A method for identifying key nodes of a drainage network oriented to prediction promotion

Abstract

This application relates to the general field of urban drainage technology, specifically to a method for identifying key nodes in a drainage network for predictive improvement. The method includes: randomly partitioning nodes in a target dataset of the target drainage network based on a preset target missing rate to obtain a subset of monitored nodes and a subset of unmonitored nodes; obtaining hydraulic state data of nodes in the monitored node subset from the target dataset, and inputting the hydraulic state data into a trained hydraulic prediction model to obtain predicted hydraulic state data of nodes in the unmonitored node subset; calculating the contribution of nodes based on the predicted hydraulic state data, and determining the global importance of nodes in the target drainage network based on the contribution; and selecting key nodes of the target drainage network at the target missing rate from the target dataset based on the global importance of nodes in the target drainage network.

Description

Technical Field

[0001] This application relates to the field of urban drainage technology, and in particular to a method for identifying key nodes in drainage networks for predictive improvement. Background Technology

[0002] Under the dual pressures of global climate change and urbanization, urban flooding is becoming increasingly frequent, posing a severe challenge to the operational efficiency of urban drainage networks. As the lifeline of urban flood control and drainage, accurate monitoring and prediction of the operational status of drainage networks are crucial for ensuring urban safety. However, due to the high cost of deploying monitoring sensors and the difficulty of maintenance, real-world monitoring networks often exhibit characteristics of being "numerous and widespread but highly sparse," failing to achieve full coverage of the network's status and creating numerous monitoring blind spots.

[0003] Currently, the technical approaches used in the industry to solve pipeline network status prediction include the use of data-driven machine learning methods, which have the potential for rapid prediction. However, their performance is highly dependent on the quality and representativeness of the input data. If real-time monitoring data for individual nodes in the pipeline network is missing, the model will struggle to accurately infer the overall status of the network, significantly reducing prediction accuracy. Therefore, a better solution for identifying key nodes in drainage pipeline networks, geared towards improved prediction, is needed to overcome these difficulties. Summary of the Invention

[0004] This application provides a method for identifying key nodes in drainage pipe networks for predictive improvement, in order to improve the accuracy of pipe network status prediction.

[0005] In a first aspect, this application provides a method for identifying key nodes in a drainage network for predictive improvement, the method comprising: Based on the preset target missing rate, the nodes in the target data of the target drainage network are randomly divided to obtain a subset of monitored nodes and a subset of unmonitored nodes. The hydraulic state data of the nodes in the monitoring node subset are obtained from the target dataset, and the hydraulic state data are input into the trained hydraulic prediction model to obtain the predicted hydraulic state data of the nodes in the unmonitored node subset. Based on the predicted hydraulic state data, the contribution of each node in the monitoring node subset to any node in the unmonitored node subset is calculated, and based on the contribution, the global importance of each node in the monitoring node subset in the target drainage network is determined. Based on the global importance of nodes in the target drainage network, key nodes of the target drainage network under the target missing rate are selected from the target dataset, so as to predict the global state of the target drainage network based on the key nodes.

[0006] Furthermore, the hydraulic prediction model is trained using structured tabular data, and the hydraulic prediction model is an XGBoost model.

[0007] Furthermore, calculating the contribution of nodes in the monitoring node subset to any node in the unmonitored node subset based on the predicted hydraulic state data includes: Based on the predicted hydraulic state data of the nodes in the unmonitored node subset at each time step, the mean hydraulic state of the unmonitored node subset at each time step is determined. Based on the tree model interpreter that matches the XGBoost model, the predicted hydraulic state data of the nodes in the unmonitored node subset and the mean hydraulic state of the unmonitored node subset at each time step are used to obtain the SHAP value matrix of the unmonitored nodes. The values ​​in the SHAP value matrix represent the contribution of the nodes in the monitored node subset to the nodes in the unmonitored node subset at each time step.

[0008] Furthermore, determining the global importance of nodes in the monitoring node subset within the target drainage network based on the contribution includes: For any node in the subset of monitoring nodes, the absolute value of the contribution is processed to obtain the processed contribution. For any node in the subset of monitoring nodes, the average of the processed contribution is taken to obtain the global importance of the node in the target drainage network.

[0009] Furthermore, given the existence of an engineering budget and a predetermined number of sensors for the target drainage network, the step of filtering out key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network, under the target missing rate, includes: Based on the global importance of the nodes in the target drainage network, the nodes contained in the target drainage network are sorted to obtain the sorting result; From the sorting results, the key nodes of the target drainage network under the target missing rate are determined, and the target missing rate is determined based on the project budget and the preset number of sensors.

[0010] Furthermore, when the target drainage network has a contribution threshold, the step of filtering out key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network at the target missing rate includes: For each node in the target drainage network, the corresponding contribution is summed to obtain the sum of the node's contribution. Based on the contribution of the nodes in the target drainage network, the nodes in the target drainage network are sorted in descending order to obtain the sorting result; For the sorting results, if the sum of the contributions of the top k-1 nodes in the target drainage network is less than a preset value, and the sum of the contributions of the top k nodes in the target drainage network is greater than or equal to a preset value, then the top k nodes in the target drainage network are regarded as the key nodes of the target drainage network under the target missing rate.

