Intelligent water conservancy centralized control center system based on intelligent on-duty body cluster and scheduling method

Abstract

The invention discloses an intelligent water conservancy centralized control center system based on an intelligent on-duty body cluster and a scheduling method, and belongs to the technical field of intelligent water conservancy and industrial internet. The system comprises a multi-source data fusion and state sensing module, a multi-agent cooperative scheduling engine, a digital twin simulation deduction module, a task decomposition and cooperative execution module and an intelligent on-duty body cluster. Intelligent on-duty bodies deployed in water conservancy facilities collect operation data in a non-intrusive visual perception and simulation peripheral mode and execute control operation, a multi-agent collaborative scheduling engine carries out cross-facility joint optimization based on global state data, and a digital twin module carries out simulation verification on a scheduling scheme and then issues and executes the scheduling scheme.

Description

Technical Field

[0001] This invention relates to the fields of smart water conservancy and industrial internet technology, and in particular to a smart water conservancy centralized control center system and scheduling method based on a smart duty system cluster. Background Technology

[0002] With the deepening of water conservancy informatization, unified monitoring and coordinated scheduling of water resources at the basin level has become a core objective of water conservancy modernization. In the water conservancy and hydropower industry, promoting unmanned or minimally staffed operations and integrated basin scheduling is a development trend, posing a serious challenge to the traditional centralized control center construction model. Chinese invention CN120338210A discloses a reservoir scheduling method based on deep learning adaptive dynamic networks and reinforcement learning. This method belongs to the field of reservoir scheduling technology. It includes acquiring real-time water level data, meteorological data, and radar image data of regional cloud cover and surface features from the reservoir, fusing these three data sources, and inputting the fused data into a pre-defined adaptive dynamic Transformer network model to obtain a prediction output. The prediction output includes a reservoir water level sequence, an inflow sequence, and a discharge sequence. A reinforcement learning state space is constructed and input into a pre-defined reinforcement learning network. The action space of the reinforcement learning network includes flood discharge, power generation flow, and ecological flow. The optimization objective of the reinforcement learning network is to maximize the cumulative discount reward, and the reward function is a weighted function of flood control reward, power generation reward, and ecological reward. The reinforcement learning network iteratively optimizes and ultimately converges to output the optimal combination of flood discharge, power generation flow, and ecological flow.

[0003] However, the aforementioned existing technologies have the following shortcomings. First, this method only targets the scheduling optimization of a single reservoir, employing a centralized deep learning and reinforcement learning architecture. It cannot perform cross-facility collaborative scheduling of multiple heterogeneous water conservancy facilities, making it difficult to meet the needs of integrated water resources management at the basin level. When coordinating the joint operation of various types of facilities such as hydropower stations, pumping stations, and sluice gates, this method is ineffective. Second, this method relies on direct data access to existing reservoir automation systems, requiring underlying equipment to have unified communication protocols such as IEC61850 and standardized data interfaces. For water conservancy facilities with different construction dates, diverse equipment brands, and numerous legacy systems, large-scale automation system upgrades are required to achieve data acquisition and remote control. These upgrades are costly, time-consuming, and involve significant risks and obstacles. Third, although the reinforcement learning decision-making process uses an adaptive dynamic Transformer network for prediction, it lacks a virtual simulation verification step in a digital twin environment. Directly issuing and executing the scheduling plan poses certain security risks, especially under extreme weather conditions such as typhoon season flood control pre-discharge scenarios, which may lead to unpredictable consequences. It cannot achieve the safety guarantee of simulation before execution. Fourth, this method adopts a centralized deep learning architecture, in which all data collection, prediction calculation and decision optimization are completed in the cloud. It has high requirements for network real-time performance and cannot guarantee the safe operation of the facility in the event of communication interruption or network delay. It also lacks the ability to perform autonomous execution at the edge.

[0004] Therefore, there is a need for a smart water conservancy control center system and method that can quickly integrate various heterogeneous water conservancy facilities in a low-cost, non-intrusive manner, support cross-facility collaborative optimization scheduling, and have simulation verification and edge autonomous execution capabilities. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a smart water conservancy centralized control center system and scheduling method based on intelligent duty unit clusters. This system deploys intelligent duty units as universal adapters and local actuators in various water conservancy facilities to construct a new centralized control system that features cloud-edge collaboration, centralized decision-making, and distributed execution, achieving low-cost, rapid deployment, high compatibility, and high intelligence for unified management and control of basin-level water conservancy facilities.

[0006] To achieve the above objectives, this invention provides a smart water conservancy centralized control center system based on an intelligent duty unit cluster. This system adopts a three-layer architecture: a cloud-based centralized control platform layer, an edge-based intelligent duty unit layer, and a secure communication layer, comprising: The multi-source data fusion and status awareness module, deployed in the cloud, is used to aggregate multi-source heterogeneous data from the intelligent duty system cluster and perform spatiotemporal alignment and feature fusion to generate a global status dataset. A multi-agent collaborative scheduling engine, deployed in the cloud, models each water conservancy facility as an agent with perception, decision-making, and execution capabilities. It employs a centralized training and decentralized execution architecture. Its global value evaluation network uses a QMIX network structure to nonlinearly mix the local values ​​of each agent, and achieves this through non-negativity constraints on the mixed network weights. Monotonicity constraint; The digital twin simulation and deduction module is deployed in the cloud and is used to perform simulation verification in a digital twin that is consistent with the actual water conservancy system before the scheduling plan is issued and executed. The simulation time scale is more than 100 times the actual time, and the scheduling plan is only allowed to be issued and executed after the simulation verification passes the security constraints. The task decomposition and collaborative execution module, deployed in the cloud, is used to parse the simulation-verified scheduling scheme into a task execution graph with dependencies and distribute it. The intelligent duty system cluster consists of edge intelligent terminals deployed next to the control rooms of various water conservancy facilities. It bypasses the existing facility control system in a non-intrusive access manner without modifying the original control loop. It is responsible for local data acquisition and control operation execution, and runs safely locally according to the established program when the cloud network is interrupted. The secure communication layer uses a VPN encrypted tunnel on the water conservancy private network to connect the cloud and the edge.

[0007] The present invention also provides a smart water conservancy scheduling method based on the above system, including: S1, data acquisition and fusion step; S2, collaborative scheduling decision step; S3, simulation verification step, in which the scheduling scheme generated in step S2 is simulated and deduced in a digital twin, and only proceeds to step S4 if the safety constraints are met; S4, task distribution step; S5, local execution and feedback step.

