Storm flood river recovery scheduling method based on digital twin reverse multi-objective optimization
By constructing a digital twin-based inverse multi-objective optimization method for river clearing scheduling, and utilizing a temporal neural network model and decoupling technology, Pareto optimal scheduling parameters are generated, solving the quantitative assessment problem in river clearing after heavy rain and achieving efficient and flexible river scheduling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING FORESTRY UNIVERSITY
- Filing Date
- 2026-03-26
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies lack accurate quantitative assessments for river clearing and scheduling after heavy rain, leading to reliance on experience in the clearing process, making it difficult to cope with extreme rainstorm events. Furthermore, traditional methods lack flexibility and resilience.
A digital twin-based inverse multi-objective optimization method is adopted. By constructing a temporal neural network model, the river water quality indicators are decoupled. A static-dynamic hybrid dual-channel LSTM network structure is used to train a rain-river prediction model and generate Pareto optimal scheduling parameters to achieve efficient scheduling of river clearing.
It enables efficient prediction and scheduling of river water quality, allows for advance planning of dredging measures, reduces dredging time, improves the accuracy and flexibility of scheduling, and adapts to extreme rainstorm events.
Smart Images
Figure CN122264572A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of river management technology, and in particular to a method for scheduling river clearing after heavy rain based on digital twin-based reverse multi-objective optimization. Background Technology
[0002] When heavy rainfall hits the ground, it first washes away accumulated vehicle exhaust deposits, tire wear particles, and atmospheric dry and wet deposits from impermeable surfaces in cities (roads, roofs). It also carries residual fertilizers and pesticides (especially total phosphorus and ammonia nitrogen) from farmland soil. Rainwater flowing into combined sewer systems also stirs up sludge that has accumulated for years at the bottom of the pipes, and in extreme rainstorms, it can trigger overflows from sewage treatment plants, discharging untreated sewage directly into rivers. These pollution sources surge into rivers within hours, causing a sharp drop in dissolved oxygen, a surge in turbidity and ammonia nitrogen concentrations, and the large amounts of oxygen-consuming substances deplete the water's oxygen levels. After a lag period of about 20 days, these substances release accumulated nutrients, triggering abnormal algal blooms.
[0003] Existing technologies for treating polluted river sections involve installing rapid physical / chemical purification devices (such as cyclone separators and flocculation sedimentation) at stormwater outlets to intercept particulate pollutants, or using methods such as oxygenation, flocculant addition, and mobile treatment vessels for emergency purification. However, current management practices by water conservancy departments generally suffer from a focus on end-of-pipe treatment rather than source prevention, reliance on subjective human judgment, and a lack of forecasting and coordinated scheduling capabilities. This results in a passive response after disasters, leading to long recovery times. Furthermore, adjustments to drainage projects are mainly concentrated on "gray projects" such as "expanding pipe diameter" and "building storage tanks," lacking flexibility and adaptability to the increasing frequency of extreme rainstorms.
[0004] With the development of automatic control technology, the optimization of river restoration schemes after rainstorms is gradually evolving from verifying the effectiveness of single measures to a systematic research paradigm of "mechanism understanding - quantitative assessment - intelligent decision-making." At the mechanistic level, research has revealed the dual impact mechanism of rainstorms on water quality, adding a temporal dimension of complexity to restoration scheduling. In terms of quantitative assessment, researchers have used methods such as average concentration per event and pollutant mass-water quantity curves to preliminarily characterize the transport patterns of pollutants during rainstorm events under certain conditions, and to quantitatively evaluate the pollution reduction effects of different scheduling measures.