[0011] Furthermore, the training process of the hydraulic prediction model includes: Obtain a training dataset of all nodes and multiple operating conditions of the sample drainage network, wherein the training dataset contains time series data of the hydraulic state of the nodes; Based on different missing rates, the nodes in the sample drainage network are randomly divided to obtain a subset of monitored nodes and a subset of unmonitored nodes. Hydraulic state data for the monitored node sample subset and the unmonitored node sample subset are determined from the hydraulic state training dataset, respectively. The hydraulic prediction model is trained based on the hydraulic state data of the sample subset of the monitored nodes and the sample subset of the unmonitored nodes to obtain the trained hydraulic prediction model.

[0012] Furthermore, the method for identifying key nodes in a drainage network for predictive improvement also includes: If the target drainage network is an already operational drainage network, an optimized layout scheme for the target drainage network is constructed based on the key nodes. The optimized layout scheme includes at least the attribute data of the key nodes and the sorting results of the target drainage network. In the case where the target drainage network is an unused drainage network, a monitoring and planning scheme for the target drainage network is constructed based on the key nodes. The optimized layout scheme includes at least the attribute data of the key nodes and the sorting results of the target drainage network.

[0013] Secondly, this application provides a key node identification device for drainage pipe networks oriented towards predictive improvement, the key node identification device for drainage pipe networks oriented towards predictive improvement includes: The partitioning module is used to randomly partition the nodes in the target dataset of the target drainage network based on a preset target missing rate, so as to obtain a subset of monitored nodes and a subset of unmonitored nodes. The prediction module is used to obtain the hydraulic state data of the nodes in the monitoring node subset from the target dataset, and input the hydraulic state data into the trained hydraulic prediction model to obtain the predicted hydraulic state data of the nodes in the unmonitored node subset. The calculation module is used to calculate the contribution of the nodes in the monitoring node subset to any node in the unmonitored node subset based on the predicted hydraulic state data, and to determine the global importance of the nodes in the monitoring node subset in the target drainage network based on the contribution. The filtering module is used to filter out key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network, under the target missing rate, so as to predict the global state of the target drainage network based on the key nodes.

[0014] Thirdly, embodiments of this application also provide a key node identification system for drainage pipe networks for predictive improvement. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the methods described above.

[0015] Fourthly, this application provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the above-described method. The computer-readable storage medium may be volatile or non-volatile.

[0016] Fifthly, this application provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement the above-described method when executing instructions stored in the memory.

[0017] Sixthly, this application provides a computer program product including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in a processor of an electronic device, the processor in the electronic device performs the above-described method.

[0018] In the embodiments of this specification, a subset of monitored nodes and a subset of unmonitored nodes are obtained through random partitioning. The hydraulic state data of the former is input into a hydraulic prediction model to obtain the predicted hydraulic state data of the latter. Then, the global importance of the nodes in the former is determined based on the predicted hydraulic state data, and the key nodes in the target drainage network are identified through global importance. This process proposes the concept of key nodes in drainage networks and a method for their identification. Since key nodes are indispensable to the target drainage network under the target missing rate, the overall state of the network can be accurately inferred by identifying key nodes. This has crucial theoretical value and practical significance for the network layout of drainage networks, improving the accuracy of hydraulic state prediction under sparse data, and realizing economical and efficient intelligent drainage management. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 The example illustrates a flowchart of a method for identifying key nodes in a drainage network for predictive improvement. Figure 2 An example of an interpretability analysis graph for node P720 is shown in the figure.

[0021] Figure 3 An example of a SHAP feature importance analysis graph (for target dimension 94) is shown in the figure.

[0022] Figure 4 An example is shown in the overall SHAP interpretability analysis diagram of a study area.

[0023] Figure 5 The diagram above illustrates a process flow diagram of a key node identification device for drainage pipe networks aimed at predictive improvement.

[0024] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0025] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0026] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0027] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A exists alone, A and B exist simultaneously, and B exists alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0028] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0029] Figure 1 A flowchart illustrating a method for identifying key nodes in a drainage network for predictive improvement according to an embodiment of this disclosure is provided. This method can be applied to a device for identifying key nodes in a drainage network for predictive improvement. The device can be a terminal device, a server, or other processing equipment. The terminal device can be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc.

[0030] In some possible implementations, this method for identifying key nodes in a predictive improvement-oriented drainage network can be implemented by a processor calling computer-readable instructions stored in memory.

[0031] like Figure 1 As shown, the method for identifying key nodes in a drainage network for predictive improvement may include: Step S11: Based on the preset target missing rate, the nodes in the target data of the target drainage network are randomly divided to obtain a subset of monitored nodes and a subset of unmonitored nodes.

[0032] The missing rate represents the proportion of nodes in the drainage network without monitoring sensors. Nodes equipped with monitoring sensors are designated as monitored nodes, while those without are designated as unmonitored nodes. Nodes in the drainage network can be strategically placed based on experience or evenly distributed throughout the network, such as at preset intervals. The type of sensor can be determined based on pre-defined hydraulic condition monitoring requirements, which may include water depth, flow rate, flow velocity, oxygen content, etc. The sensor can collect hydraulic conditions in situ or in real-time, with a sampling frequency of 3-60 times per minute.

