A power-environment coupling big data collaborative optimization decision system

Abstract

This invention relates to the field of big data management technology, specifically to a collaborative optimization decision-making system for power-environment coupled big data. The system includes a watershed coupling module, which integrates real-time acquired multi-source ecological data with a preset mechanism model to construct a watershed water-energy-food-ecology coupled model; a hierarchical decision-making module, comprising a coordination layer agent representing top-level strategic goals and an execution layer agent representing different stakeholders; a decision attribution analysis module, which analyzes the collaborative scheduling strategy to obtain key driving factors leading to preset scheduling behavior; and a counterfactual deduction optimization module, which defines counterfactual scheduling scenarios based on key driving factors, calls the watershed water-energy-food-ecology coupled model for rapid deduction, quantifies and compares the comprehensive impact of the counterfactual scheduling scenarios and the collaborative scheduling strategy on water, energy, food, and ecology at a preset time scale, and optimizes the collaborative scheduling strategy.

Description

Technical Field

[0001] This invention relates to the field of big data management technology, specifically to a collaborative optimization decision-making system for power-environment coupled big data. Background Technology

[0002] River basin cascade hydropower stations are important sources of clean energy, and their operation is crucial for ensuring power supply. Traditional big data management systems are designed and operated primarily around single or limited resource management objectives, resulting in significant technical shortcomings: First, the resource management objectives are singular and fragmented. For example, the energy management system used by power dispatching departments aims to maximize power generation or optimize economic benefits. Water management departments, focusing on flood control and water supply security, primarily concern themselves with hydrological parameters such as reservoir capacity and water level, implementing relatively static monthly or quarterly water allocation rules based on historical experience. Meanwhile, environmental and agricultural departments, as downstream stakeholders in water resource allocation, primarily rely on monitoring and assessment, such as monitoring water quality, fish activity, or remotely sensing soil moisture. However, this data cannot be fed back in real-time to the upstream hydropower dispatching decision-making loop. This fragmentation in management leads to severe misallocation and suboptimal optimization of overall river basin resources. Furthermore, existing scheduling rules for ecological protection are typically set with a fixed minimum flow threshold. Such rigid rules cannot adapt to the real and dynamic needs of ecosystems. For example, rare fish species require specific water flow pulses during their breeding season, or concentrated water replenishment is needed during critical agricultural irrigation periods. Static rules appear inefficient or even harmful in such scenarios. Essentially, this system lacks a collaborative prediction and decision-making platform capable of coupling multi-domain models. This prevents managers from conducting integrated quantitative predictions and assessments of the cascading effects of a scheduling decision across multiple dimensions, including electricity, ecology, and agriculture. Consequently, cross-departmental collaborative management remains at the level of inefficient negotiation and qualitative judgment.

[0003] Therefore, how to address the serious problems of data silos, conflicting management objectives, and rigid scheduling rules in watershed hydropower scheduling technology, and how to couple the complex mega-system of water-energy-food-ecology (four dimensions or even more dimensions) to improve the overall utilization efficiency of water resources, energy resources, and ecological environment resources, so as to meet the strategic needs of coordinated security of "water-energy-food-ecology", has become a key bottleneck in the current development of this field.

[0004] To address this, a collaborative optimization decision-making system based on power-environment coupled big data is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a collaborative optimization decision-making system that couples power and environment big data. By coupling a complex mega-system of water, energy, food, and ecology in four or more dimensions, it improves the overall utilization efficiency of water resources, energy resources, and ecological environment resources, so as to meet the strategic needs of coordinated security for "water, energy, food, and ecology".

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A collaborative optimization decision-making system based on power-environment coupled big data includes:

[0008] The watershed coupling module integrates real-time acquired multi-source ecological data with a preset mechanism model to construct a watershed water-energy-food-ecology coupling model.

[0009] The hierarchical decision-making module includes a coordination layer agent representing the top-level strategic goals and an execution layer agent representing different stakeholders. During the training of the watershed water-energy-food-ecology coupling model, the coordination layer agent dynamically adjusts the reward function of the execution layer agent, guiding the execution layer agents to engage in collaborative game and generating a collaborative scheduling strategy that balances the interests of multiple parties at different times.

[0010] The decision attribution analysis module analyzes the collaborative scheduling strategy to obtain the key driving factors that lead to the preset scheduling behavior;

[0011] The counterfactual deduction and optimization module defines counterfactual scheduling scenarios based on the key driving factors, calls the watershed water-energy-food-ecology coupling model for rapid deduction, quantifies and compares the comprehensive impact of counterfactual scheduling scenarios and collaborative scheduling strategies on water, energy, food and ecology at a preset time scale, and optimizes the collaborative scheduling strategy.

[0012] Preferably, the multi-source ecological data includes: farmland soil moisture data and vegetation index data based on satellite remote sensing, watershed flow data and sediment content data based on hydrological station monitoring, rare aquatic organism activity data based on sonar monitoring, meteorological data, and power grid load data.

[0013] The preset mechanism models include: watershed hydrological model, crop growth model and ecological evolution model.

[0014] Preferably, the process of constructing a watershed water-energy-food-ecology coupled model is as follows:

[0015] The watershed topology is established based on graph neural networks, with reservoirs, irrigation districts, ecological protection zones and power grid nodes as graph nodes, and physical connections and ecological relationships as edges;

[0016] The multi-source ecological data is assimilated into the preset mechanism model in real time using an ensemble Kalman filter algorithm, and the state parameters of the graph nodes are dynamically updated.

[0017] By learning the causal relationships between nodes through dynamic Bayesian networks, the propagation effect of scheduling behavior on the state changes of each node is quantified, forming the watershed water-energy-food-ecology coupling model with causal inference capabilities.

[0018] Preferably, the execution layer intelligent agents include power generation efficiency intelligent agents, agricultural irrigation intelligent agents, ecological protection intelligent agents, and shipping support intelligent agents, each of which has an independent objective function and reward mechanism;

[0019] The process of generating the aforementioned cooperative scheduling strategy is as follows:

[0020] A dynamic internal bidding mechanism for water resource use rights is established. Each execution-layer agent uses the water-energy-food-ecology coupling model of the basin as its training environment, calculating utility estimates for key water volumes and release timing based on its objective function and reward mechanism within this environment. The coordination-layer agent dynamically adjusts and updates the reward function of each execution-layer agent based on the top-level strategic objective, the overall state of the basin, and the utility estimates of each agent.

[0021] Through multiple rounds of iterative training, each execution layer agent, guided by the coordination layer agent, gradually converges to the Pareto optimal cooperative scheduling strategy.

[0022] Preferably, the process of dynamically adjusting and updating the reward function of each execution layer agent is as follows:

[0023] A two-way reward adjustment mechanism integrating punishment and incentives that incorporates external constraints is established. The coordination layer agent decomposes the reward function of each execution layer agent into a basic reward item, a collaborative incentive item, and a violation penalty item. The collaborative incentive item is calculated based on the utility valuation of each execution layer agent. The violation penalty item is obtained by constructing a set of hard constraints based on the top-level strategic goals and the overall state of the watershed, and building a quantitative evaluation system for the severity of violating the set of hard constraints.

[0024] As the number of training rounds increases, the weight coefficient of the co-incentive term is gradually reduced using an exponential decay function.