[0005] However, many difficulties remain in practical engineering applications. One major hurdle to overcome, from understanding the mechanisms to intelligent decision-making, is the quantitative assessment of river water quality's response to rainfall and remediation measures. Currently, river remediation after heavy rain is precisely a scenario difficult to quantify. Quantitative research requires abstracting and simplifying the research object into a mathematical model. However, this research object not only needs to be described by a large number of physical quantities, but these quantities are also strongly coupled, making it difficult to construct a model with sufficient predictive accuracy. For example, the nitrification process of ammonia nitrogen directly consumes dissolved oxygen, and the recovery efficiency of dissolved oxygen is constrained by water temperature, aeration intensity, and microbial activity. Turbidity not only reflects particulate matter content but also affects the migration paths of COD and total phosphorus through adsorption, making it difficult to completely separate the processes of "particulate matter sedimentation" and "pollutant removal." More complexly, this coupling relationship is not static but dynamically evolves with factors such as rainfall intensity, the number of preceding drought days, and sediment conditions. During the same rainstorm, when the initial scouring effect dominates, all indicators spike simultaneously, while in the middle and later stages, the difference in dilution and degradation rates may create a false sense of "decoupling." This highly nonlinear, time-varying coupling network, influenced by multiple interacting factors, often renders traditional quantitative methods based on single-indicator analysis ineffective. Any attempt to evaluate the effectiveness of a measure in isolation is highly susceptible to bias due to ignoring implicit coupling paths.
[0006] During their research on the Liangshui River in the suburbs of Beijing, the inventors discovered that for natural rivers with stable ecosystems, they can be regarded as a thin aeration tank. Unlike conventional water quality control, it is not necessary to ensure that all indicators meet the standards. Only human intervention is needed to ensure that a small number of core indicators meet the standards, while the remaining miscellaneous indicators can be automatically resolved by the system's self-purification ability. Summary of the Invention
[0007] This invention provides a method for scheduling river clearing after heavy rain based on digital twin reverse multi-objective optimization.
[0008] The technical problem to be solved is that, in the process of river clearing after heavy rain, the current clearing process still relies on experience because it is impossible to accurately obtain a quantitative assessment of the river water quality's response to rainfall and clearing measures.
[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a river clearing scheduling method based on digital twin reverse multi-objective optimization after rainstorms, wherein the intervention means used for river clearing are denoted as clearing means, and the scheduling method includes the following steps: Step 1: Quantify the clarification methods and river water quality indicators to form quantitative indicators that can be used for modeling and are feasible to operate. The quantified clarification methods are recorded as scheduling parameters. Step 2: Collect historical rainstorm-driven data, river status data, and scheduling record data of the river to be processed, and preprocess them to construct a dataset with rainstorm events as units, which includes quantified rainfall characteristic data, scheduling parameters, and river water quality indicators, denoted as the rain-river dataset. Step 3: Train a time-series neural network model that can predict dynamic changes in water quality based on the rainwater dataset, denoted as the rainwater prediction model, and decouple each input in the model; the rainwater prediction model is a mathematical model that uses scheduling parameters as design variables to obtain target variables; Step 4: Select the target variables for river clearing after heavy rain. The target variables include a primary objective and a secondary objective. Both the primary and secondary objectives have performance constraints, and the optimization of the primary objective is prioritized while satisfying the performance constraints. Select the primary and secondary objectives according to the site requirements. Step 5: Upon receiving water quality deterioration warnings or monitoring data, collect the rainfall characteristics and river water quality indicators, use the rain-river prediction model as the design function to perform reverse optimization, generate a set of Pareto optimal scheduling parameters, and formulate a scheduling plan based on the Pareto optimal scheduling parameters. The "reverse optimization" here refers to starting from the target state and working backward to derive the initial state or solution. In other words, knowing what conditions the river water quality indicators should meet, we can obtain the scheduling parameters from this target state.
[0010] Step Six: Execute the scheduling plan. After the river is cleared, update the rainwater and river prediction model based on the actual effect data of the scheduling plan.
