An Ecological Scheduling Assessment and Optimization Method for Downstream Basins of Water Conservancy and Hydropower Projects
By identifying ecologically sensitive areas using multi-source remote sensing data and image segmentation algorithms, and combining graph neural networks and reinforcement learning algorithms to simulate watershed processes, the shortcomings of traditional water conservancy and hydropower engineering scheduling methods in dynamic ecological demand response and extreme event risk prevention and control are solved, achieving high-precision ecological scheduling optimization and multi-objective collaborative balance.
Patent Information
- Application Number
- CN202611087243.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional water conservancy and hydropower project scheduling methods are difficult to dynamically respond to the ecological needs of the basin, lack real-time intelligent decision-making capabilities, and are insufficient in the prevention and control of extreme event risks.
Multi-source remote sensing data and image segmentation algorithms are used to identify and partition ecologically sensitive areas. Graph neural networks and reinforcement learning algorithms are combined to simulate watershed processes and perform causal reasoning. An adaptive ecological scheduling optimization model is constructed, and a multi-objective genetic algorithm is integrated to optimize the scheduling strategy.
It achieves high-precision automatic identification and dynamic zoning of ecologically sensitive areas, can evaluate multi-dimensional ecological indicators in real time, improves the adaptability and scientific nature of scheduling schemes, takes into account both ecological protection and economic development, and has risk prevention and control capabilities.
Smart Images

Figure CN123117587A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water conservancy and hydropower development and water ecological protection technology. Background Technology
[0002] With the continuous improvement of watershed development and utilization, water conservancy and hydropower projects play an important role in water resource allocation, flood control and disaster reduction, and water supply security. However, excessive or unreasonable scheduling and operation methods have also exacerbated the degradation of downstream river ecosystems, manifesting as a series of ecological and environmental problems such as ecological water shortage, decline in aquatic biodiversity, wetland shrinkage, and water quality deterioration. Traditional reservoir scheduling methods are mostly based on single objectives or limited water supply, power generation, and safety control, often neglecting the diverse and dynamic needs of downstream watershed ecosystems for hydrological processes, making it difficult to achieve coordinated development of ecology, society, and economy. At the same time, affected by uncertain factors such as climate change, extreme weather, and fluctuations in social water demand, watershed scheduling problems exhibit extremely high dynamic complexity and uncertainty, requiring scheduling systems to have adaptive optimization and risk response capabilities.
[0003] In recent years, with increasing attention paid to the mechanisms of eco-hydrological processes, the concept of ecological scheduling has been gradually introduced into the field of water conservancy and hydropower engineering management. Scholars have successively proposed various methods such as multi-objective optimization, ecological water demand calculation, and water environment simulation. However, most current mainstream scheduling optimization schemes are based on static, multi-stage programming, or heuristic algorithms, making it difficult to dynamically respond to the spatiotemporal variations and complex constraints of watershed ecological needs, and lacking real-time intelligent decision-making capabilities for actual operating conditions. Especially under multi-objective, multi-constraint, and nonlinear feedback conditions, a single algorithm struggles to efficiently solve for the optimal solution that balances ecological benefits, social needs, and economic interests. Furthermore, timely early warning and effective prevention and control of extreme events such as droughts and floods also require breakthroughs. Summary of the Invention
[0004] The purpose of this invention is to provide an ecological scheduling assessment and optimization method for downstream watersheds of water conservancy and hydropower projects. This method enables accurate identification of key ecological functions in downstream watersheds, quantification of dynamic demands, simulation of complex processes, and optimal scheduling decisions, thereby improving the ecological protection effectiveness and socio-economic sustainable development capacity of downstream watersheds of water conservancy and hydropower projects.
[0005] To address the aforementioned technical problems, this invention provides an ecological scheduling assessment and optimization method for downstream watersheds of water conservancy and hydropower projects, comprising the following steps: S1. Intelligent identification and zoning of ecologically sensitive areas: By integrating multi-source remote sensing datasets and image segmentation algorithms, the automatic extraction and spatial zoning of ecologically sensitive areas across the entire downstream watershed is completed. S2. Multi-objective dynamic modeling of ecological demand: Based on the zoning results of ecologically sensitive areas output in step S1, a multi-objective dynamic ecological demand model is constructed. The weights of various ecological functions are set by fuzzy comprehensive evaluation coupled hierarchical analysis method. The ecological demand of the whole region is iteratively updated based on real-time hydrological data of rainfall, runoff, and seasonal time series. S3. Watershed process simulation and causal reasoning analysis: Construct a watershed process simulation model based on graph neural network, and output the spatiotemporal evolution prediction sequence of hydrodynamic, water quality and aquatic biological response of each downstream ecological unit; build a structured causal model to identify the causal dependency links between scheduling intervention behavior and various ecological indicators. S4. Adaptive optimization of ecological scheduling scheme: The multi-objective ecological demand output in step S2 and the watershed ecological simulation prediction results output in step S3 are used as optimization objectives and constraint boundaries. The reservoir scheduling operation parameters and the time-period discharge flow process are set as decision action variables. The optimal scheduling strategy is solved iteratively by integrating reinforcement learning algorithm and evolutionary algorithm. The risk warning threshold and socio-economic water use constraints are embedded in the strategy solution stage to complete the scheme selection.
