Intelligent exhaust port dynamic regulation and control system and method based on multi-source time sequence fusion

Abstract

The invention discloses an intelligent discharge port dynamic regulation and control system and method based on multi-source time sequence fusion, and relates to the technical field of intelligent water affairs, and the system comprises a data collection module which is used for obtaining the multi-source data of a target area in real time; the space-time fusion module is used for mapping the meteorological radar data to the topological space of the drainage pipe network through a space-time alignment encoder, dynamically fusing the meteorological radar data, the surface runoff sensor data and the historical drainage data by adopting a multi-head attention mechanism, and generating a space-time fusion tensor; the prediction model module is used for constructing a TCN-LSTM hybrid model, taking the space-time fusion tensor as input and taking future drainage flow as output; and the dynamic regulation and control module is used for optimizing the gate opening degree and the pump station power control sequence in a rolling manner based on a model predictive control algorithm, and performing dual-target optimization by combining real-time electricity price data and overflow risks. According to the system, the prediction precision of the drainage flow can be improved, and double-target balance of the overflow risk and the energy consumption cost can be achieved.

Description

Technical Field

[0001] The present invention relates to the field of smart water technology, and specifically to an intelligent outlet dynamic control system and method based on multi-source time series fusion, which is used for real-time optimization control and waterlogging risk prevention and control of urban drainage systems. Background Art

[0002] With the acceleration of urbanization, urban drainage systems face multiple challenges, including frequent extreme weather events, aging pipe networks, and low control efficiency. Traditional drainage control methods, which rely primarily on manual experience or preset rules, are unable to cope with dynamically changing rainfall conditions and complex pipe network hydraulic characteristics, leading to both the risk of urban flooding and energy waste. Existing technologies have the following major problems:

[0003] 1. Single-source data utilization: Traditional methods often rely on a single data source (such as historical flow data or weather forecasts) and lack the ability to analyze the spatiotemporal correlations between multiple data sources (such as weather, runoff, and soil moisture). This makes it difficult to capture the complex coupling relationship between rainfall, runoff, and pipe network response. With the rapid development of information technology and data science, deep learning technology has been widely used in various application fields. Although some sensor-based automatic monitoring systems exist, they often issue alerts based on specific rules and thresholds, lacking sufficient flexibility and adaptability.

[0004] ② Inadequate model prediction capabilities: Traditional time series models (such as LSTM and ARIMA) cannot effectively handle the long-term spatiotemporal dependencies of meteorological data. While physical models (such as SWMM) can simulate hydraulic processes, their high computational complexity makes them difficult to meet the demands of real-time regulation. Furthermore, existing models generally lack adaptability to extreme scenarios and the ability to generalize across cities.

[0005] ③ Static control strategies: Control methods based on PID or fixed thresholds are unable to cope with dynamically changing rainfall intensity and pipe network loads, resulting in the inability to simultaneously optimize overflow risk and pump station energy consumption. Economic factors such as real-time electricity prices are also not included in control objectives.

[0006] ④ The system has weak real-time and robustness: Traditional solutions rely on centralized cloud computing, which has problems such as data transmission delays and insufficient computing power of edge devices; when faced with emergencies such as sensor anomalies or extreme rainfall, there is a lack of effective redundant control mechanisms. Summary of the Invention

[0007] In view of this, the present invention provides an intelligent outlet dynamic control system and method based on multi-source time series fusion, which solves the above problems through multi-source data fusion, hybrid prediction model, dynamic optimization algorithm and edge-cloud collaborative architecture.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] In a first aspect, an embodiment of the present invention provides an intelligent outlet dynamic control system based on multi-source time series fusion, comprising:

[0010] Data acquisition module, used to obtain multi-source data of the target area in real time, including weather radar data, surface runoff sensor data, historical drainage data and real-time electricity price data;

[0011] The spatiotemporal fusion module maps weather radar data to the drainage network topology space through a spatiotemporal alignment encoder and dynamically fuses weather radar data, surface runoff sensor data, and historical drainage data using a multi-head attention mechanism to generate a spatiotemporal fusion tensor.

[0012] The prediction model module constructs a TCN-LSTM hybrid model, taking the spatiotemporal fusion tensor as input and the future drainage flow as output. The model processes the spatiotemporal fusion tensor through a TCN layer to capture long-range spatiotemporal dependency features, and its output serves as input to the LSTM layer to model the short-term dynamic fluctuations of drainage flow. The mass conservation equation is embedded as a physical constraint during training.

[0013] The dynamic control module uses the model predictive control algorithm to rollingly optimize the gate opening and pump station power control sequence, and combines real-time electricity price data and overflow risk to perform dual-objective optimization.