[0033] To better illustrate this application and highlight its main points, the specific embodiments described in this specification are based on urban drainage networks. Those skilled in the art should understand that this specification can also be applied to other drainage networks, such as those in ship cabins and aircraft.

[0034] The target missing rate is the missing rate corresponding to the target drainage network. This specification does not specify a particular value for the target missing rate; it can be determined based on actual circumstances. For example, the target missing rate can be set between 5% and 95%. After obtaining the target missing rate, the total number of monitoring nodes can be determined based on the number of nodes in the target drainage network. Then, using a fixed random seed, monitoring nodes are randomly selected from the nodes of the drainage network using this total number of nodes. The set corresponding to the monitoring nodes is the subset of monitoring nodes, and the set corresponding to the remaining nodes in the target dataset is the subset of non-monitoring nodes.

[0035] The target dataset mentioned above may contain attribute data (such as location, number, etc.) and hydraulic state data of nodes in the target drainage network. Similar to the data source in the subsequent training dataset, the relevant data of the nodes in the target dataset can be obtained through simulation using a Storm Water Management Model (SWMM).

[0036] The random partitioning process in step S11 can be performed multiple times, ensuring that every node in the target dataset has a chance to be used as a monitoring node. For each partitioning result, subsequent steps S12-S14 are executed.

[0037] Step S12: Obtain the hydraulic state data of the nodes in the monitoring node subset from the target dataset, and input the hydraulic state data into the trained hydraulic prediction model to obtain the predicted hydraulic state data of the nodes in the unmonitored node subset.

[0038] The hydraulic prediction model is used to predict the hydraulic state data of the remaining nodes using hydraulic state data from a subset of nodes in the pipeline network. After determining the subset of monitoring nodes, the hydraulic state data of these monitoring nodes can be identified from the target dataset. Inputting this hydraulic state data into the hydraulic prediction model yields the hydraulic state data of the unmonitored nodes.

[0039] The urban drainage network stormwater management model (SWMM) can generate simulated data in batches by programmatically and automatically configuring multiple rainfall scenarios (Chicago rain patterns with different return periods and peak coefficients). As mentioned above, the generated data can not only be used to obtain hydraulic state data in the target dataset, but also to generate a training dataset for training the hydraulic prediction model. In one possible implementation, the training process of the hydraulic prediction model includes: Obtain a training dataset of all nodes and multiple operating conditions of the sample drainage network, wherein the training dataset contains time series data of the hydraulic state of the nodes; Based on different missing rates, the nodes in the sample drainage network are randomly divided to obtain a subset of monitored nodes and a subset of unmonitored nodes. Hydraulic state data for the monitored node sample subset and the unmonitored node sample subset are determined from the hydraulic state training dataset, respectively. The hydraulic prediction model is trained based on the hydraulic state data of the sample subset of the monitored nodes and the sample subset of the unmonitored nodes to obtain the trained hydraulic prediction model.

[0040] The sample drainage network is a simulated drainage network used to generate training data. It can be the same as or similar to the target drainage network, or it can be completely different from the target drainage network.

[0041] The core of the training process for the hydraulic prediction model is to train a high-performance hydraulic prediction model that can adapt to different monitoring sparsity (i.e. missing rate) using a high-fidelity training dataset of hydraulic states, so as to provide a reliable prediction benchmark for the subsequent identification of key nodes.

[0042] After obtaining the training dataset, it can be processed to obtain gradient training datasets with different missing data rates (also known as mask ratios). In one example, the missing data rate of the nodes corresponding to the sample drainage network can be set to a range, for example, from 0.05 to 0.95, with a step size of 0.05. This creates 19 different sparsity levels, corresponding to monitoring node coverage rates from 95% to 5%. At each missing rate, using a fixed random seed, all nodes in the sample drainage network are randomly divided into a subset of monitored nodes (simulated monitored nodes) and a subset of unmonitored nodes (simulated unmonitored nodes).

[0043] To enhance the generalization ability of the hydraulic prediction model to diverse node-missing scenarios, repeated experiments can be conducted using different random seeds under the same mask ratio. The construction of this gradient training dataset is not for simple data augmentation, but to systematically and comprehensively simulate various monitoring network layouts that may occur in the real world (i.e., combinations of monitoring points with different numbers and locations), and then evaluate the network information value of nodes under different monitoring coverage. This allows for the training of a surrogate model that can stably evaluate the inference of the hydraulic state of the entire network from any number of monitoring nodes under any sparse conditions.

[0044] After constructing the dataset, the data needs to be structured to obtain data directly usable by the hydraulic prediction model. Specifically, the SWMM data interface library (such as swmm_api) can be used to accurately extract the data. The process can iterate through the .out files of all simulation scenarios, thus extracting the hydraulic state time series data corresponding to the sample subsets of monitored nodes and the sample subsets of unmonitored nodes respectively. Table 1 shows the original hydraulic state (water depth) time series data. In Tables 1-3, P represents a node (point).