[0025] Preferably, the process of obtaining the key driving factor is as follows:

[0026] The Shapley additive interpretation algorithm is used to analyze the cooperative scheduling strategy and calculate the marginal contribution of each input feature in the cooperative scheduling strategy to the final decision.

[0027] Identifying key time steps and state variables in the decision-making process based on attention mechanisms;

[0028] By combining domain expert knowledge, the marginal contribution values ​​are weighted and sorted to select the top N key driving factors that affect the preset scheduling behavior, and a natural language explanation report containing the decision-making logic is generated.

[0029] Preferably, the process of optimizing the cooperative scheduling strategy is as follows:

[0030] Based on user-inputted custom scheduling parameters and the key driving factors, a counterfactual scheduling scenario is constructed, and parallel computing technology is used to perform rapid simulation and deduction in the watershed water-energy-food-ecology coupling model; the custom scheduling parameters include reservoir outflow time sequence, generator start-up and shutdown plan and ecological water replenishment scheme;

[0031] Calculate the difference in comprehensive benefits between the counterfactual scheduling scenario and the collaborative scheduling strategy at a preset time scale; identify the weaknesses and improvement space of the current collaborative scheduling strategy based on the difference in comprehensive benefits, and generate strategy improvement suggestions including optimization direction and adjustment range; feed the strategy improvement suggestions back to the hierarchical decision module to update the agent's strategy parameters and reward function configuration, and optimize the collaborative scheduling strategy.

[0032] Preferably, the calculation process for the comprehensive benefit difference is as follows:

[0033] A comprehensive impact indicator system is constructed, including power generation benefits, grain output, water supply guarantee rate, irrigation satisfaction rate, ecological flow compliance rate, navigation guarantee rate, and biodiversity index. The objective weights of each indicator in the comprehensive impact indicator system are calculated using the entropy weight method. The subjective weights of the comprehensive impact indicator system are determined using the analytic hierarchy process. Based on the objective weights and the subjective weights, a comprehensive weight coefficient is formed, and the comprehensive benefit difference is calculated.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] 1. This invention integrates multi-source ecological data, including satellite remote sensing data on soil moisture, hydrological sediment data, and sonar data on fish activity, through a watershed coupling module. This data is then assimilated in real-time with a pre-set mechanistic model. Breaking away from the traditional power load-centric scheduling approach, this invention can proactively identify soil moisture risks or irrigation needs during critical periods such as spring planting and the dry season, taking into account downstream farmland soil moisture, shipping demand, and ecological flow limits. It automatically decides on "ecological water replenishment-based power generation." Even at the expense of maximizing short-term power generation benefits, it ensures agricultural irrigation and food security, enhances the watershed's water resource service capacity for food security strategies, and significantly reduces the risks associated with subjective human decision-making through digital twin simulations. This provides a solid scientific foundation for precise, proactive, and forward-looking watershed resource management, greatly improving the timeliness and accuracy of decision-making.

[0036] 2. The hierarchical decision-making module of this invention adopts a hierarchical collaborative multi-agent game framework. A coordinating agent dynamically guides multiple target-oriented agents, such as those for power generation, agriculture, ecology, and shipping, gradually converging towards a multi-target collaborative scheduling strategy. Combining the decision attribution analysis module's decision attribution analysis with the counterfactual inference optimization module's counterfactual inference mechanism, key variables affecting specific scheduling strategies are identified, such as "reservoir flow velocity fluctuations during fish migration" or "the blocking effect of specific sediment concentrations on downstream waterways." This mechanism not only significantly improves the interpretability and transparency of the system's scheduling strategy but also enables rapid strategy adjustments based on causal drivers, enhancing the system's robustness and adaptability.

[0037] 3. This invention constructs a watershed coupling model integrating graph neural networks and dynamic Bayesian networks. It maps the causal relationships between reservoirs, power grids, irrigation districts, and ecological nodes within the watershed using digital twins. Through shared data standards and interfaces across multiple departments, it serves as a joint decision-making platform for water resources, agriculture, energy, and environmental protection departments. A comprehensive impact assessment system constructed using the entropy weight method and the analytic hierarchy process (AHP) can quantify the impact of each scheduling scheme on grain yield, electricity revenue, biodiversity, and shipping during the strategy simulation stage. This capability significantly improves the efficiency and scientific rigor of cross-domain decision-making, providing a powerful support tool for implementing the coordinated security strategy of "water-energy-food-ecology." Attached Figure Description

[0038] Figure 1 A schematic diagram of the structure of a collaborative optimization decision-making system for power-environment coupled big data provided in an embodiment of the present invention;

[0039] Figure 2 A flowchart illustrating the collaborative optimization decision-making method for power-environment coupled big data provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the working principle of the hierarchical decision-making module provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This invention proposes a collaborative optimization decision-making system based on power-environment coupled big data. By coupling a complex mega-system encompassing four or more dimensions—water, energy, food, and ecology—it improves the overall utilization efficiency of water resources, energy resources, and ecological environment resources, thereby meeting the strategic needs of coordinated security for "water-energy-food-ecology." The effectiveness of this invention will be illustrated below with two embodiments.

[0043] Example 1:

[0044] In this embodiment, the system proposed in this invention is applied to the spring ecological scheduling scenario of a cascade hydropower station group of Dam A-Dam B in a certain river basin. This scenario involves a complex collaborative decision-making problem during the spring (March-May) period, which requires ensuring power supply while also taking into account the downstream agricultural irrigation needs, the ecological flow needs during the breeding season of rare fish, and the navigation needs of the river basin's golden waterway.

[0045] Figure 1 The specific structural diagram of the system of this invention includes: a watershed coupling module, which integrates real-time acquired multi-source ecological data with a preset mechanism model to construct a watershed water-energy-food-ecology coupling model; a hierarchical decision-making module, including a coordination layer agent representing the top-level strategic goal and an execution layer agent representing different stakeholders; the coordination layer agent dynamically adjusts the reward function of the execution layer agent during the training of the watershed water-energy-food-ecology coupling model, guiding the execution layer agents to engage in cooperative game theory and generating a collaborative scheduling strategy that balances the interests of multiple parties at different times; a decision attribution analysis module, which analyzes the collaborative scheduling strategy to obtain the key driving factors that lead to the preset scheduling behavior; and a counterfactual inference optimization module, which defines counterfactual scheduling scenarios based on the key driving factors, calls the watershed water-energy-food-ecology coupling model for rapid inference, quantifies and compares the comprehensive impact of the counterfactual scheduling scenarios and the collaborative scheduling strategy on water, energy, food, and ecology at a preset time scale, and optimizes the collaborative scheduling strategy. Figure 2 This is a flowchart illustrating a collaborative optimization decision-making method for power-environment coupled big data. The following is based on... Figure 1 and Figure 2 The following explanation is provided regarding the content:

[0046] The watershed coupling module integrates real-time acquired multi-source ecological data with a preset mechanism model to construct a watershed water-energy-food-ecology coupling model.

[0047] The multi-source ecological data includes: farmland soil moisture data and vegetation index data based on satellite remote sensing, watershed flow data and sediment content data based on hydrological station monitoring, rare aquatic organism activity data based on sonar monitoring, meteorological data, and power grid load data.

[0048] The preset mechanism models include: watershed hydrological model, crop growth model and ecological evolution model.