[0014] In one embodiment, it further includes:

[0015] The safety redundancy module is used to detect the prediction confidence of the prediction model module. When the confidence is lower than the threshold, it switches to PID control and triggers manual review.

[0016] In one embodiment, the system adopts an edge-cloud collaborative architecture, including edge computing devices and a cloud management platform; wherein, the edge device deploys a lightweight TCN-LSTM model and adopts an event-driven triggering mechanism, and the cloud platform supports transfer learning and online incremental learning, and provides a visual monitoring interface.

[0017] In one embodiment, the spatiotemporal fusion module includes:

[0018] Bilinear interpolation unit, which maps the 2D grid rainfall data from the weather radar to the coordinates of the drainage network nodes;

[0019] Graph attention network unit, which is used to aggregate meteorological features and pipe network topology attributes of adjacent nodes to generate spatially correlated meteorological feature vectors;

[0020] The Transformer encoder dynamically fuses meteorological, runoff, and historical drainage data through a multi-head attention mechanism and outputs a spatiotemporal fusion tensor.

[0021] In one embodiment, in the prediction model module, the TCN-LSTM hybrid model is trained by the following steps:

[0022] In the pre-training phase, a synthetic dataset generated by SWMM simulation is used for training, covering normal rainfall, heavy rain, and pipe blockage scenarios;

[0023] During the fine-tuning phase, extreme rainfall noise data is injected to improve the model robustness through adversarial training.

[0024] In the transfer learning phase, the TCN layer weights are fixed, and the LSTM layer and fully connected layer are fine-tuned using the target city data.

[0025] In one embodiment, in the prediction model module, the loss function of the TCN-LSTM hybrid model is:

[0026]

[0027] in, represents the total loss function, is the mean square error, which measures the difference between the model prediction value and the true value; is a physical constraint term. ∇⋅(Q⋅v) in the formula is the divergence of the vector field Q⋅v, ∇ is the divergence operator; Q is the instantaneous volume flow rate of a node or pipe cross section in the drainage network, and v is the instantaneous flow velocity of the water flow in the drainage network; is the weighted penalty term for adversarial sample prediction error; α, β, is a hyperparameter.

[0028] In one embodiment, the objective optimization function of the dynamic control module is as follows:

[0029]

[0030] Where N is the total number of time steps, Forecast flow for step k, is the safe flow threshold, is the pump station power, The real-time electricity price, is the weight coefficient.

[0031] In one embodiment, the safety redundancy module generates a prediction result distribution through Bootstrap sampling, and triggers a control mode switch when the traffic prediction variance exceeds a threshold or the prediction deviation is greater than 15% for multiple consecutive times.

[0032] In one embodiment, the edge computing device compresses the model volume through channel pruning and 8-bit integer quantization technology, and activates the prediction model only when the traffic exceeds the threshold or the sensor is abnormal.

[0033] In a second aspect, an embodiment of the present invention further provides an intelligent outlet dynamic control method based on multi-source timing fusion, using the intelligent outlet dynamic control system based on multi-source timing fusion as described in any one of the first aspects, the method comprising:

[0034] S1. Real-time collection of multi-source data in the target area, including weather radar data, surface runoff sensor data, and historical drainage data, and pre-processing;

[0035] S2. Fusion of multi-source data through spatiotemporal alignment encoder and multi-head attention mechanism to generate spatiotemporal fusion tensor;

[0036] S3. Inputting the fused tensor into the constructed TCN-LSTM hybrid model to predict future drainage flow;

[0037] S4. Based on the model predictive control algorithm, the gate opening and pump station power control sequence are continuously optimized, and dynamic regulation is performed based on real-time electricity prices and overflow risks.

[0038] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following technical advantages:

[0039] 1. Efficient data utilization: Integrating meteorological, runoff, historical drainage, and real-time electricity price data, and generating high-value fusion tensors through spatiotemporal alignment and multi-head attention mechanisms, significantly improves data relevance and information dimensionality, providing comprehensive input for prediction models.

[0040] 2. Improved prediction accuracy: The TCN-LSTM hybrid model combines long-term spatiotemporal dependency capture with short-term dynamic modeling and embeds a physically constrained loss function. This ensures that prediction results conform to both data patterns and fluid mechanics principles, improving prediction accuracy.

[0041] 3. Dynamic optimization and control: Based on the MPC algorithm, the gate opening and pump station power are continuously optimized, and the dual-objective balance between overflow risk and energy consumption cost is achieved in combination with real-time electricity prices. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0043] Figure 1 This is a block diagram of the intelligent outlet dynamic control system based on multi-source timing fusion provided by the present invention.

[0044] Figure 2This is a schematic diagram of the process flow of the intelligent outlet dynamic control system based on multi-source timing fusion provided by the present invention.

