 0.061 cubic meters per second, MAPE 18.5%, R² value 0.67) and the statistical regression model (RMSE 0.048 cubic meters per second, MAPE 15.2%, R² value 0.73).

[0213] Based on the prediction results, the risk score of each node is calculated, where the node centrality is calculated based on the PageRank algorithm. The ratio of predicted traffic to historical average traffic reflects the degree of current traffic abnormality. The capacity margin is expressed as (design capacity - predicted traffic) / design capacity.

[0214] By applying the attention mechanism to adaptively calculate the weights of each factor, we found that there are obvious differences in the weight distribution under different circumstances:

[0215] Under normal rainfall conditions: centrality weight is 0.35, flow ratio weight is 0.25, and capacity margin weight is 0.40;

[0216] Under extreme rainfall conditions: centrality weight is 0.20, flow ratio weight is 0.45, and capacity margin weight is 0.35.

[0217] Risk assessment results indicate 12 high-risk areas in the region, primarily concentrated in three hotspots: the junction of the old and new urban areas, the central commercial district, and low-lying residential areas. Dynamic risk tracking indicates that risks in the central commercial district primarily occur during daytime hours on weekdays, while risks in low-lying residential areas are more pronounced during nighttime rainfall.

[0218] For the identified high-risk areas, a structural causal model containing 37 key variables was constructed, including rainfall intensity, upstream flow, pipeline congestion level, pump station operating status, etc.

[0219] Application of intervention learning analysis found that the main causes of overflow in the commercial center area were insufficient pipe diameter (intervention effect 0.62) and clogged stormwater inlets (intervention effect 0.53), rather than the previously believed excessive upstream water flow (intervention effect 0.21); overflow in low-lying residential areas was mainly caused by insufficient pump station capacity (intervention effect 0.75) and untimely scheduling (intervention effect 0.58).

[0220] The graph attention network automatically discovered two previously unrecognized causal relationships: the relationship between the underground parking entrance and the blockage of surrounding rainwater inlets (attention weight 0.47), and the relationship between the operation of the commercial plaza fountain system and pipe network pressure fluctuations (attention weight 0.39).

[0221] Counterfactual reasoning analysis showed that if the pumping station was started two hours before rainfall instead of after the rain started, the overflow risk in low-lying residential areas could be reduced by 65%; if the diameter of the key pipe section in the commercial center was increased from 600 mm to 800 mm, the overflow risk could be reduced by 78%.

[0222] Based on the above analysis results, the key parameters that need to be optimized include 12 pipe diameter adjustments, 3 pipeline slope adjustments, 2 pump station capacity expansions and 4 additional rainwater inlets.

[0223] The NSGA-III (Non-dominated Sorting Genetic Algorithm III) algorithm was used for multi-objective optimization, considering flood control capacity, construction cost, and operation and maintenance efficiency. With a population size of 200 and 500 iterations, 31 Pareto optimal solutions were obtained. Five representative solutions were selected for in-depth evaluation:

[0224] Option A: The best flood control capability, but higher cost;

[0225] Option B: lowest cost and moderate flood control capacity;

[0226] Solution C: The highest operation and maintenance efficiency, but higher costs;

[0227] Options D and E: Compromise options that balance the three indicators.

[0228] A one-dimensional hydrodynamic model was used to simulate the five schemes under rainfall conditions with a return period of 5 years, 10 years, 20 years, and 50 years. The results showed that Scheme D could still maintain system stability under a rainfall condition with a return period of 20 years, and its cost was 25% lower than that of Scheme A. Finally, Scheme D was selected for implementation.

[0229] The improvement in prediction accuracy is shown in Table 1:

[0230] Table 1: Comparison of the performance of different prediction methods on the test set

[0231]

[0232] The proposed method improves prediction accuracy by 25.8%, providing more reliable data support for drainage system design. In particular, under extreme rainfall conditions, the prediction error of traditional methods increases dramatically (MAPE reaches 31.6%), while the proposed method maintains high accuracy (MAPE is 11.3%).

[0233] The system performance and economic improvement are shown in Table 2:

[0234] Table 2: Performance comparison between the optimized drainage system and the original system

[0235]

[0236] By accurately identifying system bottlenecks and critical nodes, the optimized system achieved maximum performance improvements with minimal investment. In particular, by rationally configuring pump station start-stop strategies, the system's recovery time after rainfall was shortened, mitigating the impact of prolonged waterlogging on urban functions.

[0237] Practical applications have verified that the method proposed in this invention not only improves technical indicators, but also achieves good results in practical applications, providing an efficient and feasible new method for the optimal design of urban drainage systems.

[0238] like Figures 2 to 6 As shown in the figure, there are respectively a line graph comparing the prediction accuracy of different prediction methods under different rainfall intensities; a radar chart comparing the method of the present invention with the traditional method in multiple key performance indicators; an area graph showing the changes in overflow risk of the system before and after optimization under different rainfall recurrence periods; a scatter plot of the calculation time of the method of the present invention in drainage networks of different scales; and a bar graph showing the performance of different schemes generated by multi-objective optimization in three dimensions: flood control capacity, construction cost, and operation and maintenance efficiency.