[0049] The process of constructing a watershed water-energy-food-ecology coupled model is as follows:

[0050] The watershed topology is established based on graph neural networks, with reservoirs, irrigation districts, ecological protection zones and power grid nodes as graph nodes, and physical connections and ecological relationships as edges;

[0051] The multi-source ecological data is assimilated into the preset mechanism model in real time using an ensemble Kalman filter algorithm, and the state parameters of the graph nodes are dynamically updated.

[0052] By learning the causal relationships between nodes through dynamic Bayesian networks, the propagation effect of scheduling behavior on the state changes of each node is quantified, forming the watershed water-energy-food-ecology coupling model with causal inference capabilities.

[0053] Specifically, the process of acquiring the multi-source ecological data is as follows:

[0054] Farmland soil moisture data is calculated using shortwave infrared data and is used to monitor soil moisture content;

[0055] Vegetation index data is calculated by the ratio difference between the near-infrared band and the red band, and is used to reflect the crop growth status in major agricultural areas such as downstream plains.

[0056] The basin flow data is collected in real time from multiple major hydrological monitoring stations in the basin;

[0057] The sediment content data was obtained through a combination of turbidity sensors and manual sampling, including two indicators: suspended sediment content and bedload transport rate.

[0058] Data on the activities of rare aquatic organisms were collected by deploying 12 sets of multibeam sonar monitoring equipment within a 200-kilometer radius downstream of Dam A to monitor the activity trajectories and habitat distribution of rare aquatic organisms in real time.

[0059] Meteorological data, including rainfall, temperature, wind speed, and humidity, are obtained from the National Meteorological Center and meteorological stations within the basin.

[0060] The power grid load data is obtained in real time from the power grid dispatch center in the area where the basin is located, including data on power load demand, electricity price information, and unit operating status.

[0061] The construction of the preset mechanism model includes:

[0062] The watershed hydrological model uses the SWAT distributed hydrological model as its basic framework. Input parameters include rainfall, evaporation, land use type, and soil characteristics, while output parameters include runoff, groundwater recharge, and soil moisture. The model divides the watershed into 126 sub-watershed units, each with an area of ​​approximately 800 square kilometers.

[0063] The crop growth model is based on the DSSAT crop growth simulation model to establish growth simulation models for major crops such as rice, wheat, and rapeseed. Input parameters include temperature, rainfall, solar radiation, soil moisture, and fertilizer application, while output parameters include crop yield, water requirement, and growth stage.

[0064] An ecological evolution model was used to construct a habitat suitability model for rare fish species. Input parameters included water temperature, flow velocity, dissolved oxygen, and turbidity. The output parameter was the Habitat Suitability Index (HSI), with HSI values ​​ranging from 0 to 1, where values ​​above 0.8 indicated suitability for reproduction.

[0065] The process of constructing a watershed water-energy-food-ecology coupled model includes:

[0066] The main reservoirs in the basin are designated as reservoir nodes, the main irrigation areas as irrigation area nodes, the basin area as ecological protection zone nodes, and the main substations of the power grid in the basin as power grid nodes. The physical connections between the nodes are represented by rivers, waterways, transmission lines, and irrigation canals, while the ecological relationships are represented by food chains, migration routes, and habitat dependencies.

[0067] The graph neural network adopts a graph convolutional neural network (GCN) architecture. The node feature vector has a 64-dimensional dimension, containing state variables such as water level, flow rate, power generation, soil moisture, and biodiversity index. The edge feature vector has a 32-dimensional dimension, containing parameters such as transmission delay, influence intensity, and distance attenuation coefficient.

[0068] The ensemble Kalman filter algorithm assimilates real-time monitored multi-source ecological data into a pre-defined mechanistic model. First, a state vector X is established, containing the state parameters of each graph node, with a vector dimension of 1024. Then, an observation vector Y is constructed, containing real-time monitoring data from all sensors, with a vector dimension of 512.

[0069] The prediction steps of ensemble Kalman filtering are as follows: Based on the current system state estimate, control inputs (gate opening, unit output, etc.) and system dynamics model, a nonlinear state transition method is used to predict the system state at the next moment, while considering the influence of random disturbances in the system process.

[0070] Update steps: Based on real-time observation data and state predictions, the prediction results are corrected using Bayesian estimation methods. By calculating the observation residuals (the difference between the actual observations and the predicted observations) and the Kalman gain, a more accurate state estimate is obtained.

[0071] Dynamic Bayesian Network (DBN) is employed to learn the causal relationships between nodes and quantify the propagation impact of scheduling behavior on the state changes of each node. The DBN network structure contains three time layers: time t-1, t, and t+1. Each time layer contains the state variables of all nodes, and the directed edges between nodes represent the strength and direction of the causal relationship.

[0072] The causal relationship learning employs a structured learning algorithm. First, an initial network structure is established based on expert knowledge, containing clear physical causal relationships, such as the impact of upstream reservoir outflow on downstream river flow. Then, the Expectation-Maximization (EM) algorithm is used to learn conditional probability parameters. The conditional probability distribution of each node is represented by a Gaussian mixture model with a mixture component count of 3. Finally, a greedy search algorithm is used to optimize the network structure, and the Bayesian Information Criterion (BIC) is used as the evaluation function to balance model complexity and goodness of fit. Ultimately, the system can quantify specific causal relationships such as "an increase of 100 cubic meters per second in the outflow from Reservoir C will increase the flow velocity in river segment E by 0.15 meters per second after 48 hours, thereby increasing the habitat suitability index for rare fish species by 0.08."

[0073] By integrating heterogeneous data from multiple sources, including satellite remote sensing, hydrological monitoring, sonar detection, meteorological observation, and power grid data, and combining hydrological models, crop growth models, and ecological evolution models, a watershed digital twin with high fidelity and causal inference capabilities was constructed using graph neural network modeling, ensemble Kalman filter data assimilation, and dynamic Bayesian network causal inference techniques. This enabled real-time replication of the state of complex watershed systems and prediction of future impacts, overcoming the technical shortcomings of static and isolated traditional simulation models. It provides an accurate and reliable simulation environment for intelligent decision-making, significantly improving the foresight and accuracy of scheduling decisions.

[0074] Furthermore, the hierarchical decision-making module includes a coordination layer agent representing the top-level strategic goals and an execution layer agent representing different stakeholders. During the training of the watershed water-energy-food-ecology coupling model, the coordination layer agent dynamically adjusts the reward function of the execution layer agents, guiding collaborative game theory among the execution layer agents to generate a coordinated scheduling strategy that balances the interests of multiple parties at different times. (Refer to...) Figure 3 ;

[0075] The execution layer agents include power generation efficiency agents, agricultural irrigation agents, ecological protection agents, and shipping support agents. Each execution agent has an independent objective function and reward mechanism.

[0076] The process of generating the aforementioned cooperative scheduling strategy is as follows:

[0077] A dynamic internal bidding mechanism for water resource use rights is established. Each execution-layer agent uses the water-energy-food-ecology coupling model of the basin as its training environment, calculating utility estimates for key water volumes and release timing based on its objective function and reward mechanism within this environment. The coordination-layer agent dynamically adjusts and updates the reward function of each execution-layer agent based on the top-level strategic objective, the overall state of the basin, and the utility estimates of each agent.