[0045] Figure 3 Schematic diagram of the spatiotemporal alignment encoder provided by the present invention.

[0046] Figure 4 This is a diagram of the TCN-LSTM hybrid model architecture provided by the present invention.

[0047] Figure 5 This is a flow chart of the MPC dynamic control strategy provided by the present invention.

[0048] Figure 6 This is a diagram of the internal structure of the edge computing device provided by the present invention.

[0049] Figure 7 Flowchart of the intelligent outlet dynamic control method based on multi-source timing fusion provided by the present invention. DETAILED DESCRIPTION

[0050] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0051] The embodiment of the present invention discloses an intelligent outlet dynamic control system based on multi-source time series fusion, referring to Figure 1 As shown, including:

[0052] Data acquisition module, used to obtain multi-source data of the target area in real time, including weather radar data, surface runoff sensor data, historical drainage data and real-time electricity price data;

[0053] The spatiotemporal fusion module maps weather radar data to the drainage network topology space through a spatiotemporal alignment encoder and dynamically fuses weather radar data, surface runoff sensor data, and historical drainage data using a multi-head attention mechanism to generate a spatiotemporal fusion tensor.

[0054] The prediction model module constructs a TCN-LSTM hybrid model, taking the spatiotemporal fusion tensor as input and the future drainage flow as output. The model processes the spatiotemporal fusion tensor through a TCN layer to capture long-range spatiotemporal dependency features, and its output serves as input to the LSTM layer to model the short-term dynamic fluctuations of drainage flow. The mass conservation equation is embedded as a physical constraint during training.

[0055] The dynamic control module uses the model predictive control algorithm to rollingly optimize the gate opening and pump station power control sequence, and combines real-time electricity price data and overflow risk to perform dual-objective optimization.

[0056] The safety redundancy module is used to detect the prediction confidence of the prediction model module. When the confidence is lower than the threshold, it switches to PID control and triggers manual review.

[0057] The entire system adopts an edge-cloud collaborative architecture, including edge computing devices and a cloud management platform. The edge devices deploy a lightweight TCN-LSTM model and adopt an event-driven triggering mechanism. The cloud platform supports transfer learning and online incremental learning, and provides a visual monitoring interface.

[0058] The intelligent outlet dynamic control system based on multi-source time series fusion, provided by this invention, integrates weather radar, surface runoff sensors, and historical data, building an intelligent control system through spatiotemporal data fusion technology. A dynamic feature analysis model is used to extract key information from multi-source data. This is combined with a hybrid prediction model to predict drainage demand. Physical laws are used to optimize prediction results and enhance adaptability to extreme weather conditions. Then, based on an intelligent optimization algorithm, gate opening and pump station power are adjusted in real time, reducing the risk of flooding while also reducing energy consumption. A safety monitoring mechanism is also included to ensure stable system operation.

[0059] The following combination Figure 2 The overall architecture process of the system is described as follows: Figure 2 It reflects the system's layered architecture and data flow logic, and embodies the closed-loop design of "data → prediction → regulation → feedback".

[0060] 1. Data acquisition module, realizing multi-source data acquisition and preprocessing:

[0061] For example, by accessing the Meteorological Bureau's API, real-time spatiotemporal grid data on rainfall intensity for the next 1-6 hours is obtained with a spatial resolution of ≤1 km² and a time granularity of 5 minutes. After normalization, this data is stored in a tensor format consisting of time × longitude × latitude × rainfall intensity. A LoRaWAN wireless sensor network is deployed simultaneously to collect surface runoff rate (m³ / s) and soil saturation (%) at a frequency of once per minute. Outliers are removed and data is smoothed using a sliding window averaging method. Sensor deployment adopts a three-dimensional layout strategy: "node-region-function." Flow sensors are precisely deployed at the throat of the municipal drainage network, such as outlets, sewage pumping stations, and intersection nodes of trunk and branch networks, where runoff initially converges, to capture runoff data entering the network in real time, providing a core basis for the scientific scheduling of the drainage system. For high-risk areas for urban flooding, such as low-lying road sections, underground traffic passages, and sunken commercial plazas, a multi-dimensional monitoring network covering ground and underground spaces is constructed to achieve dynamic perception and early warning of water level changes in seconds. Surface runoff monitoring terminals are installed at urban road watershed nodes and river overflow-sensitive areas to track the migration path and convergence trend of water flow throughout the entire process. In sponge facility areas with significant soil infiltration functions, such as urban green spaces and parks, soil moisture content and groundwater level sensors are deployed to evaluate the operating efficiency of ecological storage facilities through quantitative monitoring, thereby forming a full-chain data perception system from source monitoring to terminal treatment, providing solid data support for urban flood control, drainage, and water environment management.

