AI collaborative scheduling optimization system and method for cross-energy network complementary to water and electricity

Abstract

The application discloses a water and electricity complementary cross-energy network AI collaborative scheduling optimization system and method, the system comprises a data synchronization module, a graph construction module, a collaborative perception module, a consensus generation module and a scheduling execution module; the data synchronization module is used for time alignment of hydrological data, unit operation data and power grid load data; the graph construction module is used for constructing a water and electricity double-layer graph containing a water power directed graph, an electrical undirected graph and a cross-layer coupling edge; the collaborative perception module is used for generating collaborative perception embedding of each scheduling agent; the consensus generation module is used for generating a consensus signal vector; the scheduling execution module is used for outputting scheduling actions and projecting to a physical feasible region to obtain scheduling instructions. The application realizes unified modeling of the water power relationship of cascade hydropower stations, the power grid topological relationship and the water power coupling relationship, and can improve the accuracy, robustness and engineering implementability of multi-station collaborative scheduling.

Description

Technical Field

[0001] This invention relates to the field of hydropower station dispatching technology, and in particular to an AI-based collaborative dispatching optimization system and method for hydropower complementarity across energy networks. Background Technology

[0002] With the development of new power systems, cascade hydropower stations, in addition to their conventional power generation tasks, also need to meet requirements such as peak shaving, frequency regulation, and grid load response. In actual operation, there are significant upstream and downstream hydraulic connections between hydropower stations within a river basin. The inflow and outflow processes of upstream stations affect the water level, flow rate, and power generation capacity of downstream stations. Simultaneously, after grid connection, the output of each hydropower station is influenced by changes in grid load, transmission channel capacity, and network structure constraints. Therefore, cascade hydropower scheduling is a complex optimization problem involving the coupling of hydraulic and electrical processes.

[0003] Existing technologies for scheduling cascade hydropower stations primarily focus on reservoir water level control, power generation plan allocation, or water volume optimization within a single river basin. They typically consider hydrological conditions, water balance, and unit operating boundaries, but fail to adequately address grid-side load demand and network transmission constraints, making it difficult to meet the application requirements of coordinated scheduling between hydropower station groups and the power grid. On the other hand, while some power dispatching methods consider load changes and power allocation, they often treat hydropower units as ordinary power sources, failing to fully reflect the time delay of water flow propagation, head changes, and coupling characteristics in the conversion of hydropower to electrical energy between cascade power stations. This results in models that cannot accurately reflect real-world operating patterns. Furthermore, hydrological data, unit operating data, and grid load data often come from different sources and have different sampling frequencies. Existing methods still have shortcomings in synchronizing multi-source heterogeneous data, modeling cross-network relationships, and multi-station collaborative decision-making, easily leading to local optimization but insufficient overall coordination. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned shortcomings and provide a cross-energy network AI collaborative scheduling optimization system and method for hydropower complementarity, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by this invention is: a cross-energy network AI collaborative scheduling optimization system for hydropower complementarity, applied to the collaborative scheduling of a cascade hydropower station group and the power grid in the same river basin. Each hydropower station in the hydropower station group is equipped with a scheduling intelligent agent. The system includes: The data synchronization module is used to acquire the unit equipment parameters and power grid line parameters of each hydropower station, collect hydrological data, unit operation data and power grid load data of each hydropower station and perform time alignment to obtain synchronization characteristic data; The graph construction module is used to construct a hydropower bilayer graph based on the upstream and downstream river topology and power grid line topology of each hydropower station, and to construct the synchronous feature data as the node features of the hydropower bilayer graph. The collaborative perception module is used to control each scheduling agent to perform working condition adaptive graph attention aggregation in the hydraulic directed graph and the electrical undirected graph of the hydropower double-layer graph to obtain an aggregated message, and to fuse the aggregated message with the corresponding synchronous feature data to generate a collaborative perception embedding. The consensus generation module is used to perform attention pooling on all collaborative sensing embeddings to obtain a watershed summary vector, which is then fused with the power grid demand vector to generate a consensus signal vector, which is then broadcast to each scheduling agent. The scheduling execution module is used to control the policy network of each scheduling agent to receive the corresponding collaborative perception embedding and the consensus signal vector, infer and output scheduling actions and project them onto the physical feasible domain to obtain scheduling instructions.

[0006] Preferably, the graph construction module includes a hydraulic graph construction unit, an electrical graph construction unit, a node feature extraction unit, and a cross-layer coupling unit, wherein: The hydraulic graph construction unit is used to construct a hydraulic directed graph with each hydropower station as a node and the upstream and downstream water flow transmission direction as directed edges. The attributes of each directed edge include the water flow propagation time delay from the upstream station to the downstream station and the interval attenuation coefficient. The electrical graph construction unit is used to construct an undirected electrical graph with the grid-connected nodes of each hydropower station and the substation nodes of the power grid as nodes and the transmission lines as edges. The attributes of each edge include the line reactance and the upper limit of transmission power. The node feature extraction unit is used to extract the upstream water level, tailwater level, inflow, outflow, and turbine flow as hydraulic layer node features from the synchronous feature data, and to extract the unit output and grid load as electrical layer node features. The cross-layer coupling unit is used to establish cross-layer coupling edges between the hydraulic layer nodes and electrical layer nodes corresponding to the same hydropower station, forming a hydropower two-layer diagram. The attributes of the cross-layer coupling edge include the net head of the hydropower station and the partial derivatives of the unit output with respect to the net head and the flow rate through the unit.

[0007] Preferably, the cross-layer coupling unit is configured with coupling edge attribute calculation logic, which includes: The net head is calculated based on the upstream water level, the tailwater level, and the hydraulic loss, wherein the hydraulic loss is calculated based on the flow rate through the generator and the parameters of the generator unit. Based on the turbine characteristic curve of the unit, the partial derivatives of the net head and flow rate are calculated under the operating conditions determined by the net head and flow rate, respectively, to obtain the partial derivatives of the unit output with respect to the net head and the partial derivatives of the unit output with respect to the flow rate.

[0008] Preferably, the collaborative sensing module includes a working condition coding unit, a hydraulic aggregation unit, an electrical aggregation unit, and a feature fusion unit, wherein: The operating condition coding unit is used to determine the current operating condition based on hydrological data and power grid load data and encode it into an operating condition coding vector. The operating condition coding vector is then fused with the node features of the hydraulic directed graph and the electrical undirected graph to generate hydraulic layer scaling factor and electrical layer scaling factor. The hydraulic aggregation unit is used by each scheduling agent to perform time-series alignment of the node features of adjacent nodes based on the hydraulic directed graph with the water flow propagation time delay, calculate the hydraulic layer attention coefficient and scale it with the hydraulic layer scaling factor, and use the scaled hydraulic layer attention coefficient to weighted aggregate the aligned adjacent node features to obtain their respective hydraulic layer aggregation messages. The electrical aggregation unit is used by each scheduling agent to calculate the electrical layer attention coefficient based on the electrical undirected graph and scale it with the electrical layer scaling factor, and then use the scaled electrical layer attention coefficient to weight and aggregate the features of adjacent nodes to obtain their respective electrical layer aggregation messages. The feature fusion unit is used to concatenate the hydraulic layer aggregation message and the electrical layer aggregation message with the hydraulic layer node features and electrical layer node features of the corresponding intelligent agent and then perform nonlinear mapping to generate the collaborative perception embedding of each scheduling intelligent agent.

[0009] Preferably, the operating condition coding unit is configured with operating condition coding logic, which includes: The current hydrological condition level is determined by comparing the inflow and upstream water level of each hydropower station with preset hydrological thresholds. The current load condition level is determined by comparing the power grid load with preset load thresholds. The hydrological condition level and the load condition level are individually thermally encoded and then concatenated to obtain the condition encoding vector. The mean values ​​of all node features in the hydraulic directed graph and the electrical undirected graph are respectively taken to obtain the corresponding feature mean values. The working condition encoding vector is concatenated with each feature mean value, projected and normalized to obtain the hydraulic layer scaling factor and the electrical layer scaling factor, respectively.

[0010] Preferably, the hydraulic aggregation unit is configured with hydraulic aggregation logic, which includes: For each scheduling agent, obtain the water flow propagation time delay of each upstream adjacent node in the hydraulic directed graph that points to the node where the agent is located, and backtrack the corresponding time delay at the current time to obtain the historical node features of each adjacent node, which are used as the aligned adjacent node features. The agent's own node features, aligned features of each neighboring node, and edge attributes of the corresponding directed edges are concatenated. The original attention scores of each neighboring node are calculated by nonlinear activation. The original attention scores of all neighboring nodes are normalized to obtain the hydraulic layer attention coefficient. The hydraulic layer attention coefficient is scaled using the hydraulic layer scaling factor, and the scaled hydraulic layer attention coefficient is used to perform a weighted summation of the features of each aligned adjacent node to obtain the hydraulic layer aggregated message.

[0011] Preferably, the feature fusion unit is configured with feature fusion logic, the feature fusion logic including: The hydraulic layer aggregated message, electrical layer aggregated message, hydraulic layer node features, and electrical layer node features of each scheduling agent are concatenated along the feature dimension to obtain the fused feature vector; The fused feature vectors are sequentially subjected to linear transformation and nonlinear activation to obtain the collaborative perception embedding of each scheduling agent.

[0012] Preferably, the consensus generation module includes a pooling unit, a demand encoding unit, and a consensus broadcasting unit, wherein: The pooling unit is used to calculate the attention weights corresponding to each collaborative perception embedding based on the collaborative perception embeddings of all scheduling agents, and to sum the attention weights of all collaborative perception embeddings to obtain the watershed summary vector. The demand encoding unit is used to calculate the power grid dispatch demand index based on the power grid load data and encode it into a power grid demand vector. The consensus broadcasting unit is used to concatenate the watershed summary vector and the power grid demand vector and perform nonlinear mapping to generate a consensus signal vector, and then broadcast the consensus signal vector to each scheduling agent.

[0013] Preferably, the scheduling execution module includes a strategy reasoning unit, a feasible region determination unit, and a projection unit, wherein: The policy reasoning unit is used by the policy network of each scheduling agent to concatenate the corresponding collaborative perception embedding with the consensus signal vector as input, and reason out the scheduling action, which includes the power allocation of each unit and the reservoir discharge flow. The feasible region determination unit is used to determine the physical feasible region at the current moment based on the current water level constraints, the upper and lower limits of unit output, and the minimum ecological flow. The projection unit is used to project the scheduling action onto the physical feasible domain to obtain the scheduling instruction.

[0014] Preferably, the policy networks of each scheduling agent in the policy inference unit are obtained through centralized training using a distributed execution paradigm. The system further includes a training module configured with training logic, which includes: Construct a local value network for each scheduling agent, and evaluate the local value based on the cooperative perception embedding, consensus signal vector and scheduling action of the corresponding agent. Local reward functions for each scheduling agent are constructed based on power generation revenue, water resource utilization efficiency, penalties for exceeding water level limits, and penalties for violations of ecological flow regulations. A value hybrid network is constructed. Under the condition that the global value increases monotonically with respect to each local value, the local values ​​of all agents are synthesized into the global value through the value hybrid network. The value hybrid network punishes joint actions that violate coupling constraints through Lagrange multipliers of water balance constraints and power balance constraints. During the execution phase, each policy network only receives the cooperative perception embedding and consensus signal vector of the corresponding agent for independent reasoning; The local reward at each time step is calculated based on the local reward function of each scheduling agent. The parameters of the local value network and the value mixing network are updated with the temporal difference error of the global value. The policy gradient is updated by backpropagating the global value to each policy network through the value mixing network until the temporal difference error of the global value is lower than the preset error threshold.

[0015] This invention also discloses a cross-energy network AI collaborative scheduling optimization method for hydropower complementarity, applied to the aforementioned cross-energy network AI collaborative scheduling optimization system for hydropower complementarity. The method includes the following steps: S1: Obtain the unit equipment parameters and power grid line parameters of each hydropower station, collect hydrological data, unit operation data and power grid load data of each hydropower station and perform time alignment to obtain synchronization characteristic data; S2: Construct a hydropower bilayer graph based on the upstream and downstream river topology and power grid line topology of each hydropower station, and construct the synchronous feature data as the node features of the hydropower bilayer graph; S3: Each scheduling agent performs condition-adaptive graph attention aggregation in the hydraulic directed graph and electrical undirected graph of the hydropower double-layer graph to obtain an aggregated message, and then fuses the aggregated message with the corresponding synchronization feature data to generate a collaborative perception embedding. S4: Attention pooling is performed on all collaborative perception embeddings to obtain the watershed summary vector, which is then fused with the power grid demand vector to generate a consensus signal vector, which is then broadcast to each scheduling agent. S5: The policy network of each scheduling agent receives the corresponding cooperative perception embedding and the consensus signal vector, infers and outputs scheduling actions and projects them onto the physical feasible domain to obtain scheduling instructions.

[0016] Furthermore, step S2 specifically includes: S2.1: Using each hydropower station as a node and the upstream and downstream water flow transmission direction as directed edges, construct a hydraulic directed graph. The attributes of each directed edge include the water flow propagation time delay and the interval attenuation coefficient from the upstream station to the downstream station. The water flow propagation time delay and the interval attenuation coefficient are calculated through historical hydrological data. S2.2: Construct an electrical undirected graph with the grid connection nodes of each hydropower station and the substation nodes of the power grid as nodes and the transmission lines as edges. The attributes of each edge include the line reactance and the upper limit of transmission power. S2.3: Extract the upstream water level, tailwater level, inflow, outflow, and turbine flow from the synchronous feature data as hydraulic layer node features, and extract the unit output and grid load as electrical layer node features; S2.4: Establish cross-layer coupling edges between the hydraulic layer nodes and electrical layer nodes corresponding to the same hydropower station to form a hydropower double-layer graph. The attributes of the cross-layer coupling edges include the net head of the hydropower station and the partial derivatives of the unit output with respect to the net head and the flow rate through the unit.