[0078] Through multiple rounds of iterative training, each execution layer agent, guided by the coordination layer agent, gradually converges to the Pareto optimal cooperative scheduling strategy.

[0079] The process of dynamically adjusting and updating the reward function of each execution layer agent is as follows:

[0080] A two-way reward adjustment mechanism integrating punishment and incentives that incorporates external constraints is established. The coordination layer agent decomposes the reward function of each execution layer agent into a basic reward item, a collaborative incentive item, and a violation penalty item. The collaborative incentive item is calculated based on the utility valuation of each execution layer agent. The violation penalty item is obtained by constructing a set of hard constraints based on the top-level strategic goals and the overall state of the watershed, and building a quantitative evaluation system for the severity of violating the set of hard constraints.

[0081] As the number of training rounds increases, the weight coefficient of the co-incentive term is gradually reduced using an exponential decay function.

[0082] Specifically, the objective function and reward mechanism for each executing agent are as follows:

[0083] Power Generation Benefit Agent: The objective function is to maximize power generation revenue. Input states include reservoir water level, inflow, electricity load demand, and electricity price, with a 32-dimensional state space. The action space represents the output adjustment of each generating unit, with a 12-dimensional action space corresponding to the generating units of Power Plant A. The reward function integrates power generation, electricity price revenue, and start-up and shutdown costs.

[0084] Agricultural Irrigation Agent: The objective function is to maximize the satisfaction of agricultural irrigation needs. Input states include soil moisture, crop water requirements, rainfall forecasts, and irrigation district water levels, with a 28-dimensional state space. The action space represents the water supply flow allocation for each irrigation district, with a 12-dimensional action space. The reward function integrates irrigation satisfaction rate, crop yield increase, and water shortage loss.

[0085] Ecological conservation agent: The objective function is to maximize ecosystem health. Input states include water temperature, flow rate, dissolved oxygen, fish activity density, etc., with a 24-dimensional state space. The action space is the temporal allocation of ecological flow, with an 8-dimensional action space. The reward function integrates habitat suitability index, biodiversity index, and ecological damage penalty.

[0086] The shipping support agent's objective function is to maximize shipping capacity. Input states include channel depth, current speed, vessel traffic volume, and freight demand, with a 20-dimensional state space. The action space consists of channel maintenance flow allocation, with a 6-dimensional action space. The reward function integrates navigation assurance rate, freight throughput, and channel maintenance costs.

[0087] The process of generating the aforementioned cooperative scheduling strategy is as follows:

[0088] This study simulates the competition for water resources among various stakeholders using an internal bidding mechanism for water use rights. The available water volume is auctioned off in time slots (6 hours per slot), with each execution layer agent submitting a bid for the water volume in that slot based on its own needs. Bids are calculated based on utility valuation, decision confidence, and urgency coefficients. The bidding mechanism employs the Vickrey auction rule, where the highest bidder wins the resource but only needs to pay the second-highest bid. This mechanism incentivizes agents to submit honest bids and prevents malicious price gouging.

[0089] Each agent in the execution layer uses the watershed water-energy-food-ecology coupling model as its training environment. Within this environment, it calculates utility estimates for key water volumes and release timing based on its objective function and reward mechanism. Utility estimates are calculated using the Deep Q-learning (DQN) algorithm. Each agent maintains a deep neural network Q(s, a; θ), with a four-layer fully connected structure and 128, 256, 256, and 128 hidden neurons, respectively. The network input is the current state s, and the output is the Q-value estimate for each action. The training process employs an experience replay mechanism, with a replay buffer size of 50,000 and a batch size of 64. The target network is updated every 1000 steps, with a learning rate of 0.001.

[0090] The coordination layer agent dynamically adjusts and updates the reward functions of each execution layer agent based on the top-level strategic objectives, the overall state of the watershed, and the utility estimates of each execution layer agent. The coordination layer agent employs a policy gradient algorithm with a variational autoencoder (VAE) network structure. The encoder encodes the global state into a 64-dimensional latent vector, and the decoder outputs the reward function adjustment parameters for each execution layer agent. In the spring ecological scheduling scenario, the priority of the top-level strategic objectives is set as follows: ecological protection > agricultural irrigation > shipping support > power generation benefits.

[0091] Through multiple rounds of iterative training, each execution layer agent, guided by the coordination layer agent, gradually converges to the Pareto-optimal cooperative scheduling strategy. The convergence criterion is: in 100 consecutive training rounds, the policy change amplitude of each agent is less than 0.01, and the variance of the comprehensive reward function is less than 0.001. The comprehensive reward function is the weighted sum of the rewards for each agent.

[0092] The process of dynamically adjusting and updating the reward function of each execution layer agent is as follows:

[0093] A two-way reward adjustment mechanism that integrates external constraints and penalties is established. The coordination layer agent decomposes the reward function of each execution layer agent into a basic reward item, a collaborative incentive item, and a violation penalty item.

[0094] Basic reward: Maintain the original goal orientation of each agent, with a fixed weight coefficient of 0.6.

[0095] Collaborative incentive term: Calculated based on the utility estimates of each executive layer agent. When an agent's utility estimate for a specific water resource is beneficial to the achievement of other agents' goals, the coordinating layer agent allocates a corresponding collaborative incentive reward to that agent based on the magnitude and scope of the utility estimate. The calculation process is as follows: the collaborative incentive term for agent i is equal to the weighted product of its utility estimate and the degree of utility improvement of other agents; the weights are determined based on the importance of each agent in the current period.

[0096] Penalty for breach of contract: Based on the top-level strategic objectives and the overall state of the watershed, a set of hard constraints is established, and a quantitative assessment system for the severity of violations of these hard constraints is constructed. In the spring ecological scheduling scenario, the set of hard constraints is determined according to the top-level strategic objectives as follows: ecological flow red line (not less than 3000 cubic meters per second), water supply security baseline (water supply guarantee rate of major irrigation areas not less than 95%), and power grid stability constraint (load response time not exceeding 15 minutes). The overall state of the watershed is obtained through comprehensive assessment of real-time monitored hydrological, ecological, and power data. The penalty value for breach of contract by the agent is equal to the sum of the products of the severity, duration, and damage coefficient of each breach. The severity of the breach ranges from 0 to 1 and is quantified according to the actual degree of violation of the hard constraints; the duration is in hours; and the damage coefficient is determined according to the degree of importance attached to the violation of different constraints in the top-level strategic objectives.

[0097] As the number of training rounds increases, the weight coefficient of the co-incentive term is gradually reduced using an exponential decay function.

[0098] By establishing a hierarchical collaborative multi-agent architecture and a two-way reward adjustment mechanism that integrates external constraints and penalties, intelligent negotiation and dynamic balance under multi-departmental interest conflicts are achieved. This solves the technical problem that traditional multi-objective optimization methods require preset weights and are difficult to adapt to dynamically changing management objectives. Through the internal bidding mechanism for water resource use rights and the dynamic guidance of the coordination layer agents, each execution layer agent can autonomously find the Pareto optimal collaborative scheduling strategy while ensuring key constraints. This significantly improves the intelligence level of multi-party interest coordination and the adaptability of decision-making strategies, realizing an upgrade from "hard allocation" to "intelligent negotiation" management model.