[0045] Table 1 - Example of raw time series data exported from SWMM

[0046] After integrating the extracted data into a structured two-dimensional dataset, the input monitored node matrix (X) and the unmonitored node matrix (Y) can be constructed scene-by-scene (row concatenation) and node-by-node (column concatenation) methods, respectively. The former corresponds to a subset of monitored node samples, and the latter corresponds to a subset of unmonitored node samples. Exporting both matrices to a standard format file (such as CSV) allows for direct use in training subsequent machine learning models (i.e., hydraulic prediction models). This training process can be achieved by aligning the predicted hydraulic data of the unmonitored nodes output by the hydraulic prediction model with the hydraulic data in the unmonitored node matrix. Tables 2 and 3 show the monitored node matrix and the unmonitored node matrix, respectively.

[0047] Table 2 - Example of Monitoring Node Matrix (X)

[0048] Table 3 - Example of Unmonitored Node Matrix (Y)

[0049] The hydraulic prediction model can be a machine learning model. As mentioned earlier, the data used to train the hydraulic prediction model can be the structured tabular data shown in Tables 2 and 3. Since the XGBoost decision tree model is an advanced implementation of gradient boosting decision trees, in one possible implementation, the hydraulic prediction model is an XGBoost model. When processing the structured tabular data generated by this invention, the XGBoost model generally has higher prediction accuracy and stronger generalization ability than models such as random forests. This is crucial for accurately quantifying the contribution of each node in subsequent steps, as an inaccurate baseline model may lead to distortion in subsequent contribution evaluations.

[0050] In one example, XGBoost can be repeatedly trained using the aforementioned 19 or more sparsity levels to meet the needs of large-scale simulation experiments. The efficient parallel processing capability of the XGBoost model can significantly shorten the model training cycle. Through sufficient training, the XGBoost model can learn a nonlinear mapping relationship from the hydraulic state (input) of sparsely monitored nodes to the hydraulic state (output) of other unmonitored nodes in the entire network.

[0051] In the aforementioned process, the hydraulic prediction model is systematically trained to adapt to different missing rates, thereby enabling it to efficiently, accurately, and reliably invert and predict the complex nonlinear hydraulic state of all nodes in the entire drainage network based on limited monitoring point data. This process overcomes the limitations of other machine learning algorithms (such as random forests) when handling highly sparse data and avoids the disadvantages of deep learning models (such as Crossformer or GRU) in terms of model complexity, training cost, and feature interpretability. This allows for a high-precision understanding of the entire network's hydraulic state even under the realistic condition of sparsely deployed monitoring sensors.

[0052] Evaluation metrics used during training may include the coefficient of determination (R²), mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE).

[0053] The evaluation of the hydraulic prediction model during training can be divided into two levels: first, evaluating the overall prediction performance of all monitored nodes; and second, evaluating the prediction performance of each individual unmonitored node. Optimizing the model parameters based on the evaluation metrics can bring the model performance to its optimal state. Furthermore, all evaluation results can be structured and saved, with labels such as the missing data rate of pipeline nodes and model type added, to facilitate systematic analysis of the impact of the missing data rate on the model's prediction accuracy.

[0054] Step S13: Based on the predicted hydraulic state data, calculate the contribution of the nodes in the monitoring node subset to any node in the unmonitored node subset, and determine the global importance of the nodes in the monitoring node subset in the target drainage network based on the contribution.

[0055] The contribution factor indicates the strength of the influence of a monitoring node's hydraulic state change on the hydraulic state of other unmonitored nodes in the drainage network. The global importance factor indicates the significance of any node in the monitoring process of the drainage network.

[0056] In one example, the subsequent determination of key nodes can employ a hierarchical approach. Specifically, the contribution of a monitoring node to an unmonitored node can be determined first based on the predicted hydraulic state data; then, the global importance of a monitoring node to the entire target drainage network can be determined based on its contribution. For example, the global importance can be obtained by summing the contributions corresponding to a monitoring node. The detailed process of determining this hierarchical key node can be found in subsequent embodiments and will not be elaborated here. It should be understood that for any monitoring node in the drainage network, one time step corresponds to one hydraulic state data point; for any unmonitored node in the drainage network, one time step corresponds to one predicted hydraulic state data point.

[0057] The global importance in step S13 enables this application to quantify and prioritize the information value of each monitoring point in the drainage network from a more macroscopic perspective, thereby guiding the optimized layout of the drainage network.

[0058] Step S14: Based on the global importance of nodes in the target drainage network, key nodes of the target drainage network under the target missing rate are selected from the target dataset, so as to predict the global state of the target drainage network based on the key nodes.

[0059] Among them, the key nodes are those required to infer the overall hydraulic state of the drainage network (including the target drainage network and the sample drainage network) under a certain missing rate.

[0060] Specifically, through a single random partition, the corresponding monitoring nodes can be obtained, and then the global importance of these monitoring nodes can be determined through steps S12-S13. In this way, through multiple random partitions, the global importance of all nodes in the target drainage network can be obtained. Furthermore, by ranking the global importance of the nodes in the target drainage network, the required key nodes can be determined based on the ranking results. The number of key nodes can be determined according to the number of nodes in the target drainage network and the target missing rate. Clearly, the global importance of key nodes is greater than that of other nodes.