[0011] Furthermore, in step one, the following methods are used to form quantitative indicators that can be used for modeling and are operationally feasible: Step 1.1: Analyze the operating conditions of the equipment used in the clearing process, and select the flow control parameters, time control parameters, and equipment control parameters as scheduling parameters; Step 1.2: Screen the river water quality indicators and select water quality control variables. The water quality control variables are river water quality indicators that can be quantified and respond quickly to scheduling parameters.
[0012] Furthermore, in step three, the model's output is first decoupled to form uncoupled river water quality indicators as the model's output. Then, based on the decoupled output, a model training method that can decouple the input is selected to train the model.
[0013] Furthermore, in step three, the model's output is decoupled in the following way: River water quality indicators are classified into three categories: rapid oxygen consumption indicators, slow oxygen consumption indicators, and suspended solids indicators. A representative indicator is selected from each category, and all other indicators are discarded, so that the output of the model is reduced to a state where all output parameters are uncoupled. The inputs in the model are decoupled in the following way: In step three, a static-dynamic hybrid dual-channel LSTM network structure is adopted: Channel 1: Input rainfall characteristic data and initial river state data as static features; Channel 2: Input the sequence of scheduling parameters as time-series features, inputting them sequentially at each time step, with static features incorporated into each time step.
[0014] Furthermore, the rapid oxygen consumption-related indicators include ammonia nitrogen and dissolved oxygen, with ammonia nitrogen as the representative indicator; The indicators related to slow oxygen consumption include chemical oxygen demand, total phosphorus, and chlorophyll a, with chemical oxygen demand as the representative indicator. The suspended matter-related indicators include suspended matter, turbidity, and transparency, with turbidity being the representative indicator.
[0015] Furthermore, the rainfall characteristic data in step five are either the actual data measured by the meteorological station after the rainfall or the forecast data given by the weather forecast before the rainfall.
[0016] Furthermore, during the river clarification process, the reclaimed water, diverted water, and rainwater resources used for clarification are collectively referred to as clarified water. The scheduling parameters include the scheduling volume, timing, and water quality level of the clarified water; the flocculant addition point, timing, and dosage; the aeration vessel's operating path, start time, and duration; the gate opening sequence; and the pump station start and stop times. The target variables include clarification time, clarification cost, and clarified water utilization rate. Based on the site requirements, one objective function is selected as the primary objective, and the other two are secondary objectives.
[0017] Furthermore, in step five, the objective function used for reverse optimization is a constrained vector objective function, and the NSGA-II multi-objective genetic algorithm is used to obtain the Pareto front through reverse optimization.
[0018] Furthermore, in step six, after the scheduling is completed, the actual river water quality indicators are collected. If the error between the predicted value and the measured value is greater than 15%, the new measured value is included in the training set to start the model update.
[0019] Compared with existing technologies, the post-rainstorm river clearing scheduling method based on digital twin reverse multi-objective optimization has the following advantages: In this invention, the river after a rainstorm is regarded as a thin aeration tank. The river water quality indicators are classified into three categories: rapid oxygen consumption related indicators, slow oxygen consumption related indicators, and suspended solids related indicators. A representative indicator is selected from each category, and all other indicators are discarded. This reduces the dimensionality of the model output to the point where all output parameters are uncoupled, thereby decoupling the model output and forming uncoupled river water quality indicators as the model output. Then, based on the decoupled output, a static-dynamic hybrid dual-channel LSTM network structure is adopted. Rainfall feature data and river initial state data are input into channel 1 as static features; scheduling parameter sequence is input into channel 2 as time-series features and is input sequentially at each time step. Static features are incorporated into each time step, thereby decoupling the input. The model trained in this way can reliably predict the responses of river water quality indicators most in need of human intervention to rainfall and river clearing measures, and can be used as a design function in the reverse optimization process. Then, using a constrained Pareto front as the objective function, a set of Pareto-optimal scheduling parameters is generated, and a scheduling scheme is formulated based on these parameters; thus achieving efficient scheduling in the river clearing process after heavy rain. Furthermore, it allows for obtaining rainfall characteristic data from mid-weather forecasts in advance to specify the scheduling scheme, and allows clearing measures to be implemented before rainfall, thereby further improving the clearing effect. Attached Figure Description