[0099] Furthermore, the decision attribution analysis module analyzes the collaborative scheduling strategy to obtain the key driving factors that lead to the preset scheduling behavior; the process of obtaining the key driving factors is as follows:

[0100] The Shapley additive interpretation algorithm is used to analyze the cooperative scheduling strategy and calculate the marginal contribution of each input feature in the cooperative scheduling strategy to the final decision.

[0101] Specifically, in the spring ecological scheduling scenario, the input features include: current reservoir water level, inflow, downstream river flow, soil moisture index, fish activity density, power load demand, rainfall forecast, and a total of 64 features.

[0102] The preset scheduling behaviors include ecological water replenishment power generation behavior, spring farming priority water transfer behavior, shipping support and flow enhancement behavior, power grid peak shaving behavior, and multi-objective trade-off scheduling behavior.

[0103] For different preset scheduling behaviors, the SHAP value calculation focuses on different feature subsets: for ecological water replenishment power generation, the focus is on analyzing ecologically related features such as water temperature, fish activity density, and breeding season; for spring plowing priority water diversion, the focus is on analyzing agriculturally related features such as soil moisture, crop water requirements, and rainfall forecasts; for shipping support and flow enhancement behaviors, the focus is on analyzing shipping-related features such as channel depth, sediment content, and ship traffic volume; for power grid peak shaving and peak reduction behaviors, the focus is on analyzing power-related features such as power load, electricity price changes, and power generation costs; and for multi-objective trade-off scheduling behaviors, the contribution of all 64 features is comprehensively analyzed.

[0104] The SHAP value is calculated using the TreeSHAP algorithm, which is suitable for decision tree-based models. During the calculation, SHAP values ​​are calculated for all features at each decision time step, resulting in a 64-dimensional contribution vector.

[0105] The attention mechanism of the Transformer architecture is used to identify key time steps and state variables in the decision-making process;

[0106] In the time dimension, a time attention mechanism is constructed, with the input being a continuous 72-hour state sequence and the output being the importance weight of each time step. Time steps with a weight value greater than 0.1 are marked as critical time steps.

[0107] In the state variable dimension, a feature attention mechanism is constructed, with a 64-dimensional state vector as input and importance weights for each feature as output. State variables with weights greater than 0.05 are marked as key state variables.

[0108] The marginal contribution values ​​are weighted and sorted by combining domain expert knowledge to select the top N key driving factors affecting the preset scheduling behavior, and a natural language explanation report containing decision-making logic is generated.

[0109] Specifically, the expert knowledge fusion employs the Analytic Hierarchy Process (AHP) to construct an expert judgment matrix. Based on pairwise comparisons of the importance of various factors by five watershed scheduling experts, a judgment matrix A is formed, where a_ij represents the importance ratio of factor i to factor j.

[0110] The expert weight vector is obtained by solving the characteristic equation of the judgment matrix.

[0111] The overall importance score is calculated based on the objective contribution calculated by the SHAP algorithm and the subjective importance determined by expert knowledge. A weighted fusion method is used to calculate the overall importance score of each factor.

[0112] Based on the overall score, the top 10 factors were selected as key driving factors. In the spring ecological scheduling scenario, the key driving factors for different preset scheduling behaviors typically include:

[0113] Ecological water replenishment-type power generation behavior: water temperature change rate during the breeding season of a rare fish species, fish activity density, deep water temperature of the reservoir, breeding season time window, etc.

[0114] Prioritize water diversion for spring plowing: consider factors such as soil moisture deficit in downstream farmland, crop water demand forecasts, rainfall forecast deviations, and irrigation area water levels.

[0115] The factors contributing to increased shipping capacity include: insufficient channel depth, severe siltation, number of vessels waiting to pass through locks, and urgency of cargo demand.

[0116] Peak shaving and peak regulation behavior of power grids includes: peak-to-valley difference in power load, fluctuation range of electricity price, changes in generation cost, and power grid stability indicators.

[0117] Multi-objective trade-off scheduling behavior: differences in utility valuations among departments, severity of constraint conflicts, overlap of time windows, degree of resource scarcity, etc.

[0118] The natural language interpretation report is generated using a template-based method, automatically generating interpretation text based on the type and importance of key driving factors. For example: "Interpretation of ecological water replenishment power generation behavior: The decision to increase the outflow and start the deep hole unit for ecological water replenishment is mainly driven by the following factors: (1) The water temperature rises too quickly during the breeding season of a certain rare fish. The current water temperature of 28.5℃ is close to the upper limit of 30℃ suitable for breeding (contribution 23%); (2) Sonar monitoring shows that the fish aggregation density has increased by 40% compared with the same period in previous years, and the breeding behavior is active (contribution 19%); (3) The deep water temperature is 5℃ lower than the surface water temperature. Water release through deep holes can effectively regulate the downstream water temperature (contribution 15%)."

[0119] By employing the Shapley additive interpretation algorithm, attention mechanism, and decision attribution analysis technology that integrates domain expert knowledge, the system achieves transparency and interpretability in complex AI decision-making processes. This addresses the key technical bottleneck of deep learning model "black box" decisions being difficult for managers to understand and trust. The system can accurately identify key driving factors affecting scheduling decisions and generate natural language explanation reports, providing a common understanding basis for cross-departmental collaboration. It significantly enhances the credibility and acceptability of AI-assisted decision-making and promotes the practical application and dissemination of intelligent decision-making technology in critical infrastructure management.

[0120] Furthermore, the counterfactual deduction and optimization module defines counterfactual scheduling scenarios based on the key driving factors, calls the watershed water-energy-food-ecology coupling model for rapid deduction, quantifies and compares the comprehensive impact of counterfactual scheduling scenarios and collaborative scheduling strategies on water, energy, food and ecology under a preset time scale, and optimizes the collaborative scheduling strategy.

[0121] The process of optimizing the aforementioned collaborative scheduling strategy is as follows:

[0122] Based on user-inputted custom scheduling parameters and the key driving factors, a counterfactual scheduling scenario is constructed, and parallel computing technology is used to perform rapid simulation and deduction in the watershed water-energy-food-ecology coupling model; the custom scheduling parameters include reservoir outflow time sequence, generator start-up and shutdown plan and ecological water replenishment scheme;

[0123] Among them, the reservoir outflow timing is adjusted by the user to monitor the hourly outflow of the Three Gorges Reservoir for the next 7 days; the generator start-up and shutdown plan is adjusted by the user to monitor the start-up and shutdown timing and output distribution of the 12 generator units; and the ecological water replenishment plan is set by the user to set specific ecological water replenishment periods and flow rates.

[0124] The counterfactual scenarios are constructed using the Monte Carlo method. Based on user-specified scheduling parameters, and considering the uncertainty of these parameters, 1000 scenario samples are generated. Each sample adds a normally distributed random perturbation to the user parameters, with the standard deviation set to 5% of the parameter value.

[0125] A comprehensive impact indicator system was constructed, including power generation efficiency, grain output, water supply guarantee rate, irrigation satisfaction rate, ecological flow compliance rate, navigation guarantee rate, and biodiversity index. The objective weights of each indicator in the comprehensive impact indicator system were calculated using the entropy weight method. The subjective weights of the comprehensive impact indicator system were determined using the analytic hierarchy process.