[0061] In the embodiments of this specification, a subset of monitored nodes and a subset of unmonitored nodes are obtained through random partitioning. The hydraulic state data of the former is input into a hydraulic prediction model to obtain the predicted hydraulic state data of the latter. Then, the global importance of the nodes in the former is determined based on the predicted hydraulic state data, and the key nodes in the target drainage network are identified through global importance. This process proposes the concept of key nodes in drainage networks and a method for their identification. Since key nodes are indispensable to the target drainage network under the target missing rate, the overall state of the network can be accurately inferred by identifying key nodes. This has crucial theoretical value and practical significance for the network layout of drainage networks, improving the accuracy of hydraulic state prediction under sparse data, and realizing economical and efficient intelligent drainage management.

[0062] The aforementioned contribution can be determined using interpretability tools. Besides the advantages of the XGBoost model (strong predictive performance and robustness (strong generalization ability)), another important advantage is its efficient and seamless integration with a specially optimized tree model interpreter (corresponding to the tree-based Shapley value interpretation method (TreeSHAP)). This makes the subsequent calculation of the SHAP value for each node fast and accurate, avoiding the huge computational overhead of model-agnostic interpretation methods (such as KernelSHAP). In one possible implementation, calculating the contribution of nodes in the monitored node subset to any node in the unmonitored node subset based on the predicted hydraulic state data includes: Based on the predicted hydraulic state data of the nodes in the unmonitored node subset at each time step, the mean hydraulic state of the unmonitored node subset at each time step is determined. Based on the tree model interpreter that matches the XGBoost model, the predicted hydraulic state data of the nodes in the unmonitored node subset and the mean hydraulic state of the unmonitored node subset at each time step are used to obtain the SHAP value matrix of the unmonitored nodes. The values ​​in the SHAP value matrix represent the contribution of the nodes in the monitored node subset to the nodes in the unmonitored node subset at each time step.

[0063] Among them, the tree model interpreter is a TreeExplainer specifically optimized for efficient tree models such as XGBoost. It can accurately calculate the contribution of each input feature (i.e., the hydraulic state data of the simulated monitoring nodes) in the test set to the predicted value (i.e., the hydraulic state data of the simulated unmonitored nodes), thereby generating a multidimensional Shapley Additive exPlanations (SHAP) value matrix. Based on the Shapley value theory in cooperative game theory, the SHAP value matrix can decompose the model's prediction results into the sum of the contributions of each monitoring node to the baseline prediction. As shown in Equation 1, for any sample x Predicted value at the current time step f ( x All of them can be precisely decomposed: (Formula 1) in, f ( x ) is the predicted hydraulic state data (such as water depth) of unmonitored nodes at the current time step output by the hydraulic prediction model. E [ f ( X [)] represents the hydraulic state data of the hydraulic prediction model at all time steps (number of steps X, x∈X). f ( X The mean of () can be used as a baseline or desired output. M This indicates the total number of monitoring nodes input into the hydraulic prediction model. i ( x ) is the core quantitative indicator of this manual, representing the first i monitoring nodes (e.g.) Figure 2 The current hydraulic state data (page 749) for a certain unmonitored node (e.g., Figure 2 Predicted hydraulic state data from P720 (China) E [ f ( X Push towards f ( x The contribution made by ∑M i=1 i (x ) represents the SHAP value matrix. Where, Figure 2 An example of an interpretability analysis graph for SHAP is shown in the figure.

[0064] TreeExplainer, by leveraging the tree structure of the XGBoost model, can efficiently compute these... i ( x This avoids the combinatorial explosion problem of traditional Shapley value calculation. Clearly, XGBoost in this specification cannot be arbitrarily replaced, as it is the best vehicle for achieving the objective of "evaluating node contribution" in this specification, considering its performance, efficiency, and native compatibility with tree model interpreters.

[0065] The above process generates a set of SHAP values ​​for each unmonitored node (e.g., the water depth at P720), corresponding to the contributions of all monitored nodes (e.g., P749). For example, when predicting the water depth at node P720, the SHAP contribution values ​​of all other monitored nodes (e.g., P749) are analyzed. Figure 2 and Figure 3 As shown in the figure, these SHAP values ​​form the data foundation for key node identification in this specification. Figure 3 An example of a SHAP feature importance analysis graph (for target dimension 94) is shown. Furthermore, it can also form... Figure 4 The overall SHAP interpretability analysis diagram of the study area.

[0066] The aforementioned contribution calculation process quantifies the global information value (i.e., global importance) of each monitoring node in inferring the hydraulic state of the entire pipe network by performing SHAP analysis on the trained XGBoost model. This process, by introducing Model Interpretability Technique (SHAP) into the optimized layout of drainage pipe network monitoring points, achieves a refined and additive interpretation of the contribution of individual nodes in a subset of monitoring nodes at each time step. This allows for a more intuitive understanding of the correlation strength and influence direction of hydraulic state data between different monitoring nodes, penetrating the complexity of the model. Consequently, it provides direct and quantifiable engineering guidance and decision-making basis for the optimized layout of monitoring points in urban drainage pipe networks.