[0008] Compared with existing technologies, this invention has the following advantages. First, by using a non-intrusive access method for intelligent duty agents, this invention completely eliminates the need for modifications to existing facility control systems. Adding a new facility only requires deploying one intelligent duty agent and connecting it to the network for centralized control, achieving plug-and-play rapid deployment and reducing investment cycle and cost by an order of magnitude. Second, this invention adopts a multi-agent collaborative scheduling architecture, with each water conservancy facility corresponding to one intelligent agent. Cross-facility joint optimization is achieved through a distributed policy network and global value assessment, overcoming the limitation of single reservoir scheduling methods in handling multi-facility collaboration. Third, this invention introduces a digital twin simulation and deduction stage, performing virtual verification before the actual execution of the scheduling scheme. Based on hydraulic models and equipment response models, the execution effect is predicted, effectively reducing scheduling risks under extreme conditions. Fourth, this invention adopts a cloud-edge collaborative architecture. Control commands are generated in the cloud and verified through simulation before being sent to the edge intelligent duty agents for local execution. Even when the network is interrupted, the edge devices can still operate safely according to the established program, significantly enhancing system robustness. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0010] Figure 2 This is a flowchart illustrating the scheduling method of the present invention. Detailed Implementation

[0011] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. The following examples use the intelligent centralized control transformation of six water conservancy facilities under the jurisdiction of a regional water conservancy management agency, namely the Qingshan cascade hydropower station (including Qingshan I and Qingshan II power stations), the plain drainage pumping station (including three pumping stations), and the river control gate (one), as an application scenario to illustrate the specific implementation process of the system and method of the present invention.

[0012] like Figure 1 As shown, the intelligent water conservancy control center system based on an intelligent duty agent cluster of this invention is deployed and operated using a three-layer architecture, including a cloud-based control platform layer, an edge-based intelligent duty agent layer, and a secure communication layer. The cloud-based control platform layer is deployed in the private cloud environment or public cloud platform of the regional water conservancy management center, serving as the system's central brain. It includes a multi-source data fusion and status awareness module, a multi-agent collaborative scheduling engine, a digital twin simulation and deduction module, a task decomposition and collaborative execution module, and an intelligent cockpit module, responsible for full-domain data aggregation, intelligent analysis, and collaborative scheduling. The edge-based intelligent duty agent layer consists of edge intelligent terminals deployed in the control rooms of various water conservancy facilities, forming an intelligent duty agent cluster. This cluster acts as the hands, eyes, and primary brain of the control center on-site, responsible for local data acquisition and control operation execution. The secure communication layer uses a water conservancy dedicated network VPN encrypted tunnel to connect the cloud and the edge, ensuring the security and reliability of data and command transmission.

[0013] The multi-source data fusion and status awareness module is the data infrastructure of the cloud-based centralized control platform. It is responsible for aggregating multi-source heterogeneous data from the intelligent duty system cluster and performing standardized processing, providing high-quality data support for subsequent intelligent analysis and decision-making. The module receives facility operation data from six different types of water conservancy facilities, with a rich variety of data types. This includes data on hydropower station unit output, generator head, guide vane opening, rotational speed, vibration signals, and bearing temperature; pump station pump unit operating status, motor current, inlet and outlet flow rates, inlet and outlet pool water levels; sluice gate opening, upstream and downstream water levels, hoist oil pressure, and gate inclination; as well as environmental temperature and humidity, electrical parameters of the power distribution room, and monitoring video images for each facility. The data acquisition cycle is configured differently based on the importance and rate of change of the data type. The acquisition cycle for key operating parameters such as unit output and gate opening is set to 1 to 5 seconds; the acquisition cycle for environmental parameters such as temperature and humidity is set to 30 to 60 seconds; and the acquisition frame rate for video images is set to 1 to 10 frames per second.

[0014] Spatiotemporal alignment is a core preprocessing step in the multi-source data fusion and state awareness module, used to address the inconsistencies in time and space between heterogeneous multi-source data. Since the intelligent monitoring units are deployed at water conservancy facilities in different geographical locations, clock deviations may exist, and the data reporting frequencies and acquisition cycles of different facilities are inconsistent. Directly using raw data can lead to data alignment errors and analytical biases. This module uses a network time protocol to synchronize the clocks of all intelligent monitoring units, with a synchronization cycle of 1 minute and clock deviation controlled within 50 milliseconds, ensuring the comparability of timestamps in the data reported by each facility. For data with different acquisition cycles, linear interpolation or nearest neighbor interpolation methods are used to align to a unified time grid. The time grid resolution is set to 1 second. Linear interpolation is suitable for continuously changing analog data such as water level and flow rate, while nearest neighbor interpolation is suitable for discrete state data such as pump start-up and shutdown status. The spatial correlation matrix is ​​constructed based on the topological location relationships and hydraulic connections of water conservancy facilities in the watershed. The matrix dimension is the square of the number of facilities, and the matrix elements... Indicates the first The facility and the Spatial correlation between facilities. The upstream Qingshan Hydropower Station and the downstream Pingyuan Pumping Station have a direct hydraulic connection; floodwaters flowing downstream will affect the intake water level of the downstream pumping station. Therefore, the spatial correlation between them is set to a high value of 0.8 to 0.9. The Qingshan I and Qingshan II hydropower stations, belonging to the same cascade, have an upstream-downstream hydraulic connection, with a spatial correlation set to 0.9 to 1.0. The spatial correlation between facility pairs without direct hydraulic connection is set to a low value of 0.1 to 0.3.

[0015] Feature fusion processing employs a multi-head attention mechanism to achieve deep fusion of different types of operational data. This mechanism adaptively learns the correlations between different data sources and extracts global feature representations across facilities. The operational data of each facility is encoded into feature vectors. The encoding process first normalizes the original data, unifying data with different dimensions to a standard space of zero mean and unit variance. Then, it maps the data to fixed-dimensional feature vectors through a fully connected layer. Let the dimension of the feature vector be... The data complexity ranges from 128 to 512, and in this embodiment, it is set to 256. The calculation process of the multi-head attention mechanism is as follows: Input feature matrix The dimension is ,in For the number of facilities, this embodiment The formula for calculating the multi-head attention mechanism is: , in, The query matrix is ​​derived from the input feature matrix. With query weight matrix Multiplying them together, we get the dimension as follows: ; The key matrix is ​​derived from the input feature matrix. With key weight matrix Multiplying them together, we get the dimension as follows: ; The value matrix is ​​derived from the input feature matrix. AND-weight matrix Multiplying them together, we get the dimension as follows: ; The number of attention points ranges from 4 to 8, and in this embodiment, the value is 8. To output the projection matrix, the dimension is ,in For each attention head, the value vector dimension takes the value... This embodiment The formula for calculating each attention point is: , , in, For the first The query projection matrix of each attention head, with dimensions of ; For the first The key projection matrix of each attention head, with dimension [missing information]. ; For the first The value projection matrix of each attention head has dimensions of . ; The dimension of the key vector, with values... This embodiment ; This is a scaling factor used to prevent the softmax gradient from vanishing due to excessively large dot product values.