[0062] In addition, historical outlet flow, gate opening and pump station power data with a time granularity of 5 minutes can be exported from the Supervisory Control And Data Acquisition (SCADA) system, and Z-score normalization processing can be performed after missing values are filled by linear interpolation. Electricity price data can also be obtained to finally form a standardized time series.

[0063] The system simultaneously collects data from multiple sources. For one thing, it acquires spatiotemporal grid data on future rainfall intensity from weather radar. This data provides information on the temporal and spatial distribution of rainfall, helping to predict its impact on the drainage system. Furthermore, surface runoff sensors monitor runoff rate and soil saturation in real time. These two parameters provide a direct reflection of surface water flow conditions and the soil's water storage capacity. Furthermore, it collects time-series records of historical outlet flows and gate openings, which reveal past operational patterns of the drainage system. Real-time electricity price data is also collected to optimize energy costs.

[0064] 2. The spatiotemporal fusion module realizes spatiotemporal alignment and multi-source data fusion; solves the spatiotemporal heterogeneity problem of multi-source data, and enhances the spatial correlation of input data and prediction accuracy.

[0065] To effectively integrate data from these various sources, a spatiotemporal alignment encoder is first used to map the two-dimensional grid data from weather radar to the physical topological space of the drainage network. Next, a multi-head attention mechanism is employed to dynamically calculate the importance weights of features from each data source. This mechanism deeply captures long-range dependencies between data sources and fully exploits the inherent connections between them. Ultimately, a multi-dimensional fusion tensor is generated that incorporates meteorological, runoff, and historical regulation patterns, providing valuable input for subsequent prediction models.

[0066] Specific reference Figure 3 As shown, the spatiotemporal fusion module includes:

[0067] 1) Bilinear interpolation unit, which maps the 2D grid rainfall data from the weather radar to the coordinates of the drainage network nodes;

[0068] 2) Graph Attention Network Unit, which aggregates meteorological features and pipe network topology attributes of adjacent nodes to generate spatially correlated meteorological feature vectors;

[0069] 3) Transformer encoder, which dynamically fuses meteorological, runoff, and historical drainage data through a multi-head attention mechanism and outputs a spatiotemporal fusion tensor.

[0070] For example, first construct a drainage network topology map, with outlets / pump stations as nodes and pipeline connections as edges, and store coordinate information in GeoJSON format; then use bilinear interpolation to map the meteorological grid rainfall intensity to the coordinates of each outlet node. The calculation formula is:

[0071]

[0072] in, is (grid point weights distributed in inverse proportion to distance), It represents the rainfall intensity values of the four neighboring grid points around the target node (outlet node) during the bilinear interpolation process; then, the graph attention network (GAT) is used to input the node meteorological characteristics (rainfall intensity) and edge attributes (pipeline length, slope), and a meteorological feature vector with spatial correlation is generated by calculating the attention weights between nodes; finally, the meteorological characteristics, runoff data (runoff rate, soil saturation) and historical drainage data (flow rate, gate opening) are input into the Transformer encoder, and after dynamic fusion through the multi-head attention mechanism, a spatiotemporal fusion tensor of time × node × feature dimensions is output as the input of the subsequent prediction model.

[0073] 3. Prediction model module:

[0074] Construct a TCN-LSTM hybrid model for training and optimization. In this hybrid model, TCN and LSTM adopt a series architecture, referring to Figure 4As shown in the figure, a hybrid architecture consisting of TCN and LSTM layers is constructed. The TCN layer is configured with a dilation factor d=2, a convolution kernel size k=3, and four layers of dilated causal convolution to capture the long-range spatiotemporal dependencies of meteorological data. The LSTM layer has 128 hidden units, and the model input is a spatiotemporal fusion tensor with dimensions of time steps × number of nodes × feature dimensions (for example, 12 steps of historical data × N outlet nodes × number of multi-source features). The LSTM output is mapped to the future prediction number (6 steps) through a fully connected layer, resulting in an output with dimensions of future time steps × number of nodes × number of target variables (for example, 6 steps × N nodes × 1, representing the predicted flow rate). The LSTM layer is used to model the short-term dynamic fluctuations of drainage flow.

[0075] In addition, the mass conservation equation can be embedded in the loss function as a physical constraint layer.

[0076] During the pre-training phase, SWMM simulation was used to generate 100,000 pieces of synthetic data covering scenarios such as normal rainfall, heavy rain, and pipe blockage. The SWMM model can simulate various working conditions of urban drainage systems and generate a large amount of representative data, which helps the model learn the basic laws of drainage systems.

