Urban drainage optimization management method and system based on multi-modal data

Abstract

The invention provides an urban drainage optimization management method and system based on multi-modal data, and relates to the technical field of optimization management, and the method comprises the steps: building a digital twinborn model based on the multi-modal data of an urban drainage network; based on the digital twinborn model, city drainage optimization is started; detecting key decision nodes needing user intervention decision in the urban drainage optimization process; predicting the maximum decision-making capability of the user for the key decision-making node; and assisting the user to make a decision on the key decision node at the level closest to the maximum decision ability. According to the urban drainage optimization management method and system based on the multi-modal data, the operation efficiency and stability of a drainage system are optimized, the decision-making experience of a user is improved through a personalized auxiliary strategy, and an efficient, intelligent and sustainable solution is provided for urban drainage management.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optimization management, in particular to a city drainage optimization management method and system based on multi-modal data. BACKGROUND

[0002] The optimization management of city drainage system is an important part of modern city infrastructure construction, which is directly related to the city flood control and drainage capacity, environmental sustainable development and the quality of life of residents. With the acceleration of urbanization, the drainage pipe network is facing the challenges of pipe network aging and low operation efficiency, and the traditional management method is difficult to cope with the complex and variable operation environment and real-time decision-making needs.

[0003] At present, digital twin technology provides a real-time synchronous digital mirror for the drainage system through the virtual modeling of physical entities, which can integrate multi-modal data (such as pipe parameters, flow data, rainfall information, etc.) and realize dynamic simulation of system state. However, the existing digital twin technology still has limitations in the application of city drainage optimization: first, the data integration and real-time updating capability is insufficient, which is difficult to efficiently process multi-source heterogeneous data, resulting in a large deviation between the model and reality; second, the intelligent support for optimization scheduling scheme generation is lacking, and traditional manual scheduling is difficult to cope with sudden heavy rain and other high-load scenarios; third, the key decision node detection and user decision assistance mechanism are imperfect, and users are prone to decision-making errors in complex decision-making scenarios due to information overload or time pressure. For example, existing systems rely on static rules or simple threshold judgments of key nodes, lack dynamic adaptive detection capabilities; at the same time, the user's cognitive load and decision-making ability are not fully considered in assisted decision-making, resulting in too much interference or insufficient support of assisted information.

[0004] In view of the above problems, it is urgent to provide a city drainage optimization management method based on multi-modal data, which comprehensively utilizes digital twin technology, machine learning algorithm and real-time decision support framework to realize intelligent and efficient management of drainage system. SUMMARY

[0005] The present application aims to provide a city drainage optimization management method and system based on multi-modal data to solve the problems pointed out in the background art.

[0006] In the first aspect, the present application provides a city drainage optimization management method based on multi-modal data, comprising: Based on the multi-modal data of city drainage pipe network, a digital twin model is constructed; Based on the digital twin model, city drainage optimization is started; Detecting the key decision nodes that need user intervention in the city drainage optimization process; Predicting the maximum decision-making ability of the user for the key decision nodes; assist the user to make decisions on the critical decision nodes at a level closest to the maximum decision-making capacity.

[0007] Optionally, the multi-modal data based on the urban drainage pipe network is used to construct a digital twin model, including: Based on the digital twin engine, a digital twin model is constructed according to the multi-modal data of the urban drainage pipe network.

[0008] Optionally, the digital twin model is used to start urban drainage optimization, including: Based on the pre-trained urban drainage optimization model, the digital twin model is used to start urban drainage optimization.

[0009] Optionally, the critical decision nodes requiring user intervention in the urban drainage optimization process are detected, including: Based on the critical decision node detection rule, the critical decision nodes requiring user intervention in the urban drainage optimization process are detected.

[0010] Optionally, the maximum decision-making capacity of the user for the critical decision nodes is predicted, including: Based on the pre-trained maximum decision-making capacity prediction model, the maximum decision-making capacity of the user for the critical decision nodes is predicted according to the user's urban drainage optimization decision history.

[0011] Optionally, the user is assisted to make decisions on the critical decision nodes at a level closest to the maximum decision-making capacity, including: Based on the predicted maximum decision-making capacity of the user for the critical decision nodes, the decision time limit and the assistance density are dynamically matched; The time period from the current time to the decision deadline within the decision time limit is divided into multiple time windows according to a window order dynamic weight function; For each of the time windows, when the latest current time enters the time window, the following operations are performed: According to the decision-making progress of the user currently making decisions on the critical decision nodes, the decision-making capacity of the user in the previous time window, the predicted maximum decision-making capacity of the user for the critical decision nodes, and the window order of the time window, the capacity exertion target in the time window is dynamically calculated and determined through decision-making capacity exertion trajectory modeling; The historical decision behavior sequence of the user in the current time window based on the digital twin model is obtained in real time; Based on the historical decision behavior sequence and a preset decision assistance trigger rule library, it is determined whether the user triggers the decision assistance process; If the decision assistance process is triggered, the real-time decision situation awareness engine accurately determines an assistance scheme and a latest assistance time limit based on the historical decision behavior sequence, the remaining time of the current time window, and the ability exertion target; For each time node before the latest assistance time limit, the following operations are performed: Based on the incremental information of the change of the historical decision behavior sequence to the current time node, the real-time constraint influence coefficient of the assistance logic of the assistance scheme on the user decision ability exertion is calculated through incremental information sensitivity analysis; Based on the window order of the time window and the remaining time of the current time window, the constraint influence threshold is dynamically matched through an adaptive threshold decay rule; When the real-time constraint influence coefficient is lower than the constraint influence threshold, the assistance scheme is optimized based on the real-time constraint influence coefficient through reinforcement learning to obtain an optimized assistance scheme, wherein the reinforcement learning adopts a Q-learning algorithm to minimize the deviation of the constraint influence coefficient and the threshold; The optimized assistance scheme is pushed to the user in real time to guide the user decision ability to continuously approach the ability exertion target.

[0012] Optionally, the setting of the decision time limit is based on a user decision pressure threshold model to promote maximum exertion of decision ability, and the setting of the assistance density is based on a user cognitive load tolerance model to avoid excessive interference or insufficient support of the assistance behavior on the decision process.

[0013] Optionally, the weight coefficient of the time window is larger for a later window order to adapt to the improvement of the tolerance of assistance interference at the end of decision.

[0014] Optionally, the decision ability exertion trajectory modeling adopts a nonlinear regression algorithm to fit the user historical decision ability release curve.

[0015] In a second aspect, an embodiment of the present application provides a city drainage optimization management system based on multi-modal data, comprising: A construction module for constructing a digital twin model based on multi-modal data of a city drainage network; An optimization module for starting city drainage optimization based on the digital twin model; A detection module for detecting key decision nodes requiring user intervention in the city drainage optimization process; A prediction module for predicting the maximum decision ability of the user for the key decision nodes; An assistance module for assisting the user to make decisions for the key decision nodes at a level closest to the maximum decision ability.

[0016] The present application has the following beneficial effects: By integrating digital twin technology, machine learning algorithms, and real-time decision support frameworks, the management efficiency and intelligent level of the urban drainage system are significantly improved. The high-precision digital twin model constructed based on multi-modal data realizes real-time synchronous simulation of the drainage system, with model accuracy far exceeding traditional static models, ensuring the reliability and accuracy of the optimized scheduling scheme; through the pre-trained urban drainage optimization model combined with deep reinforcement learning algorithms, dynamic scheduling schemes are generated, taking into account efficiency and energy saving; the introduction of key decision node detection rules and maximum decision-making ability prediction models dynamically identifies high-risk scenarios and quantifies user decision-making ability, significantly improving decision-making accuracy; through dynamic matching of decision-making time limit and auxiliary density, as well as auxiliary scheme optimization based on reinforcement learning, the system avoids information overload while ensuring that user decision-making ability approaches the optimal level, shortening user response time and improving decision-making accuracy. Overall, this scheme not only optimizes the operation efficiency and stability of the drainage system, but also improves user decision-making experience through personalized assistance strategies, providing an efficient, intelligent, and sustainable solution for urban drainage management.