[0067] To elevate the understanding from individual predictions to the global model behavior, this specification further calculates the global importance of each monitoring node based on its contribution. In one possible implementation, determining the global importance of nodes in the subset of monitoring nodes within the target drainage network based on their contribution includes: For any node in the subset of monitoring nodes, the absolute value of the contribution is processed to obtain the processed contribution. For any node in the subset of monitoring nodes, the average of the processed contribution is taken to obtain the global importance of the node in the target drainage network.

[0068] Specifically, this can be achieved by averaging the absolute values ​​of the SHAP values ​​(i.e., contribution) of each monitored node (such as P749 and all other monitored nodes in the network) across all time steps and across all unmonitored nodes (such as P720, P751, and all other unmonitored nodes in the network). In one example, global importance can be determined using Formula 2: (Formula 2) Among them, Global Importance j It is every monitoring node j The global importance, where N is the number of time steps. K This is the total number of unmonitored nodes. j,n,k Indicates monitoring node j At time step n For undetected nodes k Predicted SHAP value (contribution).

[0069] As shown in Formula 2, by taking the absolute value of the SHAP value and averaging it, the comprehensive information value of the monitoring node for inferring the hydraulic state of the entire target drainage network can be quantified. This process does not need to distinguish whether the contribution is positive or negative, because global importance represents the magnitude of the overall influence of the monitoring node on the model prediction.

[0070] After obtaining the global importance of nodes in the target drainage network, a ranking and visualization analysis of the global importance of nodes can be performed. Specifically, all monitoring nodes can be sorted in descending order based on the calculated global importance. Subsequently, intuitive visualization charts can be generated, such as SHAP Summary Plots and Feature Importance Bar Plots (e.g., ...). Figure 3 As shown in the diagrams, these charts clearly demonstrate the relative importance of each node to the predictions of the hydraulic prediction model. These charts not only provide ranking information but also assist in analyzing the distribution of importance among different nodes and the specific ways in which they influence predictions (e.g., the SHAP summary diagram can show how the magnitude of the eigenvalue affects the sign of the Shapley value).

[0071] After ranking the global importance of nodes in the target drainage network, the final implementation stage can proceed, namely, adaptive screening and engineering output of key nodes. This specification provides two screening rules adapted to different engineering needs, based on a descending order of node global importance, balancing budget constraints and monitoring benefits. In one possible implementation, given an engineering budget and a preset number of sensors for the target drainage network, the step of screening key nodes of the target drainage network at the target missing rate from the target dataset based on the global importance of nodes in the target drainage network includes: Based on the global importance of the nodes in the target drainage network, the nodes contained in the target drainage network are sorted to obtain the sorting result; From the sorting results, the key nodes of the target drainage network under the target missing rate are determined, and the target missing rate is determined based on the project budget and the preset number of sensors.

[0072] Specifically, nodes in the target drainage network can first be sorted in descending order based on their global importance. Then, considering the project deployment budget and sensor quantity limitations, the missing rate can be determined. The top m nodes from the sorted results can then be selected as critical nodes based on this missing rate. The global importance of the critical nodes in the target drainage network is greater than that of the other nodes in the network. This method is suitable for new pipeline monitoring planning scenarios with a clear budget.

[0073] In another possible implementation, if a contribution threshold exists for the target drainage network, the step of filtering key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network, under the target missing rate, includes: For each node in the target drainage network, the corresponding contribution is summed to obtain the sum of the node's contribution. Based on the contribution of the nodes in the target drainage network, the nodes in the target drainage network are sorted in descending order to obtain the sorting result; For the sorting results, if the sum of the contributions of the top k-1 nodes in the target drainage network is less than a preset value, and the sum of the contributions of the top k nodes in the target drainage network is greater than or equal to a preset value, then the top k nodes in the target drainage network are regarded as the key nodes of the target drainage network under the target missing rate.

[0074] Specifically, this can be achieved by accumulating the contribution (i.e., the sum of the contributions of a subset of nodes in the target drainage network, ∑mt=1I). (t) The sum of the contributions of nodes in the target drainage network, ∑M i=1I i The ratio of the two factors determines the set of key nodes with the minimum number of nodes. In one example, Equation 3 can be used to calculate the cumulative contribution: (Formula 3) Where m represents the number of critical nodes. M I represents the total number of monitoring nodes input into the hydraulic prediction model. (t) Represents a node t Contribution I i Represents a node i The contribution level, when the cumulative contribution level C m When the preset threshold (e.g., 90%) is reached for the first time, the first m nodes are the set of key nodes that meet the information requirements of the whole domain monitoring with the fewest number of nodes. This can achieve the coverage of most pipeline status information with the fewest number of sensors, and is suitable for the quality improvement and optimization of existing pipeline monitoring networks.

[0075] Cumulative contribution C m The criteria for determining whether the preset threshold is reached for the first time can be: in the descending sorting results, the cumulative contribution of the top m-1 nodes is less than the preset threshold, and the cumulative contribution of the top m nodes is greater than or equal to the preset threshold; or the sum of the contributions of the nodes in the top k-1 target drainage network is less than a preset value (this preset value can be determined based on the preset threshold and ∑M i=1 I i The product is determined), and the sum of the contributions of the top k nodes in the target drainage network is greater than or equal to a preset value.