[0239] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A method for optimizing the design of urban drainage networks based on a BIM model, characterized in that: include: Collect multi-source hydrological data, pre-process them through spatiotemporal data heterogeneous fusion technology, and generate a comprehensive digital representation of the drainage system; Based on comprehensive digital representation, the drainage system is modeled as a spatiotemporal graph structure. A deep learning model combining a temporal graph convolutional network and a gated recurrent unit is applied to analyze the dynamic response characteristics of the system under different time periods and rainfall conditions, and to predict the carrying capacity of the drainage system. Based on the prediction results, an overflow risk assessment model is constructed to identify key influencing factors through the attention mechanism, quantify the node risk and identify high-risk areas; In the process of building the overflow risk assessment model, a node risk scoring function is defined. The node risk score consists of three parts, including: Node centrality reflects the topological importance of a node in the network; The ratio of predicted traffic to historical average traffic reflects the degree of current traffic anomaly; Capacity margin, that is, the ratio of the difference between the design capacity and the predicted flow to the design capacity; Based on the output of the risk assessment model, a fault model based on a causal graph is constructed. Through intervention learning and counterfactual reasoning, the causal chain of the overflow event is identified, achieving a transition from correlation analysis to causal understanding. Based on the comprehensive results of risk assessment and causal analysis, the BIM model was used to optimize the drainage system parameters, and the optimization effect was verified through numerical simulation.

2. The urban drainage network optimization design method based on the BIM model according to claim 1 is characterized in that: The preprocessing of multi-source hydrological data includes: Clean the raw data through outlier detection and processing, missing value filling and data format unification; The data with different temporal resolutions are unified to the same time step through resampling technology, and the data with different coordinate systems and spatial resolutions are mapped to a unified spatial reference system through geographic coordinate transformation and spatial interpolation methods; The key features reflecting the characteristics of the drainage system are extracted, and the features from different sources are concatenated to form a unified feature vector.

3. The urban drainage network optimization design method based on the BIM model according to claim 1 is characterized in that: The fault model based on the cause-effect diagram includes the following steps: Based on the physical characteristics of the drainage system and expert knowledge, a structural causal model of the system state variables is constructed; Estimating causal effects between variables through intervention learning, distinguishing between correlation and causation; Automatically discover potential hidden causal relationships in data using graph attention networks; Conduct counterfactual analysis of overflow events that have occurred to identify key intervention points.

4. The method for optimizing the design of urban drainage networks based on a BIM model according to claim 1, characterized in that: The optimization of drainage system parameters using the BIM model includes the following steps: Establish a mapping relationship between BIM model parameters and risk factors to identify key parameters that affect system performance; Construct a multi-objective optimization model that considers flood control capacity, construction cost, and operation and maintenance efficiency; The non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem and obtain the Pareto optimal solution set; The performance of the optimization scheme under extreme working conditions is verified through numerical simulation, and the best scheme is selected for implementation.

5. The method for optimizing the design of urban drainage networks based on a BIM model according to claim 4 is characterized in that: The objective function of the multi-objective optimization model includes: The flood control capability objective function represents maximizing the negative value of the maximum risk score of all nodes in the system at all time steps; The construction cost objective function represents the sum of the construction costs of all facilities; The operation and maintenance efficiency objective function represents the sum of the maintenance cost of the facility and the complexity of the interaction between facilities.

6. The urban drainage network optimization design method based on the BIM model according to claim 1 is characterized in that: The deep learning model is trained to predict the carrying capacity of the drainage system using the following steps: The historical data is divided into training set and validation set in a ratio of 7:3; Use training data to train models, including graph convolutional networks, gated recurrent units, and multilayer perceptrons; Evaluate model performance on the validation set and use early stopping strategy to avoid overfitting; The best performing model is saved for prediction.

7. The method for optimizing the design of urban drainage networks based on a BIM model according to claim 1, characterized in that: The calculation process of the counterfactual reasoning includes: Infer the value of exogenous variables based on observed data and structural causal models; A hypothetical intervention is imposed on the target variable, changing its value from the actual observed value to the hypothesized value; Calculate the impact of intervention spread on other variables based on the causal graph structure; Calculate the predicted value of the target outcome variable under the intervention condition; Compare the differences between actual observations and counterfactual predicted values ​​to evaluate the intervention effect.

8. The method for optimizing the design of urban drainage networks based on a BIM model according to claim 1, characterized in that: The process of modeling the drainage system as a spatiotemporal graph structure includes: Extract the spatial coordinates and connection relationships of drainage facilities from the BIM model; Attach time series characteristic data to each node, including historical flow, water level and rainfall; A time-varying adjacency matrix is ​​constructed, and the edge weights are determined based on the physical parameters of the pipe diameter and slope.

9. An urban drainage network optimization design system based on BIM model, characterized in that: A method for optimizing the design of an urban drainage network based on a BIM model according to any one of claims 1 to 8, comprising: Multi-source data fusion module, used to collect and fuse multi-source hydrological data to generate a comprehensive digital representation of the drainage system; The spatiotemporal graph modeling module is used to construct a spatiotemporal graph structure based on comprehensive digital representations and use deep learning models to predict the dynamic carrying capacity of the system; Risk quantification module, used to assess overflow risks based on prediction results and identify key influencing factors and high-risk areas; Causal analysis module, used to build a fault causal model and identify the causal relationship of overflow events through intervention learning and counterfactual reasoning; The optimization and verification module is used to optimize BIM parameters based on risk and causal analysis results and verify the effectiveness of the optimization scheme.

Images