[0017] Further, in step S2.4, the calculation process of the properties of the cross-layer coupling edge includes: The net head is calculated based on the upstream water level, the tailwater level, and the hydraulic loss, wherein the hydraulic loss is calculated based on the flow rate through the generator and the parameters of the generator unit. Based on the turbine characteristic curve of the unit, the partial derivatives of the net head and flow rate are calculated under the operating conditions determined by the net head and flow rate, respectively, to obtain the partial derivatives of the unit output with respect to the net head and the partial derivatives of the unit output with respect to the flow rate.

[0018] Furthermore, step S3 specifically includes: S3.1: Determine the current operating condition based on hydrological data and power grid load data and encode it into an operating condition encoding vector. Then, fuse the operating condition encoding vector with the node features of the hydraulic directed graph and the electrical undirected graph to generate hydraulic layer scaling factor and electrical layer scaling factor. S3.2: Each scheduling agent, based on the hydraulic directed graph, performs time-series alignment of the node features of adjacent nodes with the water flow propagation time delay, calculates the hydraulic layer attention coefficient and scales it with the hydraulic layer scaling factor, and uses the scaled hydraulic layer attention coefficient to weighted aggregate the aligned adjacent node features to obtain their respective hydraulic layer aggregation messages. S3.3: Each scheduling agent calculates the electrical layer attention coefficient based on the electrical undirected graph and scales it using the electrical layer scaling factor. The scaled electrical layer attention coefficient is used to weight and aggregate the features of adjacent nodes to obtain their respective electrical layer aggregation messages. S3.4: After concatenating the hydraulic layer aggregation message and the electrical layer aggregation message with the hydraulic layer node features and electrical layer node features of the corresponding intelligent agents, perform nonlinear mapping to generate the collaborative perception embedding of each scheduling intelligent agent.

[0019] Furthermore, step S3.1 specifically includes: S3.1.1: Determine the current hydrological condition level by comparing the inflow and upstream water level of each hydropower station with the preset hydrological thresholds; determine the current load condition level by comparing the power grid load with the preset load thresholds. S3.1.2: After performing unique thermal coding on the hydrological condition level and the load condition level respectively, the two are concatenated to obtain the condition coding vector; S3.1.3: Take the mean value of all node features of the hydraulic directed graph and the electrical undirected graph respectively to obtain the corresponding feature mean value. Concatenate the working condition encoding vector with each feature mean value, project and normalize it to obtain the hydraulic layer scaling factor and the electrical layer scaling factor respectively.

[0020] Furthermore, step S3.2 specifically includes: S3.2.1: For each scheduling agent, obtain the water flow propagation time delay of the directed edges corresponding to the upstream adjacent nodes of the node where the agent is located in the hydraulic directed graph, and obtain the historical node features of each adjacent node by backtracking the corresponding time delay at the current time, and use them as the aligned adjacent node features. S3.2.2: The agent's own node features, aligned features of each neighboring node, and edge attributes of the corresponding directed edges are concatenated. The original attention scores of each neighboring node are calculated by nonlinear activation. The original attention scores of all neighboring nodes are normalized to obtain the hydraulic layer attention coefficient. S3.2.3: Scale the hydraulic layer attention coefficient with the hydraulic layer scaling factor, and use the scaled hydraulic layer attention coefficient to perform a weighted summation of the features of each adjacent node after alignment to obtain the hydraulic layer aggregated message.

[0021] Furthermore, step S3.4 specifically includes: The hydraulic layer aggregated message, electrical layer aggregated message, hydraulic layer node features, and electrical layer node features of each scheduling agent are concatenated along the feature dimension to obtain the fused feature vector; The fused feature vectors are sequentially subjected to linear transformation and nonlinear activation to obtain the collaborative perception embedding of each scheduling agent.

[0022] Furthermore, step S4 specifically includes: S4.1: Calculate the attention weights corresponding to each collaborative sensing embedding based on the collaborative sensing embeddings of all scheduling agents, and sum the attention weights of all collaborative sensing embeddings to obtain the watershed summary vector; S4.2: Calculate the power grid dispatch demand index based on the power grid load data and encode it as a power grid demand vector; S4.3: After concatenating the basin summary vector and the power grid demand vector, perform nonlinear mapping to generate a consensus signal vector, and broadcast the consensus signal vector to each scheduling agent.

[0023] Furthermore, step S5 specifically includes: S5.1: The policy network of each scheduling agent concatenates the corresponding cooperative perception embedding with the consensus signal vector as input, and infers and outputs scheduling actions, including the power allocation of each unit and the reservoir discharge flow. S5.2: Determine the physical feasible region at the current moment based on the current water level constraints, the upper and lower limits of unit output, and the minimum ecological flow. S5.3: Project the scheduling action onto the physical feasible domain to obtain a scheduling instruction.

[0024] Furthermore, the policy networks of each scheduling agent are obtained through centralized training using a distributed execution paradigm. The training process includes the following steps: A1: Construct a local value network corresponding to each scheduling agent. The local value network evaluates the local value based on the cooperative perception embedding, consensus signal vector and scheduling action of the corresponding agent. A2: Construct local reward functions for each scheduling agent based on power generation revenue, water resource utilization efficiency, water level violation penalties, and ecological flow violation penalties; A3: Construct a value hybrid network. Under the condition that the global value increases monotonically with respect to each local value, the local values ​​of all agents are synthesized into the global value through the value hybrid network. The value hybrid network punishes joint actions that violate coupling constraints through the Lagrange multipliers of water balance constraints and power balance constraints. A4: During the execution phase, each policy network only receives the cooperative perception embedding and consensus signal vector of the corresponding agent for independent reasoning; A5: Calculate the local reward at each time step based on the local reward function of each scheduling agent, update the parameters of the local value network and the value mixing network with the time difference error of the global value, and backpropagate the global value to each policy network to update the policy gradient until the time difference error of the global value is lower than the preset error threshold.

[0025] Beneficial effects of this invention: 1. This invention establishes a hydroelectric bilayer graph, which includes a hydraulic directed graph, an electrical undirected graph, and cross-layer coupling relationships, thereby achieving unified modeling of upstream and downstream water flow transmission relationships, power grid topology transmission relationships, and hydropower-to-electricity conversion relationships. Compared to the existing technology that models the hydraulic system and the power system separately, this invention can more comprehensively and accurately reflect the coupling mechanism between hydropower station groups and the power grid, thus improving the overall modeling capability and collaborative optimization capability under complex scheduling scenarios.

[0026] 2. This invention aligns hydrological data, generator operation data, and grid load data over time, and combines adaptive collaborative sensing, global consensus generation, and a physical feasible domain projection mechanism. This enables each scheduling agent to make independent decisions based on its local operating status while also responding to changes in the overall basin status and grid demand. Ultimately, it outputs scheduling commands that meet water level constraints, generator output constraints, and ecological flow constraints. Compared to existing technologies, this invention effectively improves the accuracy, robustness, consistency, and engineering feasibility of multi-station collaborative scheduling. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the structure of the cross-energy network AI collaborative scheduling optimization system for hydropower complementarity provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the cross-energy network AI collaborative scheduling optimization method for hydropower complementarity provided in an embodiment of the present invention. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0029] Example 1: Please see Figure 1 One embodiment of the present invention provides a cross-energy network AI collaborative scheduling optimization system for hydropower complementarity, applied to the collaborative scheduling of a group of cascade hydropower stations in the same river basin and the power grid. Each hydropower station in the group is equipped with a scheduling intelligent agent. The system includes: The data synchronization module is used to acquire the unit equipment parameters and power grid line parameters of each hydropower station, collect hydrological data, unit operation data and power grid load data of each hydropower station and perform time alignment to obtain synchronization characteristic data; The graph construction module is used to construct a hydropower bilayer graph based on the upstream and downstream river topology and power grid line topology of each hydropower station, and to construct the synchronous feature data as the node features of the hydropower bilayer graph. The collaborative perception module is used to control each scheduling agent to perform working condition adaptive graph attention aggregation in the hydraulic directed graph and the electrical undirected graph of the hydropower double-layer graph to obtain an aggregated message, and to fuse the aggregated message with the corresponding synchronous feature data to generate a collaborative perception embedding. The consensus generation module is used to perform attention pooling on all collaborative sensing embeddings to obtain a watershed summary vector, which is then fused with the power grid demand vector to generate a consensus signal vector, which is then broadcast to each scheduling agent. The scheduling execution module is used to control the policy network of each scheduling agent to receive the corresponding collaborative perception embedding and the consensus signal vector, infer and output scheduling actions and project them onto the physical feasible domain to obtain scheduling instructions.

[0030] The graph construction module includes a hydraulic graph construction unit, an electrical graph construction unit, a node feature extraction unit, and a cross-layer coupling unit, wherein: The hydraulic graph construction unit is used to construct a hydraulic directed graph with each hydropower station as a node and the upstream and downstream water flow transmission direction as directed edges. The attributes of each directed edge include the water flow propagation time delay from the upstream station to the downstream station and the interval attenuation coefficient. The electrical graph construction unit is used to construct an undirected electrical graph with the grid-connected nodes of each hydropower station and the substation nodes of the power grid as nodes and the transmission lines as edges. The attributes of each edge include the line reactance and the upper limit of transmission power. The node feature extraction unit is used to extract the upstream water level, tailwater level, inflow, outflow, and turbine flow as hydraulic layer node features from the synchronous feature data, and to extract the unit output and grid load as electrical layer node features. The cross-layer coupling unit is used to establish cross-layer coupling edges between the hydraulic layer nodes and electrical layer nodes corresponding to the same hydropower station, forming a hydropower two-layer diagram. The attributes of the cross-layer coupling edge include the net head of the hydropower station and the partial derivatives of the unit output with respect to the net head and the flow rate through the unit.

[0031] The collaborative sensing module includes a working condition coding unit, a hydraulic aggregation unit, an electrical aggregation unit, and a feature fusion unit, wherein: The operating condition coding unit is used to determine the current operating condition based on hydrological data and power grid load data and encode it into an operating condition coding vector. The operating condition coding vector is then fused with the node features of the hydraulic directed graph and the electrical undirected graph to generate hydraulic layer scaling factor and electrical layer scaling factor. The hydraulic aggregation unit is used by each scheduling agent to perform time-series alignment of the node features of adjacent nodes based on the hydraulic directed graph with the water flow propagation time delay, calculate the hydraulic layer attention coefficient and scale it with the hydraulic layer scaling factor, and use the scaled hydraulic layer attention coefficient to weighted aggregate the aligned adjacent node features to obtain their respective hydraulic layer aggregation messages. The electrical aggregation unit is used by each scheduling agent to calculate the electrical layer attention coefficient based on the electrical undirected graph and scale it with the electrical layer scaling factor, and then use the scaled electrical layer attention coefficient to weight and aggregate the features of adjacent nodes to obtain their respective electrical layer aggregation messages. The feature fusion unit is used to concatenate the hydraulic layer aggregation message and the electrical layer aggregation message with the hydraulic layer node features and electrical layer node features of the corresponding intelligent agent and then perform nonlinear mapping to generate the collaborative perception embedding of each scheduling intelligent agent.

[0032] The consensus generation module includes a pooling unit, a demand encoding unit, and a consensus broadcasting unit, wherein: The pooling unit is used to calculate the attention weights corresponding to each collaborative perception embedding based on the collaborative perception embeddings of all scheduling agents, and to sum the attention weights of all collaborative perception embeddings to obtain the watershed summary vector. The demand encoding unit is used to calculate the power grid dispatch demand index based on the power grid load data and encode it into a power grid demand vector. The consensus broadcasting unit is used to concatenate the watershed summary vector and the power grid demand vector and perform nonlinear mapping to generate a consensus signal vector, and then broadcast the consensus signal vector to each scheduling agent.

[0033] The scheduling execution module includes a strategy reasoning unit, a feasible region determination unit, and a projection unit, wherein: The policy reasoning unit is used by the policy network of each scheduling agent to concatenate the corresponding collaborative perception embedding with the consensus signal vector as input, and reason out the scheduling action, which includes the power allocation of each unit and the reservoir discharge flow. The feasible region determination unit is used to determine the physical feasible region at the current moment based on the current water level constraints, the upper and lower limits of unit output, and the minimum ecological flow. The projection unit is used to project the scheduling action onto the physical feasible domain to obtain the scheduling instruction.

[0034] Example 2: Please see Figure 2 The present invention provides an embodiment of an AI-based collaborative scheduling optimization method for hydropower complementarity across energy networks, applied to the aforementioned AI-based collaborative scheduling optimization system for hydropower complementarity across energy networks. The configuration strategy includes: S1: Obtain the unit equipment parameters and power grid line parameters of each hydropower station, collect hydrological data, unit operation data and power grid load data of each hydropower station and perform time alignment to obtain synchronization characteristic data.