[0016] Attention weights are dynamically adjusted based on data timeliness and facility priority, enabling the model to adaptively focus on more important and newer information. Timeliness weights The decay rate decreases as the difference between the data timestamp and the current time increases, and the decay function takes an exponential form: , in, The attenuation coefficient ranges from 0.01 to 0.1, and in this embodiment, it is set to 0.05. This is the difference between the data timestamp and the current time, in seconds. hour, This indicates that the data is completely fresh; when seconds, This indicates that the data is significantly outdated. Facility priority weight. The facilities are pre-configured based on their importance in flood control, power generation, and water supply, and stored as static parameters in the system configuration file. The priority weights for flood control gates and drainage pumping stations are set to 1.8 to 2.0; the priority weight for the main power generation station, Qingshan Level 1 Hydropower Station, is set to 1.5 to 1.8; the priority weight for auxiliary power generation station, Qingshan Level 2 Hydropower Station, is set to 1.2 to 1.5; and the priority weight for ordinary facilities is set to 1.0. After feature fusion processing, a standardized global status dataset is generated. This dataset is encapsulated in a unified JSON format and includes fields such as timestamps, facility identifiers, data types, values, and quality indicators. It is stored in the distributed time-series database InfluxDB, supporting millisecond-level real-time queries and historical backtracking up to one year, with query latency controlled within 100 milliseconds.

[0017] The multi-agent collaborative scheduling engine is the core decision-making component of the system. It is responsible for generating cross-facility collaborative scheduling schemes based on global state data, achieving joint optimization of multiple objectives such as power generation efficiency, flood control safety, ecological flow, and collaborative efficiency. This engine employs a multi-agent deep reinforcement learning framework, modeling each water conservancy facility as an agent with perception, decision-making, and execution capabilities. Global optimal scheduling is achieved through collaborative learning and information exchange among agents. In this embodiment, there are six agents, corresponding to the agents of Qingshan Level 1 Hydropower Station, Qingshan Level 2 Hydropower Station, Pingyuan No. 1 Pumping Station, Pingyuan No. 2 Pumping Station, Pingyuan No. 3 Pumping Station, and the river control gate.

[0018] Constructing the state space is the primary task of a multi-agent cooperative scheduling engine, requiring the transformation of raw runtime data into a state representation suitable for reinforcement learning algorithms. An intelligent agent at time Local observation It consists of four parts: 1) the real-time operating parameter vector of the facility, with dimensions determined by the facility type: 15-20 dimensions for hydropower stations, 10-15 dimensions for pumping stations, and 8-12 dimensions for sluice gates; 2) recent historical state sequences, containing the trajectory of key parameter changes over the past 30 minutes, sampled at 5-minute intervals to obtain data from 6 historical moments, with the flattened dimension being 6 times that of the real-time parameters; 3) meteorological forecast data, including rainfall, wind speed, and temperature forecasts for the next 24 hours, with a dimension of 24; and 4) the correlation status of upstream and downstream facilities, selecting key parameters of facilities with a correlation degree greater than 0.5 based on the spatial correlation matrix, with the dimension determined by the number of associated facilities. In summary, the observation dimension of the hydropower station agent is 100-130, the pumping station agent is 70-100, and the sluice gate agent is 50-80. Global State The state space is a collection of local observations from all agents, constructed by stitching together the data. Its dimension is the sum of the dimensions of each agent's observations, and in this embodiment, it is approximately 500 to 600 dimensions. The state space also includes the future time-period water level sequence and flow sequence output by the Transformer prediction network, with a prediction duration of 30 to 120 minutes, a time step of 15 minutes, and a prediction dimension equal to the number of monitoring sections multiplied by the number of time steps.

[0019] The definition of the action space must balance scheduling flexibility and equipment safety constraints, providing sufficient degrees of freedom to achieve optimization goals while avoiding exceeding the safe operating range of the equipment. Actions of an agent The specific content and value range are determined based on the facility type. The actions of the hydropower station's intelligent agent include the output adjustment of each unit, with an adjustment range of 20% to 20% of the negative rated power and a step size of 1% to 5% of the rated power, preferably 2%. For the Qingshan I-level hydropower station with an installed capacity of 50MW, the single-step adjustment is 1MW; for the Qingshan II-level hydropower station with an installed capacity of 30MW, the single-step adjustment is 0.6MW. The actions of the pumping station's intelligent agent include start / stop commands for each pump unit and the speed adjustment of the variable frequency pumps. The start / stop commands are binary variables (0 or 1), and the speed adjustment range is 10% to 10% of the negative rated speed and 2% of the rated speed. The actions of the sluice gate's intelligent agent include adjusting the opening of each gate, with an adjustment range from 15% of the negative full opening to 15% of the positive full opening. The adjustment step size for the gate opening is 2% to 10% of the full opening, preferably 2%. The control gate has 5 gates, each with a full opening height of 8 meters and a single-step adjustment amount of 0.16 meters. (Combined actions) It consists of the actions of six intelligent agents, and the total dimension is determined according to the number of controllable devices in each facility. In this embodiment, it is about 15 to 25 dimensions.

[0020] Distributed policy networks employ a centralized training and decentralized execution architecture, a classic paradigm in multi-agent reinforcement learning. During training, global information is used to learn collaborative policies; during execution, each agent makes independent decisions based solely on local observations, balancing learning and execution efficiency. Global value evaluation network To evaluate the global value of joint state-action pairs, a QMIX network structure is employed. This is a value decomposition-based multi-agent reinforcement learning method that decomposes the global action value into a nonlinear combination of the local action values ​​of each agent. The formula for calculating the global value evaluation network is as follows: , in, This is a set of network parameters, including the local Q-network parameters of each agent and the hybrid network parameters. For the first The local action value of an agent is calculated by the local Q-network based on local observations and actions. As a hybrid network, a hypernetwork architecture is used to generate hybrid weights, with the global state as the input to the hypernetwork. The output is a hybrid weight matrix. The design of the hybrid network satisfies the monotonicity constraint, i.e. To ensure that the global value monotonically increases with respect to the local value of each agent, this constraint is achieved by imposing a non-negativity constraint on the hybrid weights. The Local Q-Network adopts a three-layer fully connected structure, with the input dimension being the local observation dimension plus the action dimension, the number of hidden layer neurons being 256 and 128 respectively, the output dimension being 1, and the activation function being ReLU.

[0021] Local policy network Deployed in each agent, it outputs a probability distribution or deterministic action for scheduling based on local observations. For discrete actions such as pump start / stop, a classification strategy network outputs the selection probability of each action: , in, For the first Parameters of the agent policy network, This is the forward computation function of the policy network. For continuous actions such as output adjustment and opening adjustment, the mean and standard deviation of the action are output by the Gaussian policy network: , in, and These are the mean and standard deviation functions of a Gaussian distribution, calculated from the two output heads of the policy network. The policy network employs a three-layer fully connected structure with 256 and 128 neurons in the hidden layers, using ReLU activation. The output layer selects either Softmax activation or linear activation based on the action type.