[0020] Figure 1 This is a flowchart of the post-rainstorm river clearing scheduling method based on digital twin reverse multi-objective optimization in this invention. Figure 2 This is a structural diagram of the LSTM model in this invention; Figure 3 This is a flowchart for the multi-objective optimization in step five. Detailed Implementation
[0021] Taking the proposed solution for the project of controlling rainstorm pollution and the post-rain multi-source synergistic restoration mechanism for urban rivers in Beijing (BJST20250204Q) as an example, such as Figure 1 As shown, a river clearing scheduling method based on digital twin inverse multi-objective optimization is proposed for post-rainstorm river clearing. The intervention measures used for river clearing are denoted as clearing measures. The scheduling method includes the following steps: Step 1: Quantify the clarification methods and river water quality indicators to form quantitative indicators that can be used for modeling and are feasible to operate. The quantified clarification methods are recorded as scheduling parameters. Common methods for restoring river clarity include adding flocculants, using aeration boats, diluting water, and altering the river's flow distribution. River water quality indicators are very complex; for simplicity, they are generally described using water quality grades, each corresponding to several specific indicators and some apparent characteristics.
[0022] Step 2: Collect historical rainstorm-driven data, river status data, and scheduling record data of the river to be processed, and preprocess them to construct a dataset with rainstorm events as units, which includes quantified rainfall characteristic data, scheduling parameters, and river water quality indicators, denoted as the rain-river dataset. A rainstorm event is considered a unit of processing. This refers to all events that occur around a rainfall event, from the start of the rainfall to the implementation of flood control measures and the restoration of river flow. If flood control measures are implemented before the rainfall (e.g., based on a weather forecast and a scheduling plan), the time period is extended backward to the point when the flood control measures are first implemented.
[0023] Step 3: Train a time-series neural network model that can predict dynamic changes in water quality based on the rain-river dataset, denoted as the rain-river prediction model, and decouple the inputs in the model; the rain-river prediction model is a mathematical model that uses scheduling parameters as design variables to obtain target variables; The mathematical model here can be viewed as a generalized function, taking rainfall characteristic data, scheduling parameters, and some initial values (such as initial values of river water quality) as input, and outputting the changes in river water after rainfall and the use of re-cleaning methods.
[0024] Step 4: Select the target variables for river clearing after the rainstorm. The target variables include the primary objective and the secondary objective. Both the primary and secondary objectives have performance constraints, and the optimization of the primary objective should be prioritized while meeting the performance constraints. Select the primary and secondary objectives according to the site requirements. The performance constraints here refer to the hard requirements imposed on the target variable on-site. For example, municipal authorities may require the target variable to be cleared within 12 hours, or there may be a budget limit. Which target variable is the primary objective depends on which one is prioritized on-site.
[0025] like Figure 3 As shown, step five: When receiving water quality deterioration warnings or monitoring data, collect the rainfall characteristics data and river water quality indicators, use the rain-river prediction model as the design function to perform reverse optimization, generate a set of Pareto optimal scheduling parameters, and formulate a scheduling plan based on the Pareto optimal scheduling parameters. The scheduling parameters here, such as when, where, and how much flocculant to use, still differ slightly from the actual implementation. They need to be refined based on the on-site material reserves and personnel situation to form a specific plan.
[0026] Step Six: Execute the scheduling plan. After the river is cleared, update the rainwater and river prediction model based on the actual effect data of the scheduling plan.
[0027] This approach takes into account the limited amount of data typically available for model training, which can impact model accuracy. For example, this embodiment collects rainfall-driven data, river status data, and scheduling records for the Liangshui River from 2023 to 2025. The digital twin model is trained using data from the first two years, and its accuracy is validated using data from the last year. Supplementing the model with new data can improve its performance.