[0035] Specifically, in the scenario of coordinated dispatch of a cascade hydropower station group and the power grid, hydrological processes, unit operating status, and power grid load changes exhibit significant multi-source heterogeneity and temporal scale differences. For example, hydrological data is typically updated at the minute or hour level, reflecting the slow changes in the inflow process; unit operating data may be collected at the second level or higher, reflecting the rapid dynamic characteristics of unit regulation; and power grid load data is affected by user-side electricity consumption behavior, exhibiting characteristics of both intraday fluctuations and random disturbances. If the above multi-source data is directly fused and modeled, it is easy to introduce temporal misalignment due to inconsistent time bases, resulting in distorted feature representation and affecting the accuracy of graph modeling and collaborative perception. Therefore, this embodiment first uniformly acquires and standardizes the multi-source data, and on this basis, constructs strictly time-aligned synchronous feature data to provide a consistent data input foundation for subsequent steps.

[0036] In this embodiment, the unit equipment parameters include, but are not limited to, the unit's rated capacity, minimum and maximum output constraints, turbine efficiency curves, start-up and shutdown time constants, and ramp rates, which are used to characterize the unit's physical operating boundaries and energy conversion characteristics. The power grid line parameters include line reactance, transmission capacity limits, and network topology connections, which reflect the transmission constraints of electrical energy in the power grid. These parameters, as static information, are uniformly organized during the initialization phase and linked with dynamic data to establish a correlation index, ensuring semantic consistency between different data sources.

[0037] Furthermore, the hydrological data includes indicators reflecting the state of hydropower resources, such as upstream water level, tailrace water level, and inflow. The unit operation data includes indicators reflecting the energy conversion process, such as unit output, flow through the unit, and start-up / shutdown status. The grid load data includes indicators reflecting the characteristics of electricity demand-side characteristics, such as load demand at each node and its changing trends. For the aforementioned dynamic data, this embodiment first identifies and processes outliers. For example, it removes abrupt changes caused by sensor malfunctions through threshold filtering or sliding window statistical methods, and fills in missing data using interpolation or historical averages to improve data integrity and reliability.

[0038] In the time alignment process, this embodiment uses a unified scheduling time step as the reference time axis, mapping data with different sampling frequencies onto this reference time axis. Specifically, for high-frequency data, downsampling is performed through time window aggregation methods (such as mean, maximum, or weighted average) to make it consistent with the scheduling time scale; for low-frequency data, upsampling is performed using linear interpolation or keeping the most recent value to complete the time series. Furthermore, to avoid the impact of physical processes such as water flow propagation time delay and grid response delay on data consistency, this embodiment introduces a delay compensation mechanism in the alignment process. Based on historical statistics or physical models, the water flow propagation time between different hydropower stations and the grid-side response time delay are estimated, and the relevant data are time-shifted and corrected to ensure that each feature has physical consistency at a unified time point.

[0039] Furthermore, after time alignment, various features are normalized or standardized to eliminate the influence between different units. Simultaneously, the processed multi-source features are organized according to hydropower station nodes to form structured synchronous feature data, where each time step corresponds to a joint feature vector containing hydrological status, unit status, and grid load information.

[0040] Through the above data acquisition and time alignment process, high-quality fusion of multi-source heterogeneous data under a unified time benchmark is achieved, enabling the constructed hydropower dual-layer map to accurately reflect the real operating status of each hydropower station at the same time and its coupling relationship with the power grid, thereby providing a reliable data foundation for collaborative perception and scheduling decisions.

[0041] S2: Construct a hydropower bilayer graph based on the upstream and downstream river topology and power grid line topology of each hydropower station, and construct the synchronous feature data as the node features of the hydropower bilayer graph.

[0042] Specifically, in the scenario of coordinated dispatch of a cascade hydropower station group and the power grid, the transmission process of water energy in the river and the transmission process of electrical energy in the power grid exhibit significantly different topological structures and mechanisms: water flow exhibits a directional, step-by-step transmission relationship along the river, while electrical energy in the power grid forms a mesh transmission structure through multi-node interconnection. Therefore, this embodiment constructs a hydro-electric dual-layer graph to express the hydraulic and electrical relationships in a unified graph structure, thereby achieving structured modeling of the water-electric coupling system.

[0043] In this embodiment, a hydraulic directed graph is first constructed based on the upstream and downstream river channel connections between each hydropower station. Specifically, each hydropower station is used as a graph node, and directed edges are established according to the actual flow direction in the river channel, so that upstream hydropower stations point to downstream hydropower stations, thereby depicting the water transfer path in the basin. Simultaneously, edge attributes reflecting the water flow propagation characteristics are introduced onto the directed edges, such as water flow propagation time delay and interval attenuation coefficient, to describe the time delay and energy loss that exist during water flow transmission in the river channel. This ensures that the hydraulic relationships are not only reflected in the topological connections but also in the dynamic process characteristics.

[0044] Furthermore, an undirected electrical graph is constructed based on the power transmission structure on the grid side. Specifically, the grid-connected nodes of each hydropower station and related substation nodes are used as graph nodes, and undirected edges are established according to the transmission line connection relationships to characterize the bidirectional flow characteristics of electrical energy in the power grid. Attributes such as line reactance and transmission capacity upper limits are set on the undirected edges to reflect the impedance characteristics and capacity constraints during power transmission, enabling the electrical network to truly represent the physical limitations of power grid operation.

[0045] After constructing the hydraulic and electrical diagrams, cross-layer relationships are further established between the hydraulic and electrical layer nodes corresponding to the same hydropower station, thus forming a unified hydropower two-layer diagram structure. Through this cross-layer relationship, the states of the same hydropower station in the hydraulic and electrical processes can be mapped to each other, realizing the explicit expression of the physical process of water energy conversion into electrical energy in the diagram structure, providing a foundation for collaborative modeling.

[0046] In terms of node feature construction, this embodiment maps synchronous feature data in layers according to hydraulic and electrical layers. Specifically, features reflecting hydrodynamic state, such as upstream water level, tailrace water level, inflow, outflow, and turbine flow, are mapped to hydraulic layer nodes to characterize the real-time status of each hydropower station during the hydrological process; features reflecting power operation status, such as unit output and grid load, are mapped to electrical layer nodes to characterize the supply and demand relationship of each node in the power grid.

[0047] Through the above-mentioned construction process of hydraulic directed graph, electrical undirected graph and cross-layer association, a hydroelectric double-layer graph structure that can simultaneously depict water flow transmission, power transmission and hydroelectric coupling relationship is formed, and synchronous feature data is embedded into the corresponding nodes, so that the graph structure can reflect the actual system operation status at both the topology level and the feature level.

[0048] S3: Each scheduling agent performs condition-adaptive graph attention aggregation in the hydraulic directed graph and electrical undirected graph of the hydropower double-layer graph to obtain an aggregated message, and then fuses the aggregated message with the corresponding synchronization feature data to generate a collaborative perception embedding.

[0049] Specifically, during the coordinated operation of a cascade hydropower station group and the power grid in a river basin, there are both strong temporal dependencies based on water flow transmission and spatial coupling relationships based on power grid interconnection among different hydropower stations. Furthermore, the importance and intensity of these relationships vary significantly under different operating conditions (such as high-water season, low-water season, high-load or low-load conditions). Therefore, this embodiment introduces an operating condition-adaptive graph attention aggregation mechanism, enabling nodes to dynamically select the most relevant neighborhood information under different operating environments, thus achieving a refined modeling of the hydropower coupling characteristics.

[0050] In this embodiment, the current operating state is first characterized based on hydrological and power grid load information from the synchronous feature data, and an operating condition coding vector is constructed. This operating condition coding vector is used to characterize the current hydrological-load combination state of the system, such as high water level and high load, low water level and low load, or other transitional states. Based on this, the operating condition coding vector is fused with node features to obtain scaling factors modulated for hydraulic and electrical relationships respectively, enabling attention calculation to dynamically adjust the importance weights of neighborhood information according to the current operating condition.

[0051] Furthermore, a time-constrained attention aggregation operation is performed on the hydraulic directed graph. Since there is a propagation time lag between upstream and downstream flows, the influence of upstream nodes on downstream nodes does not occur at the same time step. Therefore, before aggregating neighborhood information, the features of upstream nodes are first backtracked based on the propagation time lag of the corresponding directed edges to obtain historical features aligned with the current time. Based on this, the target node's own features, the aligned upstream node features, and the corresponding edge attributes are jointly modeled, and the influence weight of each upstream node on the current node is calculated using an attention mechanism. This weight reflects the contribution of different upstream water flows to the current hydropower station's operating state, thus achieving adaptive modeling of the hydraulic transmission process. Subsequently, the attention weights, scaled by operating conditions, are used to weighted sum the features of each upstream node to obtain the hydraulic layer aggregation message.

[0052] Simultaneously, attention aggregation based on topology association is performed in the undirected electrical graph. Since electrical energy can be transmitted along multiple paths in the power grid, and the electrical coupling strength between different nodes is influenced by both network structure and load distribution, a weighted model of adjacent node features is implemented during aggregation using an attention mechanism. Specifically, the target node features are interactively calculated with the features of each adjacent node to obtain the original attention score, which is then modulated based on line attributes to reflect the electrical connection strength. Normalization is then applied to obtain the electrical layer attention coefficients, and a scaling factor is introduced for dynamic adjustment. This ensures that information from key nodes receives higher weight under high load conditions, thereby enhancing the ability to perceive power grid bottlenecks or critical channels. Finally, the weighted summation yields the electrical layer aggregated message.

[0053] After obtaining the hydraulic and electrical aggregation messages, they are further fused with the original synchronization feature data of the corresponding nodes to construct a collaborative perception embedding that comprehensively represents the node state. Specifically, the hydraulic aggregation information, electrical aggregation information, and the node's own hydrological and electrical features are concatenated, and feature transformation is performed through nonlinear mapping. This ensures that the fused representation includes both information passed from the neighborhood and retains the node's own local state, thus forming a unified expression of the hydroelectric coupling operation state.

[0054] Through the aforementioned adaptive dual-graph attention aggregation mechanism, a collaborative perception process capable of dynamically adjusting information flow paths is essentially constructed. During the hydrological-dominated operation phase, the aggregation weight of the hydraulic layer is relatively enhanced, making the system more focused on the relationship between incoming water and water distribution; while during the grid load-dominated operation phase, the aggregation weight of the electrical layer is relatively increased, making the system more focused on the balance of power supply and demand and network constraints. This adaptive adjustment mechanism makes the information expression more accurate under different operating conditions, effectively improving the responsiveness and robustness of scheduling decisions to complex coupling relationships.

[0055] S4: Attention pooling is performed on all collaborative perception embeddings to obtain the watershed summary vector, which is then fused with the power grid demand vector to generate a consensus signal vector, which is then broadcast to each scheduling agent.

[0056] Specifically, during the coordinated scheduling of a cascade hydropower station group in a river basin, although the collaborative sensing embeddings of each node can fully reflect the local hydrological status, electrical operating status, and neighboring interaction information, the representation of a single node still has locality and is difficult to directly depict the global operating situation across the entire basin. Therefore, this embodiment globally aggregates the collaborative sensing embeddings of all nodes to extract summary information that can represent the overall state of the basin, and on this basis, introduces grid demand-side information to construct a unified scheduling consensus expression.

[0057] In this embodiment, attention-based pooling is first performed on all collaborative sensing embeddings. Unlike simple averaging or summing, attention pooling dynamically assigns weights based on the importance of each node in its current operating state, thereby highlighting the impact of key hydropower stations on the overall operation. Specifically, feature transformation is performed on each node embedding to calculate its corresponding attention score, and the attention scores of all nodes are normalized to obtain the weight coefficients for each node. Subsequently, all collaborative sensing embeddings are weighted and summed using the weight coefficients to obtain a basin summary vector, which comprehensively reflects the overall characteristics of water resource distribution, power supply and demand, and coupling relationships within the basin.

[0058] Furthermore, to ensure that dispatching decisions not only reflect the supply-side status but also respond to changes in grid demand, this embodiment constructs a grid demand vector based on grid load data. This demand vector can be encoded from information such as the current load level, load change trends, and load distribution at key nodes, and is used to characterize the grid's electricity demand structure and dispatching objective orientation at the current moment.

[0059] During the fusion process, the basin summary vector and the power grid demand vector are concatenated, and feature transformation is performed through nonlinear mapping to obtain the consensus signal vector. This vector simultaneously contains overall basin state information and power grid demand constraint information, and can serve as a unified global scheduling reference.

[0060] Through the aforementioned attention pooling and feature fusion processes, a mapping mechanism from local perception to global cognition is essentially constructed. When node contributions differ significantly, the attention mechanism automatically strengthens the information representation of key nodes, making the global summary more representative. When grid load changes drastically, the introduction of the demand vector can guide the global representation to adjust towards meeting load demand, thereby achieving a dynamic balance between the supply and demand sides.

[0061] S5: The policy network of each scheduling agent receives the corresponding cooperative perception embedding and the consensus signal vector, infers and outputs scheduling actions and projects them onto the physical feasible domain to obtain scheduling instructions.

[0062] Specifically, in the coordinated scheduling of a cascade hydropower station group and the power grid in a river basin, scheduling decisions need to simultaneously consider the local operating status of hydropower stations and the global supply and demand balance. Therefore, this embodiment combines collaborative sensing embedding with consensus signal vectors as input, enabling scheduling decisions to possess both a fine-grained perception capability of local states and adherence to global collaborative optimization objectives, thereby achieving consistent decision-making among multiple stakeholders.