[0022] The design of the global reward function is crucial for multi-objective collaborative optimization. It requires comprehensive consideration of multiple objectives, including power generation efficiency, flood control safety, ecological flow, and collaborative efficiency, and dynamic adjustment of the weights of each objective based on the current operating conditions. The global reward function defined in this invention comprises four sub-items: , in, This is a reward item for power generation efficiency; a positive value represents the revenue from power generation. For flood control safety rewards, negative values ​​indicate penalties for flood risk. This is an ecological traffic reward item; a negative value indicates a penalty for ecological violations. This is a reward for collaborative efficiency; negative values ​​represent penalties for collaborative misalignment. , , , These are the corresponding weight coefficients, satisfying... .

[0023] The power generation benefit incentive is calculated based on the total power generation and the grid connection price for the current period, encouraging agents to maximize power generation revenue while meeting security constraints. , in, This embodiment is a collection of hydropower station facilities, including two power stations: Qingshan Level 1 and Qingshan Level 2. For a moment The on-grid electricity price is expressed in yuan per kilowatt-hour, and is determined according to the peak-valley-flat pricing policy. The peak price is 0.65 yuan, the normal price is 0.45 yuan, and the off-peak price is 0.25 yuan. For the first The power generation coefficient of the hydropower station, taking into account both turbine efficiency and generator efficiency, ranges from 8.5 to 9.8. In this embodiment, the value is 9.2 for the Qingshan I-1 power station and 8.8 for the Qingshan II-2 power station. The power generation flow rate is expressed in cubic meters per second. The head for power generation is measured in meters and is calculated by subtracting the downstream water level from the upstream water level. The overall efficiency of the unit is defined as a value ranging from 0.85 to 0.95. The time step is in hours.

[0024] The flood control safety reward is calculated based on the deviation between the water level at key monitoring sections and the warning water level, and a secondary penalty is used to impose stronger penalties for exceeding the warning level: , in, This embodiment is a set of key monitoring sections, including six sections: upstream of the Qingshan Level 1 Dam, upstream of the Qingshan Level 2 Dam, upstream of the river control gate, and the intake pools of three pumping stations. For the first Each cross-section at time The real-time water level, in meters; The warning water level for this section shall be determined by the water resources authority. The cross-sectional weighting coefficient is determined based on the flood control importance of the cross-section. The value is 2.0 for the upstream reservoir dam section, 1.5 for the downstream river channel section, and 1.0 for the pump station intake section. When the real-time water level is lower than the warning level, When the function value is 0, no penalty is applied; when the real-time water level exceeds the warning level, the penalty value is proportional to the square of the exceedance, and the more severe the exceedance, the greater the penalty.

[0025] The ecological flow incentive is calculated based on the flow guarantee level of downstream ecological sections, encouraging agents to maintain appropriate downstream flow to preserve river ecological health. , in, This embodiment is a set of ecological cross sections, including two ecologically sensitive cross sections: the downstream of the Qingshan Second-Level Dam and the downstream of the control gate. For the first Each cross-section at time Real-time flow rate, in cubic meters per second; The target ecological flow for this section is determined based on the research results on river ecological water demand. During the flood season, the value is taken as 30% of the multi-year average flow, and during the non-flood season, the value is taken as 15% of the multi-year average flow. The cross-sectional weighting coefficient is 1.5 for cross-sections with high ecological sensitivity and 1.0 for general cross-sections. The penalty value adopts a relative deviation form to eliminate the influence of differences in flow rate at different cross-sections.

[0026] The collaborative efficiency reward is used to encourage coordination and cooperation among agents and to penalize uncoordinated scheduling actions. , in, For intelligent agents and The coordination coefficient between facilities is determined based on the hydraulic correlation between the facilities, and its value is consistent with the spatial correlation matrix. For intelligent agents At any moment The change in motion is the difference between the current action and the action at the previous moment. To determine the expected agent based on the cooperation rules The amount of change in actions taken, such as when an upstream power station increases water discharge, it is expected that the downstream pumping station will correspondingly increase its drainage capacity. This represents the maximum value of the change in motion, used for normalization.

[0027] The weighting coefficients are adaptively adjusted based on current hydrological and meteorological conditions through a dynamic attention mechanism, enabling the system to automatically switch scheduling modes according to operating conditions. A conditional feature vector is defined. The system has 12 dimensions and includes information such as the current water level deviation (the relative deviation between the water level at each section and the warning water level), meteorological warning level (0 for no warning, 1 for blue warning, 2 for yellow warning, 3 for orange warning, and 4 for red warning), electricity load demand (the ratio of current load to predicted load), and seasonal factors (1 for the flood season and 0 for the non-flood season). The formula for calculating the weighting coefficient is: , in, The weight adjustment matrix has dimensions of . ; The bias vector has a dimension of 4; both are trainable parameters, initialized through supervised learning using historical scheduling data and expert experience. Under the flood control mode during the flood season, when the meteorological warning level is greater than or equal to 2, The value ranges from 0.5 to 0.8, becoming the dominant target; in the dry season power generation mode, when the water level is sufficient and there is no weather warning, The value ranges from 0.4 to 0.6; during ecologically sensitive periods, such as the fish breeding season, The value range is from 0.3 to 0.5.

[0028] The policy optimization employs a multi-agent proximal policy optimization algorithm, a highly efficient and stable policy gradient method that avoids performance degradation by limiting the policy update magnitude. The advantage function is defined as follows: , in, This is a discount factor, ranging from 0.95 to 0.99. In this example, it is set to 0.97, indicating the importance attached to future rewards. Let be the state value function, estimated by the value network. The policy gradient update formula is: , in, The importance sampling ratio represents the ratio of the probability of the new and old policies on the current state action pair; The pruning parameter ranges from 0.1 to 0.2; in this embodiment, it is set to 0.15 to limit the policy update magnitude. During training, an experience replay buffer stores historical interaction data with a capacity of 200,000 records. For each update, 512 records are randomly sampled uniformly from the buffer to calculate the gradient. The Adam optimizer is used for parameter updates with a learning rate of 0.0003.

[0029] The digital twin simulation module is a security verification component of the system, responsible for testing the effectiveness and security of scheduling schemes in a virtual environment, implementing a security mechanism of simulation before execution. This module constructs a digital twin highly consistent with the actual water conservancy system, integrating river hydraulics models, hydraulic structure flow models, and hydraulic machinery response models, enabling rapid simulation of the execution effects of scheduling schemes. The simulation timescale can be more than 100 times larger than the actual time, meaning that one minute of simulation calculation can simulate 100 minutes of actual operation, providing ample verification time for scheduling decisions.