[0017] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and attained by means of the instrumentalities particularly pointed out in the written description and claims hereof.

[0018] The technical solutions of the present application will be further described in detail below with the help of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application, and are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 A flowchart of the urban drainage optimization management method based on multi-modal data in the embodiment of the present application; Figure 2 A schematic diagram of the urban drainage optimization management system based on multi-modal data in the embodiment of the present application. DETAILED DESCRIPTION

[0020] The preferred embodiments of the present application will be described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and do not limit the present application.

[0021] Figure 1 A flowchart of the urban drainage optimization management method based on multi-modal data in the embodiment of the present application is shown in FIG. Figure 1 The method comprises: S1. Construct a digital twin model based on multimodal data of the urban drainage network. Step S1 specifically includes the following sub-steps: S11. Based on the digital twin engine, construct a digital twin model according to the multimodal data of the urban drainage network.

[0022] In step S11, the digital twin model refers to the virtual modeling of the physical entities of the urban drainage network using a digital twin engine, generating a digital mirror that is synchronized in real time with the real drainage system. This mirror is used to simulate and optimize the operational status of the drainage system. The digital twin engine is a software platform integrating multimodal data processing, real-time simulation, and dynamic updates. It can process multimodal data such as sensor data, Geographic Information System (GIS) data, and historical operational data. Multimodal data includes physical parameters of the drainage network (such as pipe diameter, slope, and material), real-time flow data (such as flow rate per cubic meter per second monitored by flow meters), environmental data (such as rainfall and soil moisture), and historical operational records (such as drainage volume over the past 24 hours). The process of building the digital twin model first acquires real-time data from the sensor network of the urban drainage network through a data acquisition interface. For example, it acquires pipe flow rate through ultrasonic flow meters and rainfall intensity through rain sensors. Subsequently, the digital twin engine utilizes 3D modeling techniques (such as OpenGL-based rendering algorithms) to map the physical structure of the pipe network into a virtual 3D model, and simulates the dynamic behavior of the drainage system using fluid dynamics simulation algorithms (such as numerical solutions to the Navier-Stokes equations). Dynamic updates to the model are achieved through real-time data streams (such as sensor data transmitted via the MQTT protocol), updated every minute to ensure minimal deviation between the model and the actual drainage system. Model accuracy is evaluated using root mean square error (RMSE), with a target RMSE of less than 0.05 m³ / s. Physical parameters are extracted from pipe network design drawings and GIS databases; real-time flow data is collected from sensors deployed at key nodes in the pipe network; environmental data is obtained from meteorological station APIs; and historical operation records are extracted from the urban drainage management system database.

[0023] In one implementation example, in a drainage optimization project in a coastal city, a digital twin engine was deployed on a high-performance computing server (equipped with 128GB of memory and an NVIDIA A100 GPU) to process multimodal data covering a 500-kilometer drainage network across the city. The engine first extracts the physical parameters of the network from a GIS database, such as the main pipe diameter being 1.2 meters and the slope being 0.5%. It then collects flow data in real time using ultrasonic flow meters deployed at 100 key nodes (sampling frequency of 1Hz, data transmitted to the engine in JSON format via the MQTT protocol). Simultaneously, it obtains hourly rainfall data from the local weather station API (e.g., current rainfall intensity of 10 mm / h) and extracts drainage records for the past 30 days from the city's drainage management database. The engine uses OpenGL to render the 3D structure of the network, generating a virtual model containing 5000 pipe nodes, and calculates the water flow velocity within the pipes (accurate to 0.01 m / s) using fluid dynamics simulation algorithms. The model is updated every minute via data stream, synchronizing with the pipeline network status in real time. For example, during a rainstorm, if the flow rate at a certain node increases from 2 m³ / s to 5 m³ / s, the model updates synchronously and verifies the simulation error (RMSE is 0.03 m³ / s). In this way, the digital twin model provides a high-precision virtualization environment for subsequent drainage optimization.

[0024] S2. Based on the aforementioned digital twin model, begin optimizing urban drainage. Step S2 specifically includes the following sub-steps: S21. Based on the pre-trained urban drainage optimization model, urban drainage optimization begins according to the digital twin model.

[0025] In step S21, the urban drainage optimization model is a pre-trained model based on machine learning and optimization algorithms. It generates an optimized scheduling scheme for the drainage network based on the output data of the digital twin model, aiming to maximize drainage efficiency and reduce overflow risk. The optimization model takes the real-time status of the digital twin model (such as pipe flow and node water level) and environmental inputs (such as rainfall prediction) as inputs. The outputs include pump station operating power (unit: kilowatts), valve opening degree (range: 0-100%), and drainage path allocation ratio (accurate to 0.1%). The model is pre-trained using deep reinforcement learning algorithms (such as DeepQ-Network, DQN). The training data comes from historical drainage data and simulated rainfall scenarios. The reward function is defined as a weighted sum of drainage efficiency (water volume discharged per second) and overflow loss (absolute value of overflow), with weights of 0.7 and 0.3, respectively. The optimization process obtains the current network status by calling the digital twin model's API interface in real time, for example, obtaining the real-time water level of a node (accurate to 0.01 meters) through a RESTful API. The optimization objective of the model is to keep the overflow rate below 1% while ensuring that the pumping station's energy consumption does not exceed 80% of its rated power. Real-time status data is obtained from the digital twin model's API, rainfall prediction data is extracted from the weather station's API, historical training data is extracted from the urban drainage database, and model parameters (such as neural network weights) are loaded from pre-trained model files.

[0026] For example, in the aforementioned coastal city drainage optimization project, the urban drainage optimization model was deployed on the same high-performance server. The pre-trained model was trained based on drainage data from the past five years and 1000 simulated rainfall scenarios (generated using the Monte Carlo method). The reward function combined drainage efficiency (drainage rate of 10 m³ / s) and overflow loss (target below 0.1 m³ / s). During a rainstorm event, the optimization model obtained real-time data from the digital twin model via a RESTful API. For example, the water level at a main pipeline node was 1.5 meters, and the rainfall forecast showed a rainfall intensity of 15 mm / h for the next hour. After calculation, the model output an optimization plan: adjust the power of pump station No. 3 to 500 kW (75% of rated power), set the main valve opening to 80%, and allocate 50% of the drainage volume to the backup drainage path. The optimization plan was sent to the pump station control system in real time (via the Modbus protocol). After execution, the overflow rate was detected to have dropped to 0.8%, and the energy consumption remained at 78% of the rated power. By continuously accessing the real-time status of the digital twin model, the model's dynamic adjustment scheduling scheme is optimized to ensure the stable operation of the drainage system under high load.

[0027] S3. Detect key decision nodes requiring user intervention in the urban drainage optimization process. Step S3 specifically includes the following sub-steps: S31. Based on the key decision node detection rules, detect the key decision nodes that require user intervention in the process of urban drainage optimization.