[0076] The above process, through sorting, can scientifically identify key nodes with the highest information value, thereby providing direct and quantifiable engineering guidance and decision-making basis for optimizing the layout of monitoring points in urban drainage pipe networks. It solves the core pain points of traditional methods, such as monitoring point layout relying on experience, lack of quantitative basis, and low monitoring efficiency, and provides direct engineering guidance for optimizing the layout of drainage pipe network monitoring.

[0077] Once the key nodes are identified, engineering applications can be implemented. These key nodes can directly support at least two types of engineering scenarios. In one possible implementation, the method for identifying key nodes in drainage pipe networks for predictive improvement further includes: If the target drainage network is an already operational drainage network, an optimized layout scheme for the target drainage network is constructed based on the key nodes. The optimized layout scheme includes at least the attribute data of the key nodes and the sorting results of the target drainage network. In the case where the target drainage network is an unused drainage network, a monitoring and planning scheme for the target drainage network is constructed based on the key nodes. The optimized layout scheme includes at least the attribute data of the key nodes and the sorting results of the target drainage network.

[0078] The attribute data can include node IDs, locations, and other data. The ranking results can include structured information such as the global importance score, ranking position, and / or cumulative contribution percentage of each node. This attribute data and ranking results can form a standardized list of recommended key monitoring nodes, along with a corresponding full-node importance ranking table and visual analysis report.

[0079] Specifically, there are two main engineering scenarios: One approach is to optimize the operation and maintenance of existing pipe networks (i.e., drainage pipe networks already in use). After receiving the optimized layout plan, key inspections, maintenance, and calibrations of the deployed sensors at critical nodes in the optimized layout plan can be carried out to ensure the reliability and stability of core data sources and maximize the information efficiency of the existing monitoring network.

[0080] Another approach is the planning and construction of new smart drainage networks (i.e., drainage networks that are not yet in use). After receiving the monitoring plan, sensors can be deployed at key nodes in the monitoring plan to achieve a scientific layout of the monitoring network from the source. This allows for accurate perception and prediction of the hydraulic status of the entire network at the lowest possible deployment cost, which perfectly aligns with the core invention objective of this invention: "Identification of key nodes in drainage networks for prediction improvement".

[0081] In the above process, by deploying sensors at key nodes, the monitoring benefits can be maximized at the lowest cost, providing high-quality data input for subsequent early warning and prediction models, thereby accurately inferring the overall status of the pipeline network and improving the accuracy of early warning and prediction.

[0082] The present invention also provides a key node identification device for drainage pipe network for predictive improvement. Figure 5 A block diagram of a drainage network key node identification device for predictive improvement according to an embodiment of this specification is shown. This device can be a terminal device, a server, or other processing equipment. The terminal device can be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, vehicle-mounted device, wearable device, etc.

[0083] In some possible implementations, this predictive improvement-oriented key node identification device for drainage networks can be implemented by a processor calling computer-readable instructions stored in memory.

[0084] like Figure 5 As shown, the drainage network key node identification device 50 for predictive improvement may include: The partitioning module 51 is used to randomly partition the nodes in the target data of the target drainage network based on a preset target missing rate, so as to obtain a subset of monitored nodes and a subset of unmonitored nodes. Prediction module 52 is used to obtain hydraulic state data of nodes in the monitoring node subset from the target dataset, and input the hydraulic state data into the trained hydraulic prediction model to obtain predicted hydraulic state data of nodes in the unmonitored node subset. The calculation module 53 is used to calculate the contribution of the nodes in the monitoring node subset to any node in the unmonitored node subset based on the predicted hydraulic state data, and to determine the global importance of the nodes in the monitoring node subset in the target drainage network based on the contribution. The filtering module 54 is used to filter out key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network, under the target missing rate, so as to predict the global state of the target drainage network based on the key nodes.

[0085] This application also provides a key node identification system for drainage pipe networks for predictive improvement. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any of the methods described above.

[0086] This invention is now complete.

[0087] In summary, in the embodiments of this specification, a subset of monitored nodes and a subset of unmonitored nodes are obtained through random partitioning. The hydraulic state data of the former is input into the hydraulic prediction model to obtain the predicted hydraulic state data of the latter. Then, the global importance of the nodes in the former is determined based on the predicted hydraulic state data, and the key nodes in the target drainage network are identified through global importance. This process proposes the concept of key nodes in drainage networks and a method for their identification. Since key nodes are indispensable to the target drainage network under the target missing rate, the overall state of the network can be accurately inferred by identifying key nodes. This has crucial theoretical value and practical significance for the network layout of drainage networks, improving the accuracy of hydraulic state prediction under sparse data, and realizing economical and efficient intelligent drainage management.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying key nodes in a drainage network for predictive improvement, characterized in that, The method for identifying key nodes in a drainage network for predictive improvement includes: Based on the preset target missing rate, the nodes in the target data of the target drainage network are randomly divided to obtain a subset of monitored nodes and a subset of unmonitored nodes. The hydraulic state data of the nodes in the monitoring node subset are obtained from the target dataset, and the hydraulic state data are input into the trained hydraulic prediction model to obtain the predicted hydraulic state data of the nodes in the unmonitored node subset. Based on the predicted hydraulic state data, the contribution of each node in the monitoring node subset to any node in the unmonitored node subset is calculated, and based on the contribution, the global importance of each node in the monitoring node subset in the target drainage network is determined. Based on the global importance of nodes in the target drainage network, key nodes of the target drainage network under the target missing rate are selected from the target dataset, so as to predict the global state of the target drainage network based on the key nodes.