[0030] The hydraulic simulation engine is the core of the digital twin simulation module, providing numerical solutions for unsteady river flows based on the Saint-Venant equations. The continuity equation for unsteady flow in a one-dimensional open channel expresses the conservation of water volume, and the momentum equation expresses the conservation of momentum; together, they form a complete set of governing equations. , , in, The cross-sectional area of ​​the water passage is measured in square meters and is calculated from the cross-sectional shape and water level. The flow rate is expressed in cubic meters per second. Time, in seconds; The distance traveled is in meters. This indicates either inflow or outflow, measured in cubic meters per second per meter. Positive values ​​indicate inflow, and negative values ​​indicate outflow. The acceleration due to gravity is taken as 9.81 meters per second squared. This refers to the water level, measured in meters. The energy gradient is calculated using Manning's formula. ,in The roughness coefficient, The hydraulic radius is denoted as . The equations are discretized using a Preissmann four-point implicit difference scheme, which has unconditional stability and allows for a larger time step. The spatial step is determined based on the severity of river topographic changes: 200-500 meters for sections with gentle topographic changes and 50-100 meters for sections with dramatic changes. The time step is adaptively adjusted according to the Courant number stability condition, ranging from 1 to 60 seconds, with a typical value of 10 seconds. Boundary conditions include: the upstream inflow boundary uses the measured or predicted inflow hydrograph; the downstream water level boundary uses the water level-discharge relationship curve or the measured water level hydrograph; the internal boundary describes the flow characteristics of hydraulic structures, with weir flow or orifice flow formulas used for sluice gates and unit characteristic curves used for hydropower stations.

[0031] The equipment response simulator is used to simulate the dynamic response process of hydraulic machinery, including the gate opening and closing process and the output change process of the hydraulic turbine unit. The dynamic model of the gate opening and closing process is described by a first-order inertial element: , in, This represents the gate opening degree, with a value ranging from 0 to 1, where 0 indicates fully closed and 1 indicates fully open. The target opening degree is determined by the scheduling instructions; The opening and closing time constant depends on the gate size and the hoisting capacity. The time constant of the control gate ranges from 60 seconds to 300 seconds; in this embodiment, it is set to 120 seconds, meaning it takes approximately 5 minutes for the gate to go from fully closed to fully open. The hydraulic turbine output response model also uses a first-order inertial element. , in, The power output of the generating unit is measured in megawatts (MW). The amount of effort exerted to achieve the target is determined by the dispatch instructions; The output response time constant depends on the dynamic characteristics of the governor and turbine, and ranges from 5 to 30 seconds. In this embodiment, the value is 15 seconds for the Qingshan I-level power station and 20 seconds for the Qingshan II-level power station. The flow rate change during the pump start-up and shutdown process is described using an S-curve to more realistically reflect the actual start-up and shutdown process. , in, The stable operating flow rate is determined by the pump set characteristic curve; This is the upward slope coefficient, with a value ranging from 0.1 to 0.5; The half-response time point is the moment when the flow rate reaches half of its stable value.

[0032] The expected performance indicators from the simulation are used to quantitatively evaluate the effectiveness of the scheduling plan, including the following four categories of indicators. The first category is the highest water level in the downstream river channel. Calculate the highest water level at all monitoring sections during the simulation period, with the following constraints: ,in The water level is limited for flood control. The second category is total power generation. The cumulative power generation of all hydropower stations during the simulation period, expressed in megawatt-hours. The third category is the ecological baseflow guarantee rate. Calculate the proportion of time periods that meet ecological flow requirements out of the total time periods, with the following constraints: The fourth category is scheduling response time. The time from the issuance of the scheduling plan to the achievement of the target status by each facility is subject to the following constraints: ,in The preset response time limit is 30 minutes for regular dispatch and 10 minutes for emergency dispatch.

[0033] The scenario generator is used to stress test scheduling schemes and evaluate their safety and robustness under extreme conditions. Based on a historical database of extreme conditions and the Monte Carlo sampling method, various meteorological and hydrological scenarios are generated. Scenario parameters include inflow processes, rainfall processes, and downstream water level boundaries, with sampling distributions determined based on the statistical characteristics of historical data. Peak inflow flows are sampled using a Pearson III distribution with a return period covering 5 to 100 years; cumulative rainfall is sampled using a log-normal distribution, with a typical rainfall pattern used for temporal distribution; and downstream water level boundaries are sampled using a normal disturbance from the historical highest water level. Each scheduling scheme is evaluated using Monte Carlo simulations across 200 scenarios, and the mean, standard deviation, and extreme quantiles of each indicator are statistically analyzed. When any simulation indicator in any scenario exceeds the safety threshold, violation scenario information is recorded, including scenario parameters, violation indicators, and violation magnitude, and fed back to the multi-agent collaborative scheduling engine to guide the strategy towards a safer direction.

[0034] The task decomposition and collaborative execution module is responsible for transforming the simulation-verified scheduling scheme into executable operational tasks and coordinating the execution process of various intelligent duty units. It serves as a bridge connecting cloud-based decision-making and edge execution. This module first analyzes the scheduling scheme, extracting the target state parameters of each facility, including target unit output, target gate opening, target pump operation combination, and target achievement time. Then, it compares the target state with the current state, calculates the state difference, and generates incremental adjustment tasks.

[0035] The task execution graph uses a directed acyclic graph (DAG) structure to represent the dependencies between tasks and is the core data structure for task scheduling. Nodes in the graph represent atomic operation tasks. Each task includes a task identifier (unique number), target facility (facility code), operation type (output regulation, opening regulation, start / stop control, etc.), parameter value (target value or regulation amount), and expected execution time (start time and duration). Directed edges represent the dependencies between tasks, and edge weights represent the minimum time interval constraint, i.e., the shortest waiting time required after the preceding task is completed. The determination of task dependencies follows these rules: upstream facility water release tasks must be executed before downstream facility water storage tasks to avoid safety hazards caused by a sudden rise in downstream water levels; multiple operation tasks of the same facility are executed in the order determined according to equipment safety procedures, for example, the gate lock must be released before opening a sluice gate; power generation tasks involving grid dispatch must be coordinated with the grid dispatch center before execution to ensure the safe and stable operation of the grid; flood control emergency tasks take priority over regular operation tasks.

[0036] Critical path analysis is used to determine parallel execution strategies for tasks and optimize the overall execution time of the scheduling scheme. The length of the critical path in the task execution graph is calculated; the critical path is the longest path from the start node to the end node, determining the shortest completion time of the scheduling scheme. Groups of tasks that can be executed in parallel are identified, meaning tasks not on the same dependency chain can be executed simultaneously. For tasks without dependencies, a parallel delivery strategy is adopted, with multiple intelligent agents simultaneously receiving tasks and starting execution, minimizing the total execution time. For tasks with dependencies, subsequent tasks are automatically triggered after the preceding task is successfully executed and confirmed. Triggering conditions include the preceding task's status changing to successful completion and the waiting time meeting the minimum interval constraint. Task execution status is updated in real time, including four states: pending execution, executing, successful completion, and execution failure. State transitions are driven by feedback messages from intelligent agents. Failed tasks are handled according to a preset strategy: retryable tasks are automatically re-delivered, with a maximum of 3 retries; non-retryable tasks are reported for manual processing, and subsequent tasks dependent on that task are suspended.