[0028] In step S31, key decision nodes refer to special scenarios requiring user intervention in urban drainage optimization, such as overflow risk exceeding 5% or pump station power approaching its rated upper limit (90%). Key decision node detection rules are a predefined set of conditions, setting thresholds based on pipeline status parameters (e.g., water level, flow rate), environmental parameters (e.g., rainfall intensity), and system operating status (e.g., pump station power). The detection rules determine whether trigger conditions are met by real-time monitoring of the digital twin model's output data, such as water level exceeding 80% of pipeline capacity (obtained from the digital twin model API, unit: meters) or rainfall intensity exceeding 20 mm / h (obtained from the meteorological station API). The detection process uses a rule engine (e.g., Drools) for real-time evaluation, with rule priority ordered according to risk level, and high-risk conditions (e.g., overflow risk) having the highest priority. Detection results are output as key decision node identifiers (e.g., "Node N1 Overflow Risk"), and related parameters (e.g., water level 2.0 meters, flow rate 5 m³ / s) are recorded. Pipeline status parameters are obtained from the API of the digital twin model, environmental parameters are extracted from the meteorological station API, and system operating status is collected from the pump station control system via the Modbus protocol.

[0029] In one implementation example, in a coastal city drainage optimization scenario, key decision node detection rules are deployed on the rule engine Drools, containing 50 rules, such as "When the water level at a certain node exceeds 1.8 meters and the rainfall intensity is greater than 15 mm / h, trigger the overflow risk decision node." During a rainstorm, the rule engine obtains a water level of 1.9 meters and a flow rate of 6 m³ / s at a certain node through a digital twin model API, and a rainfall intensity of 18 mm / h from a weather station API. After evaluation, the rule engine triggers the "Node N1 Overflow Risk" decision node, generates an identifier, and records relevant parameters (water level 1.9 meters, flow rate 6 m³ / s, rainfall intensity 18 mm / h). The detection results are pushed to the user's decision interface in real time via the WebSocket protocol, prompting the user to intervene in the decision-making process, such as whether to adjust the allocation ratio of backup drainage paths. The detection process is executed every 10 seconds to ensure timely capture of key decision nodes.

[0030] S4. Predict the user's maximum decision-making ability for the key decision-making node. Step S4 specifically includes the following sub-steps: S41. Based on the pre-trained maximum decision-making capability prediction model, predict the user's maximum decision-making capability for the key decision nodes according to the user's urban drainage optimization decision-making history.

[0031] In step S41, the maximum decision-making ability prediction model is a machine learning-based model used to predict a user's maximum decision-making ability at key decision nodes. It is defined as a quantitative indicator (range: 0-100, unit: dimensionless) of the optimal decision a user can make given time and information conditions. Decision-making ability is quantified through the user's historical decision-making behavior (such as decision response time and decision accuracy). Accuracy is based on the deviation between historical decision results and the optimal solution (e.g., a drainage path allocation deviation of less than 5% is considered correct). The model uses a random forest algorithm. Input features include the user's historical decision response time (unit: seconds, extracted from user interface logs), decision accuracy (calculated from historical decision records), complexity of key decision nodes (based on the number of node parameters and risk level score, range: 0-10), and environmental factors (such as rainfall intensity, unit: mm / h). The model is pre-trained using a training set (containing at least 1000 user decision records) and predicts the output as the maximum decision-making ability value (e.g., 85). Historical decision data is extracted from the user operation log database, node complexity is calculated using a rule engine, and environmental factors are obtained from the weather station API.

[0032] For example, in a coastal city project, the maximum decision-making ability prediction model is based on a random forest algorithm. The training data includes 5000 decision records from 100 users over the past year, with features including response time (average 3 seconds), decision accuracy (average 90%), and node complexity (average 5). When an overflow risk decision node is triggered, the model extracts historical data for a user from user logs (response time 2.8 seconds, accuracy 92%), obtains the node complexity (score 7) from the rule engine, and obtains the rainfall intensity (20 mm / h) from the weather station API. The model predicts that the user's maximum decision-making ability at the current node is 88. The prediction result is pushed to the user interface via WebSocket as a basis for subsequent auxiliary decisions. The model re-evaluates after each key decision node is triggered to ensure that the prediction result matches the user's current state.

[0033] S5. Assist the user in making decisions at key decision nodes at a level closest to their maximum decision-making ability. Step S5 specifically includes the following sub-steps: S51. Based on the predicted maximum decision-making ability of users at key decision nodes, dynamically match the decision time limit and the auxiliary density; the decision time limit is set based on the user decision pressure threshold model to promote the maximization of decision-making ability, and the auxiliary density is set based on the user cognitive load tolerance model to avoid excessive interference or insufficient support of auxiliary behavior on the decision-making process.

[0034] In step S51, the decision time limit refers to the maximum allowed time (in seconds) for a user to complete a key decision node. It is dynamically determined using a user decision stress threshold model to maximize the user's decision-making ability under moderate pressure. Based on psychological research, the model sets the stress threshold at 80%-90% of the user's maximum decision-making ability (e.g., when the maximum decision-making ability is 88, the stress threshold is 70-79). The decision time limit is calculated as follows: first, the maximum decision-making ability (output from step S41, range: 0-100) is obtained; then, the optimal decision time range is calculated using the stress threshold model (based on a Sigmoid function mapping) (e.g., ability 88 corresponds to 10-15 seconds). The auxiliary density refers to the frequency of providing auxiliary information per second (in seconds). It is determined using a user cognitive load tolerance model based on cognitive load theory. The tolerance is set at the user's maximum information processing rate (e.g., 5 times / second) to avoid excessive auxiliary information leading to cognitive overload or insufficient auxiliary information leading to inadequate support. The auxiliary density is calculated as follows: based on the maximum decision-making ability, node complexity (obtained from the rule engine, range: 0-10), and user historical response time (extracted from user logs, unit: seconds), the optimal density is calculated using a linear regression model (e.g., a density of 2 times / second corresponds to a capability of 88, complexity of 7, and response time of 3 seconds). The maximum decision-making ability is obtained from step S41, node complexity is output from the rule engine, historical response time is extracted from the user operation log database, and stress threshold and cognitive load tolerance are loaded from the pre-trained model parameters.

[0035] The user decision-making stress threshold model is constructed based on the stress-performance relationship in psychology (Yerkes-Dodson Law), which states that moderate stress can improve decision-making performance, while excessive or insufficient stress reduces performance. The model aims to map the user's maximum decision-making ability (range 0-100, obtained from step S41) to the optimal decision-making time limit (unit: seconds), keeping the stress within 80%-90% of the maximum decision-making ability. First, a training dataset is collected, including the user's historical decision data (extracted from user operation logs, including decision nodes, response times, decision results, etc.) and the corresponding maximum decision-making ability. Next, the Sigmoid function is used as the mapping model because it can effectively describe the non-linear relationship between ability and time. The Sigmoid function has the form: T = a / (1 + e^(-b*(Cc))), where T is the decision-making time limit, C is the maximum decision-making ability, and a, b, and c are model parameters, determined by non-linear regression fitting on the training data (e.g., 1000 decision records). The initial parameter values ​​can be set to a = 20 (maximum time limit range), b = 0.1 (control curve steepness), and c = 50 (median capability), and optimized using gradient descent. After training, the model is deployed on a server, and the pre-trained parameters are stored in a database. During runtime, the maximum decision capability (e.g., 88) is input, and the model outputs the decision time limit (e.g., 12 seconds) corresponding to the stress threshold (e.g., 75), which is pushed to the user interface via WebSocket.