[0126] Among them, power generation benefit is the total power generation multiplied by the average grid-connected electricity price; grain output is the total regional grain output calculated based on the crop growth model, taking into account the three main crops of rice, wheat and corn, and the output is affected by factors such as irrigation water volume, soil moisture and temperature; water supply guarantee rate is the degree to which urban water supply demand is met; irrigation satisfaction rate is the degree to which agricultural irrigation demand is met; ecological flow compliance rate is the proportion of time during which ecological flow is not lower than the minimum threshold, which is set at 3000 cubic meters per second; navigation guarantee rate is the proportion of time during which the channel depth meets the navigation requirements, which is 3.5 meters; biodiversity index is calculated based on the Shannon diversity index.

[0127] The objective weight calculation process is as follows: standardize the data of each indicator to eliminate the influence of dimensions; calculate the objective weight of each indicator based on the entropy weight method.

[0128] The calculation process of subjective weights is as follows: First, a hierarchical structure is constructed, with the target layer being "comprehensive evaluation of scheduling schemes" and the criterion layer consisting of 7 evaluation indicators; then, watershed management experts compare the importance of the indicators pairwise, and a judgment matrix is ​​constructed using the 1-9 scaling method; finally, the eigenvector corresponding to the largest eigenvalue of the judgment matrix is ​​solved, and the weight vector is calculated.

[0129] In the spring ecological scheduling scenario, the subjective weights obtained by the analytic hierarchy process are: ecological flow compliance rate (0.25), irrigation satisfaction rate (0.20), biodiversity index (0.18), power generation benefit (0.15), shipping guarantee rate (0.12), water supply guarantee rate (0.06), and grain yield (0.04).

[0130] Based on the objective weights and the subjective weights, a comprehensive weight coefficient is formed, and the difference in comprehensive benefits between the counterfactual scheduling scenario and the collaborative scheduling strategy under a preset time scale is calculated.

[0131] Based on the comprehensive benefit difference, the weaknesses and improvement space of the current collaborative scheduling strategy are identified, and strategy improvement suggestions containing optimization directions and adjustment ranges are generated. The strategy improvement suggestions are fed back to the hierarchical decision-making module to update the agent's strategy parameters and reward function configuration, and optimize the collaborative scheduling strategy.

[0132] Among these, weak links are identified through comparative analysis of indicator dimensions with negative benefit differences. For example, if counterfactual scenarios perform better in terms of ecological flow compliance rates, it indicates that the current strategy is inadequate in ecological protection.

[0133] The improvement recommendations are generated based on the magnitude and distribution of the benefit difference, and specific adjustment recommendations are generated, including: parameter adjustment recommendations, time series optimization recommendations, and coordination mechanism adjustment recommendations.

[0134] The improvement suggestions are converted into agent parameter update instructions, including:

[0135] Policy network parameter update: Based on the policy gradient algorithm, the policy network parameters of each agent are adjusted according to the improvement direction. The learning rate is set to 0.0001, and the update steps are 100 steps.

[0136] Reward function configuration update: Adjust the reward function weights for each agent based on improvement suggestions. The adjustment range is limited to within 10% to avoid system oscillation.

[0137] Constraint Updates: The thresholds of hard constraints are dynamically adjusted based on actual operational conditions. For example, based on the latest ecological monitoring data, the lower limit of ecological flow is adjusted from 3000 cubic meters per second to 3200 cubic meters per second.

[0138] A new round of counterfactual analysis will be conducted to verify the updated strategy and evaluate its effectiveness. If the overall benefit improvement exceeds 5%, the strategy update will be accepted; otherwise, the original strategy will be rolled back, and the improvement will be adjusted before trying again.

[0139] By constructing counterfactual scheduling scenarios, establishing a multi-dimensional comprehensive impact indicator system, integrating objective and subjective weights in the evaluation method, and implementing a closed-loop feedback mechanism for strategy improvement suggestions, the system achieves a shift from a "passive response" to an "active learning" decision-making model. This overcomes the technical limitations of traditional decision-making systems that lack self-improvement capabilities. The system can continuously optimize scheduling strategies based on simulation results, automatically generate specific strategy adjustment suggestions through weak link identification and improvement space analysis, and feed them back to the intelligent agent module, forming a self-evolving intelligent decision-making closed loop. This significantly improves the system's learning ability and long-term decision quality.

[0140] This embodiment constructs a collaborative optimization decision-making system comprising a watershed coupling module, a multi-agent decision-making module, a decision attribution analysis module, and a counterfactual inference optimization module. This achieves a fundamental shift from the traditional "departmental segmentation, experience-based decision-making, and ex-post adjustment" to "system coupling, intelligent decision-making, and proactive optimization." It breaks down data silos and decision-making barriers between departments such as water resources, power, agriculture, and ecology, establishing a technical framework for integrated collaborative scheduling of "water-energy-food-ecology." This provides systematic technical support for the overall optimal allocation of watershed resources and strategic security, significantly improving the scientific nature and execution efficiency of complex scheduling problems.

[0141] Example 2:

[0142] In Example 1, the system proposed in this invention successfully couples a complex mega-system of four or more dimensions—water, energy, food, and ecology—improving the overall utilization efficiency of water resources, energy resources, and ecological environment resources to meet the strategic needs of coordinated security for "water-energy-food-ecology." To further verify the effectiveness of this invention, this application also includes a coordinated optimization decision-making process for autumn ecological scheduling in another watershed.

[0143] The watershed coupling module integrates real-time acquired multi-source ecological data with a preset mechanism model to construct a watershed water-energy-food-ecology coupling model.

[0144] The multi-source ecological data includes: farmland soil moisture data and vegetation index data based on satellite remote sensing, watershed flow data and sediment content data based on hydrological station monitoring, rare aquatic organism activity data based on sonar monitoring, meteorological data, and power grid load data.

[0145] The preset mechanism models include: watershed hydrological model, crop growth model and ecological evolution model.

[0146] The process of constructing a watershed water-energy-food-ecology coupled model is as follows:

[0147] The watershed topology is established based on graph neural networks, with reservoirs, irrigation districts, ecological protection zones and power grid nodes as graph nodes, and physical connections and ecological relationships as edges;

[0148] The multi-source ecological data is assimilated into the preset mechanism model in real time using an ensemble Kalman filter algorithm, and the state parameters of the graph nodes are dynamically updated.

[0149] By learning the causal relationships between nodes through dynamic Bayesian networks, the propagation effect of scheduling behavior on the state changes of each node is quantified, forming the watershed water-energy-food-ecology coupling model with causal inference capabilities.

[0150] The graph neural network adopts a spatiotemporal adaptive graph convolutional architecture, including:

[0151] A multi-level adjacency matrix learning mechanism is established to dynamically adjust the connection weights and influence ranges between graph nodes based on seasonal changes in watershed hydrological conditions and extreme weather events. The multi-level adjacency matrix includes a physical adjacency matrix, an ecological association matrix, and a functional dependency matrix.

[0152] A spatiotemporally decoupled graph convolutional layer is constructed to model the temporal evolution characteristics and spatial propagation characteristics of nodes respectively. The temporal attention mechanism is used to capture the state change patterns at different time scales, and the spatial attention mechanism is used to identify key propagation paths and influencing nodes.

[0153] An adaptive graph structure update module is set up. Based on the historical change trend of node states and the strength of mutual influence, a reinforcement learning method is used to automatically optimize the topology of the graph, so as to realize the dynamic evolution modeling of watershed topological relationships.