[0077] During the fine-tuning phase, adversarial training was performed by injecting extreme rainfall noise (increasing its intensity by 200% and its duration by 50%) to improve model robustness. The adversarial training process first involves data mixing, mixing normal data with adversarial samples in a 7:3 ratio to construct a diverse training dataset. Subsequently, the loss function was adjusted and optimized. Based on the original mean squared error (MSE) loss function, which incorporates physical constraints, an adversarial loss term was introduced to form a new overall loss function:

[0078]

[0079] in, represents the total loss function, is the mean square error, which measures the difference between the model prediction value and the true value; is a physical constraint term. ∇⋅(Q⋅v) in the formula computes the divergence of the vector field Q⋅v. ∇ serves as the core of the divergence operator (the nabla operator is used to construct the divergence operation). Q is the instantaneous volume flow rate at a node or pipe cross section in the drainage network, and v is the instantaneous flow velocity of the water in the drainage network. is the weighted penalty term for adversarial sample prediction error; α, β, is a hyperparameter.

[0080] During the fine-tuning phase, an adversarial training mechanism was introduced to improve the model's robustness by adding extreme rainfall noise data. While extreme rainfall is rare in reality, it has a significant impact on drainage systems. This approach allows the model to better cope with various complex rainfall scenarios.

[0081] Finally, we enter the adversarial training iteration stage. In each round of training, we first use normal data to update the model parameters, and then perform backpropagation calculations through adversarial samples to further optimize the model's adaptability to extreme events, thereby improving the model's robustness and generalization performance.

[0082] This model combines data-driven insights with physical laws to improve predictive interpretability and adaptability to extreme scenarios. Furthermore, the model supports cross-city transfer learning, employing an architecture that shares feature extractors and independently fine-tunes classifiers. This allows for rapid adaptation to the specific drainage systems of different cities by leveraging drainage data from other cities, improving the model's generalization capabilities.

[0083] Transfer learning is an effective method for improving model performance across different data domains. The following details its application in a drainage system model. First, the weights of the TCN layer are fixed. Because the TCN layer acts as a universal feature extractor, it has already learned important spatiotemporal patterns, such as rainfall-runoff correlations, during pre-training. To prevent target domain data (e.g., data from a new city) from corrupting these universal features, we freeze the TCN layer parameters. Next, the LSTM and fully connected layers are retrained using historical drainage data from the target city, such as flow rates and gate openings. During training, hyperparameters are fine-tuned, with the learning rate set to 1e-4, a lower value than during pre-training. Adam is used as the optimizer, and training is performed for 50 epochs.

[0084] To better achieve domain adaptation, two strategies can be employed. First, feature alignment uses the Maximum Mean Divergence (MMD) loss to effectively reduce the difference in feature distribution between the source domain (synthetic data) and the target domain (real data). Second, incremental learning uses elastic weight consolidation (EWC) technology on the cloud platform to prevent the loss of knowledge learned in the source domain during training in the target domain. Through these steps and strategies, transfer learning can help the model achieve better performance on drainage system data from different cities.

[0085] 4. Dynamic control module and safety redundancy module, used to implement model predictive control (MPC) and safety control;

[0086] Reference Figure 5As shown in Figure 1, the prediction model input is the traffic forecast results for the next N steps (e.g., 6 steps, 5 minutes each). This enables dynamic real-time control and balances safety and economy.

[0087] First, with the goal of minimizing the overflow risk (assuming the maximum flow Q_max = 1200m³ / h, the overflow threshold is set to 0.85Q_max = 1020m³ / h) and energy consumption costs, the IPOPT nonlinear optimization library is used to rolling optimize the gate opening and pump station power control sequence for the next N time steps every 5 minutes; the MPC algorithm will continuously adjust the control strategy based on the output of the prediction model to adapt to the dynamic changes of the drainage system. When the prediction confidence is lower than the preset threshold, the system will automatically switch to PID (Proportional-Integral-Derivative) control and trigger the manual review mechanism. This is to ensure the safe and stable operation of the drainage system when the prediction results are unreliable. For example, 100 sets of prediction results are generated through Bootstrap sampling and the flow prediction variance is calculated. , such as when When the prediction error exceeds 50 or three consecutive predictions have a deviation greater than 15% (corresponding to a confidence level less than 90%), the safety redundancy mechanism is triggered. The system automatically switches to PID control mode and sends an alert to the operation and maintenance personnel, thus ensuring dynamic optimization of control while establishing a dual closed-loop safety protection system. This mechanism establishes a dynamic balance between intelligent decision-making and safe operation by quantifying prediction uncertainty and seamlessly switching between control modes.