[0036] The user cognitive load tolerance model, based on cognitive load theory, aims to determine the optimal auxiliary density (unit: times / second) to avoid information overload or underload. The model comprehensively considers maximum decision-making ability (obtained from S41), node complexity (obtained from the rule engine, ranging from 0-10), and historical user response time (extracted from logs). During construction, training data is first collected, containing 1000 user interaction records (including decision nodes, complexity, response time, auxiliary frequency, and user feedback). A linear regression model is used, in the form: D = w1C + w2N + w3*R + b, where D is the auxiliary density, C is the maximum decision-making ability, N is the node complexity, R is the historical response time, w1, w2, and w3 are weights, and b is the bias. The training data is fitted using the least squares method to obtain the weights (e.g., w1 = 0.02, w2 = 0.15, w3 = -0.1, b = 0.5). To ensure model adaptability, a regularization term (L2 regularization) is added to prevent overfitting. The trained model is deployed on a server. During runtime, it takes input parameters (e.g., capability 88, complexity 7, response time 2.8 seconds) and outputs the optimal auxiliary density (e.g., 2.5 times / second). Based on this, the system pushes auxiliary information (e.g., pop-up suggestions) every 0.4 seconds through the user interface, ensuring that the information frequency is moderate and does not interfere with operation.

[0037] The above approach uses the Sigmoid function and linear regression to implement the two models, respectively, to ensure a dynamic match between decision-making time and auxiliary density, taking into account the actual needs of users' decision-making ability and cognitive load.

[0038] For example, in a coastal city drainage optimization project, when the "Node N1 Overflow Risk" decision node is detected, step S41 predicts the user's maximum decision-making ability to be 88. The decision pressure threshold model (based on the Sigmoid function, with pre-trained parameters stored on the server) calculates a pressure threshold of 75, corresponding to a decision time limit of 12 seconds. This is pushed to the user interface via WebSocket, prompting "Please complete the decision within 12 seconds." Simultaneously, the cognitive load tolerance model (based on linear regression, with training data including 1000 user interaction records) calculates an assistance density of 2.5 times / second based on the maximum decision-making ability of 88, node complexity of 7 (obtained from the rule engine), and the user's historical response time of 2.8 seconds (extracted from logs). During implementation, the system pushes assistance information (such as a suggestion to adjust the valve opening to 85%) to the user interface every 0.4 seconds, ensuring a moderate assistance frequency. After receiving the decision time limit prompt, the user completes the decision within 12 seconds, and the assistance information is displayed in a pop-up window to avoid interfering with user operation. In this way, the system effectively promotes the user's decision-making ability while avoiding cognitive overload.

[0039] S52. Divide the time period from the current time to the decision deadline within the decision time limit into multiple time windows according to the window order dynamic weight function; the later the window order, the larger the weight coefficient of the time window, so as to adapt to the increased tolerance for auxiliary interference in the late stage of decision-making.

[0040] In step S52, a time window refers to a period within the decision-making timeframe divided into multiple sub-intervals to assist the user's decision-making in stages. The duration (in seconds) of each window is determined by a dynamic weighting function based on the window order to accommodate the higher tolerance for auxiliary interference towards the end of the decision-making process. The dynamic weighting function based on the window order is based on an exponential growth model, defined as a weight coefficient that increases with the window order (from 1 to N, where N is the total number of windows), with a weight coefficient range of 0.1-1.0, and higher weights for windows at the end (e.g., the Nth window has a weight of 1.0). The time window division method is as follows: First, obtain the decision-making timeframe (from step S51, in seconds). Based on the preset number of windows N (e.g., 5, based on empirical values), calculate the duration of each window (e.g., 12 seconds divided into 5 windows with weight coefficients of 0.1, 0.2, 0.4, 0.7, and 1.0, respectively, with the duration allocated proportionally according to the weight). The duration calculation formula is: Window time = Decision-making timeframe × Weight coefficient / Total weights. The total weight is obtained by summing the weight coefficients of all windows (e.g., 0.1 + 0.2 + 0.4 + 0.7 + 1.0 = 2.4). After window division, the start and end times (in seconds, accurate to 0.01 seconds) of each window are recorded in system memory for use in subsequent steps. The decision time limit is obtained from step S51, the number of windows and weight coefficients are loaded from pre-configured parameters, and the current time is obtained from the system clock (via the Linux system call gettimeofday, accurate to microseconds).

[0041] Continuing with the previous example, the decision time limit is 12 seconds, the system presets 5 windows, and the weight coefficients are configured as 0.1, 0.2, 0.4, 0.7, and 1.0, with a total weight of 2.4. The dynamic weight function for window order calculates the duration of each window: first window time = 12 × 0.1 / 2.4 = 0.5 seconds, second window time = 12 × 0.2 / 2.4 = 1 second, third window time = 12 × 0.4 / 2.4 = 2 seconds, fourth window time = 12 × 0.7 / 2.4 = 3.5 seconds, and fifth window time = 12 × 1.0 / 2.4 = 5 seconds. After division, the time windows are [0, 0.5 seconds], [0.5 seconds, 1.5 seconds], [1.5 seconds, 3.5 seconds], [3.5 seconds, 7 seconds], and [7 seconds, 12 seconds]. The start and end times are precisely recorded using the system clock (called by gettimeofday). The system displays the current time window on the user interface (e.g., "Currently in window 3, 9 seconds remaining") and passes this window information to subsequent steps for dynamic decision support. The segmentation process ensures a longer window period at the end of the decision-making process, accommodating users' higher tolerance for assistance in the later stages of the decision-making process.

[0042] S53. For each time window, when the latest current time enters the time window, perform the following operations: S531. Based on the user's current decision-making progress on key decision nodes, the user's decision-making ability demonstrated in previous time windows, the predicted maximum decision-making ability of the user on key decision nodes, and the window order of the time window, the ability performance target within the time window is dynamically calculated and determined through decision performance trajectory modeling; the decision performance trajectory modeling adopts a nonlinear regression algorithm to fit the user's historical decision performance release curve.

[0043] In step S531, the capability realization target refers to the level of decision-making ability that the user should achieve within the current time window (range: 0-100, unit: dimensionless). This is dynamically calculated through decision-making capability realization trajectory modeling to ensure that the user's decision-making ability gradually approaches the maximum decision-making ability. Decision progress is defined as the percentage of decisions the user has currently completed (range: 0-100%, extracted from user interface interaction logs, e.g., 50% completion of the valve opening adjustment operation). The realized decision-making ability is the user's cumulative decision-making ability within previous time windows (range: 0-100, calculated from user operation logs, e.g., a weighted average of the number of correct decisions). The maximum decision-making ability is obtained from step S41 (e.g., 88). The window order is the sequence number of the current time window (from 1 to N, obtained from step S52). The decision-making capability realization trajectory modeling uses a nonlinear regression algorithm (such as multinomial regression). The input features include decision progress, realized decision-making ability, maximum decision-making ability, and window order, and the output is the capability realization target. The modeling process first extracts the decision progress in real time from the user interface logs (by parsing user clicks and inputs, e.g., clicking the "Adjust Valve" button is counted as 10% progress). The realized decision capability is calculated from historical operation records (e.g., the proportion of correct decisions multiplied by a weighting coefficient). The window order is obtained from step S52. A nonlinear regression model is pre-trained based on historical decision data (at least 1000 records), fitting the user's capability release curve to predict the current window's capability target (e.g., a capability target of 70). Decision progress is extracted from the user interface logs (collected in real time via WebSocket), realized decision capability is calculated from operation logs, maximum decision capability is obtained from step S41, and the window order is obtained from step S52.