[0028] In step one, the following methods are used to generate quantitative indicators that can be used for modeling and are feasible to operate: Step 1.1: Analyze the operating conditions of the equipment used in the clearing process, and select the flow control parameters, time control parameters, and equipment control parameters as scheduling parameters; The specific cleaning methods to be used on-site are taken here; not all of them need to be included.
[0029] Step 1.2: Screen the river water quality indicators and select water quality control variables. The water quality control variables are river water quality indicators that can be quantified and respond quickly to scheduling parameters.
[0030] Rapid response here refers to the output responding quickly to the input. A response time on the order of hours is typically considered rapid; for example, the suspended solids content begins to decrease within one hour after the use of flocculants. The reason for choosing rapidly responding river water quality indicators is that in this scenario, indicators with slow responses are usually heavily coupled, not directly related, and unclear. Their inclusion would not only fail to improve the model's prediction accuracy but would actually reduce it to some extent.
[0031] In step three, the model output is first decoupled to form uncoupled river water quality indicators as the model output. Then, based on the decoupled output, a model training method that can decouple the input is selected to train the model.
[0032] The inputs are relatively loosely coupled (the scheduling parameters themselves do not restrict each other), but the outputs are heavily coupled. For example, the ammonia nitrogen and dissolved oxygen levels are heavily coupled, affecting the modeling. Therefore, we first decouple the outputs.
[0033] In step three, the model output is decoupled in the following way: River water quality indicators are classified into three categories: rapid oxygen consumption indicators, slow oxygen consumption indicators, and suspended solids indicators. A representative indicator is selected from each category, and all other indicators are discarded, so that the output of the model is reduced to a state where all output parameters are uncoupled. Direct decoupling is difficult, complex, and ineffective. Here, we use model dimensionality reduction to achieve decoupling, which is simple and effective. The three representative indicators after dimensionality reduction are uncoupled. The reason we can monitor them to ensure the effectiveness of the remediation measures is that we treat the river as a dilute aeration tank (meaning low pollutant content and relatively easy replenishment of dissolved oxygen). As long as dissolved oxygen is sufficient, other pollutants (except suspended solids) washed in by heavy rain will quickly self-clean up. Therefore, we only need to ensure sufficient dissolved oxygen and that suspended solids do not exceed the standard. However, dissolved oxygen cannot be directly used as a representative indicator. If dissolved oxygen is high enough but oxygen-consuming factors are not eliminated, it will quickly drop. Therefore, we indirectly judge dissolved oxygen by monitoring oxygen-consuming factors. As long as oxygen-consuming factors are eliminated, dissolved oxygen will quickly and automatically rise again.
[0034] The inputs in the model are decoupled in the following way: like Figure 2 As shown, in step three, a static-dynamic hybrid dual-channel LSTM (Long Short-Term Memory) network structure is used: Channel 1: Input rainfall characteristic data and initial river state data as static features; Channel 2: Input the sequence of scheduling parameters as time-series features, inputting them sequentially at each time step, with static features incorporated into each time step.
[0035] The decoupling here is primarily to avoid a coupling relationship where rainfall must precede clearing measures, since rainfall characteristic data can be obtained in advance from mid-weather forecasts. By dividing the data into two channels, and treating rainfall characteristic data and initial river state data as static features rather than as time-series features, the requirement of rainfall preceding clearing measures can be avoided, allowing clearing measures to precede rainfall.
[0036] Rapid oxygen consumption-related indicators include ammonia nitrogen and dissolved oxygen, with ammonia nitrogen being the representative indicator. Ammonia nitrogen (NH3-N) refers to the nitrogen in water in the form of free ammonia (NH3) and ammonium ions (NH4+). + Nitrogen exists in the form of ammonia nitrogen. It was chosen as a representative indicator of rapid oxygen consumption because the nitrification process of ammonia nitrogen (converting to nitrate under aerobic conditions) directly and rapidly consumes dissolved oxygen—theoretically, oxidizing 1 mg of ammonia nitrogen requires approximately 4.6 mg of dissolved oxygen. And ammonia nitrogen is the most important rapidly oxygen-consuming substance in river water.