[0063] In this embodiment, the collaborative sensing embedding corresponding to each node is first concatenated with the consensus signal vector to form a joint feature representation containing local state information and global constraint information. Based on this, a policy network performs a nonlinear mapping on this joint feature, outputting continuous or discrete scheduling actions. These scheduling actions include, but are not limited to, power allocation schemes for each generating unit and reservoir discharge flow regulation strategies, directly reflecting the optimal control decision under current operating conditions.

[0064] Furthermore, this embodiment introduces a physical feasible region constraint processing procedure after the action is generated, performing constraint mapping on the original scheduling action. Specifically, firstly, based on the current reservoir operating status and equipment constraints, the feasible region of the scheduling action is determined. This region is jointly defined by multiple constraints, including water balance constraints, unit output upper and lower limit constraints, and ecological constraints. Subsequently, the original action output by the strategy network is mapped to this feasible region. Through the above projection process, the safety and executability of the scheduling decision are ensured, while avoiding the strategy distortion problem caused by forced truncation, enabling a smooth transition of the output result between the continuous optimization space and the actual operating space.

[0065] By fusing collaborative sensing information with global consensus information and combining it with a physical feasible region projection mechanism, this embodiment constructs a closed-loop process from perception to decision-making and then to constraint correction. When local state changes are significant, the strategy can quickly respond to individual operational needs; when there is a global supply-demand imbalance, consensus signals can guide each node to adjust towards the overall optimal direction; and in the actual execution phase, feasible region constraints ensure that all scheduling actions meet the actual engineering requirements.

[0066] For example, in this embodiment, a method for coordinated scheduling of a three-stage cascade hydropower station group and a regional power grid in the same river basin is given as an example to illustrate the implementation details of the scheme in this application within a single scheduling cycle.

[0067] First, the current scheduling cycle is set to 15 minutes, and the basin includes upstream first-level hydropower station A, midstream second-level hydropower station B, and downstream third-level hydropower station C. After time alignment, the synchronization characteristics data for the current scheduling cycle are as follows: Power Station A has a dam inlet water level of 285.2m, a tailrace water level of 241.0m, an inflow of 520 cubic meters per second, an outflow of 500 cubic meters per second, a flow through the generator of 460 cubic meters per second, and a unit output of 180MW; Power Station B has a dam inlet water level of 233.8m, a tailrace water level of 198.4m, an inflow of 475 cubic meters per second, an outflow of 460 cubic meters per second, a flow through the generator of 420 cubic meters per second, and a unit output of 152MW; Power Station C has a dam inlet water level of 176.5m, a tailrace water level of 145.1m, an inflow of 448 cubic meters per second, an outflow of 440 cubic meters per second, a flow through the generator of 405 cubic meters per second, and a unit output of 138MW. Meanwhile, the current total load of the regional power grid is 560MW, and the loads of the three grid-connected nodes adjacent to A, B, and C are 210MW, 185MW, and 165MW, respectively. Further settings include: the water flow propagation time delay from A to B is 2 scheduling steps, with an interval attenuation coefficient of 0.97; the water flow propagation time delay from B to C is 1 scheduling step, with an interval attenuation coefficient of 0.98; and the upper limits of transmission power for the corresponding grid-connected lines of A, B, and C are 220MW, 180MW, and 160MW, respectively.

[0068] Subsequently, a hydraulic directed graph is constructed based on the aforementioned river topology, establishing two directed edges: A to B and B to C. An electrical undirected graph is constructed based on the grid topology, connecting A to substation node S1, B to substation node S2, and C to substation node S3. Undirected edges are also established between S1 and S2, and between S2 and S3, based on line interconnection relationships. On this basis, the upstream water level, tailrace water level, inflow, outflow, and turbine flow are mapped to hydraulic layer node characteristics, while the unit output and grid load are mapped to electrical layer node characteristics. Furthermore, cross-layer coupling edges are established between hydraulic layer nodes and electrical layer nodes corresponding to the same power station. Taking power station B as an example, based on the upstream water level of 233.8m, the tailrace water level of 198.4m, and the hydraulic loss of 2.1m calculated from the turbine flow, the net head can be expressed as: Net head = Upstream water level - Tailrace water level - Hydraulic loss, i.e., the net head is approximately 33.3m. By combining the turbine characteristic curve of the unit at the current operating point, the partial derivative of the unit output with respect to the net head can be further calculated to be approximately 3.8 MW / m, and the partial derivative of the unit output with respect to the flow rate is approximately 0.21 MW / (cubic meter per second). Thus, the cross-layer coupling edge property can be obtained to characterize the local sensitivity of the process of converting water energy into electrical energy.

[0069] Furthermore, the operating condition code is determined based on the current hydrological and power grid load conditions. For example, a high hydrological operating condition is defined as an inflow exceeding 500 cubic meters per second or a dam upstream water level exceeding a corresponding threshold, and a high load operating condition is defined as a total power grid load exceeding 530 MW. Therefore, within this scheduling cycle, the entire basin can be classified as a high hydrological-high load combined operating condition, and this combined operating condition is encoded into an operating condition coding vector. Then, the mean values ​​of all node characteristics in the hydraulic layer and the mean values ​​of all node characteristics in the electrical layer are calculated separately, concatenated with the operating condition coding vector, and then projected and normalized to obtain the hydraulic layer scaling factor and the electrical layer scaling factor. For example, the hydraulic layer scaling factor is 1.18, and the electrical layer scaling factor is 1.12, indicating that within this scheduling cycle, the impact of upstream hydraulic inflow and the impact on key electrical nodes are both enhanced.

[0070] Subsequently, dual-map collaborative sensing is performed on each power station. Taking power station B as an example, in the hydraulic layer, based on the propagation time delay from A to B being two scheduling steps, the node features of power station A in the first two scheduling steps are extracted from the historical sequence as aligned upstream features. Assuming the inflow rate of A at the corresponding time after backtracking is 508 cubic meters per second and the water level in front of the dam is 284.6 m, the aligned upstream features, the current hydraulic features of B itself, and the attributes of the A-to-B edge are input into the attention calculation process to obtain the original attention score. After normalization, if the hydraulic layer attention coefficient of B to A is 0.64, then after modulation by the hydraulic layer scaling factor of 1.18, the corresponding hydraulic influence weight is further enhanced, thus obtaining the hydraulic layer aggregated message of B. At the same time, in the electrical layer, B interacts with adjacent substation nodes and adjacent lines using features. Let the weight of the original attention score between node B and node S2 after normalization be 0.58, and the combined influence weights of the paths associated with S1 and S3 be 0.21 and 0.21 respectively. Then, after modulation with an electrical layer scaling factor of 1.12, the electrical layer aggregated message of B is obtained. Next, the hydraulic layer aggregated message, electrical layer aggregated message, and B's own hydraulic layer node features and electrical layer node features are concatenated, and then subjected to linear transformation and nonlinear activation to obtain the collaborative sensing embedding of B in this scheduling cycle. The same process is applied to A and C to obtain the collaborative sensing embeddings for all power stations.

[0071] Furthermore, attention pooling is applied to the collaborative sensing embedding of all power plants to generate a basin summary vector. For example, if the attention weights for A, B, and C are 0.37, 0.34, and 0.29 respectively, it indicates that in this scheduling cycle, A (with a larger upstream inflow) and B (which plays a mid-level regulating role) contribute significantly to the global state. Simultaneously, grid demand indicators are extracted from regional grid load data, such as the current total load of 560MW, the average load growth rate of the last two scheduling steps of 2.8%, and the relatively high local load at node S2, and these are encoded into a grid demand vector. Then, the basin summary vector and the grid demand vector are concatenated, and a consensus signal vector is obtained through nonlinear mapping. This consensus signal vector can be understood as an overall scheduling guidance expression for the current moment, including moderate upstream power generation, coordinated midstream regulation, and downstream consideration of both ecological and local power supply.

[0072] After obtaining the collaborative sensing embedding and consensus signal vectors of each power station, the scheduling action reasoning continues. For example, suppose the original scheduling actions obtained by A, B, and C in this scheduling cycle are as follows: A suggests a unit output of 192MW and a discharge flow of 515 cubic meters per second; B suggests a unit output of 167MW and a discharge flow of 455 cubic meters per second; and C suggests a unit output of 150MW and a discharge flow of 430 cubic meters per second. At this point, the physical feasible domain for the current scheduling cycle needs to be constructed in conjunction with physical constraints. Assume that the upper and lower limits of unit output at stations A, B, and C under the current water level conditions are 160MW to 200MW, 130MW to 160MW, and 120MW to 145MW, respectively; the minimum ecological flow rates at each station are 420 cubic meters per second, 390 cubic meters per second, and 400 cubic meters per second, respectively; and that, constrained by the upper limit of line transmission power, the maximum allowable power output of station B does not exceed 160MW, and the maximum allowable power output of station C does not exceed 145MW. This shows that both Bilibili and C-site's original output actions exceeded the limits to varying degrees.

[0073] Subsequently, the original scheduling actions were projected onto the physically feasible domain. For station A, the original actions of 192MW and 515 cubic meters per second were both within the allowable range, so they were directly retained. For station B, the original output of 167MW exceeded the current allowable limit of 160MW, so its projection was corrected to 160MW, while the discharge flow of 455 cubic meters per second remained unchanged. For station C, the original output of 150MW exceeded the allowable limit of 145MW, so it was corrected to 145MW, while the discharge flow of 430 cubic meters per second was checked to meet the minimum ecological flow constraint. The final scheduling instructions for this scheduling cycle were: station A output 192MW, discharge flow 515 cubic meters per second; station B output 160MW, discharge flow 455 cubic meters per second; station C output 145MW, discharge flow 430 cubic meters per second. At this time, the total output of the three stations is 497MW, and the remaining load gap can be made up by other conventional power sources or backup regulation resources in the region. Meanwhile, the increased power generation of Station A helps to utilize the higher water inflow from upstream, while Stations B and C maintain coordinated output under the premise of meeting line constraints, unit constraints and ecological constraints, thereby ensuring that the overall dispatch result not only responds to the grid demand, but also meets the operating boundaries of the cascade hydropower stations.

[0074] As can be seen from the above examples, within the same scheduling cycle, this method can complete the entire process of multi-source heterogeneous data alignment, hydropower dual-layer graph construction, operating condition adaptive collaborative perception, global consensus generation, and constrained scheduling action correction. This enables the scheduling results to simultaneously reflect the upstream and downstream water flow transmission relationship, the power demand relationship of the power grid, and the hydropower coupling characteristics, thus achieving unified optimization decision-making for the coordinated operation scenario of cascade hydropower station groups and the power grid.

[0075] The specific process of S2 is as follows: S2.1: Using each hydropower station as a node and the upstream and downstream water flow transmission direction as directed edges, construct a hydraulic directed graph. The attributes of each directed edge include the water flow propagation time delay and the interval attenuation coefficient from the upstream station to the downstream station. The water flow propagation time delay and the interval attenuation coefficient are calculated through historical hydrological data.

[0076] Specifically, water flow is transmitted step by step from upstream to downstream along the river channel, exhibiting obvious directionality and temporal dependence. Modeling using only undirected or static relationships is insufficient to accurately reflect the delay effects and energy changes during water transfer. Therefore, this embodiment constructs a hydraulically directed graph to structurally represent the water flow propagation path and its dynamic characteristics.

[0077] In this embodiment, each hydropower station is abstracted as a graph node. Based on the actual river topology, directed edges are established from upstream nodes to downstream nodes according to the direction of water flow, thus forming a network structure that conforms to the hydrological transmission law. Furthermore, key physical attributes are introduced into the directed edges to enhance the expressive power of the graph structure.

[0078] Among them, the water flow propagation time lag is used to describe the time required for water flow to be transmitted from the upstream hydropower station to the downstream hydropower station. This parameter can be obtained by performing time series correlation analysis on the historical inflow and the downstream response flow. The interval attenuation coefficient is used to characterize the water loss caused by factors such as evaporation, leakage and regulation during the propagation process. It can be estimated based on historical flow comparison or water balance relationship.

[0079] S2.2: Construct an electrical undirected graph with the grid connection nodes of each hydropower station and the substation nodes of the power grid as nodes and the transmission lines as edges. The attributes of each edge include the line reactance and the upper limit of transmission power.

[0080] Specifically, during power grid operation, electrical energy exhibits multi-path, bidirectional flow characteristics between different nodes, and the coupling relationships between different nodes are mainly determined by the electrical connection structure. Modeling based solely on simple topological relationships is insufficient to reflect the physical constraints and network characteristics during power transmission. Therefore, this embodiment constructs an electrical undirected graph to provide a unified expression of the power grid structure and its constraints.

[0081] In this embodiment, the grid-connected nodes of each hydropower station and related substation nodes are used as graph nodes. Undirected edges are established based on the actual transmission line connection relationships to depict the electrical connection state between nodes in the power grid. Since electrical energy can flow bidirectionally between connected nodes, an undirected edge form is adopted to conform to the physical characteristics of the power system.

[0082] Furthermore, key electrical properties are introduced on the undirected edges. Among them, line reactance describes the impedance characteristics of the transmission line to current, which can be calculated based on the line parameters and reflects the ease of power transmission between different nodes; the upper limit of transmission power represents the maximum power allowed to pass through the line under safe operating conditions, which can be determined by the line's rated capacity or operating procedures, and is used to characterize the transmission capacity constraints of the power grid.