[0006] Step S1 includes the following steps: S11. Simultaneously acquire high-resolution satellite remote sensing imagery and UAV aerial ground imagery, and uniformly perform spatial positioning and geometric registration on both types of imagery; integrate multiple environmental data such as surface reflectance, multi-temporal vegetation index, water index, and surface temperature to construct a multimodal spatiotemporal three-dimensional tensor group. ,in For regional height, For area width, For multi-source feature dimensions, For the time dimension; collect data from biodiversity survey sites in the watershed, and construct an image segmentation model to train a supervised label set; S12. A variant of the spatiotemporal embedding U-Net model is adopted as the backbone network for image segmentation. The backbone network incorporates a spatiotemporal attention module and a multi-scale feature fusion module. The spatiotemporal attention module calculates the cross-temporal feature similarity weights. The calculation formula is: In the formula, For encoder number Layer, Time ,Location eigenvectors, This refers to the vector dot product operation. S13. Based on the Gaussian kernel density estimation coupled spatial adaptive threshold method, the pixel sensitivity probability map output by the model is divided into regions. The regions obtained are fixedly divided into three categories: core protection zone, transition buffer zone, and general zone. S14. Introduce graph convolution algorithm, and combine the topological relationship of watershed river network system, topography elevation, and ecological connectivity channels to perform post-processing optimization on the partition vector boundary to eliminate broken and fragmented patches.
[0007] In step S12, the spatiotemporal embedding U-Net variant model employs a dual-branch composite loss function, which is composed of a weighted superposition of a boundary-aware Dice loss branch and a pixel-level cross-entropy loss branch. The formula for calculating the boundary-aware Dice loss is as follows: In the formula, For the first One pixel predicted value, For the first The actual labeled value of each pixel. To prevent the denominator from being zero, a constant, This represents the total number of pixels in the image.
[0008] Step S2 includes the following steps: S21. Construct a hierarchical framework for hierarchical analysis, which is divided into three layers from top to bottom: target layer, criteria layer, and scheme layer. The criteria layer is fixed with four categories of criteria: species habitat, wetland function, ecosystem service, and human water use. The pairwise judgment matrix of the criteria layer is constructed using the expert scoring method. After consistency verification, the matrix is normalized and solved to obtain the static basic weights of each criterion. S22. Construct a triangular fuzzy membership function for each type of ecological unit. Binding suitable hydrological threshold ranges for each ecological indicator The piecewise expression for the triangular fuzzy membership function is: In the formula, These are real-time hydrological index observations. For the first The lower limit threshold of each ecological indicator For the first The upper limit threshold for each ecological indicator; time Next Ecological target demand The calculation formula is: In the formula, For the first The weight of each ecological goal criterion The total number of watershed zones, For the first Partition to the first Contribution weight to ecological goals for Time of the first Partition corresponding to the first The hydrological state variables of the target For the corresponding triangular fuzzy membership function; S23. Introduce a data envelopment analysis model, treating each watershed zone and each preset scheduling scheme as an independent decision-making unit, and calculate the relative input-output efficiency of ecological goal satisfaction for different units; the data envelopment analysis objective function is: In the formula, As the weight of ecological output indicators, As the weight of water resource input indicators, To achieve ecological goals, For adjustable water resource input, the basic weights are corrected using the entropy weight method, and under extreme hydrological scenarios, fuzzy data envelopment analysis is switched to complete adaptive parameter correction.