[0037] The intelligent duty unit cluster serves as the edge execution layer of the system, responsible for non-intrusive facility access and local control. This is the key innovation that distinguishes this invention from traditional centralized control systems. Each intelligent duty unit is deployed in the control room of the water conservancy facility, working in parallel with the existing automation system. It interacts with the facility's central control computer through visual perception and simulated peripheral operation, without requiring any modification or intrusion into the original system.

[0038] The hardware of the edge intelligent terminal consists of three main parts: an industrial-grade embedded computer, a video acquisition unit, and an operation execution unit. The industrial-grade embedded computer is the computing core of the intelligent terminal, employing a multi-core processor based on Intel or ARM architecture with a clock speed of 1.5GHz to 2.5GHz, 8GB to 16GB of memory, and industrial-grade solid-state drives with capacities ranging from 128GB to 512GB. The computer supports edge AI inference acceleration and can be optionally equipped with an NVIDIA Jetson series GPU module or an Intel Movidius neural computing stick to accelerate deep learning inference during interface parsing. The operating system uses Ubuntu Linux or a customized real-time operating system to ensure system stability and real-time performance. The video acquisition unit is responsible for capturing the display screen of the facility's central control computer. There are two implementation methods: The first uses a video capture card, directly connected to the central control computer's display output port via HDMI or VGA. The capture card supports resolutions from 1920×1080 to 3840×2160 pixels, frame rates from 30 to 60 frames per second, and latency controlled within 20 milliseconds. The second method uses a high-definition industrial camera, aimed at the central control computer's display screen. This is suitable for scenarios where a display output port cannot be accessed. The camera resolution is 1920×1080 pixels, with a frame rate of 30 frames per second, requiring image correction algorithms to eliminate perspective distortion. The operation execution unit uses a USB-HID emulator, connected to the facility's central control computer via a physical USB interface, simulating standard keyboard and mouse signals. The central control computer recognizes it as a normal input device, requiring no driver installation. The USB-HID emulator supports the following operations: mouse movement (absolute coordinates and relative displacement), mouse buttons (left button, right button, middle button, scroll wheel), keyboard buttons (104-key standard keyboard), key combinations (Ctrl, Alt, Shift, etc.), and text input (ASCII characters and Unicode characters).

[0039] The interface parsing unit intelligently analyzes the acquired human-machine interface images to extract structured equipment status data. The interface parsing process includes four steps: image preprocessing, control detection, text recognition, and data extraction. Image preprocessing denoises, enhances, and corrects the original images to improve the accuracy of subsequent recognition. Control detection uses the YOLO series target detection model to identify functional control areas in the interface, including common control types such as numerical display boxes, status indicators, buttons, input boxes, drop-down menus, and sliders. The detection model is trained using a large number of labeled industrial control interface images, achieving a detection accuracy of over 98% and an average inference time of 15 milliseconds per frame. Text recognition performs optical character recognition on the numerical display box area, using a CRNN (Convolutional Recurrent Neural Network) model to recognize numbers, letters, and Chinese characters. The recognition model is trained separately for three common display formats: seven-segment displays, LCD displays, and graphical interfaces, achieving a recognition accuracy of over 99.5%. Status indicators determine the working status using a color classifier, employing the HSV color space for color feature extraction, supporting four common indicator colors: red, green, yellow, and gray, with a classification accuracy of over 99%. Data extraction matches the recognition results with the interface template, establishes a correspondence between control positions and data semantics, and outputs structured key-value pairs. The parsing results are encapsulated according to a unified data model in JSON format, containing fields such as device identifier, parameter name, parameter value, unit, timestamp, and quality flag, and are reported to the cloud platform via a secure communication network.

[0040] The operation execution unit generates analog input signals based on control commands issued by the cloud platform, driving the central control computer to execute corresponding operations. Control commands are encapsulated in JSON format, containing fields such as command identifier, operation type, target control, parameter value, and timeout. Operation types include: mouse movement (specifying target coordinates), mouse click (specifying click location and key), mouse double-click (specifying double-click location), mouse drag (specifying start and end points), keyboard key press (specifying key codes), text input (specifying input strings), and combined operations (specifying operation sequences). The execution process employs a visual feedback verification mechanism. After each operation is executed, the interface image is re-captured, and interface parsing confirms whether the operation was successfully implemented. The verification logic is designed according to the operation type: after clicking a button, it checks if the button state changes; after inputting a value, it checks if the display box updates to the target value; after adjusting a slider, it checks if the position reaches the target position. If the operation is ineffective, a retry is performed, with an appropriate delay (100 to 500 milliseconds) added before the retry to allow the interface to refresh. The maximum number of retries is 3. All operations are fully recorded, including operation timestamps, operation types, operation parameters, execution results, screenshots, number of retries, etc. The logs are stored locally and synchronized to the cloud platform for post-event auditing and problem tracing.

[0041] The intelligent cockpit module serves as the system's human-machine interface, providing dispatchers with panoramic situational awareness and convenient operation access, making it a core tool for dispatch and command. This module utilizes WebGL technology to construct a 3D visualization scene, developed using the Cesium or Three.js graphics engine, supporting access from mainstream browsers without requiring a dedicated client. The 3D scene integrates multi-source spatial data, including Geographic Information System (GIS) maps (providing basic base maps of topography, water systems, and administrative divisions), facility BIM models (providing detailed 3D models of hydropower plant buildings, pumping station structures, and sluice gate structures), real-time data (dynamically labeled digitally overlaid on the model), and early warning information (presented with prominent icons and color codes). The 3D scene displays all water conservancy facilities within the basin according to their actual geographical locations and scale, supporting multiple perspective operations such as panning, zooming, rotating, and tilting. The level of detail in the facility models dynamically adjusts according to the viewing distance, displaying detailed models at close range and simplified models at a distance to ensure rendering performance. Real-time data, including core parameters such as water level, flow rate, output, and opening degree, is overlaid on key locations of the facility models in the form of digital labels, and the data is updated every second. The changing trends of key parameters are visually presented through color coding: green indicates the normal range, yellow indicates approaching the limit, and red indicates exceeding the limit. Warning information is pushed to dispatchers using eye-catching dynamic icons and audio-visual alerts. Warning levels are indicated by four levels: blue, yellow, orange, and red. Warning content includes abnormal events such as water level exceeding limits, equipment failure, communication interruption, and task timeout.

[0042] The intelligent cockpit module also provides an interpretable display of scheduling decisions, helping dispatchers understand the decision-making logic of the AI ​​system and enhancing the trust and efficiency of human-machine collaboration. After the multi-agent collaborative scheduling engine generates a scheduling plan, the cockpit automatically displays the decision-making basis and expected effects of the plan, mainly including four aspects: current hydrological and meteorological condition analysis, showing recent water level change trends, weather forecast curves, and electricity load forecasts in chart form; target weight configuration, showing the current weights and adjustment reasons for the four targets—power generation benefits, flood control safety, ecological flow, and collaborative efficiency—in pie chart form; suggested actions for each facility, showing the adjustment direction, adjustment range, and target status of each facility in list form; and simulation prediction results, showing the predicted water level and flow changes at each key section after implementing the scheduling plan in time-series curve form, compared with safety limits. Dispatchers can view the detailed reasoning chain of the decision-making process, and clicking on any facility expands to display the facility's intelligent agent's local observations, action selection basis, and expected benefits. The cockpit also supports manual intervention and plan adjustments; dispatchers can directly modify target parameters or issue control commands through point-and-click operations. Manually modified parameters are transmitted in real time to the digital twin simulation module for effect prediction. The cockpit simultaneously updates the simulation prediction curves to help dispatchers assess the impact of manual intervention. After confirming the manual adjustment plan, the task decomposition and collaborative execution module distributes and executes tasks according to the same process as the automated plan.