2. The method for identifying key nodes in a drainage network for predictive improvement according to claim 1, characterized in that, The hydraulic prediction model is trained using structured tabular data and is an XGBoost model.

3. The method for identifying key nodes in a drainage network for predictive improvement according to claim 2, characterized in that, The step of calculating the contribution of nodes in the monitoring node subset to any node in the unmonitored node subset based on the predicted hydraulic state data includes: Based on the predicted hydraulic state data of the nodes in the unmonitored node subset at each time step, the mean hydraulic state of the unmonitored node subset at each time step is determined. Based on the tree model interpreter that matches the XGBoost model, the predicted hydraulic state data of the nodes in the unmonitored node subset and the mean hydraulic state of the unmonitored node subset at each time step are used to obtain the SHAP value matrix of the unmonitored nodes. The values ​​in the SHAP value matrix represent the contribution of the nodes in the monitored node subset to the nodes in the unmonitored node subset at each time step.

4. The method for identifying key nodes in a drainage network for predictive improvement according to claim 1, characterized in that, The step of determining the global importance of nodes in the monitoring node subset within the target drainage network based on the contribution degree includes: For any node in the subset of monitoring nodes, the absolute value of the contribution is processed to obtain the processed contribution. For any node in the subset of monitoring nodes, the average of the processed contribution is taken to obtain the global importance of the node in the target drainage network.

5. The method for identifying key nodes in a drainage network for predictive improvement according to claim 1, characterized in that, Given a project budget and a predetermined number of sensors for the target drainage network, the process of selecting key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network, under the target missing rate, includes: Based on the global importance of the nodes in the target drainage network, the nodes contained in the target drainage network are sorted to obtain the sorting result; From the sorting results, the key nodes of the target drainage network under the target missing rate are determined, and the target missing rate is determined based on the project budget and the preset number of sensors.

6. The method for identifying key nodes in a drainage network for predictive improvement according to claim 1, characterized in that, When the target drainage network has a contribution threshold, the step of filtering out key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network at the target missing rate includes: For each node in the target drainage network, the corresponding contribution is summed to obtain the sum of the node's contribution. Based on the contribution of the nodes in the target drainage network, the nodes in the target drainage network are sorted in descending order to obtain the sorting result; For the sorting results, if the sum of the contributions of the top k-1 nodes in the target drainage network is less than a preset value, and the sum of the contributions of the top k nodes in the target drainage network is greater than or equal to a preset value, then the top k nodes in the target drainage network are regarded as the key nodes of the target drainage network under the target missing rate.

7. The method for identifying key nodes in a drainage network for predictive improvement according to claim 1, characterized in that, The training process of the hydraulic prediction model includes: Obtain a training dataset of all nodes and multiple operating conditions of the sample drainage network, wherein the training dataset contains time series data of the hydraulic state of the nodes; Based on different missing rates, the nodes in the sample drainage network are randomly divided to obtain a subset of monitored nodes and a subset of unmonitored nodes. Hydraulic state data for the monitored node sample subset and the unmonitored node sample subset are determined from the hydraulic state training dataset, respectively. The hydraulic prediction model is trained based on the hydraulic state data of the sample subset of the monitored nodes and the sample subset of the unmonitored nodes to obtain the trained hydraulic prediction model.

8. The method for identifying key nodes in a drainage network for predictive improvement according to claim 1, characterized in that, The method for identifying key nodes in a drainage network for predictive improvement also includes: If the target drainage network is an already operational drainage network, an optimized layout scheme for the target drainage network is constructed based on the key nodes. The optimized layout scheme includes at least the attribute data of the key nodes and the sorting results of the target drainage network. In the case where the target drainage network is an unused drainage network, a monitoring and planning scheme for the target drainage network is constructed based on the key nodes. The optimized layout scheme includes at least the attribute data of the key nodes and the sorting results of the target drainage network.

9. A key node identification device for drainage pipe networks for predictive improvement, characterized in that, The drainage network key node identification device for predictive improvement includes: The partitioning module is used to randomly partition the nodes in the target dataset of the target drainage network based on a preset target missing rate, so as to obtain a subset of monitored nodes and a subset of unmonitored nodes. The prediction module is used to obtain the hydraulic state data of the nodes in the monitoring node subset from the target dataset, and input the hydraulic state data into the trained hydraulic prediction model to obtain the predicted hydraulic state data of the nodes in the unmonitored node subset. The calculation module is used to calculate the contribution of the nodes in the monitoring node subset to any node in the unmonitored node subset based on the predicted hydraulic state data, and to determine the global importance of the nodes in the monitoring node subset in the target drainage network based on the contribution. The filtering module is used to filter out key nodes of the target drainage network from the target dataset based on the global importance of nodes in the target drainage network, under the target missing rate, so as to predict the global state of the target drainage network based on the key nodes.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 8.

Images