[0037] Dissolved oxygen (DO) refers to the amount of molecular oxygen dissolved in water, usually expressed in mg / L. It is fundamental to the survival of aquatic organisms and the self-purification capacity of water bodies. However, dissolved oxygen itself is not suitable as a representative indicator.
[0038] The indicators related to slow oxygen consumption include chemical oxygen demand, total phosphorus, and chlorophyll a, with chemical oxygen demand as the representative indicator. Chemical Oxygen Demand (COD) refers to the amount of oxygen consumed (mg / L) when a water sample is treated with a strong oxidant (such as potassium dichromate) under certain conditions. It reflects the total amount of reducing substances (mainly organic matter) in water that can be chemically oxidized. COD represents the total organic load in water, but its oxygen consumption process is relatively slow—microorganisms need a considerable amount of time to degrade it. Compared to the "acute oxygen consumption" of ammonia nitrogen, COD reflects "chronic oxygen pressure." Furthermore, COD is correlated with total phosphorus and chlorophyll a (e.g., organic phosphorus is a component of COD, and algal death also contributes to COD), thus it can be used as a proxy for these indicators.
[0039] Total phosphorus (TP) refers to the total amount of phosphorus in water in all its forms, including dissolved and particulate organic and inorganic phosphorus compounds. Phosphorus itself does not directly consume oxygen, but it is a limiting nutrient for algal growth. After algae proliferate, they consume dissolved oxygen through respiration and decomposition after death; this oxygen consumption has a significant lag effect. Chlorophyll a refers to the concentration of chlorophyll a in water, commonly measured in micrograms per liter (μg / L), and sometimes expressed as milligrams per cubic meter (mg / m³). Chlorophyll a is a photosynthetic pigment found in the cells of all photosynthetic phytoplankton (algae), and its content is often used as a proxies for algal biomass. Chlorophyll a directly reflects the current abundance of algae in the water. A large presence of algae means that oxygen is produced during the day through photosynthesis, consumed at night through respiration, and further decomposed by microorganisms after death, consuming oxygen. This oxygen-consuming process also exhibits significant time lag and periodicity.
[0040] The indicators related to suspended matter include suspended solids, turbidity, and transparency, with turbidity being the representative indicator.
[0041] Suspended solids (SS) refer to the solid matter remaining on a water sample after it has passed through a 0.45 μm filter membrane and been dried to constant weight at 103-105 °C. It reflects the total amount of suspended solid particles in the water. While classified as suspended solids, it is not a representative indicator. SS is a direct measurement of particulate matter mass, but requires laboratory analysis and cannot be monitored online in real time.
[0042] Turbidity refers to the degree to which suspended particles in water impede the transmission of light, and is usually expressed in turbidity units (NTU). It is a measure of the optical properties of water. Its role in classification: It is chosen as a representative indicator for suspended solids (SS). Reasons include: High correlation: Turbidity and SS are usually highly linearly correlated (correlation coefficient R² can reach above 0.85-0.95); Engineering measurability: Turbidity sensors can achieve real-time online monitoring, making them ideal for real-time input to digital twin models; Physical significance: Changes in turbidity can quickly respond to the effects of rainfall washout and regulatory measures (such as flocculant addition).
[0043] Transparency refers to the clarity of water, usually measured using the Seiler disk method, which is the depth (cm or m) to which a white disk disappears in the water. Its role in classification: it is categorized as suspended matter, but is not a representative indicator. Transparency is negatively correlated with turbidity, but measurements are greatly affected by subjective factors and cannot be performed online in real time; therefore, it is used as a supplementary indicator rather than a representative one.