[0088] The dynamic control strategy uses a collaborative architecture of MPC and safety redundancy mechanisms to achieve intelligent control of the drainage system. Model predictive control (MPC) constructs a multi-objective optimization function by rolling-optimizing the gate opening and pump station power sequence for the next N time steps (5-30 minutes):

[0089]

[0090] Where N is the total number of time steps, Forecast flow for step k, is the safe flow threshold, is the pump station power, The real-time electricity price, The system updates the control sequence based on the actual sensor data every 5 minutes and eliminates the prediction deviation through the feedback correction mechanism.

[0091] Reference Figure 6As shown in the figure, the intelligent outlet dynamic control system based on multi-source time series fusion provided by the present invention adopts an edge-cloud collaborative layered architecture to achieve efficient collaboration during its specific implementation. It involves three hardware components: a data acquisition terminal, an edge computing device, and a cloud management platform that work together. The edge computing device improves operational efficiency and saves resources through model compression and event-driven triggers. The cloud management platform supports transfer learning and online incremental learning, and provides a visual monitoring interface, realizing a closed-loop intelligent system of "real-time decision-making at the edge - continuous optimization in the cloud", improving the overall operational efficiency and management convenience of the system.

[0092] Specifically, on the edge device side, based on the NVIDIA Jetson Xavier hardware platform equipped with the TensorRT acceleration engine, the TCN-LSTM model is compressed to 15MB through channel pruning (retaining 80% of key channels) and 8-bit integer quantization technology, reducing the model's computational complexity and storage requirements, and improving the model's operating efficiency on edge devices.

[0093] The event-driven prediction mechanism is triggered only when the flow rate is greater than 800m³ / h or the sensor data is abnormal; the cloud management platform, for example, uses elastic weight consolidation (EWC) at 1 a.m. every day. =100) algorithm performs online incremental learning to update the model, while providing a real-time monitoring interface: the blue curve shows the outlet flow prediction value, and the red curve shows the actual value. High-risk nodes with flow rates greater than 1020m³ / h are highlighted for early warning, and the prediction confidence level is dynamically annotated with green / yellow / red labels, forming a closed-loop intelligent system of "real-time decision-making at the edge and continuous optimization in the cloud."

[0094] In this embodiment, edge computing devices utilize lightweight model deployment (through channel pruning and 8-bit quantization to optimize the TCN-LSTM model), equipped with event-driven triggers (activating the prediction model only when flow exceeds a threshold or sensor anomalies occur) and a safety redundant controller (implementing MPC and PID switching logic). The cloud management platform supports transfer learning and online incremental learning (using elastic weight solidification technology to prevent catastrophic forgetting). A visual monitoring interface displays outlet status, prediction curves, and control instructions in real time, supporting manual intervention. This system, through a layered architecture design, achieves the coordinated operation of data collection, edge intelligence, and cloud optimization, ensuring efficient and stable operation of the drainage system.

[0095] System testing and verification ensure the effectiveness of the technology through a multi-dimensional verification system: the offline test is based on a dataset of 32 rainfall events (including 7 flooding records) in Shenzhen Futian District during the flood season in 2022. The results show that the MSE is reduced by 32% compared with the traditional LSTM model, and the overflow prediction accuracy of heavy rain scenarios reaches 89.7%; the online pilot is deployed in Singapore PUB drainage system. The 2023 rainy season operation data shows that the incidence of flooding is reduced by 22%, the average daily energy consumption of the pump station is reduced by 17%, and the energy consumption optimization rate during peak and valley electricity price periods is 41%; in the fault recovery test, when simulating communication interruption, the edge end automatically switches to local PID control, and the overflow risk of key outlets is still controlled within 5%, verifying the robustness and practicality of the system under complex working conditions.

[0096] In addition, system maintenance and updates adopt a preventive maintenance strategy and a dynamic model iteration mechanism: monthly sensor accuracy calibration is performed (an automatic calibration process is triggered when the error is greater than 5%) and silt at the outlet is cleared to ensure the reliability of data collection and physical equipment; drainage characteristic data of new cities (such as Rotterdam) are collected every quarter, and model parameters are updated through transfer learning technology to achieve cross-regional drainage characteristic adaptation and continuously optimize system performance.

[0097] The intelligent outlet dynamic control system based on multi-source time series fusion provided by the present invention has the following application scenarios:

[0098] To prevent and control urban flooding risks, the system integrates real-time weather radar data (which provides information on the temporal and spatial distribution of rainfall over the next 1-6 hours), surface runoff sensor data (covering runoff rate and soil saturation), and historical drainage data (including flow rate and gate openings), to dynamically predict drainage needs and precisely adjust outlet gate openings and pump station power, thereby mitigating the risk of overflows caused by extreme rainfall. In heavy rain scenarios, a physically constrained TCN-LSTM hybrid model predicts flow surge trends and proactively triggers pump station power optimization strategies. This has yielded significant results, reducing the incidence of flooding by 22% in measured cases.