[0009] Step S3 includes the following steps: S31. Discretize the downstream watershed into graph nodes according to sub-river segments, wetland patches, and sub-watershed units. Use natural river channels, artificial water conveyance pipelines, and water body connectivity sections as directed edges between nodes to construct a watershed topology graph that is dynamically updated with hydrological time series. ; S32. Collect real-time inflow, regional precipitation, watershed evapotranspiration, reservoir discharge, river level, and regional species abundance data from multiple dimensions. After standardization and normalization, these data are spliced together to form attribute tensors specific to each node. ; S33. A watershed simulation network is constructed using time-series graph convolutional units superimposed with gated recursive units. GRU or LSTM are selected as the gated recursive units. The network node state update formula is: In the formula, for Time Node Latent state vector, For nodes The set of all adjacent nodes, for Time Node To node edge weights, For the characteristic transformation function, It is the ReLU activation function. The network can be trained parameters; the feature space interaction and temporal state update of the neighboring nodes are completed layer by layer, and the predicted sequence of water level, dissolved oxygen and benthic organism density for future multi-step length is output. S34. Construct a causal directed acyclic graph, where the variable nodes include reservoir scheduling behavior, meteorological driving inputs, watershed hydrological response, and ecological evaluation indicators; perform counterfactual reasoning based on do-interventional factors, and calculate the average causal effect (ACE) of scheduling operations on the target ecological indicator using the following formula: In the formula, As an ecological response indicator, For scheduling operations, a binary intervention variable, To estimate the causal interference, we quantify the contribution percentage of each variable.
[0010] Step S4 includes the following steps: S41. Encode the flow rate and gate opening parameters for each time period within the complete scheduling cycle into decision chromosomes, and use the NSGA-II multi-objective genetic algorithm to perform population iteration, outputting a Pareto optimal scheduling candidate solution set under ecological and economic dimensions. S42. Construct an Actor-Critic architecture deep reinforcement learning network, adding an entropy regularization constraint term to the network; use reservoir water level, real-time inflow, ecological satisfaction, and risk warning signals as a concatenated state vector. A continuous action space is constructed using flow regulation parameters within a continuous interval; the Bellman error loss calculation formula for the Critic network is: The gradient update formula for the Actor network policy is: In the formula, For the current value network, For soft target value networks, This is the total reward value for a single step. As a reward discount factor, For policy networks; S43. Store all candidate scheduling schemes output by the genetic algorithm into the reinforcement learning experience replay pool, and iteratively update the model parameters of the Actor policy network and the Critic evaluation network based on the policy gradient method. S44. The probability distribution of extreme hydrological events, water quality safety threshold, total urban and rural water supply red line, and flood control water level are all embedded as hard constraints into the optimization process. Dispatch strategies that do not meet the constraints are eliminated, and the compliant optimal dispatching scheme is output.
[0011] Step S4 involves constructing a multi-objective comprehensive reward function, the expression of which is: In the formula, For the first Ecological goals and benefits For the first Socioeconomic benefits, For dynamic weights, For extreme event risk loss function, As a risk penalty coefficient, The scheduling strategy needs to be optimized.
[0012] In step S33, each layer of the temporal graph convolutional unit is set with an independent trainable neighborhood feature transformation weight matrix. After the neighborhood node features are mapped by the weight matrix, they are aggregated to the central node to complete the spatial information transmission.
[0013] In step S41, the NSGA-II algorithm maintains population diversity through fast non-dominated sorting and crowding distance calculation.
[0014] After the partitioning in step S1 is completed, the incremental transfer learning process is started. The spatiotemporal embedding U-Net variant model is fine-tuned based on the latest remote sensing images and field ecological monitoring sampling data. The model update cycle is synchronized with the collection and update frequency of remote sensing data and field monitoring data.
[0015] Compared with existing technologies, this invention significantly improves the scientific rigor and precision of ecological scheduling in downstream watersheds of water conservancy and hydropower projects by integrating remote sensing intelligent identification, dynamic ecological demand modeling, watershed spatiotemporal process simulation and causal reasoning analysis, and an adaptive ecological scheduling optimization method that combines reinforcement learning and evolutionary algorithms. On the one hand, it achieves high-precision automatic identification and dynamic zoning of ecologically sensitive areas, providing a scientific basis for watershed ecological protection and functional zoning. On the other hand, it can dynamically quantify ecological needs based on multi-objective, multi-temporal, and multi-source data, enabling real-time assessment and adjustment of multi-dimensional ecological, biological, and hydrological indicators, thus enhancing the adaptability of scheduling schemes to complex ecosystems. Furthermore, by introducing graph neural networks and causal reasoning, it achieves spatiotemporal linkage, efficient simulation, and identification of key causal links in the watershed's ecological physical processes, providing solid support for scientific decision-making.