[0044] The rainfall characteristic data in step five are the actual measured data from the meteorological station after the rainfall, or the forecast data given by the weather forecast before the rainfall.
[0045] This allows the application of remediation measures to precede the actual start of rainfall (and the model in this embodiment does not require the application of remediation measures to be later than rainfall), thereby achieving better results (early prevention generally greatly shortens the remediation time).
[0046] In the process of river clarification, the reclaimed water, externally diverted water, and rainwater resources used for clarification are collectively referred to as clarified water. The scheduling parameters include the scheduling volume, timing, and water quality level of clarified water; the flocculant dosing point, timing, and dosage; the aeration vessel's operating path, start time, and duration; the gate opening sequence; and the pump station start and stop times. The target variables include clarification time, clarification cost, and clarified water utilization rate. Based on the site requirements, one objective function is selected as the primary objective, and the other two are secondary objectives.
[0047] Here, some of the target variables are calculated from the scheduling parameters, while others can be obtained directly from the model output. For example, once the scheduling parameters are known, there are many mature cost accounting methods to calculate the cost of this process. The utilization rate of reclaimed water can also be easily calculated from the model output (the amount of water that meets the standards and can be reused after scheduling divided by the amount of reclaimed water scheduled). The reclaiming time can be directly obtained from the model output (the time from the start of scheduling to the river water quality meeting the standards).
[0048] In step five, the objective function used for reverse optimization is a constrained vector objective function, and the NSGA-II multi-objective genetic algorithm is used to obtain the Pareto front through reverse optimization.
[0049] The vector objective function is constructed as follows: Re-cleaning time T: It is generally desirable to keep it as short as possible, so minimize T(x). Re-cleaning cost C: It is generally desirable to keep it as low as possible, minimizing C(x). Reclaimed water utilization rate U: Generally, it is desirable to be as high as possible, maximizing U(x). Thus, the overall vector objective function can be written as: Where x is the decision vector (such as model hyperparameters, network structure, etc.). It is the decision-making space.
[0050] In addition to some strict requirements on the target variable in the field, the constraints here include the following: Water availability constraints (water consumption cannot exceed the available water supply), river safety flow constraints (river flow during scheduling cannot exceed the maximum / minimum allowable flow), and water quality compliance constraints (river water quality during scheduling cannot be lower than the local minimum allowable level).
[0051] In step six, after the scheduling is completed, the actual river water quality indicators are collected. If the error between the predicted value and the measured value is greater than 15%, the new measured value is included in the training set to start the model update.
[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for scheduling river clearing after heavy rain based on digital twin-based inverse multi-objective optimization, wherein the intervention measures used for river clearing are denoted as clearing measures, characterized in that: The scheduling method includes the following steps: Step 1: Quantify the clarification methods and river water quality indicators to form quantitative indicators that can be used for modeling and are feasible to operate. The quantified clarification methods are recorded as scheduling parameters. Step 2: Collect historical rainstorm-driven data, river status data, and scheduling record data of the river to be processed, and preprocess them to construct a dataset with rainstorm events as units, which includes quantified rainfall characteristic data, scheduling parameters, and river water quality indicators, denoted as the rain-river dataset. Step 3: Train a time-series neural network model that can predict dynamic changes in water quality based on the rainwater dataset, denoted as the rainwater prediction model, and decouple each input in the model; the rainwater prediction model is a mathematical model that uses scheduling parameters as design variables to obtain target variables; Step 4: Select the target variables for river clearing after heavy rain. The target variables include a primary objective and a secondary objective. Both the primary and secondary objectives have performance constraints, and the optimization of the primary objective is prioritized while satisfying the performance constraints. Select the primary and secondary objectives according to the site requirements. Step 5: Upon receiving water quality deterioration warnings or monitoring data, collect the rainfall characteristics and river water quality indicators, use the rain-river prediction model as the design function to perform reverse optimization, generate a set of Pareto optimal scheduling parameters, and formulate a scheduling plan based on the Pareto optimal scheduling parameters. Step Six: Execute the scheduling plan. After the river is cleared, update the rainwater and river prediction model based on the actual effect data of the scheduling plan.