[0154] By employing a spatiotemporally adaptive graph convolutional architecture, a multi-level adjacency matrix learning mechanism, and a graph structure adaptive update module, dynamic evolution modeling of watershed topology relationships is achieved. This addresses the technical shortcomings of traditional static graph neural network modeling, which cannot adapt to seasonal changes and extreme weather events in watershed systems. By capturing temporal evolution features and spatial propagation features through spatiotemporally decoupled graph convolutional layers, and automatically optimizing the graph topology using reinforcement learning methods, the dynamic adaptability and prediction accuracy of complex watershed system modeling are significantly improved, providing stronger technical support for addressing climate change and extreme events.

[0155] Furthermore, the hierarchical decision-making module includes a coordination layer agent representing the top-level strategic goals and an execution layer agent representing different stakeholders. During the training of the watershed water-energy-food-ecology coupling model, the coordination layer agent dynamically adjusts the reward function of the execution layer agent, guiding the execution layer agents to engage in collaborative game and generating a collaborative scheduling strategy that balances the interests of multiple parties at different times.

[0156] The execution layer agents include power generation efficiency agents, agricultural irrigation agents, ecological protection agents, and shipping support agents. Each execution agent has an independent objective function and reward mechanism.

[0157] The process of generating the aforementioned cooperative scheduling strategy is as follows:

[0158] A dynamic internal bidding mechanism for water resource use rights is established. Each execution-layer agent uses the water-energy-food-ecology coupling model of the basin as its training environment, calculating utility estimates for key water volumes and release timing based on its objective function and reward mechanism within this environment. The coordination-layer agent dynamically adjusts and updates the reward function of each execution-layer agent based on the top-level strategic objective, the overall state of the basin, and the utility estimates of each agent.

[0159] Through multiple rounds of iterative training, each execution layer agent, guided by the coordination layer agent, gradually converges to the Pareto optimal cooperative scheduling strategy.

[0160] The dynamic internal bidding mechanism for water resource use rights adopts a multi-round progressive game auction approach, including:

[0161] Establish time-based differentiated bidding rules, setting different starting prices and increments for water resources in different time periods based on the scarcity of water resources, the urgency of demand, and strategic importance; and adopting different bidding strategies for the flood season, normal water season, and dry season.

[0162] A mechanism for forming intelligent agent alliances is established, allowing execution layer intelligent agents with complementary needs to form temporary alliances under specific conditions to conduct joint bidding. The success rate of bidding is improved through resource sharing and risk sharing within the alliance. Alliance revenue is distributed according to the contribution and risk-bearing ratio of each intelligent agent.

[0163] A bidding behavior learning and prediction module is constructed. Each execution layer agent observes the historical bidding behavior of other agents, uses an adversarial generative network to predict the bidding strategies of other agents, and adjusts its own bidding decisions based on the prediction results to form a dynamic game equilibrium.

[0164] By employing a multi-round progressive game auction approach, time-segmented differentiated bidding rules, an agent alliance formation mechanism, and a bidding behavior learning and prediction module, the intelligent upgrade of water resource allocation from simple bidding to complex game theory has been achieved. This overcomes the technical limitations of traditional resource allocation methods, which lack dynamic game theory capabilities and collaborative mechanisms. Through adversarial generative networks to predict competitors' strategies and an alliance revenue distribution mechanism, each agent can formulate the optimal bidding strategy in a complex game environment, significantly improving the fairness and efficiency of resource allocation and enabling intelligent coordination and win-win development among multiple stakeholders under resource scarcity conditions.

[0165] The process of dynamically adjusting and updating the reward function of each execution layer agent is as follows:

[0166] A two-way reward adjustment mechanism integrating punishment and incentives that incorporates external constraints is established. The coordination layer agent decomposes the reward function of each execution layer agent into a basic reward item, a collaborative incentive item, and a violation penalty item. The collaborative incentive item is calculated based on the utility valuation of each execution layer agent. The violation penalty item is obtained by constructing a set of hard constraints based on the top-level strategic goals and the overall state of the watershed, and building a quantitative evaluation system for the severity of violating the set of hard constraints.

[0167] As the number of training rounds increases, the weight coefficient of the co-incentive term is gradually reduced using an exponential decay function.

[0168] Furthermore, the decision attribution analysis module analyzes the collaborative scheduling strategy to obtain the key driving factors that lead to the preset scheduling behavior; the process of obtaining the key driving factors is as follows:

[0169] The Shapley additive interpretation algorithm is used to analyze the cooperative scheduling strategy and calculate the marginal contribution of each input feature in the cooperative scheduling strategy to the final decision.

[0170] Identifying key time steps and state variables in the decision-making process based on attention mechanisms;

[0171] By combining domain expert knowledge, the marginal contribution values ​​are weighted and sorted to select the top N key driving factors that affect the preset scheduling behavior, and a natural language explanation report containing the decision-making logic is generated.

[0172] Furthermore, the counterfactual deduction and optimization module defines counterfactual scheduling scenarios based on the key driving factors, calls the watershed water-energy-food-ecology coupling model for rapid deduction, quantifies and compares the comprehensive impact of counterfactual scheduling scenarios and collaborative scheduling strategies on water, energy, food and ecology under a preset time scale, and optimizes the collaborative scheduling strategy.

[0173] Based on user-inputted custom scheduling parameters and the key driving factors, a counterfactual scheduling scenario is constructed, and parallel computing technology is used to perform rapid simulation and deduction in the watershed water-energy-food-ecology coupling model; the custom scheduling parameters include reservoir outflow time sequence, generator start-up and shutdown plan and ecological water replenishment scheme;

[0174] Calculate the difference in comprehensive benefits between the counterfactual scheduling scenario and the collaborative scheduling strategy at a preset time scale; identify the weaknesses and improvement space of the current collaborative scheduling strategy based on the difference in comprehensive benefits, and generate strategy improvement suggestions including optimization direction and adjustment range; feed the strategy improvement suggestions back to the hierarchical decision module to update the agent's strategy parameters and reward function configuration, and optimize the collaborative scheduling strategy.

[0175] The feedback optimization process for the proposed strategy improvement adopts an adaptive multi-objective evolutionary strategy, including:

[0176] Establish a strategy improvement effect evaluation mechanism, track the changing trends of various indicators before and after strategy adjustment, and use a sliding time window statistical method to evaluate the actual effect of improvement suggestions; when the improvement effect is lower than the expected threshold, the strategy rollback and re-optimization process will be automatically triggered.

[0177] A hierarchical gradient feedback adjustment strategy is set up, and three feedback modes of different intensities are adopted according to the urgency and scope of the strategy improvement: rapid fine-tuning, medium-amplitude adjustment, and large-scale reconstruction. Among them, rapid fine-tuning is suitable for small parameter optimization, medium-amplitude adjustment is suitable for local strategy correction, and large-scale reconstruction is suitable for systemic strategy changes.

[0178] A multi-objective Pareto front dynamic search algorithm is constructed, which considers the trade-offs of multiple conflicting objectives during the policy optimization process. By maintaining and updating the Pareto optimal solution set, it provides decision-makers with diversified policy selection options. An elite preservation strategy is adopted to ensure that excellent solutions already discovered are not lost during the optimization process.