[0043] like Figure 2 As shown, the intelligent water conservancy scheduling method of the present invention includes five main steps, namely, data acquisition and fusion step S1, collaborative scheduling decision step S2, simulation verification step S3, task distribution step S4, and local execution feedback step S5. The five steps are executed in a loop to form a complete scheduling closed loop of perception, decision, verification, execution, and feedback.

[0044] Step S1 is the data acquisition and fusion step, the starting point of the scheduling closed loop, responsible for building full-domain status awareness capabilities. Six edge intelligent terminals in the intelligent duty system cluster continuously run interface acquisition and parsing tasks. Each terminal extracts operational data from the display screen of the facility's central control computer at a set acquisition cycle (key parameter 1 to 5 seconds, general parameter 30 to 60 seconds). The video acquisition unit captures high-definition images, and the interface parsing unit sequentially performs image preprocessing, control detection, text recognition, and data extraction, outputting structured status data. The extracted data, after local caching and compression, is reported to the cloud platform through a VPN encrypted tunnel, with the reporting cycle consistent with the acquisition cycle. After receiving the reported data from the six facilities, the multi-source data fusion and status awareness module first performs timestamp verification, checking whether the deviation between the data timestamp and the server time is within the allowable range (50 milliseconds). Data exceeding this range undergoes timestamp correction or is marked as low-quality data. Then, outlier filtering is performed, using a sliding window statistical method to detect numerical jumps and persistent anomalies. Abnormal data is interpolated for repair or removed. Next, the aforementioned spatiotemporal alignment method is used to perform temporal alignment and spatial correlation processing on the multi-source data. A multi-head attention mechanism is then employed for feature fusion to generate a global state dataset. This global state dataset is written to a time-series database and simultaneously pushed to the state cache of the multi-agent collaborative scheduling engine. The data update cycle is 1 to 10 seconds, dynamically adjusted according to system load and network conditions to ensure that the global state held by the cloud platform is highly synchronized with the actual facility state, with a synchronization delay controlled within 5 seconds.

[0045] Step S2 is the collaborative scheduling decision-making step, the core of the scheduling closed loop, responsible for generating cross-facility collaborative scheduling schemes. The multi-agent collaborative scheduling engine periodically reads the latest global state data from the state buffer. The decision cycle is configured according to scheduling needs, ranging from 15 to 30 minutes in normal operation mode and 1 to 5 minutes in emergency response mode. After reading the state data, the engine first constructs a state space for multi-agent reinforcement learning, converting the global state data into local observation vectors and global state vectors for each agent. The conversion process includes feature selection, dimension mapping, and normalization. Then, each agent's local policy network performs forward inference based on local observation information, outputting the probability distribution (discrete actions) or Gaussian distribution parameters (continuous actions) of scheduling actions, and sampling specific actions. The action combination of the six agents forms a joint action vector. The global value evaluation network evaluates the value of the joint state-action pairs, outputting a Q-value as a measure of action quality. The global reward function calculates immediate rewards based on four objectives: power generation efficiency, flood control safety, ecological flow, and collaborative efficiency. The weights of each objective are adaptively determined by a dynamic attention mechanism based on current conditions. The policy network updates its parameters based on reward signals and the advantage function, employing a proximal policy optimization algorithm to limit the update magnitude and ensure training stability. After multiple iterations, the policy converges to near-optimal, and inference is performed during online runtime to obtain the current optimal action. The decision result output is a cross-facility collaborative scheduling scheme, encapsulated in JSON format, containing information such as the sequence of target operating parameters for each facility in future scheduling cycles, the target achievement time for each parameter, scheduling mode identifier, and priority identifier.

[0046] Step S3 is the simulation verification step, a crucial safety assurance link in the scheduling closed loop, responsible for pre-verifying the effectiveness and security of the scheduling scheme in a virtual environment. After receiving the scheduling scheme output by the multi-agent collaborative scheduling engine, the digital twin simulation module first analyzes the scheme, extracting the target state parameters and execution timing of each facility. Then, it initializes the simulation environment, using the current global state data as the initial simulation conditions and converting the scheduling scheme into simulation boundary conditions. The hydraulic simulation engine initiates the solution of the Saint-Venant equations, employing the Preissmann four-point implicit difference scheme for time progression, iteratively solving for the spatial distribution of water level and flow within each time step. The equipment response simulator synchronously simulates the gate opening and closing and unit output changes, using the equipment state as the internal boundary conditions of the hydraulic simulation. The scenario generator generates various disturbance scenarios based on current weather forecasts and historical data to stress test the scheduling scheme. After the simulation calculation is completed, various expected performance indicators are statistically analyzed, including the highest downstream river level, total power generation, ecological baseflow guarantee rate, and scheduling response time. Each indicator is compared with the safety constraints: if all indicators meet the constraints, the scheduling scheme passes verification and proceeds to step S4; if any indicator exceeds the constraints, the violation information is recorded and fed back to the multi-agent collaborative scheduling engine. Upon receiving the feedback, the engine adds the violation scenario information as negative samples to the training data, adjusts the policy parameters to move towards a safer direction, regenerates the scheduling scheme, and resubmits it for verification. This iterative process continues until the scheme passes verification or reaches the maximum number of iterations (usually 5). If verification still fails after reaching the maximum number of iterations, a manual intervention process is triggered, pushing the scheme and violation information to the intelligent cockpit for review and modification by the scheduling personnel.

[0047] Step S4 is the task distribution step, which is the instruction issuance stage of the scheduling closed loop. It is responsible for converting the scheduling scheme into executable tasks and distributing them to the edge devices. After receiving the verified scheduling scheme, the task decomposition and collaborative execution module first performs scheme parsing, extracting the target state parameters of each facility, including the target output of the unit, the target opening degree of the gate, and the target operating combination of the pump group. Then, it compares the target state with the current state read from the state buffer, calculates the state difference, and generates atomic operation tasks for incremental adjustment. Each atomic task corresponds to a specific control operation, and the task granularity is determined according to the equipment characteristics: the granularity of the unit output adjustment task is 1MW, the granularity of the gate opening adjustment task is 0.5m, and the pump group start-up and shutdown task is the start or stop of a single pump. After the tasks are generated, a directed acyclic graph (DAG) task execution graph is constructed based on the logical and temporal dependencies between tasks. The dependencies are determined according to hydraulic connections and safety regulations. The critical path analysis algorithm is used to calculate the critical path length of the task execution graph and identify task groups that can be executed in parallel. Tasks are distributed to corresponding smart duty units via a VPN encrypted tunnel, employing reliable transmission protocols (such as MQTT QoS2) to ensure accurate delivery of task instructions to the target device. Tasks without dependencies can be issued in parallel, with multiple smart duty units receiving tasks simultaneously; tasks with dependencies are issued sequentially according to their dependency order, with subsequent tasks automatically triggered after the successful execution of a preceding task. Detailed logs are recorded throughout the task distribution process, including task identifier, target facility, issuance time, and confirmation time.