2. The method for river clearing and scheduling after heavy rain based on digital twin inverse multi-objective optimization according to claim 1, characterized in that: In step one, the following methods are used to generate quantitative indicators that can be used for modeling and are feasible to operate: Step 1.1: Analyze the operating conditions of the equipment used in the clearing process, and select the flow control parameters, time control parameters, and equipment control parameters as scheduling parameters; Step 1.2: Screen the river water quality indicators and select water quality control variables. The water quality control variables are river water quality indicators that can be quantified and respond quickly to scheduling parameters.
3. The method for river clearing and scheduling after heavy rain based on digital twin inverse multi-objective optimization according to claim 2, characterized in that: In step three, the model output is first decoupled to form uncoupled river water quality indicators as the model output. Then, based on the decoupled output, a model training method that can decouple the input is selected to train the model.
4. The method for post-rainstorm river clearing scheduling based on digital twin inverse multi-objective optimization according to claim 3, characterized in that: In step three, the model output is decoupled in the following way: River water quality indicators are classified into three categories: rapid oxygen consumption indicators, slow oxygen consumption indicators, and suspended solids indicators. A representative indicator is selected from each category, and all other indicators are discarded, so that the output of the model is reduced to a state where all output parameters are uncoupled. The inputs in the model are decoupled in the following way: In step three, a static-dynamic hybrid dual-channel LSTM network structure is adopted: Channel 1: Input rainfall characteristic data and initial river state data as static features; Channel 2: Input the sequence of scheduling parameters as time-series features, inputting them sequentially at each time step, with static features incorporated into each time step.
5. The method for post-rainstorm river clearing scheduling based on digital twin inverse multi-objective optimization according to claim 4, characterized in that: The rapid oxygen consumption-related indicators include ammonia nitrogen and dissolved oxygen, with ammonia nitrogen as the representative indicator. The indicators related to slow oxygen consumption include chemical oxygen demand, total phosphorus, and chlorophyll a, with chemical oxygen demand as the representative indicator. The suspended matter-related indicators include suspended matter, turbidity, and transparency, with turbidity being the representative indicator.
6. The method for river clearing and scheduling after heavy rain based on digital twin inverse multi-objective optimization according to claim 1, characterized in that: The rainfall characteristic data in step five are the actual measured data from the meteorological station after the rainfall, or the forecast data given by the weather forecast before the rainfall.
7. The method for river clearing and scheduling after heavy rain based on digital twin inverse multi-objective optimization according to claim 1, characterized in that: In the process of river clarification, the reclaimed water, diverted water, and rainwater resources used for clarification are collectively referred to as clarified water. The scheduling parameters include the scheduling volume, timing, and water quality level of clarified water; the flocculant addition point, timing, and dosage; the aeration vessel operation path, start time, and duration; the gate opening sequence; and the pump station start and stop times. The target variables include clarification time, clarification cost, and clarified water utilization rate. Based on the site requirements, one objective function is selected as the primary objective, and the other two are secondary objectives.
8. The method for river clearing and scheduling after heavy rain based on digital twin inverse multi-objective optimization according to claim 7, characterized in that: In step five, the objective function used for reverse optimization is a constrained vector objective function, and the NSGA-II multi-objective genetic algorithm is used to obtain the Pareto front through reverse optimization.
9. The method for post-rainstorm river clearing scheduling based on digital twin inverse multi-objective optimization according to claim 1, characterized in that: In step six, after the scheduling is completed, the actual river water quality indicators are collected. If the error between the predicted value and the measured value is greater than 15%, the new measured value is included in the training set to start the model update.