[0083] S2.3: Extract the upstream water level, tailwater level, inflow, outflow, and turbine flow from the synchronous feature data as hydraulic layer node features, and extract the unit output and grid load as electrical layer node features.

[0084] In this embodiment, for the hydraulic process, the upstream water level, tailrace water level, inflow, outflow, and turbine flow are extracted as hydraulic layer node features. Among them, the upstream water level and tailrace water level reflect the head conditions and determine the hydropower conversion potential; the inflow and outflow reflect the water volume balance; and the turbine flow is directly related to the generator's power generation process. These features collectively characterize the hydrodynamic state and regulation capacity of the hydropower station at the current moment.

[0085] Furthermore, for the electrical process, generator output and grid load are extracted as electrical layer node features. Generator output reflects the power level supplied by the hydropower station to the grid, while grid load reflects the electricity consumption on the demand side; together, they determine the supply-demand balance of the power system. By mapping these features to electrical layer nodes, the operational role of each node in the grid and its impact on overall power dispatch can be effectively expressed.

[0086] S2.4: Establish cross-layer coupling edges between the hydraulic layer nodes and electrical layer nodes corresponding to the same hydropower station to form a hydropower double-layer graph. The attributes of the cross-layer coupling edges include the net head of the hydropower station and the partial derivatives of the unit output with respect to the net head and the flow rate through the unit.

[0087] The calculation process for the properties of the cross-layer coupling edge includes: The net head is calculated based on the upstream water level, the tailwater level, and the hydraulic loss, wherein the hydraulic loss is calculated based on the flow rate through the generator and the parameters of the generator unit. Based on the turbine characteristic curve of the unit, the partial derivatives of the net head and flow rate are calculated under the operating conditions determined by the net head and flow rate, respectively, to obtain the partial derivatives of the unit output with respect to the net head and the partial derivatives of the unit output with respect to the flow rate.

[0088] Specifically, the essence of a hydropower system lies in the conversion of water energy into electrical energy. This process depends on both hydrodynamic conditions and the operating status of the generating units, serving as a crucial link between the hydraulic and electrical layers. Without explicit modeling of this coupling relationship, effective transmission of hydraulic and electrical information becomes difficult, leading to information fragmentation between the two layers. Therefore, this embodiment establishes cross-layer coupling edges between the hydraulic and electrical nodes of the same hydropower station to provide a unified expression of the water-to-electricity conversion process.

[0089] In this embodiment, for each hydropower station, its hydraulic layer nodes and electrical layer nodes are connected one-to-one to form cross-layer edges, which are used to characterize the influence of hydrodynamic state on electrical output. Furthermore, key physical properties reflecting the hydro-electric conversion characteristics are introduced on the cross-layer edges to enhance the expressive power of the coupling relationship.

[0090] Among them, net head, as a core indicator describing the availability of hydropower, is calculated based on the upstream water level, the tailrace water level, and hydraulic losses. Specifically, the difference between the upstream water level and the tailrace water level reflects the theoretical head, while hydraulic losses consider the energy loss caused by friction, local losses, and other factors during the flow through the generator. This loss can be estimated based on the flow rate through the generator and the unit equipment parameters, thus obtaining the actual net head that can be used for power generation.

[0091] Furthermore, to characterize the sensitivity of unit output to changes in key variables, the partial derivatives of unit output with respect to net head and flow rate are introduced as coupling edge attributes. Specifically, based on the turbine characteristic curves, at the operating points corresponding to the current net head and flow rate, the partial derivatives of the output function with respect to net head and flow rate are calculated respectively, thereby obtaining the partial derivatives of unit output with respect to net head and flow rate.

[0092] By introducing the head of water and its related partial derivatives, the cross-layer coupling edge can not only represent the connection between hydraulics and electricity, but also characterize the strength and sensitivity of the energy conversion process, thereby providing refined support for cross-layer information transmission and collaborative modeling, and improving the ability to express the hydroelectric coupling mechanism.

[0093] The specific steps for S3 are as follows: S3.1: Determine the current operating condition based on hydrological data and power grid load data and encode it into an operating condition encoding vector. Then, fuse the operating condition encoding vector with the node features of the hydraulic directed graph and the electrical undirected graph to generate the hydraulic layer scaling factor and the electrical layer scaling factor.

[0094] Specifically, in the process of coordinated hydropower dispatch, different hydrological conditions and load levels correspond to different operational strategy focuses. For example, during the high-water season, the emphasis is on water consumption, while during the low-water season, the focus is on water resource constraints; under high load conditions, the focus is on power supply capacity, while under low load conditions, the focus is on economic operation. Therefore, this embodiment constructs a working condition coding vector to achieve an explicit representation of the current operating environment.

[0095] In this embodiment, based on hydrological data such as inflow and dam front water level, as well as power grid load data, the current operating status is classified into different levels, and the hydrological conditions and load conditions are combined and coded to obtain an operating condition coding vector. This vector is used to uniformly describe the water-power operating environment at the current moment.

[0096] Furthermore, the operating condition coding vector is fused with node features to modulate information representation at different levels. Specifically, the node features of the hydraulic directed graph and the electrical undirected graph are statistically summarized (e.g., the mean represents the overall state), and then concatenated with the operating condition coding vector before mapping and normalization to obtain the hydraulic layer scaling factor and the electrical layer scaling factor.

[0097] The scaling factor is used to dynamically adjust the attention weights during feature aggregation, adaptively enhancing key features or suppressing secondary information under different operating conditions. For example, when hydrological constraints are strong, the weight of hydraulically relevant features is increased; when the load pressure is high, the influence of electrically relevant features is strengthened.

[0098] S3.2: Each scheduling agent, based on the hydraulic directed graph, performs time-series alignment of the node features of adjacent nodes with the water flow propagation time delay, calculates the hydraulic layer attention coefficient and scales it with the hydraulic layer scaling factor, and uses the scaled hydraulic layer attention coefficient to weighted aggregate the aligned adjacent node features to obtain their respective hydraulic layer aggregation messages.

[0099] Specifically, in a hydraulic directed graph, there is a significant time-delay coupling relationship between upstream and downstream hydropower stations, and the impact of upstream water flow on downstream nodes is usually delayed. Directly aggregating node features at the same moment can easily lead to information misalignment and reduce the accuracy of hydrodynamic process modeling. Therefore, this embodiment first performs time-backtracking processing on the features of upstream adjacent nodes based on the water flow propagation time delay corresponding to the directed edges, aligning them with the current node's time, thereby obtaining physically consistent input features.

[0100] After temporal alignment is completed, the target node's own features, the aligned upstream node features, and the corresponding edge attributes are jointly represented. An attention mechanism is then used to calculate the attention coefficient of each upstream node to the current node. This attention coefficient reflects the differences in the contribution of different water sources to the current operating state, thereby enabling adaptive filtering of key upstream information.

[0101] Furthermore, a hydraulic layer scaling factor is introduced to adjust the attention coefficient, enabling it to dynamically change with the current operating conditions. When there are large fluctuations in inflow or strong hydrological constraints, the influence of key upstream nodes is enhanced; when hydrological changes are relatively stable, the interference caused by local fluctuations is weakened, thereby improving the stability of the aggregation results.

[0102] Finally, the alignment features of each upstream node are weighted and summed using the scaled attention coefficients to obtain the hydraulic layer aggregation message. This message integrates temporally aligned inflow information and neighborhood structure relationships, effectively characterizing the water flow transmission process within the watershed and providing accurate hydrodynamic feature input for collaborative sensing.

[0103] S3.3: Each scheduling agent calculates the electrical layer attention coefficient based on the electrical undirected graph and scales it using the electrical layer scaling factor. The scaled electrical layer attention coefficient is then used to weight and aggregate the features of adjacent nodes to obtain their respective electrical layer aggregated messages.

[0104] In this embodiment, the features of the target node and its neighboring nodes are first combined and represented, and then combined with corresponding edge attributes (such as line reactance) for feature interaction to calculate the original attention score of each neighboring node. This score reflects the influence strength of the neighboring nodes on the target node under the current power grid conditions; for example, nodes with low reactance or high load usually have higher weights. Subsequently, the attention scores of all neighboring nodes are normalized to obtain the electrical layer attention coefficient.

[0105] Furthermore, an electrical layer scaling factor is introduced to adjust the attention coefficient, enabling it to dynamically change according to the current load conditions. Under high load or tight network constraints, information transmission between key nodes and important channels is strengthened; under low load, the impact of local fluctuations is reduced, thereby improving the stability and robustness of the aggregation results.

[0106] Finally, the scaled electrical layer attention coefficients are used to perform a weighted summation of the features of neighboring nodes to obtain the electrical layer aggregated message. The electrical layer aggregated message integrates the supply and demand information of neighboring nodes and the characteristics of the network structure, effectively reflecting the power distribution and coupling relationship on the power grid side, and providing an accurate electrical feature representation for collaborative sensing.

[0107] S3.4: After concatenating the hydraulic layer aggregation message and the electrical layer aggregation message with the hydraulic layer node features and electrical layer node features of the corresponding intelligent agents, perform nonlinear mapping to generate the collaborative perception embedding of each scheduling intelligent agent.

[0108] Specifically, generating the collaborative perception embedding of each scheduling agent includes: concatenating the hydraulic layer aggregated message, electrical layer aggregated message, hydraulic layer node features, and electrical layer node features of each scheduling agent along the feature dimension to obtain a fused feature vector; and sequentially performing linear transformation and nonlinear activation on the fused feature vector to obtain the collaborative perception embedding of each scheduling agent.

[0109] In this embodiment, each node has obtained local structural information reflecting the water flow transmission relationship and the power grid coupling relationship. However, a hydropower system is essentially a cross-physical domain coupled system, and relying solely on information from either the hydraulic or electrical side is insufficient to fully characterize its operating state. For example, modeling based solely on hydraulic information may ignore the constraints of power grid load changes on unit output; while relying solely on electrical information cannot reflect the limitations of inflow conditions on power generation capacity. Therefore, this embodiment integrates multi-source information through feature fusion to construct a collaborative perception representation with cross-domain semantics.

[0110] In this embodiment, various features are first uniformly organized by concatenating the hydraulic layer aggregation message, the electrical layer aggregation message, and the original hydraulic and electrical features of the corresponding nodes in a preset order to form a fused feature vector. The hydraulic layer aggregation message mainly reflects the upstream water inflow and the impact of hydrodynamic propagation, the electrical layer aggregation message reflects the power grid topology and the supply and demand relationship in the neighborhood, while the original node features retain the current node's real-time operating status.

[0111] After obtaining the fused feature vector, a multi-level linear transformation is used to project the features onto the vector, mapping the high-dimensional spliced ​​features to a unified embedding space. Subsequently, a nonlinear expression capability is introduced through a nonlinear activation function (such as ReLU, Sigmoid, or other activation forms) to characterize the complex nonlinear features in the hydropower coupling relationship, such as the nonlinear impact of head changes on power output and the nonlinear response of grid load fluctuations to power distribution.

[0112] Through the aforementioned splicing and nonlinear mapping process, the resulting collaborative sensing embedding comprehensively reflects the operational status of nodes in the hydraulic domain, electrical domain, and cross-domain coupling relationships. The collaborative sensing embedding not only includes local node information but also integrates neighborhood propagation information and global structural influences, providing a unified and highly expressive feature foundation for consensus signal generation and scheduling decisions.

[0113] The specific steps of S3.1 are as follows: S3.1.1: The current hydrological condition level is determined by comparing the inflow rate and upstream water level of each hydropower station with the preset hydrological thresholds. The current load condition level is determined by comparing the power grid load with the preset load thresholds.

[0114] In this embodiment, the hydrological status is first assessed based on two key hydrological indicators: inflow rate and upstream water level. The inflow rate reflects the scale and trend of incoming water, while the upstream water level reflects the reservoir's storage status and regulation capacity. By comparing these indicators with preset hydrological thresholds, the current hydrological status can be divided into multiple levels, such as typical operating conditions like low water level and low inflow, low water level and high inflow, high water level and low inflow, and high water level and high inflow. The hydrological thresholds can be set based on historical statistical data or operational experience to distinguish operating ranges under different water resource conditions.

[0115] Furthermore, on the load side, the current power demand status is classified by comparing grid load data with preset load thresholds. For example, load levels can be divided into low load, medium load, and high load states to reflect the degree of grid supply and demand tension and dispatch pressure. The load thresholds can be determined based on historical load curves or grid operation specifications, giving load classification practical engineering significance.

[0116] After independently determining hydrological and load conditions, these two conditions can be combined to form a joint condition expression, thus comprehensively describing the current water-electricity coupling environment. For example, under high water level and high load conditions, priority should be given to releasing hydropower to meet electricity demand; while under low water level and low load conditions, the emphasis should be on water resource reserves and operational economics. Through this hierarchical and combined approach, differentiated treatment strategies can be adopted for different operating scenarios.

[0117] S3.1.2: The hydrological condition level and the load condition level are individually thermally encoded and then concatenated to obtain the condition encoding vector.

[0118] In this embodiment, the hydrological condition levels are first encoded using one-hot encoding. Specifically, for a preset hydrological level category (such as several discrete levels), a unique binary vector representation is assigned to each category, where only the corresponding position is 1 and the rest are 0, thus forming a vector representation of the hydrological condition.

[0119] Furthermore, the load condition levels are also processed using one-hot encoding to obtain a vector representation of the load condition. This vector reflects the current load level range of the power grid and can be clearly distinguished from other load conditions in the feature space.