[0179] By employing an adaptive multi-objective evolutionary strategy, a strategy improvement effect evaluation mechanism, a hierarchical gradient feedback adjustment strategy, and a multi-objective Pareto front dynamic search algorithm, a technological leap from single-objective optimization to multi-objective collaborative optimization has been achieved. This solves the technical challenges of traditional strategy optimization methods being unable to handle multiple conflicting objectives simultaneously and lacking adaptive adjustment capabilities. By maintaining the Pareto optimal solution set and an elite preservation strategy, it can provide decision-makers with diverse strategy selection options, while supporting strategy rollback and re-optimization. This significantly improves the intelligence and robustness of strategy adjustment, ensuring continuous optimization capabilities and decision quality in complex and ever-changing environments.

[0180] Furthermore, the calculation process for the comprehensive benefit difference is as follows:

[0181] A comprehensive impact indicator system is constructed, including power generation benefits, grain output, water supply guarantee rate, irrigation satisfaction rate, ecological flow compliance rate, navigation guarantee rate, and biodiversity index. The objective weights of each indicator in the comprehensive impact indicator system are calculated using the entropy weight method. The subjective weights of the comprehensive impact indicator system are determined using the analytic hierarchy process. Based on the objective weights and the subjective weights, a comprehensive weight coefficient is formed, and the comprehensive benefit difference is calculated.

[0182] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative optimization decision-making system based on power-environment coupled big data, characterized in that, include: The watershed coupling module integrates real-time acquired multi-source ecological data with a preset mechanism model to construct a watershed water-energy-food-ecology coupling model. The hierarchical decision-making module includes a coordination layer agent representing the top-level strategic goals and an execution layer agent representing different stakeholders. During the training of the watershed water-energy-food-ecology coupling model, the coordination layer agent dynamically adjusts the reward function of the execution layer agent, guides the execution layer agents to engage in collaborative game and generates a collaborative scheduling strategy that balances the interests of multiple parties at different times. The decision attribution analysis module analyzes the collaborative scheduling strategy to obtain the key driving factors that lead to the preset scheduling behavior; the process of obtaining the key driving factors is as follows: The Shapley additive interpretation algorithm is used to analyze the cooperative scheduling strategy and calculate the marginal contribution of each input feature in the cooperative scheduling strategy to the final decision. Identifying key time steps and state variables in the decision-making process based on attention mechanisms; The marginal contribution values ​​are weighted and sorted by combining domain expert knowledge to select the top N key driving factors affecting the preset scheduling behavior, and a natural language explanation report containing decision-making logic is generated. The counterfactual deduction and optimization module defines counterfactual scheduling scenarios based on the key driving factors, calls the watershed water-energy-food-ecology coupling model for deduction, quantifies and compares the comprehensive impact of counterfactual scheduling scenarios and collaborative scheduling strategies on water, energy, food and ecology under a preset time scale, and optimizes the collaborative scheduling strategy.

2. The collaborative optimization decision-making system for power-environment coupled big data as described in claim 1, characterized in that: The multi-source ecological data includes: farmland soil moisture data and vegetation index data based on satellite remote sensing, watershed flow data and sediment content data based on hydrological station monitoring, rare aquatic organism activity data based on sonar monitoring, meteorological data, and power grid load data. The preset mechanism models include: watershed hydrological model, crop growth model and ecological evolution model.

3. The collaborative optimization decision-making system for power-environment coupled big data as described in claim 1, characterized in that: The process of constructing a watershed water-energy-food-ecology coupled model is as follows: The watershed topology is established based on graph neural networks, with reservoirs, irrigation districts, ecological protection zones and power grid nodes as graph nodes, and physical connections and ecological relationships as edges; The multi-source ecological data is assimilated into the preset mechanism model in real time using an ensemble Kalman filter algorithm, and the state parameters of the graph nodes are dynamically updated. By learning the causal relationships between nodes through dynamic Bayesian networks, the propagation effect of scheduling behavior on the state changes of each node is quantified, forming the watershed water-energy-food-ecology coupling model with causal inference capabilities.

4. The collaborative optimization decision-making system for power-environment coupled big data as described in claim 1, characterized in that: The execution layer agents include power generation efficiency agents, agricultural irrigation agents, ecological protection agents, and shipping support agents. Each execution agent has an independent objective function and reward mechanism. The process of generating the aforementioned collaborative scheduling strategy is as follows: A dynamic internal bidding mechanism for water resource use rights is established. Each execution-layer agent uses the water-energy-food-ecology coupling model of the basin as its training environment, calculating utility estimates for key water volumes and release timing based on its objective function and reward mechanism within this environment. The coordination-layer agent dynamically adjusts and updates the reward function of each execution-layer agent based on the top-level strategic objective, the overall state of the basin, and the utility estimates of each agent. Through multiple rounds of iterative training, each execution layer agent, guided by the coordination layer agent, gradually converges to the Pareto optimal cooperative scheduling strategy.

5. The collaborative optimization decision-making system for power-environment coupled big data as described in claim 4, characterized in that: The process of dynamically adjusting and updating the reward function of each execution layer agent is as follows: A two-way reward adjustment mechanism integrating punishment and incentives that incorporates external constraints is established. The coordination layer agent decomposes the reward function of each execution layer agent into a basic reward item, a collaborative incentive item, and a violation penalty item. The collaborative incentive item is calculated based on the utility valuation of each execution layer agent. The violation penalty item is obtained by constructing a set of hard constraints based on the top-level strategic goals and the overall state of the watershed, and building a quantitative evaluation system for the severity of violating the set of hard constraints. As the number of training rounds increases, the weight coefficient of the co-incentive term is gradually reduced using an exponential decay function.

6. The collaborative optimization decision-making system for power-environment coupled big data as described in claim 1, characterized in that: The process of optimizing the aforementioned collaborative scheduling strategy is as follows: Based on user-inputted custom scheduling parameters and the key driving factors, a counterfactual scheduling scenario is constructed, and parallel computing technology is used to perform rapid simulation and deduction in the watershed water-energy-food-ecology coupling model; the custom scheduling parameters include reservoir outflow time sequence, generator start-up and shutdown plan and ecological water replenishment scheme; Calculate the difference in comprehensive benefits between the counterfactual scheduling scenario and the collaborative scheduling strategy at a preset time scale; identify the weaknesses and improvement space of the current collaborative scheduling strategy based on the difference in comprehensive benefits, and generate strategy improvement suggestions including optimization direction and adjustment range; feed the strategy improvement suggestions back to the hierarchical decision module to update the agent's strategy parameters and reward function configuration, and optimize the collaborative scheduling strategy.

7. The collaborative optimization decision-making system for power-environment coupled big data as described in claim 6, characterized in that: The calculation process for the comprehensive benefit difference is as follows: a comprehensive impact indicator system is constructed, including power generation benefit, grain output, water supply guarantee rate, irrigation satisfaction rate, ecological flow compliance rate, navigation guarantee rate, and biodiversity index; the objective weight of each indicator in the comprehensive impact indicator system is calculated using the entropy weight method. The subjective weights of the comprehensive impact index system are determined by the analytic hierarchy process; a comprehensive weight coefficient is formed based on the objective weights and the subjective weights, and the comprehensive benefit difference is calculated.

Images