[0048] Step S5 is the local execution and feedback step, which is the end point of the scheduling closed loop and the trigger point for a new cycle. It is responsible for completing the control operation and feeding back the execution result. After receiving the task instruction, the intelligent duty unit first performs task verification, checking the legality of the task format and the validity of the parameter range. After the verification is passed, the task is added to the local task queue. The operation execution unit retrieves the task from the task queue and generates the corresponding analog input signal sequence according to the task type. For button click operations, the execution unit sends a mouse movement signal (moves to the center coordinates of the button) and a mouse click signal (presses and releases the left button) in sequence; for numerical input operations, the execution unit sends a mouse movement signal (moves to the coordinates of the input box), a mouse click signal (clicks to activate the input box), and a keyboard signal (inputs digits or characters in sequence). The analog signals are sent to the facility's central control computer through the USB-HID interface. After receiving the signals, the central control computer executes the corresponding response. After the operation is executed, the video acquisition unit re-acquires the interface image, and the interface parsing unit extracts the current state of the operation target area and compares it with the expected state to verify whether the operation has been successfully implemented. If the operation is successful, the task status is updated to "successfully completed," and an execution result feedback message is generated. If the operation fails, a retry or reporting is determined based on the reason for the failure. During retry, an appropriate delay is added before re-executing the operation sequence, with a maximum of 3 retries. The execution result feedback message includes information such as task identifier, execution status, execution time, target achievement value, and interface screenshots, and is reported to the cloud platform via a secure communication network. The task decomposition and collaborative execution module updates the task status based on the execution feedback. Successfully completed tasks are removed from the execution graph, triggering the issuance of subsequent dependent tasks. The multi-source data fusion and status awareness module continuously collects the facility status after execution and compares it with the expected status to verify whether the scheduling effect has achieved the expected goal. The execution results and effect data are stored as historical experience in the experience replay buffer for continuous learning and strategy optimization of the multi-agent collaborative scheduling engine, realizing the self-evolution of scheduling capabilities, completing a complete scheduling loop, and preparing for the next cycle.

[0049] The embodiments of the present invention are not limited to the specific embodiments described above. Those skilled in the art can make various equivalent changes or substitutions based on the technical solutions of the present invention, and all such changes or substitutions should be included within the protection scope of the present invention.

Claims

1. A smart water conservancy centralized control center system based on an intelligent duty system cluster, characterized in that: It adopts a three-layer architecture consisting of a cloud-based centralized control platform layer, an edge-based intelligent duty system layer, and a secure communication layer, including: The multi-source data fusion and status awareness module, deployed in the cloud, is used to aggregate multi-source heterogeneous data from the intelligent duty system cluster and perform spatiotemporal alignment and feature fusion to generate a global status dataset. A multi-agent collaborative scheduling engine, deployed in the cloud, models each water conservancy facility as an agent with perception, decision-making, and execution capabilities. It employs a centralized training and decentralized execution architecture. Its global value evaluation network uses a QMIX network structure to nonlinearly mix the local values ​​of each agent, and achieves this through non-negativity constraints on the mixed network weights. Monotonicity constraint; The digital twin simulation and deduction module is deployed in the cloud and is used to perform simulation verification in a digital twin that is consistent with the actual water conservancy system before the scheduling plan is issued and executed. The simulation time scale is more than 100 times the actual time, and the scheduling plan is only allowed to be issued and executed after the simulation verification passes the security constraints. The task decomposition and collaborative execution module, deployed in the cloud, is used to parse the simulation-verified scheduling scheme into a task execution graph with dependencies and distribute it. The intelligent duty system cluster consists of edge intelligent terminals deployed next to the control rooms of various water conservancy facilities. It bypasses the existing facility control system in a non-intrusive access manner without modifying the original control loop. It is responsible for local data acquisition and control operation execution, and runs safely locally according to the established program when the cloud network is interrupted. The secure communication layer uses a VPN encrypted tunnel on the water conservancy private network to connect the cloud and the edge.

2. The system according to claim 1, characterized in that, The multi-source data fusion and state awareness module adopts a multi-head attention mechanism to achieve deep fusion of different types of operational data, with the number of attention heads ranging from 4 to 8.

3. The system according to claim 1, characterized in that, The global reward function of the multi-agent collaborative scheduling engine includes four sub-items: power generation efficiency, flood control safety, ecological flow, and collaborative efficiency. The weight of each sub-item is adaptively adjusted according to the current hydrological and meteorological conditions through a dynamic attention mechanism.

4. The system according to claim 1, characterized in that, The hydraulic simulation engine of the digital twin simulation and derivation module performs numerical solutions for unsteady river flow based on the Saint-Venant equations and uses the Preissmann four-point implicit difference scheme for discretization.

5. The system according to claim 1, characterized in that, The digital twin simulation and deduction module also includes a scene generator, which is used to generate various meteorological and hydrological scenarios for stress testing based on historical extreme working condition databases and Monte Carlo sampling methods.

6. The system according to claim 1, characterized in that, It also includes an intelligent cockpit module, which uses WebGL technology to build 3D visualization scenes and provide interpretable displays of scheduling decisions.

7. A smart water conservancy scheduling method based on the system described in any one of claims 1 to 6, characterized in that, include: S1, Data Acquisition and Fusion Steps; S2, Collaborative Scheduling Decision-Making Steps; S3, Simulation Verification Step: The scheduling scheme generated in step S2 is simulated and deduced in the digital twin first, and only proceeds to step S4 if the safety constraints are met. S4, Task distribution steps; S5, Local execution and feedback steps.

8. The method according to claim 7, characterized in that, In step S2, the first An intelligent agent at time Local observation It consists of four parts: the facility's real-time operating parameter vector, recent historical status sequence, meteorological forecast data, and the correlation status of upstream and downstream facilities.

9. The method according to claim 7, characterized in that, In step S3, Monte Carlo simulation evaluation is performed on each scheduling scheme in 200 scenarios. When the simulation index in any scenario exceeds the safety threshold, the violation scenario information is recorded and fed back to step S2 to change the safety constraints.

10. The method according to claim 7, characterized in that, In step S5, when the cloud network is interrupted, the edge intelligent duty unit can still operate safely locally according to the established procedure.

Images