[0099] In the drainage system energy efficiency optimization process, a model predictive control (MPC) algorithm is used to continuously optimize the control sequence, combining real-time electricity prices with pump station energy consumption, to achieve a dual-objective balance between overflow risk and energy costs. For example, by intelligently allocating pump station operating power during peak and off-peak electricity price periods, the pilot application achieved an energy consumption optimization rate of 41%. Furthermore, edge computing devices use event-driven triggers to activate the predictive model only when flow exceeds a threshold or a sensor anomaly occurs, reducing redundant calculations and conserving energy and computing resources.

[0100] To address extreme weather and complex scenarios, adversarial training techniques inject extreme rainfall noise (such as a rainstorm pattern with a 200% increase in intensity) to improve the model's robustness in predicting once-in-a-century rainfall events, ensuring stable system operation under abnormal conditions. A safety redundancy mechanism monitors prediction confidence in real time (based on Bootstrap sampling variance). When the confidence falls below a threshold, it automatically switches to PID control and triggers manual review, establishing a dual closed-loop protection system of "intelligent prediction + safety backup."

[0101] For rapid cross-regional deployment and migration adaptation, the system supports transfer learning technology. By sharing feature extractors and independently fine-tuning classifiers, it leverages drainage data from other cities to quickly adapt the characteristics of the target city's pipeline network (for example, migrating from Shenzhen to Rotterdam), reducing customized development costs. The cloud platform uses Elastic Weight Consolidation (EWC) technology for online incremental learning, continuously optimizing model performance to adapt to seasonal rainfall patterns in different regions.

[0102] Intelligent operations and maintenance, along with real-time monitoring, are equally crucial. A visual cloud-based management interface dynamically displays outlet flow forecast curves, actual operating data, and control instructions. Prediction confidence levels are indicated by red, yellow, and green labels, assisting operations personnel in making quick decisions. Preventive maintenance strategies, such as automatic sensor calibration (triggered when errors exceed 5%) and regular cleaning of outlet sediment, are also supported to ensure reliable data collection and equipment operation.

[0103] Based on the same inventive concept, the present invention also provides a method for dynamic regulation of intelligent drainage outlets based on multi-source time series fusion. Using the intelligent dynamic regulation system of drainage outlets based on multi-source time series fusion of the above embodiment, this method provides an efficient solution for urban drainage management and complies with industrial safety standards. Figure 7 As shown, the method specifically includes:

[0104] S1. Real-time collection of multi-source data in the target area, including weather radar data, surface runoff sensor data, and historical drainage data, and pre-processing;

[0105] S2. Fusion of multi-source data through spatiotemporal alignment encoder and multi-head attention mechanism to generate spatiotemporal fusion tensor;

[0106] S3. Inputting the spatiotemporal fusion tensor into the constructed TCN-LSTM hybrid model to predict future drainage flow;

[0107] S4. Based on the model predictive control algorithm, the gate opening and pump station power control sequence are continuously optimized, and dynamic regulation is performed based on real-time electricity prices and overflow risks.

[0108] In the embodiment of the present invention, steps S1 and S2, by integrating weather radar, surface runoff sensor and historical data, and using spatiotemporal alignment encoder and multi-head attention mechanism to perform multi-source data fusion, can fully explore the intrinsic connection between different data sources, generate a fusion tensor containing multi-dimensional information, provide more valuable data for the prediction model, improve data utilization efficiency, and accurately capture the complex relationship between rainfall-runoff-pipeline network response.

[0109] In step S3, a TCN-LSTM hybrid model, combined with dilated causal convolution and a gating mechanism, captures the long-range spatiotemporal dependencies of meteorological data and the short-term dynamic fluctuations of drainage flow. Furthermore, the mass conservation equation is embedded as a physical constraint layer, ensuring that the prediction results conform to both data-driven principles and fluid dynamics. This enhances the model's interpretability and generalization capabilities, effectively improving prediction accuracy.

[0110] A multi-stage training framework is adopted during the training process, including pre-training based on the SWMM synthetic dataset, adversarial training with extreme rainfall noise data, and a modular architecture design that supports cross-city transfer learning. This enables the model to learn the basic laws of drainage systems, enhance robustness to abnormal working conditions, and quickly adapt to the characteristic differences of drainage systems in different cities, reducing the cost of customized development.

[0111] Step S4 implements dynamic optimization and control. Based on the MPC algorithm, the system continuously optimizes gate openings and pump station power control sequences for multiple future time steps, minimizing overflow risk and energy costs. This optimization is combined with economic factors such as real-time electricity prices to achieve dynamic optimization and control of the drainage system. Furthermore, a prediction confidence assessment and safety redundancy mechanism are implemented. If prediction results are unreliable, the system automatically switches to PID control and triggers manual review, ensuring safe and stable operation of the system.