[0120] After completing the one-heat encoding of the two types of operating conditions, the hydrological operating condition vector and the load operating condition vector are concatenated along the feature dimension to obtain a unified operating condition encoding vector. This vector contains information from both the hydrological and load sides, and can comprehensively describe the current hydro-electric coupling operating environment. Through the concatenation operation, operating condition information from different sources is expressed in the same feature space, which facilitates unified processing and fusion.

[0121] S3.1.3: Take the mean value of all node features of the hydraulic directed graph and the electrical undirected graph respectively to obtain the corresponding feature mean value. Concatenate the working condition encoding vector with each feature mean value, project and normalize it to obtain the hydraulic layer scaling factor and the electrical layer scaling factor respectively.

[0122] In this embodiment, the features of all nodes in the hydraulic directed graph and the electrical undirected graph are summarized and processed respectively. By averaging the features of each node in the same dimension, the average features of the hydraulic layer and the average features of the electrical layer are obtained. The feature averages are used to reflect the average level of the entire basin in terms of hydrological or electrical status at the current moment, such as the overall water level, average flow rate, or overall load level, thus forming a compact expression of the global state.

[0123] Furthermore, the operating condition encoding vector is concatenated with the mean features of the hydraulic layer and the mean features of the electrical layer, respectively, to form a joint feature representation that includes operating condition information and global state information. This fusion method enables subsequent calculations to not only perceive discrete operating condition categories but also incorporate the continuous changing characteristics of the current actual operating state, thereby improving the precision of the adjustment factors.

[0124] After obtaining the joint features, a linear projection is used to transform them, mapping the concatenated high-dimensional features to the target space to generate modulation parameters for the corresponding level. Subsequently, the projection results are normalized to ensure the output is within a stable numerical range, avoiding excessively large or small values ​​during attention calculation.

[0125] Finally, the hydraulic layer scaling factor and the electrical layer scaling factor are obtained respectively. These scaling factors are used to modulate the attention coefficients, enabling different layers to dynamically adjust the information weight distribution under different operating conditions and global states. For example, when the overall inflow is high, the hydraulic layer scaling factor enhances the focus on upstream information; when the overall load is high, the electrical layer scaling factor enhances the influence of key nodes.

[0126] The specific steps of S3.2 are as follows: S3.2.1: For each scheduling agent, obtain the water flow propagation time delay of the directed edges corresponding to the upstream adjacent nodes of the node where the agent is located in the hydraulic directed graph, and obtain the historical node features of each adjacent node by backtracking the corresponding time delay at the current time, and use them as the aligned adjacent node features.

[0127] Specifically, in the hydrodynamic coupling process of a cascade hydropower station, the impact of upstream water on downstream nodes exhibits a significant time delay, determined by river length, flow velocity, and storage conditions. Ignoring this propagation delay and directly using upstream node features at the same moment for modeling leads to information misalignment, causing downstream nodes to receive inputs inconsistent with the actual physical process, thus reducing the model's ability to characterize the flow transmission patterns. Therefore, this embodiment explicitly introduces the flow propagation delay to perform temporal alignment processing on neighborhood features.

[0128] In this embodiment, for each node, all upstream adjacent nodes pointing to that node in the hydraulic directed graph are traversed, and the water flow propagation time delay parameter recorded by the corresponding directed edge is obtained. This parameter reflects the time interval during which the water flow is transmitted from the upstream node to the current node, and is a key physical quantity in the hydraulic coupling relationship.

[0129] Furthermore, using the current scheduling time as a benchmark, a backtracking operation is performed on the time series features of upstream nodes based on the propagation delays corresponding to each directed edge. Specifically, for a given upstream node, its propagation delay is shifted from the current time step towards the historical direction, and the node features at the corresponding time point are extracted as the actual impact state of that node on the current time. If the backtracking time does not fall exactly on the discrete time step, linear interpolation can be used for approximate estimation to ensure the continuity and accuracy of the features.

[0130] After completing the above backtracking process, the upstream node features are aligned with the current node in time, meaning these features accurately reflect the inflow status affecting the node at the current moment. Compared to the unaligned original features, this processing method effectively eliminates the time misalignment problem, making the subsequent feature aggregation process more consistent with the physical laws of water flow propagation.

[0131] S3.2.2: The agent's own node features, aligned features of each adjacent node, and edge attributes of the corresponding directed edges are concatenated. The original attention scores of each adjacent node are calculated by nonlinear activation. The original attention scores of all adjacent nodes are normalized to obtain the hydraulic layer attention coefficient.

[0132] In this embodiment, the current node's own features, the aligned features of its upstream neighboring nodes, and the edge attributes of the corresponding directed edges are first concatenated to form a joint feature representation. The current node's own features provide a reference for its current operating state, the upstream node features reflect the incoming water information, and the edge attributes characterize the strength and dynamics of the hydraulic connections between nodes.

[0133] Furthermore, the concatenated joint features are input into a nonlinear mapping function (such as a multilayer perceptron) to calculate the original attention score for each upstream node. The original attention score measures the relative importance of each upstream node to the current node, and its magnitude reflects the intensity of the node's influence under the current hydrodynamic environment.

[0134] After obtaining the raw attention scores of all upstream nodes, they are normalized. Typically, a normalization function maps all scores to a weight distribution that sums to 1, thus obtaining the hydraulic layer attention coefficient. This normalization process ensures the comparability of the influence weights of different nodes and avoids the instability caused by differences in numerical scales in subsequent calculations.

[0135] S3.2.3: Scale the hydraulic layer attention coefficient with the hydraulic layer scaling factor, and use the scaled hydraulic layer attention coefficient to perform a weighted summation of the features of each adjacent node after alignment to obtain the hydraulic layer aggregated message.

[0136] In this embodiment, the hydraulic layer attention coefficient and the corresponding hydraulic layer scaling factor are scaled element-by-element or adjusted globally, so that the attention distribution can dynamically change according to the current hydrological state. This scaling process is essentially a redistribution of the influence weights of different upstream nodes, giving higher weights to nodes that are more representative under the current conditions, while relatively suppressing the weights of nodes with weaker influence or higher noise.

[0137] Furthermore, the scaled hydraulic layer attention coefficients are used to perform a weighted summation of the time-aligned upstream node features to obtain the hydraulic layer aggregated message. In this process, the feature contributions of each upstream node are superimposed according to their weights, thus forming a comprehensive input with physical meaning for the current node. The hydraulic layer aggregated message not only includes the intensity differences of incoming water information but also incorporates the effects of water flow propagation paths and time delays.

[0138] Through the aforementioned attention calculation-scaling modulation-weighted aggregation processing flow, the hydraulic layer aggregated message can adaptively adjust the information contribution structure under different hydrological conditions, retaining key upstream information while suppressing irrelevant interference, thereby more accurately depicting the dynamic transmission characteristics of water flow in the basin.

[0139] The specific steps for S4 are as follows: S4.1: Calculate the attention weights corresponding to each collaborative sensing embedding based on the collaborative sensing embeddings of all scheduling agents, and sum the attention weights of all collaborative sensing embeddings to obtain the watershed summary vector.

[0140] Specifically, in the scenario of coordinated scheduling of multiple hydropower stations, the coordinated perception embeddings corresponding to each scheduling agent characterize local information under different spatial locations, different hydropower station operating states, and power grid interaction relationships. Therefore, this embodiment introduces an attention weighting mechanism to adaptively evaluate the importance of each coordinated perception embedding, thereby achieving effective aggregation of global information.

[0141] In this embodiment, the original attention scores of each co-perceptual embedding are first calculated using a mapping function based on the collaborative perceptual embeddings of each scheduling agent. This mapping function can employ a combination of linear transformation and nonlinear activation to ensure comparability of different embeddings within a unified feature space. Subsequently, all original attention scores are normalized to obtain the attention weights corresponding to each collaborative perceptual embedding, ensuring that the sum of the weights equals 1. This guarantees the stability and interpretability of the overall information fusion process.

[0142] Furthermore, based on the aforementioned attention weights, all collaborative sensing embeddings are weighted and summed to obtain the basin summary vector. In this process, collaborative sensing embeddings with larger weights contribute more to the final result, thus highlighting the impact of key hydropower stations under the current dispatch conditions; while embeddings with smaller weights are appropriately weakened to reduce the interference of redundant or noisy information on the global representation. This basin summary vector comprehensively reflects the overall characteristics of the hydrological state, unit operating status, and grid load coupling relationship within the basin.

[0143] S4.2: Calculate the grid dispatch demand index based on the grid load data and encode it as a grid demand vector.

[0144] In this embodiment, the grid-side load data is first preprocessed, including time alignment, outlier removal, and normalization, to ensure consistency between data from different time scales and sources. Subsequently, several key indicators are extracted from the load sequence, including the current load level, short-term load change rate, historical periodic characteristics (such as intraday peak-to-valley difference), and load fluctuation intensity, to comprehensively reflect the grid operating status.

[0145] Furthermore, based on the extracted multidimensional load characteristics, these characteristics are integrated into a unified grid dispatch demand index through a preset mapping relationship or nonlinear transformation function. This index not only characterizes the grid's immediate demand for power generation but also implicitly reflects peak-shaving pressure and regulation flexibility requirements, thus providing more instructive input information for subsequent dispatch decisions.

[0146] After the demand indicators are calculated, they are vectorized and encoded. Specifically, scalar or multidimensional indicators can be mapped to a high-dimensional feature space through linear transformation combined with a nonlinear activation function, forming a power grid demand vector, which has good expressive power and learnability in the feature space.

[0147] S4.3: After concatenating the basin summary vector and the power grid demand vector, perform nonlinear mapping to generate a consensus signal vector, and broadcast the consensus signal vector to each scheduling agent.

[0148] In this embodiment, the basin summary vector and the power grid demand vector are first concatenated along the feature dimension to form a joint feature representation. The basin summary vector reflects the overall operating status and spatial coupling relationship of the hydropower station group within the basin, while the power grid demand vector reflects the current load level and dispatch pressure of the power system. Through this concatenation operation, supply and demand information are aligned and expressed in the same feature space, thereby establishing a unified semantic foundation for hydropower coordination.

[0149] Furthermore, the concatenated joint features are input into a nonlinear mapping function for feature transformation. This mapping process typically includes two stages: linear projection and nonlinear activation. Linear projection is used to perform weighted fusion and dimensionality compression of features from different sources, enabling the model to automatically learn the coupling relationship between supply and demand information. Nonlinear activation is used to enhance expressive power, enabling the model to characterize complex nonlinear interactions, such as scheduling constraints under combined high load and low water level conditions.

[0150] After completing the nonlinear mapping, a consensus signal vector is obtained. This vector can be regarded as a unified expression of the current global operating state, which simultaneously integrates the overall characteristics of the basin and the grid dispatching requirements, and can reflect the supply and demand matching relationship and potential dispatching directions.

[0151] Subsequently, the consensus signal vector is broadcast to each scheduling agent, enabling each agent to rely not only on its own and local neighborhood information when making subsequent decisions, but also to obtain globally consistent scheduling guidance information. In this way, each node maintains local autonomy while achieving a coordinated response to the global goal, thereby avoiding the overall performance degradation caused by local optima.

[0152] The specific steps for S5 are as follows: S5.1: The policy network of each scheduling agent concatenates the corresponding collaborative perception embedding with the consensus signal vector as input, and infers and outputs scheduling actions, including the power output allocation of each unit and the reservoir discharge flow.

[0153] In this embodiment, the collaborative perception embedding and consensus signal vectors corresponding to each scheduling agent are first concatenated along the feature dimension to form a joint input vector. The joint input vector simultaneously includes the node's own state, neighborhood interaction information, and watershed-level scheduling objectives, enabling the decision-making process to comprehensively consider multi-source information in a unified feature space.

[0154] Furthermore, the aforementioned joint input vector is fed into the policy mapping function for inference calculation. This mapping process can consist of multi-layer linear transformations and nonlinear activation functions, used to extract decision-related information from the high-dimensional feature space and output the corresponding scheduling actions. Through nonlinear mapping, the policy can characterize complex hydropower coupling relationships, such as the constraints of water level changes on power output allocation and the impact of grid load changes on unit response.

[0155] At the output layer, a scheduling action vector is generated, which mainly includes two aspects: first, the output allocation of each unit, used to determine the power generation of different units at the current moment; and second, the reservoir discharge flow, used to control the water release and subsequent hydrodynamic processes. These two variables together determine the process of converting hydropower into electrical energy and its impact on the river basin and power grid.

[0156] S5.2: Determine the physical feasible region at the current moment based on the current water level constraints, the upper and lower limits of unit output, and the minimum ecological flow.

[0157] In this embodiment, the water level constraint range is first determined based on the current reservoir operating status. Specifically, the allowable discharge flow range is limited according to the water level upstream of the dam and the design operating water level range (including normal storage level, dead water level, and flood control limit level, etc.) to ensure reservoir operation safety while taking into account flood control and water storage needs. If the current water level is close to the upper limit, the accumulation of inflow into the reservoir needs to be limited and the discharge appropriately increased; if it is close to the lower limit, the outflow needs to be reduced to maintain water reserves.