[0044] For example, in the "Node N1 Overflow Risk" decision-making scenario in a coastal city, the current time is in the third time window ([1.5 seconds, 3.5 seconds]). The system extracts the decision progress from the user interface log as 40% (the user has adjusted the valve opening to 50%, with a target of 85%), calculates the realized decision capability of the first two windows as 30 (based on two correct decisions, each with a weight of 15), the maximum decision capability as 88 (obtained from S41), and the window order as 3 (obtained from S52). A nonlinear regression model (pre-trained on 1000 user decision records and stored on the server) fits the capability release curve, with input features {progress: 40%, realized capability: 30, maximum capability: 88, window order: 3}, and output capability realization target as 65. The calculation process is completed through a high-performance server (taking less than 0.1 seconds), and the result is pushed to the user interface via WebSocket, displaying "Current window target capability: 65". The system adjusts subsequent auxiliary strategies based on this target to ensure that the user's decision capability gradually improves. In this way, the ability to perform its target accurately reflects the user's current state and window characteristics, providing a basis for subsequent assistance.

[0045] S532. Real-time acquisition of the user's historical decision-making behavior sequence based on the digital twin model within the current time window.

[0046] In step S532, the historical decision-making behavior sequence refers to the sequence of operation records generated by the user's interaction with the digital twin model within the current time window. This sequence is used to analyze user decision-making patterns and trigger auxiliary processes. Decision-making behaviors include user actions on the interface (such as clicking buttons or inputting parameters) and decision results (such as adjusting the pump station power to 500kW). Each behavior record includes the operation type (e.g., "adjust valve"), operation parameters (e.g., valve opening 80%), and a timestamp (accurate to 0.01 seconds). The data is acquired in real-time through the user interface (using the WebSocket protocol, data format is key-value pairs, e.g., {"Operation": "Adjust valve", "Parameter": "80%", "Time": "1.6 seconds"}) and correlated with the output of the digital twin model (e.g., the adjusted water level change is 1.8 meters, obtained from the model API). The sequence is stored in a memory cache (e.g., Redis), and its length is dynamically adjusted according to the duration of the time window (e.g., approximately 200 records are stored in a 2-second window, with a sampling frequency of 100Hz). Operational data is collected in real time from the user interface via WebSocket, the digital twin model output is obtained via RESTful API, and the timestamp is extracted from the system clock (gettimeofday call).

[0047] For example, in the third time window ([1.5 seconds, 3.5 seconds]), the system collects user interaction data in real time via WebSocket. For instance, the user clicks the "Adjust Valve" button at 1.6 seconds, inputs an opening of 60%, and confirms the operation at 1.8 seconds. The system obtains the adjusted water level as 1.85 meters and the flow rate as 5.5 m³ / s from the digital twin model API. The behavior sequence is recorded as [{"Operation": "Adjust Valve", "Parameter": "60%", "Time": "1.6 seconds", "Water Level": "1.85 meters"}, {"Operation": "Confirm", "Time": "1.8 seconds"}], and stored in the Redis cache (key: "User ID_Window 3", TTL: window duration 2 seconds). The sequence is updated every 0.01 seconds to ensure that all user operations are captured. After analyzing the sequence, the system finds that the user's adjustment speed is slow (only one adjustment is completed within 2 seconds), providing data support for subsequent auxiliary triggers.

[0048] S533. Based on the historical decision-making behavior sequence and the preset decision-aid triggering rule base, dynamically determine whether the user triggers the decision-aid process.

[0049] In step S533, the decision assistance trigger rule base is a predefined set of conditions used to determine whether a user needs assistance, based on the characteristics of historical decision behavior sequences (such as operation frequency and decision accuracy) and the current time window status. Trigger rules include: operation frequency below a threshold (e.g., less than 1 time / second, defined as the number of operations per second, calculated from the behavior sequence), decision accuracy below a threshold (e.g., below 80%, calculated by comparing with the optimal solution of the digital twin model), or decision progress lagging (e.g., progress is 40% lower than expected, obtained from S531). The rule base is implemented through a rule engine (such as Drools) and contains at least 100 rules, prioritized according to risk level (e.g., lagging progress has higher priority than low operation frequency). The determination process is as follows: features are extracted from the behavior sequence (operation frequency is calculated by dividing the count of operation records in the sequence by the window time; accuracy is calculated by comparing the user's decision with the model's optimal solution), compared with the rule base thresholds, and if any condition is met (e.g., operation frequency of 0.5 times / second is lower than the threshold of 1 time / second), the assistance process is triggered. The behavior sequence is obtained from S532, the rule threshold is loaded from the pre-configured rule base, the decision progress is obtained from S531, and the optimal solution is obtained from the digital twin model API.

[0050] For example, in the third time window, the behavior sequence shows a user operation frequency of 0.5 times / second (1 operation within 2 seconds), a decision accuracy rate of 75% (adjusted to 60%, the model's optimal solution is 85%), and a decision progress of 40% (expected 50%, obtained from S531). The rule engine loads the rule base (containing 120 rules), evaluates them, and finds that the operation frequency is below the threshold of 1 time / second, triggering the "low operation frequency" rule and determining that an auxiliary process is needed. The trigger signal is pushed to the system via WebSocket and recorded as "User ID Window 3 Triggered Assistance". The rule engine evaluates every 0.1 seconds to ensure timely response to changes in user behavior. In this way, the system accurately identifies scenarios where users need assistance, providing a foundation for subsequent steps.

[0051] S534. If the decision support process is triggered, the support scheme and the latest support time limit are accurately determined by the real-time decision situation awareness engine based on the historical decision behavior sequence, the remaining time of the current time window and the capability performance target.

[0052] In step S534, the real-time decision-making situational awareness engine is a machine learning-based analysis module used to generate auxiliary plans and the latest auxiliary time limit based on historical decision-making behavior sequences, the remaining time in the current time window, and the capability utilization target. An auxiliary plan is defined as a set of operations containing specific suggestions (e.g., "suggest adjusting the valve opening to 85%)", and the latest auxiliary time limit is the deadline for executing the assistance (in seconds, accurate to 0.01 seconds). The engine uses the Gradient Boosting Tree (GBDT) algorithm. Input features include the operation frequency of the behavior sequence (calculated from S532, in times / second), decision accuracy (calculated from S533, range: 0-100%), remaining time (calculated from the difference between the system clock and the window end time, in seconds), and capability utilization target (obtained from S531, range: 0-100). The output is an auxiliary plan (e.g., a list of suggested operations) and the latest auxiliary time limit (e.g., 0.5 seconds before the window ends). The calculation process is as follows: features are extracted from the behavior sequence, combined with the remaining time and capability objective, and the optimal assistance scheme is predicted using the GBDT model (trained based on 1000 historical assistance records, with the reward function being the capability improvement rate). The latest assistance time limit is then generated through linear interpolation based on the remaining time (e.g., with 2 seconds remaining, an interpolation coefficient of 0.25 results in a time limit of 1.5 seconds). The behavior sequence is obtained from S532, the remaining time is calculated from the system clock (gettimeofday) and the window end time (S52), the capability objective is obtained from S531, and the model parameters are loaded from the pre-trained file.

[0053] The specific construction process of the real-time decision-making situational awareness engine includes: First, collecting a training dataset containing 1000 historical auxiliary records. Each record includes input features (operation frequency of the behavior sequence, unit: times / second, calculated from S532; decision accuracy, range 0-100%, calculated from S533; remaining time, unit: seconds, calculated from the difference between the system clock and the window end time; capability performance target, range 0-100, obtained from S531) and output labels (auxiliary scheme, such as "suggest adjusting valve opening to 85%"; latest auxiliary time limit, unit: seconds). The GBDT algorithm is used to build the model because it can effectively handle nonlinear relationships and multidimensional features. The model input feature vector is {operation frequency, accuracy, remaining time, capability target}, and the output is the auxiliary scheme (operation suggestion list) and the latest auxiliary time limit. During training, the reward function is defined as the capability improvement magnitude (based on the capability change before and after historical decisions), and the model hyperparameters (such as tree depth and learning rate) are optimized through cross-validation (5-fold). To generate the latest auxiliary time limit, the model combines the remaining time and uses a linear interpolation method (time limit = current time + remaining time × interpolation coefficient, the coefficient is optimized through regression, with a typical value of 0.25). After training, the model parameters are stored as a pre-trained file and deployed on the server. During runtime, the engine obtains behavioral sequence features from S532, calculates the remaining time from the system clock (gettimeofday) and S52, obtains the capability target from S531, inputs it into the GBDT model, and outputs an auxiliary plan (e.g., "It is recommended to adjust the valve opening to 85% and increase the pump station power to 550kW") and the latest auxiliary time limit (e.g., 3 seconds). The results are pushed to the user interface via WebSocket and displayed as a pop-up window, re-evaluated every 0.1 seconds to dynamically adapt to user behavior, ensuring accurate assistance and reasonable time limits.