[0016] Ultimately, by combining reinforcement learning and multi-objective genetic algorithms to collaboratively optimize scheduling decisions, and taking into account ecological protection, economic development, and risk management, a better ecological scheduling scheme can be output, promoting the sustainable use of water resources in the basin, the healthy restoration of the ecosystem, and the improvement of the quality of the living environment. It has broad application prospects and significant comprehensive benefits.
[0017] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description
[0018] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.
[0019] Figure 1 This is a flowchart illustrating at least one embodiment of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details are presented in the embodiments of this invention to facilitate a better understanding of this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this invention. The embodiments can be combined with and referenced by each other without contradiction.
[0021] Example 1 like Figure 1 The method for ecological scheduling assessment and optimization in downstream watersheds of water conservancy and hydropower projects, as shown, includes the following steps: S1. Intelligent Identification and Zoning of Ecologically Sensitive Areas: Integrating multi-source remote sensing datasets and image segmentation algorithms, this system automatically extracts and spatially partitions ecologically sensitive areas across the entire downstream watershed. By employing remote sensing and image segmentation for automatic zoning instead of manual on-site boundary surveying, it achieves comprehensive coverage without blind spots and simultaneously outputs standardized spatial boundaries. This provides a precise spatial basis for subsequent regionally differentiated ecological water demand calculations, avoiding the problems of blurred boundaries and lag caused by manual zoning.
[0022] S2. Multi-objective Dynamic Modeling of Ecological Demand: Based on the zoning results of ecologically sensitive areas output in step S1, a multi-objective dynamic ecological demand model is constructed. Fuzzy comprehensive evaluation coupled with hierarchical analysis is used to set the weights of various ecological functions. The overall ecological demand is iteratively updated based on real-time hydrological data of rainfall, runoff, and seasonal time series. Using the zoning results as the modeling basis, differentiated water demand accounting is achieved for different regions. Simultaneously, hydrological time series data is introduced to dynamically update the demand values. This differs from traditional fixed ecological water demand quotas and can adapt to differentiated hydrological and ecological demands during flood season, dry season, and the reproductive season of flora and fauna.
[0023] S3. Watershed Process Simulation and Causal Reasoning Analysis: Construct a watershed process simulation model based on graph neural networks, outputting spatiotemporal evolution prediction sequences of hydrodynamic, water quality, and aquatic biological responses in downstream ecological units; build a structured causal model to identify causal dependency links between scheduling intervention behaviors and various ecological indicators. GNN simulation is naturally adapted to the topological connectivity of the watershed river network, and can synchronously simulate upstream and downstream linked hydrological and ecological changes; causal reasoning distinguishes between correlation and actual scheduling impact, accurately locating key control measures, and avoiding control deviations caused by relying solely on data correlation to formulate scheduling plans.
[0024] S4. Adaptive Optimization of Ecological Scheduling Scheme: Using the multi-objective ecological demands output in step S2 and the watershed ecological simulation prediction results output in step S3 as optimization objectives and constraint boundaries, reservoir scheduling operation parameters and time-period discharge flow processes are set as decision-making action variables. Reinforcement learning and evolutionary algorithms are integrated to iteratively solve for the optimal scheduling strategy. Risk warning thresholds and socio-economic water use constraints are embedded in the strategy solution stage to complete scheme selection. Pre-existing ecological demands and ecological predictions are embedded as hard constraints in the optimization process, taking into account multiple constraints such as ecological red lines, water supply, and flood control. The fusion of the two algorithms balances global optimization and real-time dynamic adaptation, addressing the shortcomings of single algorithms in balancing multiple objectives and poor adaptability to extreme hydrological scenarios.
[0025] This establishes a complete closed-loop process of "spatial identification—dynamic water demand calculation—watershed ecological simulation—intelligent scheduling optimization," placing ecological constraints at the forefront of the entire process from the top-level process architecture. The entire basic framework does not restrict specific models or algorithm details, adapting to downstream watersheds of large, medium, and small-scale water conservancy and hydropower projects. It can support routine ecological scheduling management in stable hydrological years, and can also complete scheduling prediction and control operations for special scenarios such as extreme weather and sudden pollution. While ensuring the stable operation of the reservoir's basic functions of flood control, water supply, and power generation, the entire process systematically coordinates the diverse needs of species habitat, wetland maintenance, and watershed ecosystem services, achieving a synergistic balance among the ecological, social, and economic development goals of the watershed, and comprehensively improving the sustainable development and utilization level of water resources and the overall health of the river ecosystem.