[0112] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0113] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. Intelligent outlet dynamic control system based on multi-source time series fusion, characterized by: include: Data acquisition module, used to obtain multi-source data of the target area in real time, including weather radar data, surface runoff sensor data, historical drainage data and real-time electricity price data; The spatiotemporal fusion module maps weather radar data to the drainage network topology space through a spatiotemporal alignment encoder and dynamically fuses weather radar data, surface runoff sensor data, and historical drainage data using a multi-head attention mechanism to generate a spatiotemporal fusion tensor. The prediction model module constructs a TCN-LSTM hybrid model, taking the spatiotemporal fusion tensor as input and the future drainage flow as output. The model processes the spatiotemporal fusion tensor through a TCN layer to capture long-range spatiotemporal dependency features, and its output serves as input to the LSTM layer to model the short-term dynamic fluctuations of drainage flow. The mass conservation equation is embedded as a physical constraint during training. The dynamic control module uses the model predictive control algorithm to rollingly optimize the gate opening and pump station power control sequence, and combines real-time electricity price data and overflow risk to perform dual-objective optimization.

2. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 1 is characterized in that: Also includes: The safety redundancy module is used to detect the prediction confidence of the prediction model module. When the confidence is lower than the threshold, it switches to PID control and triggers manual review.

3. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 1 is characterized in that: The system adopts an edge-cloud collaborative architecture, including edge computing devices and a cloud management platform; the edge devices deploy a lightweight TCN-LSTM model and adopt an event-driven triggering mechanism, and the cloud platform supports transfer learning and online incremental learning, and provides a visual monitoring interface.

4. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 1 is characterized in that: The spatiotemporal fusion module includes: Bilinear interpolation unit, which maps the 2D grid rainfall data from the weather radar to the coordinates of the drainage network nodes; Graph attention network unit, which is used to aggregate meteorological features and pipe network topology attributes of adjacent nodes to generate spatially correlated meteorological feature vectors; The Transformer encoder dynamically fuses meteorological, runoff, and historical drainage data through a multi-head attention mechanism and outputs a spatiotemporal fusion tensor.

5. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 1 is characterized in that: In the prediction model module, the TCN-LSTM hybrid model is trained through the following steps: In the pre-training phase, a synthetic dataset generated by SWMM simulation is used for training, covering normal rainfall, heavy rain, and pipe blockage scenarios; During the fine-tuning phase, extreme rainfall noise data is injected to improve the model robustness through adversarial training. In the transfer learning phase, the TCN layer weights are fixed, and the LSTM layer and fully connected layer are fine-tuned using the target city data.

6. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 1 is characterized in that: In the prediction model module, the loss function of the TCN-LSTM hybrid model is: ； in, represents the total loss function, is the mean square error, which measures the difference between the model prediction value and the true value; is a physical constraint term. ∇⋅(Q⋅v) in the formula is the divergence of the vector field Q⋅v, ∇ is the divergence operator; Q is the instantaneous volume flow rate of a node or pipe cross section in the drainage network, and v is the instantaneous flow velocity of the water flow in the drainage network; is the weighted penalty term for adversarial sample prediction error; α, β, is a hyperparameter.

7. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 1 is characterized in that: The target optimization function of the dynamic control module is as follows: ； Where N is the total number of time steps, Forecast flow for step k, is the safe flow threshold, is the pump station power, is the real-time electricity price, is the weight coefficient.

8. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 2 is characterized in that: The safety redundancy module generates a prediction result distribution through Bootstrap sampling, and triggers control mode switching when the flow prediction variance exceeds a threshold or the prediction deviation is greater than 15% for multiple consecutive times.

9. The intelligent outlet dynamic control system based on multi-source time series fusion according to claim 3 is characterized in that: The edge computing device compresses the model volume through channel pruning and 8-bit integer quantization technology, and activates the prediction model only when the traffic exceeds the threshold or the sensor is abnormal.

10. The intelligent outlet dynamic control method based on multi-source time series fusion is characterized by: Using the intelligent outlet dynamic control system based on multi-source time series fusion according to any one of claims 1 to 9, the method includes: S1. Real-time collection of multi-source data in the target area, including weather radar data, surface runoff sensor data, and historical drainage data, and pre-processing; S2. Fusion of multi-source data through spatiotemporal alignment encoder and multi-head attention mechanism to generate spatiotemporal fusion tensor; S3. Inputting the fused tensor into the constructed TCN-LSTM hybrid model to predict future drainage flow; S4. Based on the model predictive control algorithm, the gate opening and pump station power control sequence are continuously optimized, and dynamic regulation is performed based on real-time electricity prices and overflow risks.

Images