[0054] For example, in the third time window (2 seconds remaining, 3.5 seconds ending), the behavior sequence shows an operation frequency of 0.5 times / second, an accuracy of 75%, and a capability target of 65 (from S531). The real-time decision-making situational awareness engine (based on GBDT, with 1000 auxiliary records as training data) takes the following features as input: {operation frequency: 0.5, accuracy: 75, remaining time: 2, capability target: 65}. The output auxiliary solution is "It is recommended to adjust the valve opening to 85% and increase the pump station power to 550kW," with a maximum auxiliary time limit of 3 seconds (2 seconds remaining × interpolation coefficient 0.25 + current time 1.5 seconds). The solution is pushed to the user interface via WebSocket, displayed as a pop-up suggestion, prompting "Please adjust the valve to 85% in 3 seconds." The engine re-evaluates every 0.1 seconds to ensure the solution adapts to the user's latest behavior. In this way, the system provides accurate auxiliary suggestions and sets reasonable time limits.

[0055] S535. For each time point before the latest auxiliary time limit, perform the following operations: S5351. Based on the incremental information of the changes in the historical decision-making behavior sequence to the current time node, calculate the real-time constraint influence coefficient of the auxiliary logic of the auxiliary scheme on the user's decision-making ability through incremental information sensitivity analysis.

[0056] In step S5351, the real-time constraint influence coefficient refers to the degree to which the auxiliary logic of the auxiliary scheme restricts the user's decision-making ability (range: 0-1, 0 represents no constraint, 1 represents complete restriction). It is calculated through incremental information sensitivity analysis to ensure that the auxiliary scheme does not excessively interfere with the user's autonomous decision-making. Incremental information is defined as the change in the sequence of behaviors from the previous time node to the current time node (e.g., adding a new operation record, such as "adjust the valve to 70%)", extracted from the behavior sequence (S532). The incremental information sensitivity analysis uses a feature importance analysis method (such as SHAP value calculation). Input features include incremental operation frequency (number of new operations divided by the time interval, unit: times / second), incremental accuracy (deviation between the new decision and the optimal solution, range: 0-100%), and the suggestion strength of the auxiliary scheme (based on the number and complexity of suggested operations, range: 0-10, obtained from S534). The calculation process is as follows: incremental information is extracted from the behavior sequence (by comparing the records of the current and previous time points), incremental features (operation frequency, accuracy) are calculated, and combined with the suggestion strength, the constraint influence coefficient (e.g., 0.3) is output through the SHAP value model (pre-trained on 1000 user interaction records). The behavior sequence is obtained from S532, the time point is obtained from the system clock (gettimeofday, accurate to 0.01 seconds), the suggestion strength is parsed from the S534 auxiliary scheme, and the SHAP model parameters are loaded from the pre-trained file.

[0057] The specific implementation process of incremental information sensitivity analysis includes: First, collecting a training dataset containing 1000 user interaction records. Each record includes input features (incremental operation frequency, unit: times / second, calculated by dividing the number of new operations by the time interval; incremental accuracy, range 0-100%, calculated by the deviation between the new decision and the optimal solution; suggestion strength, range 0-10, based on the number of operations and complexity analysis of the S534 auxiliary scheme) and output labels (constraint influence coefficient, range 0-1). The SHAP value model (based on the XGBoost interpreter) is used for feature importance analysis because it can effectively quantify the contribution of each feature to the output. During training, the model uses the interaction records as input, and the optimization objective is to minimize the mean squared error between the predicted constraint influence coefficient and the actual influence. Hyperparameters (such as regularization coefficients) are adjusted through 5-fold cross-validation. After training, the SHAP model parameters are stored as a pre-trained file and deployed on a server. During runtime, the system extracts incremental information from the behavior sequence (S532) (by comparing the current time node with the record of the previous time node, e.g., adding "adjust valve to 70%" from 1.6 seconds to 1.8 seconds), calculates incremental features (operation frequency = 1 time / 0.2 seconds = 5 times / second, accuracy = 80% based on 85% deviation from the optimal solution), and parses the suggestion strength from S534 (e.g., 2 suggestions correspond to a strength of 5). These features are input into the SHAP model, outputting constraint influence coefficients (e.g., 0.25), with computation time controlled within 0.05 seconds. The results are stored in a memory cache (key: "user ID + window + time node") for subsequent auxiliary scheme optimization. The system feeds back the coefficients to the decision module via WebSocket to ensure that the auxiliary suggestions adapt to user behavior and maintain decision freedom.

[0058] For example, in the third time window, the current time node is 1.8 seconds (obtained from the system clock), and the behavior sequence shows a new operation "adjust the valve to 70%" added from 1.6 seconds to 1.8 seconds. The incremental information is extracted as {operation frequency: 1 time / 0.2 seconds = 5 times / second, accuracy: 80% (valve 70% deviation from optimal 85%)}, and the auxiliary scheme suggestion strength is 5 (parsed from S534, containing 2 suggestions). The SHAP value model input features {incremental operation frequency: 5, incremental accuracy: 80%, suggestion strength: 5}, and outputs a constraint influence coefficient of 0.25, indicating that the auxiliary scheme has a low constraint on the user's decision-making ability. The calculation takes 0.05 seconds, and the result is recorded in the memory cache (key: "User ID, window 31.8 seconds"). The system uses this coefficient to determine the suitability of the auxiliary scheme, providing a basis for subsequent optimization. In this way, the system accurately quantifies the impact of the auxiliary logic, ensuring the user's decision-making freedom.

[0059] S5352. Based on the window order of the time window and the remaining time of the current time window, the threshold is dynamically matched and constrained by the adaptive threshold decay rule.

[0060] In step S5352, the constraint impact threshold refers to the maximum allowable constraint (range: 0-1) on the user's decision-making ability by the auxiliary scheme. This is dynamically matched using an adaptive threshold decay rule to accommodate a higher tolerance for auxiliary interference at the end of the decision-making process. The window order (obtained from S52, range: 1-N) reflects the position of the time window within the decision timeframe, and the remaining time (in seconds) is the difference between the window end time (S52) and the current time (system clock). The adaptive threshold decay rule is based on an exponential decay model, defined as the threshold increasing as the window order and remaining time decrease (e.g., 0.8 for the 5th window, 0.3 for the 1st window) to allow for higher interference later. The calculation method is as follows: based on the window order and remaining time, the threshold is calculated using an exponential function (parameters α = 0.5, β = 0.1, determined through pre-training), with the formula: Threshold = Baseline Threshold + α × exp(-β × Remaining Time / Window Order). The baseline threshold is set to 0.2 (an empirical value). The window order is obtained from S52, the remaining time is calculated from the system clock (gettimeofday) and the window end time, and the model parameters are loaded from the pre-training configuration.