[0158] Furthermore, based on the operating characteristics of each unit, upper and lower limits for unit output are determined. Specifically, each unit has a corresponding minimum stable output and maximum rated output range under different head conditions, which are constrained by the unit's efficiency curve and operational safety limitations. During scheduling, the unit output must fall within this range, while also considering dynamic constraints such as unit start-up and shutdown states and ramp-up rates to avoid frequent fluctuations affecting the equipment. Simultaneously, a minimum ecological flow constraint is introduced to ensure the ecological needs of the downstream river channel. This constraint stipulates that the minimum discharge flow from the reservoir at any given time must not be lower than a preset threshold, thereby maintaining the basic ecological environment of the river channel and the sustainability of water resource utilization.

[0159] After considering the above constraints, the water level constraint, the upper and lower limits of unit output constraint, and the ecological flow constraint are expressed in a unified manner to construct the physical feasible region at the current moment. This feasible region can be represented as a multi-dimensional constraint space, where each dimension corresponds to control variables such as unit output and downstream flow, and its boundary is jointly determined by the above constraints.

[0160] S5.3: Project the scheduling action onto the physical feasible domain to obtain a scheduling instruction.

[0161] In this embodiment, the scheduling action output by the policy network is first represented as a multi-dimensional vector, where each dimension corresponds to control variables such as unit output and reservoir discharge flow. Then, this action vector is compared with the constructed physical feasible region to determine whether it falls within the feasible region. If the original action satisfies all constraints, it can be directly used as a scheduling instruction; if there is an out-of-bounds situation, constraint correction is required.

[0162] Furthermore, for actions exceeding the feasible region, a projection method is used to map them to the boundary of the feasible region or its nearest neighbor. Specifically, based on the principle of minimum distance, the out-of-bounds dimension can be truncated or scaled while keeping the action as close as possible to the original policy output. For example, when the unit output exceeds the maximum allowable value, it is limited to the upper limit boundary; when the outflow is lower than the ecological flow requirement, it is increased to the minimum threshold. Simultaneously, when multiple feasible solutions exist, the solution closest to the original action can be selected as the final output, thereby preserving the decision-making intent of the policy network to the greatest extent while ensuring that the result meets physical constraints. This approach is equivalent to establishing a balance between the optimal policy and executable constraints, ensuring that the model output is both optimization-oriented and conforms to engineering reality.

[0163] Through the above projection process, the original scheduling actions are transformed into scheduling instructions that meet physical constraints, so that the entire decision-making chain forms a closed-loop constraint from feature perception to action execution, effectively improving the security, stability and practical application value of the scheduling results.

[0164] Furthermore, the policy networks of each scheduling agent are obtained through centralized training using a distributed execution paradigm, the training including: A1: Construct a local value network corresponding to each scheduling agent. The local value network evaluates the local value based on the corresponding agent's collaborative perception embedding, consensus signal vector, and scheduling action.

[0165] In this embodiment, the collaborative perception embedding, consensus signal vector, and corresponding scheduling actions of each scheduling agent are fused together as input for local value assessment. The collaborative perception embedding reflects the local state and neighborhood information of the node in the hydraulic-electric coupling network, the consensus signal vector provides global scheduling guidance at the watershed level, and the scheduling actions reflect the specific decision-making behavior of the current node.

[0166] Furthermore, the aforementioned fused input is mapped through a process consisting of linear transformation and a nonlinear activation function to calculate the corresponding local value. This local value is used to measure the merits of the current scheduling action under a given state, and its value reflects the contribution of the action to the local operational effect and the overall goal. For example, an action that meets the grid demand while maintaining a reasonable water level change has a higher value; conversely, if an action leads to water waste or violates the scheduling trend, its value is relatively low.

[0167] A2: Construct local reward functions for each scheduling agent based on power generation revenue, water resource utilization efficiency, water level violation penalties, and ecological flow violation penalties.

[0168] In this embodiment, a power generation revenue index is first introduced to measure the economic benefits of the current dispatch action in terms of power output. This revenue can be calculated based on unit output and electricity price or equivalent revenue coefficient, and its value reflects the direct revenue generated by power generation per unit time.

[0169] Furthermore, water resource utilization efficiency indicators are introduced to measure the degree of utilization of hydropower converted into electrical energy. For example, assessments can be based on the power generation per unit volume of water or the overall efficiency, thereby encouraging higher energy conversion efficiency under limited water resources and avoiding ineffective water discharge or inefficient operation.

[0170] Meanwhile, to ensure the safe operation of the reservoir, a penalty for exceeding water level limits is introduced. When the actual water level exceeds the preset safe range (e.g., above the upper limit or below the lower limit), a penalty is applied to the reward function. This penalty can be weighted according to the severity of the exceedance, so that the more severe the exceedance, the greater the penalty, thereby guiding the scheduling strategy to proactively avoid unsafe operating conditions.

[0171] In addition, to ensure ecological and environmental needs are met, a penalty for violations of ecological flow regulations is introduced. When the outflow falls below the minimum ecological flow threshold, the current scheduling action is penalized. This penalty is typically set as a high-priority constraint to ensure that ecological needs are fully met.

[0172] After considering all the above indicators, power generation revenue and water resource utilization efficiency are used as positive incentives, while exceeding water level limits and violating ecological flow regulations are used as negative penalties. A local reward function is constructed through a weighted summation method. The weights of each item can be set according to the actual scheduling objectives to balance the relationship between economic efficiency, resource efficiency, and safety.

[0173] A3: Construct a value hybrid network. Under the condition that the global value increases monotonically with respect to each local value, the local values ​​of all agents are synthesized into the global value through the value hybrid network. The value hybrid network punishes joint actions that violate coupling constraints through Lagrange multipliers of water balance constraints and power balance constraints.

[0174] In this embodiment, the local values ​​of each scheduling agent are first processed as a unified input and then fused through a multi-layer mapping structure to generate a global value. During this process, a monotonicity constraint is applied, ensuring that the global value is a monotonically increasing function with respect to any local value. This ensures that an improvement in the local performance of a particular agent does not lead to a decrease in the overall value.

[0175] Furthermore, during the value fusion process, each local value is jointly mapped with global contextual information (such as basin status or consensus signals), ensuring that the global value not only depends on local performance but also reflects the overall system operating environment. Simultaneously, to ensure that the scheduling results meet the coupling constraints of the hydropower system, this embodiment introduces a penalty mechanism based on Lagrange multipliers during the global value construction process. Specifically, for violations of water balance or power balance constraints in joint scheduling actions, corresponding penalty terms are introduced and added to the global value. The water balance constraint ensures the conservation relationship between reservoir inflows, outflows, and water storage changes, while the power balance constraint ensures the matching relationship between total power generation and grid load demand.

[0176] In practical calculations, the weights of the Lagrange multipliers are dynamically adjusted based on the degree of constraint violation, so that the penalty is stronger when the degree of violation is greater, thereby guiding the strategy to gradually approach the feasible solution space that satisfies the constraints. Preferably, the Lagrange multipliers can be adaptively updated, so that they are continuously adjusted during training to balance the relationship between objective optimization and constraint satisfaction.

[0177] A4: During the execution phase, each policy network only receives the cooperative perception embedding and consensus signal vector of the corresponding agent for independent reasoning.

[0178] In this embodiment, each scheduling agent receives only two types of input information during execution: first, a corresponding collaborative perception embedding, which integrates its own state, hydraulic and electrical neighborhood information, and cross-layer coupling features, reflecting the node's operating state in the local environment; and second, a consensus signal vector, which originates from global information aggregation and reflects the overall operating state of the basin and the demand orientation of the power grid. By combining these two types of input, each node can gain indirect knowledge of the global environment without directly accessing the internal information of other nodes.

[0179] Furthermore, each policy network performs independent reasoning based on the above inputs and outputs the corresponding scheduling action. Since the collaborative perception embedding already contains neighborhood propagation information and the consensus signal vector provides global consistency constraints, even under independent decision-making conditions, the nodes can still maintain behavioral consistency and avoid conflicting or uncoordinated scheduling results.

[0180] A5: Calculate the local reward at each time step based on the local reward function of each scheduling agent, update the parameters of the local value network and the value mixing network with the time difference error of the global value, and backpropagate the global value to each policy network to update the policy gradient until the time difference error of the global value is lower than the preset error threshold.

[0181] In this embodiment, at each time step, the corresponding local reward value is first calculated based on the local reward function of each scheduling agent. This reward value comprehensively reflects the immediate effects of the current scheduling action in terms of economy, water resource utilization efficiency, and constraint satisfaction, and serves as the basis signal for subsequent value updates. Simultaneously, the values ​​corresponding to the current state and the state at the next time step are correlated to form a temporal value transfer relationship.

[0182] Furthermore, a temporal difference error is constructed based on the global value. Specifically, the global value at the current moment is compared with the global value at the next moment after the current reward is discounted, resulting in the temporal difference error. This error reflects the deviation between the current value assessment and the actual return, and is a key indicator for measuring the accuracy of the model's predictions.

[0183] After obtaining the temporal difference error, this error is used to update the parameters of the local value network and the value mixing network. Through the backpropagation mechanism, the local value assessment gradually approaches the true return, while optimizing the value mixing process so that the global value can more accurately reflect the overall effect of the synergistic interaction of each node. During the update process, the monotonicity constraint of the global value on the local value is maintained, thereby ensuring the stability of the learning process.

[0184] Furthermore, by using global value as a unified optimization objective, gradient information is backpropagated to each policy network through a value hybrid network to update the policy parameters. This process enables the policy networks to directly optimize towards increasing global value when adjusting their action outputs, thereby achieving collaborative learning among multiple agents rather than independent optimization.

[0185] As the training process continues, when the temporal difference error of the global value gradually decreases and falls below the preset error threshold, it indicates that the value assessment and the actual reward are basically consistent, and the policy network tends to converge. At this point, it can be considered that the obtained scheduling policy can achieve better overall performance in the current hydropower coupling environment.

[0186] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A cross-energy network AI collaborative scheduling optimization system for hydropower complementarity, applied to the collaborative scheduling of a cascade hydropower station group and the power grid within the same river basin, wherein each hydropower station in the hydropower station group is equipped with a scheduling intelligent agent, characterized in that the system... include: The data synchronization module is used to acquire the unit equipment parameters and power grid line parameters of each hydropower station, collect hydrological data, unit operation data and power grid load data of each hydropower station and perform time alignment to obtain synchronization characteristic data; The graph construction module is used to construct a hydropower bilayer graph based on the upstream and downstream river topology and power grid line topology of each hydropower station, and to construct the synchronous feature data as the node features of the hydropower bilayer graph. The collaborative perception module is used to control each scheduling agent to perform working condition adaptive graph attention aggregation in the hydraulic directed graph and the electrical undirected graph of the hydropower double-layer graph to obtain an aggregated message, and to fuse the aggregated message with the corresponding synchronous feature data to generate a collaborative perception embedding. The consensus generation module is used to perform attention pooling on all collaborative sensing embeddings to obtain a watershed summary vector, which is then fused with the power grid demand vector to generate a consensus signal vector, which is then broadcast to each scheduling agent. The scheduling execution module is used to control the policy network of each scheduling agent to receive the corresponding collaborative perception embedding and the consensus signal vector, infer and output scheduling actions and project them onto the physical feasible domain to obtain scheduling instructions.

2. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity as described in claim 1, characterized in that, The graph construction module includes a hydraulic graph construction unit, an electrical graph construction unit, a node feature extraction unit, and a cross-layer coupling unit, wherein: The hydraulic graph construction unit is used to construct a hydraulic directed graph with each hydropower station as a node and the upstream and downstream water flow transmission direction as directed edges. The attributes of each directed edge include the water flow propagation time delay from the upstream station to the downstream station and the interval attenuation coefficient. The electrical graph construction unit is used to construct an undirected electrical graph with the grid-connected nodes of each hydropower station and the substation nodes of the power grid as nodes and the transmission lines as edges. The attributes of each edge include the line reactance and the upper limit of transmission power. The node feature extraction unit is used to extract the upstream water level, tailwater level, inflow, outflow, and turbine flow as hydraulic layer node features from the synchronous feature data, and to extract the unit output and grid load as electrical layer node features. The cross-layer coupling unit is used to establish cross-layer coupling edges between the hydraulic layer nodes and electrical layer nodes corresponding to the same hydropower station, forming a hydropower two-layer diagram. The attributes of the cross-layer coupling edge include the net head of the hydropower station and the partial derivatives of the unit output with respect to the net head and the flow rate through the unit.

3. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity as described in claim 2, characterized in that, The cross-layer coupling unit is configured with coupling edge attribute calculation logic, which includes: The net head is calculated based on the upstream water level, the tailwater level, and the hydraulic loss, wherein the hydraulic loss is calculated based on the flow rate through the generator and the parameters of the generator unit. Based on the turbine characteristic curve of the unit, the partial derivatives of the net head and flow rate are calculated under the operating conditions determined by the net head and flow rate, respectively, to obtain the partial derivatives of the unit output with respect to the net head and the partial derivatives of the unit output with respect to the flow rate.

4. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 1, characterized in that, The collaborative sensing module includes a working condition coding unit, a hydraulic aggregation unit, an electrical aggregation unit, and a feature fusion unit, wherein: The operating condition coding unit is used to determine the current operating condition based on hydrological data and power grid load data and encode it into an operating condition coding vector. The operating condition coding vector is then fused with the node features of the hydraulic directed graph and the electrical undirected graph to generate hydraulic layer scaling factor and electrical layer scaling factor. The hydraulic aggregation unit is used by each scheduling agent to perform time-series alignment of the node features of adjacent nodes based on the hydraulic directed graph with the water flow propagation time delay, calculate the hydraulic layer attention coefficient and scale it with the hydraulic layer scaling factor, and use the scaled hydraulic layer attention coefficient to weighted aggregate the aligned adjacent node features to obtain their respective hydraulic layer aggregation messages. The electrical aggregation unit is used by each scheduling agent to calculate the electrical layer attention coefficient based on the electrical undirected graph and scale it with the electrical layer scaling factor, and then use the scaled electrical layer attention coefficient to weight and aggregate the features of adjacent nodes to obtain their respective electrical layer aggregation messages. The feature fusion unit is used to concatenate the hydraulic layer aggregation message and the electrical layer aggregation message with the hydraulic layer node features and electrical layer node features of the corresponding intelligent agent and then perform nonlinear mapping to generate the collaborative perception embedding of each scheduling intelligent agent.

5. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 4, characterized in that, The operating condition coding unit is configured with operating condition coding logic, which includes: The current hydrological condition level is determined by comparing the inflow and upstream water level of each hydropower station with preset hydrological thresholds. The current load condition level is determined by comparing the power grid load with preset load thresholds. The hydrological condition level and the load condition level are individually thermally encoded and then concatenated to obtain the condition encoding vector. The mean values ​​of all node features in the hydraulic directed graph and the electrical undirected graph are respectively taken to obtain the corresponding feature mean values. The working condition encoding vector is concatenated with each feature mean value, projected and normalized to obtain the hydraulic layer scaling factor and the electrical layer scaling factor, respectively.

6. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 4, characterized in that, The hydraulic aggregation unit is configured with hydraulic aggregation logic, which includes: For each scheduling agent, obtain the water flow propagation time delay of each upstream adjacent node in the hydraulic directed graph that points to the node where the agent is located, and backtrack the corresponding time delay at the current time to obtain the historical node features of each adjacent node, which are used as the aligned adjacent node features. The agent's own node features, aligned features of each neighboring node, and edge attributes of the corresponding directed edges are concatenated. The original attention scores of each neighboring node are calculated by nonlinear activation. The original attention scores of all neighboring nodes are normalized to obtain the hydraulic layer attention coefficient. The hydraulic layer attention coefficient is scaled using the hydraulic layer scaling factor, and the scaled hydraulic layer attention coefficient is used to perform a weighted summation of the features of each aligned adjacent node to obtain the hydraulic layer aggregated message.

7. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 4, characterized in that, The feature fusion unit is configured with feature fusion logic, which includes: The hydraulic layer aggregated message, electrical layer aggregated message, hydraulic layer node features, and electrical layer node features of each scheduling agent are concatenated along the feature dimension to obtain the fused feature vector; The fused feature vectors are sequentially subjected to linear transformation and nonlinear activation to obtain the collaborative perception embedding of each scheduling agent.

8. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 1, characterized in that, The consensus generation module includes a pooling unit, a demand encoding unit, and a consensus broadcasting unit, wherein: The pooling unit is used to calculate the attention weights corresponding to each collaborative perception embedding based on the collaborative perception embeddings of all scheduling agents, and to sum the attention weights of all collaborative perception embeddings to obtain the watershed summary vector. The demand encoding unit is used to calculate the power grid dispatch demand index based on the power grid load data and encode it into a power grid demand vector. The consensus broadcasting unit is used to concatenate the watershed summary vector and the power grid demand vector and perform nonlinear mapping to generate a consensus signal vector, and then broadcast the consensus signal vector to each scheduling agent.

9. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 1, characterized in that, The scheduling execution module includes a strategy reasoning unit, a feasible region determination unit, and a projection unit, wherein: The policy reasoning unit is used by the policy network of each scheduling agent to concatenate the corresponding collaborative perception embedding with the consensus signal vector as input, and reason out the scheduling action, which includes the power allocation of each unit and the reservoir discharge flow. The feasible region determination unit is used to determine the physical feasible region at the current moment based on the current water level constraints, the upper and lower limits of unit output, and the minimum ecological flow. The projection unit is used to project the scheduling action onto the physical feasible domain to obtain the scheduling instruction.

10. The cross-energy network AI collaborative scheduling and optimization system for hydropower complementarity according to claim 9, characterized in that, The policy networks of each scheduling agent in the policy inference unit are obtained through centralized training using a distributed execution paradigm. The system also includes a training module configured with training logic, which includes: Construct a local value network for each scheduling agent, and evaluate the local value based on the cooperative perception embedding, consensus signal vector and scheduling action of the corresponding agent. Local reward functions for each scheduling agent are constructed based on power generation revenue, water resource utilization efficiency, penalties for exceeding water level limits, and penalties for violations of ecological flow regulations. A value hybrid network is constructed. Under the condition that the global value increases monotonically with respect to each local value, the local values ​​of all agents are synthesized into the global value through the value hybrid network. The value hybrid network punishes joint actions that violate coupling constraints through Lagrange multipliers of water balance constraints and power balance constraints. During the execution phase, each policy network only receives the cooperative perception embedding and consensus signal vector of the corresponding agent for independent reasoning; The local reward at each time step is calculated based on the local reward function of each scheduling agent. The parameters of the local value network and the value mixing network are updated with the temporal difference error of the global value. The policy gradient is updated by backpropagating the global value to each policy network through the value mixing network until the temporal difference error of the global value is lower than the preset error threshold.

11. A method for AI-assisted scheduling optimization across energy networks for hydropower complementarity, applied to the AI-assisted scheduling optimization system for hydropower complementarity across energy networks as described in any one of claims 1 to 10, characterized in that, The method includes the following steps: S1: Obtain the unit equipment parameters and power grid line parameters of each hydropower station, collect hydrological data, unit operation data and power grid load data of each hydropower station and perform time alignment to obtain synchronization characteristic data; S2: Construct a hydropower bilayer graph based on the upstream and downstream river topology and power grid line topology of each hydropower station, and construct the synchronous feature data as the node features of the hydropower bilayer graph; S3: Each scheduling agent performs condition-adaptive graph attention aggregation in the hydraulic directed graph and electrical undirected graph of the hydropower double-layer graph to obtain an aggregated message, and then fuses the aggregated message with the corresponding synchronization feature data to generate a collaborative perception embedding. S4: Attention pooling is performed on all collaborative perception embeddings to obtain the watershed summary vector, which is then fused with the power grid demand vector to generate a consensus signal vector, which is then broadcast to each scheduling agent. S5: The policy network of each scheduling agent receives the corresponding cooperative perception embedding and the consensus signal vector, infers and outputs scheduling actions and projects them onto the physical feasible domain to obtain scheduling instructions.

12. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 11, characterized in that, Step S2 specifically includes: S2.1: Using each hydropower station as a node and the upstream and downstream water flow transmission direction as directed edges, construct a hydraulic directed graph. The attributes of each directed edge include the water flow propagation time delay and the interval attenuation coefficient from the upstream station to the downstream station. The water flow propagation time delay and the interval attenuation coefficient are calculated through historical hydrological data. S2.2: Construct an electrical undirected graph with the grid connection nodes of each hydropower station and the substation nodes of the power grid as nodes and the transmission lines as edges. The attributes of each edge include the line reactance and the upper limit of transmission power. S2.3: Extract the upstream water level, tailwater level, inflow, outflow, and turbine flow from the synchronous feature data as hydraulic layer node features, and extract the unit output and grid load as electrical layer node features; S2.4: Establish cross-layer coupling edges between the hydraulic layer nodes and electrical layer nodes corresponding to the same hydropower station to form a hydropower double-layer graph. The attributes of the cross-layer coupling edges include the net head of the hydropower station and the partial derivatives of the unit output with respect to the net head and the flow rate through the unit.

13. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 12, characterized in that, In step S2.4, the calculation process of the properties of the cross-layer coupling edge includes: The net head is calculated based on the upstream water level, the tailwater level, and the hydraulic loss, wherein the hydraulic loss is calculated based on the flow rate through the generator and the parameters of the generator unit. Based on the turbine characteristic curve of the unit, the partial derivatives of the net head and flow rate are calculated under the operating conditions determined by the net head and flow rate, respectively, to obtain the partial derivatives of the unit output with respect to the net head and the partial derivatives of the unit output with respect to the flow rate.

14. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 11, characterized in that, Step S3 specifically includes: S3.1: Determine the current operating condition based on hydrological data and power grid load data and encode it into an operating condition encoding vector. Then, fuse the operating condition encoding vector with the node features of the hydraulic directed graph and the electrical undirected graph to generate hydraulic layer scaling factor and electrical layer scaling factor. S3.2: Each scheduling agent, based on the hydraulic directed graph, performs time-series alignment of the node features of adjacent nodes with the water flow propagation time delay, calculates the hydraulic layer attention coefficient and scales it with the hydraulic layer scaling factor, and uses the scaled hydraulic layer attention coefficient to weighted aggregate the aligned adjacent node features to obtain their respective hydraulic layer aggregation messages. S3.3: Each scheduling agent calculates the electrical layer attention coefficient based on the electrical undirected graph and scales it using the electrical layer scaling factor. The scaled electrical layer attention coefficient is used to weight and aggregate the features of adjacent nodes to obtain their respective electrical layer aggregation messages. S3.4: After concatenating the hydraulic layer aggregation message and the electrical layer aggregation message with the hydraulic layer node features and electrical layer node features of the corresponding intelligent agents, perform nonlinear mapping to generate the collaborative perception embedding of each scheduling intelligent agent.

15. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 14, characterized in that, Step S3.1 specifically includes: S3.1.1: Determine the current hydrological condition level by comparing the inflow and upstream water level of each hydropower station with the preset hydrological thresholds; determine the current load condition level by comparing the power grid load with the preset load thresholds. S3.1.2: After performing unique thermal coding on the hydrological condition level and the load condition level respectively, the two are concatenated to obtain the condition coding vector; S3.1.3: Take the mean value of all node features of the hydraulic directed graph and the electrical undirected graph respectively to obtain the corresponding feature mean value. Concatenate the working condition encoding vector with each feature mean value, project and normalize it to obtain the hydraulic layer scaling factor and the electrical layer scaling factor respectively.

16. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 14, characterized in that, Step S3.2 specifically includes: S3.2.1: For each scheduling agent, obtain the water flow propagation time delay of the directed edges corresponding to the upstream adjacent nodes of the node where the agent is located in the hydraulic directed graph, and obtain the historical node features of each adjacent node by backtracking the corresponding time delay at the current time, and use them as the aligned adjacent node features. S3.2.2: The agent's own node features, aligned features of each neighboring node, and edge attributes of the corresponding directed edges are concatenated. The original attention scores of each neighboring node are calculated by nonlinear activation. The original attention scores of all neighboring nodes are normalized to obtain the hydraulic layer attention coefficient. S3.2.3: Scale the hydraulic layer attention coefficient with the hydraulic layer scaling factor, and use the scaled hydraulic layer attention coefficient to perform a weighted summation of the features of each adjacent node after alignment to obtain the hydraulic layer aggregated message.

17. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 14, characterized in that, Step S3.4 specifically includes: The hydraulic layer aggregated message, electrical layer aggregated message, hydraulic layer node features, and electrical layer node features of each scheduling agent are concatenated along the feature dimension to obtain the fused feature vector; The fused feature vectors are sequentially subjected to linear transformation and nonlinear activation to obtain the collaborative perception embedding of each scheduling agent.

18. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 11, characterized in that, Step S4 specifically includes: S4.1: Calculate the attention weights corresponding to each collaborative sensing embedding based on the collaborative sensing embeddings of all scheduling agents, and sum the attention weights of all collaborative sensing embeddings to obtain the watershed summary vector; S4.2: Calculate the power grid dispatch demand index based on the power grid load data and encode it as a power grid demand vector; S4.3: After concatenating the basin summary vector and the power grid demand vector, perform nonlinear mapping to generate a consensus signal vector, and broadcast the consensus signal vector to each scheduling agent.

19. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 11, characterized in that, Step S5 specifically includes: S5.1: The policy network of each scheduling agent concatenates the corresponding cooperative perception embedding with the consensus signal vector as input, and infers and outputs scheduling actions, including the power allocation of each unit and the reservoir discharge flow. S5.2: Determine the physical feasible region at the current moment based on the current water level constraints, the upper and lower limits of unit output, and the minimum ecological flow. S5.3: Project the scheduling action onto the physical feasible domain to obtain a scheduling instruction.

20. The cross-energy network AI collaborative scheduling optimization method for hydropower complementarity according to claim 19, characterized in that, The policy networks of each scheduling agent are obtained through centralized training using a distributed execution paradigm. The training process includes the following steps: A1: Construct a local value network corresponding to each scheduling agent. The local value network evaluates the local value based on the cooperative perception embedding, consensus signal vector and scheduling action of the corresponding agent. A2: Construct local reward functions for each scheduling agent based on power generation revenue, water resource utilization efficiency, water level violation penalties, and ecological flow violation penalties; A3: Construct a value hybrid network. Under the condition that the global value increases monotonically with respect to each local value, the local values ​​of all agents are synthesized into the global value through the value hybrid network. The value hybrid network punishes joint actions that violate coupling constraints through the Lagrange multipliers of water balance constraints and power balance constraints. A4: During the execution phase, each policy network only receives the cooperative perception embedding and consensus signal vector of the corresponding agent for independent reasoning; A5: Calculate the local reward at each time step based on the local reward function of each scheduling agent, update the parameters of the local value network and the value mixing network with the time difference error of the global value, and backpropagate the global value to each policy network to update the policy gradient until the time difference error of the global value is lower than the preset error threshold.

Images