[0061] For example, in the third time window (order 3, end time 3.5 seconds), the current time is 1.8 seconds, and the remaining time is 1.7 seconds. The adaptive threshold decay rule (α = 0.5, β = 0.1) calculates the threshold as 0.2 + 0.5 × exp(-0.1 × 1.7 / 3) ≈ 0.673. The calculation result is recorded in the memory cache (key: "User ID Window 31.8 seconds_threshold") and passed to subsequent steps. The system recalculates the threshold every 0.01 seconds to ensure dynamic adaptation. For example, in the fifth window (order 5, remaining time 0.5 seconds), the threshold increases to 0.695, reflecting a higher tolerance in later stages. In this way, the system ensures that the threshold matches the decision-making stage, avoiding excessive constraints in the early stages or insufficient constraints in later stages.

[0062] S5353. When the real-time constraint influence coefficient is lower than the constraint influence threshold, the auxiliary scheme is optimized by reinforcement learning based on the real-time constraint influence coefficient to obtain the optimized auxiliary scheme. The reinforcement learning optimization adopts the Q-learning algorithm to minimize the deviation between the constraint influence coefficient and the threshold.

[0063] In step S5353, reinforcement learning optimization refers to optimizing the auxiliary scheme using the Q-learning algorithm when the constraint influence coefficient (obtained from S5351, range: 0-1) is lower than the constraint influence threshold (obtained from S5352), generating a more suitable auxiliary suggestion. The goal is to minimize the deviation between the constraint influence coefficient and the threshold. The Q-learning algorithm is based on state-action pairs. The state includes the constraint influence coefficient, remaining time (from S5352), and capability performance target (from S531). The action includes adjusting the suggestion content of the auxiliary scheme (such as reducing the number of suggestions or reducing the suggestion intensity). The reward function is defined as the absolute value of the deviation between the constraint influence coefficient and the threshold (target is less than 0.1). The optimization process is as follows: initialize the Q-table (based on 1000 simulation training), select an action according to the current state (e.g., reduce 1 suggestion), update the Q value (learning rate α = 0.1, discount factor γ = 0.9), and generate the optimized auxiliary scheme (e.g., simplify the suggestion from "adjust the valve to 85%, power 550kW" to "adjust the valve to 85%"). The constraint impact coefficients are obtained from S5351, the remaining time and capability objectives are obtained from S5352 and S531, and the Q-table is loaded from the pre-training file.

[0064] The specific construction process of the Q-learning algorithm includes: First, defining the state space as a triple S = {C, T, G}, where C represents the constraint influence coefficient (ranging from 0 to 1, obtained from S5351, reflecting the degree of constraint of the auxiliary scheme on the system), T represents the remaining time (in seconds, obtained from S5352, representing the remaining time of the decision window), and G represents the capability realization target (ranging from 0 to 100, obtained from S531, representing the system performance target). The action space A includes operations that adjust the auxiliary scheme, such as "reducing the number of suggestions" (reducing one suggestion each time) or "reducing the suggestion strength" (strength values ​​ranging from 0 to 10, decreasing by 1 each time). The reward function is defined as R = -|C - θ|, where θ is the constraint influence threshold (obtained from S5352, representing the maximum allowable constraint influence), and the reward objective is to make the deviation |C - θ| less than 0.1. The Q-table is a two-dimensional table that stores the expected cumulative reward of state-action pairs, with an initial value of zero, generated through 1000 simulations and stored in a pre-training file in JSON format. During training, an ε-greedy strategy is employed (ε = 0.1, meaning a 10% probability of randomly selecting an action for exploration and a 90% probability of selecting the action with the highest Q value to utilize existing knowledge). Q-value updates follow the formula Q(s, a) ← Q(s, a) + α[R + γmaxa'Q(s', a') - Q(s, a)], where s is the current state (i.e., {C, T, G}), a is the current action, s' is the new state after executing the action, a' is all possible actions in the new state, α = 0.1 is the learning rate (controlling the speed of Q-value updates), γ = 0.9 is the discount factor (measuring the importance of future rewards), and R is the current reward. In practical applications, the system loads the pre-trained Q-table, obtains the current state (e.g., {C = 0.25, T = 1.7, G = 65}), selects an action using the ε-greedy strategy (e.g., "reducing the number of suggestions"), and generates an optimized auxiliary scheme (e.g., simplifying "adjust the valve to 85%, power 550kW" to "adjust the valve to 85%). Subsequently, the system recalculates the constraint influence coefficient of the optimized solution (denoted as C', e.g., 0.20) and evaluates the deviation (e.g., |C'-θ|=|0.20-0.45|=0.25). The optimized solution is stored in a memory cache (key: "User ID Window 1.8-second optimization solution") and pushed to subsequent processes via WebSocket. The entire optimization process takes less than 0.05 seconds to ensure real-time performance. The Q table is periodically updated with new data to adapt to the dynamic environment, thereby efficiently generating concise and less disruptive auxiliary solutions.

[0065] For example, at time 1.8 seconds, the constraint influence coefficient is 0.25 (from S5351), and the constraint influence threshold is 0.45 (from S5352). Since 0.25 < 0.45, reinforcement learning optimization is triggered. The Q-learning algorithm (α = 0.1, γ = 0.9, Q-table based on 1000 simulations) takes the state {constraint influence: 0.25, remaining time: 1.7, capability target: 65} as input, selects the action "reduce the number of suggestions", and optimizes the auxiliary solution from "adjust the valve to 85%, power 550kW" to "adjust the valve to 85%". The constraint influence coefficient of the optimized solution is recalculated to 0.20, with a deviation of 0.25, meeting the target. The optimized solution is recorded in the memory cache (key: "user ID window 31.8 seconds_optimized solution") and pushed to subsequent steps. The optimization process takes 0.05 seconds, ensuring real-time performance. In this way, the system generates a simpler auxiliary solution, reducing interference with the user's decision-making.

[0066] S5354. Push the optimized auxiliary solution to the user in real time to guide the user's decision-making ability to continuously approach the target of the ability.

[0067] In step S5354, pushing the optimized assistance plan refers to displaying the optimized suggestions to the user in real time through the user interface (obtained from S5353) to guide the user's decision-making ability to approach the capability performance target (from S531). The push method uses the WebSocket protocol to ensure a latency of less than 0.1 seconds. The display format is a pop-up window or highlighted text, including suggested actions (e.g., "Adjust the valve to 85%)" and a brief explanation (e.g., "This can reduce the risk of overflow"). The push frequency is based on the assistance density (obtained from S51, unit: times / second). After each push, the user's response status is recorded (e.g., whether the suggestion was executed, extracted from the user interface log). The optimized assistance plan is obtained from S5353, the assistance density is obtained from S51, the user response status is collected from the user interface via WebSocket, and the capability target is obtained from S531.

[0068] For example, at time 1.8 seconds, the optimized assistance plan is "Adjust the valve to 85%" (from S5353), with an assistance density of 2.5 times / second (from S51). The system pushes a pop-up suggestion via WebSocket, "Please adjust the valve to 85% to reduce overflow risk," with a delay of 0.08 seconds. The user executes the suggestion at 1.9 seconds, adjusting the valve to 85%, and the system records the response status as "executed" in the interface log. After the push, the decision progress is updated to 60% (extracted from the user interface), and the capability is close to the target of 65% (from S531). The system continuously monitors the user's response, pushing a suggestion every 0.4 seconds to ensure that the user's decision-making ability is gradually improved. In this way, the system effectively assists the user in making decisions.

[0069] This invention significantly improves the management efficiency and intelligence level of urban drainage systems by integrating digital twin technology, machine learning algorithms, and a real-time decision support framework. A high-precision digital twin model built based on multimodal data enables real-time synchronous simulation of the drainage system, with model accuracy far exceeding traditional static models, ensuring the reliability and accuracy of optimized scheduling schemes. A pre-trained urban drainage optimization model, combined with deep reinforcement learning algorithms, generates dynamic scheduling schemes that balance efficiency and energy saving. The introduction of key decision node detection rules and a maximum decision-making capacity prediction model dynamically identifies high-risk scenarios and quantifies user decision-making capabilities, significantly improving decision accuracy. Through dynamic matching of decision time limits and auxiliary density, and reinforcement learning-based auxiliary scheme optimization, the system avoids information overload while ensuring user decision-making capabilities approach optimal levels, shortening user response time, and improving decision accuracy. Overall, this solution not only optimizes the operational efficiency and stability of the drainage system but also enhances the user decision-making experience through personalized auxiliary strategies, providing an efficient, intelligent, and sustainable solution for urban drainage management.

[0070] More importantly, this invention provides efficient and accurate support for users at key decision-making nodes in urban drainage optimization by dynamically matching decision-making time limits and assistance density, refining time window division, and optimizing assistance schemes in real time. Specifically, firstly, by setting dynamic decision-making time limits and assistance density based on the user's maximum decision-making ability (S51), combined with psychology and cognitive load theory, it ensures that users can fully utilize their decision-making abilities under moderate pressure, while avoiding information overload or insufficient support; secondly, the dynamic weighting of time windows (S52) adapts to the higher assistance tolerance at the end of the decision-making process, optimizing the timing of assistance; thirdly, through nonlinear regression modeling and real-time decision behavior analysis (S531-S532), it dynamically calculates the target of capability utilization and captures user behavior patterns, ensuring that the assistance scheme is highly matched with the user's current state; finally, through incremental information sensitivity analysis and reinforcement learning optimization (S5351-S5354), it adjusts the assistance scheme in real time to minimize interference with the user's autonomous decision-making, while guiding the user to gradually approach the capability target through low-latency push notifications. This layered and progressive auxiliary mechanism significantly improves decision-making efficiency and accuracy, reduces overflow risk, and demonstrates high real-time performance and adaptability in practical applications (such as drainage optimization in coastal cities), providing strong support for the intelligent management of complex drainage systems.

[0071] Figure 2 A schematic diagram of an urban drainage optimization management system based on multimodal data is provided for embodiments of this application, such as... Figure 2 As shown, the system includes: Module 100 is used to build a digital twin model based on multimodal data of urban drainage pipe networks; The optimization module 200 is used to start optimizing urban drainage based on the digital twin model; The detection module 300 is used to detect key decision nodes that require user intervention in the process of urban drainage optimization. Prediction module 400 predicts the user's maximum decision-making ability for the key decision-making nodes; The auxiliary module 500 is used to assist the user in making decisions on key decision nodes at a level that is closest to the maximum decision-making ability.

[0072] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for optimizing urban drainage management based on multimodal data, characterized in that: include: A digital twin model is constructed based on multimodal data of urban drainage pipe networks; Based on the aforementioned digital twin model, urban drainage optimization was initiated. Detect key decision-making nodes that require user intervention in the process of urban drainage optimization; Predict the user's maximum decision-making ability at the key decision-making nodes; It assists users in making decisions at key decision points at a level that is closest to their maximum decision-making ability.

2. The urban drainage optimization management method based on multimodal data as described in claim 1, characterized in that, The construction of a digital twin model based on multimodal data from urban drainage networks includes: Based on the digital twin engine, a digital twin model is constructed using multimodal data from the urban drainage network.

3. The urban drainage optimization management method based on multimodal data as described in claim 1, characterized in that, The process of optimizing urban drainage based on the digital twin model includes: Based on the pre-trained urban drainage optimization model, urban drainage optimization begins according to the digital twin model.

4. The urban drainage optimization management method based on multimodal data as described in claim 1, characterized in that, The key decision-making nodes requiring user intervention in the process of optimizing urban drainage include: Based on the key decision node detection rules, key decision nodes that require user intervention in the process of urban drainage optimization are detected.

5. The urban drainage optimization management method based on multimodal data as described in claim 1, characterized in that, The prediction of a user's maximum decision-making ability at the key decision-making node includes: Based on a pre-trained maximum decision-making capability prediction model, the maximum decision-making capability of a user at the key decision nodes is predicted according to the user's urban drainage optimization decision-making history.

6. The urban drainage optimization management method based on multimodal data as described in claim 1, characterized in that, The method of assisting users in making decisions at key decision nodes at a level closest to their maximum decision-making ability includes: Based on the predicted maximum decision-making ability of users at key decision nodes, the decision time limit and the density of assistance are dynamically matched. The time period from the current time to the decision deadline within the decision time limit is divided into multiple time windows according to the window order dynamic weight function; For each of the aforementioned time windows, when the latest current time enters that time window, perform the following operations: Based on the user's current decision-making progress at key decision nodes, the user's decision-making ability demonstrated in previous time windows, the predicted maximum decision-making ability of the user at key decision nodes, and the window order of the time window, the ability performance target within the time window is dynamically calculated and determined through decision performance trajectory modeling. Real-time acquisition of the user's historical decision-making behavior sequence based on the digital twin model within the current time window; Based on the historical decision-making behavior sequence and the preset decision-assistance trigger rule base, dynamically determine whether the user triggers the decision-assistance process; If the decision support process is triggered, the support scheme and the latest support time limit are accurately determined by the real-time decision situation awareness engine based on the historical decision behavior sequence, the remaining time of the current time window, and the capability performance target. For each time point before the latest auxiliary time limit, perform the following operations: Based on the incremental information of the changes in the historical decision-making behavior sequence to the current time node, the real-time constraint influence coefficient of the auxiliary logic of the auxiliary scheme on the user's decision-making ability is calculated through incremental information sensitivity analysis. Based on the window order of the time window and the remaining time of the current time window, the threshold is dynamically matched and influenced by the constraint through the adaptive threshold decay rule. When the real-time constraint influence coefficient is lower than the constraint influence threshold, the auxiliary scheme is optimized by reinforcement learning based on the real-time constraint influence coefficient to obtain the optimized auxiliary scheme. The reinforcement learning optimization adopts the Q-learning algorithm to minimize the deviation between the constraint influence coefficient and the threshold. The optimized assistance plan is pushed to the user in real time to guide the user's decision-making ability to continuously approach the target of the capability.

7. The urban drainage optimization management method based on multimodal data as described in claim 6, characterized in that, The decision-making time limit is set based on a user decision-making pressure threshold model to maximize decision-making ability, while the auxiliary density is set based on a user cognitive load tolerance model to avoid excessive interference or insufficient support from auxiliary behaviors in the decision-making process.

8. The urban drainage optimization management method based on multimodal data as described in claim 6, characterized in that, The later the window is in the time window order, the larger its weight coefficient, in order to accommodate the increased tolerance for auxiliary interference at the end of the decision-making process.

9. The urban drainage optimization management method based on multimodal data as described in claim 6, characterized in that, The trajectory modeling of decision-making ability is achieved by using a nonlinear regression algorithm to fit the user's historical decision-making ability release curve.

10. A city drainage optimization management system based on multimodal data, characterized in that, include: The building module is used to construct digital twin models based on multimodal data of urban drainage networks; An optimization module is used to begin optimizing urban drainage based on the digital twin model. The detection module is used to detect key decision-making nodes that require user intervention in the process of urban drainage optimization. The prediction module predicts the user's maximum decision-making ability at the key decision nodes. An auxiliary module is used to assist users in making decisions at key decision nodes at a level that is closest to the maximum decision-making ability.

Images