[0026] Example 2 Based on Example 1, step S1 includes the following steps: S11. Simultaneously acquire 3m resolution high-resolution satellite remote sensing imagery and 5cm resolution UAV aerial ground imagery, and uniformly complete spatial positioning and geometric registration in the WGS84 coordinate system; integrate surface reflectance, NDVI / EVI vegetation index, MNDWI water index, and surface temperature to construct a multimodal spatiotemporal three-dimensional tensor set. Where H represents the vertical pixel height of the image, W represents the horizontal pixel width, D represents the environmental feature dimension, and T represents the number of time-series images. 200–1000 biodiversity sampling points are deployed in the field to record fish and wetland vegetation abundance, constructing a model supervision label set. Multi-source remote sensing fusion constructs a four-dimensional spatiotemporal tensor, simultaneously incorporating spatial, temporal, and multi-dimensional environmental features, enabling accurate capture of seasonal vegetation and water body changes. Field biological sampling serves as supervision labels, aligning remote sensing images with actual ecological conditions and significantly reducing ecological identification errors caused by simple image segmentation.
[0027] S12. A spatiotemporal embedding U-Net variant model is used as the segmentation backbone network, with a built-in spatiotemporal attention module and a multi-scale feature fusion module. The spatiotemporal attention weight calculation formula is as follows: Spatiotemporal attention weight calculation formula: In the formula, Let i be the feature vector of the encoder at layer l, time t, and position i. This is a vector inner product operation. Spatiotemporal attention weights enable cross-temporal image feature association calculations, automatically amplifying the weights of ecologically sensitive areas with drastic seasonal fluctuations, and improving the segmentation accuracy of the boundary between flood season and dry season; multi-scale feature fusion simultaneously takes into account both the overall pattern of large wetlands and the detailed identification of micro-habitats for small fish spawning.
[0028] The network employs a two-branch composite loss function, and the boundary-aware Dice loss is calculated using the following formula: In the formula, For pixel prediction sensitivity probability, To label the true value in the field, To prevent the denominator from being zero, N represents the total number of pixels in the image. Dice loss is specifically optimized for patch boundary segmentation, accurately identifying micro-core protected areas such as riverbeds, shallows, and scattered wetlands.
[0029] S2 includes the following steps: S21. A three-layered AHP (Aspect-Based Hierarchy) framework is constructed. The objective layer aims to maximize watershed ecosystem services. The criteria layer sets four categories of criteria: species habitat, wetland function, ecosystem services, and human livelihood water use. Expert scoring constructs a pairwise judgment matrix, and after consistency verification, the static basic weights of each criterion are normalized and solved. The four-layered criteria cover both ecological and livelihood dimensions. Consistency verification avoids logical contradictions in weights, quantitatively distinguishes the importance of multiple ecological needs, and avoids arbitrary assignment based on subjective experience.
[0030] S22. Construct triangular fuzzy membership functions to bind the hydrological suitability intervals of each ecological unit. Piecewise expressions: Piecewise expression of triangular fuzzy membership function: The triangular membership function quantifies the gradual relationship between hydrological indicators and ecological suitability, and can accurately characterize the degree of ecological satisfaction under different gradients of mild water shortage, suitability, and water surplus.
[0031] Formula for calculating the demand for the j-th ecological objective at time t: In the formula, The ecological target criteria weights are given by m, where m is the total number of watershed zones. The weighting of ecological objectives is determined by the zoning. Let t be the hydrological observation value of the zone at time t. This formula integrates three variables: zone weight, ecological target weight, and real-time hydrological membership, to achieve automatic calculation of dynamic ecological water demand by region and time period, supporting refined zoned and time-based management of the scheduling scheme.
[0032] S23. Construct a DEA efficiency measurement model, with the objective function as follows: Water resource input To meet the requirements of investment and ecological goals To calculate the relative efficiency of scheduling schemes for each region, the entropy weight method is used to correct weights in conventional hydrological scenarios, while the fuzzy DEA model is switched to correct parameters in extreme drought and flood scenarios. The DEA model quantifies the relative ecological efficiency of water resource utilization under different regions and scheduling schemes. The entropy weight method objectively corrects the subjective weight bias of AHP, and the fuzzy DEA adapts to the uncertainty of indicators under extreme hydrological conditions, thus improving the robustness of the model across all scenarios.
[0033] Step S3 includes the following steps: S31. Discretize the watershed into sub-river segments, wetland patches, and sub-watershed graph nodes, and use natural river channels and water conveyance pipelines as directed edges to construct a time-series dynamic topology graph. The graph structure replicates the natural connectivity topology of a watershed, which can be flexibly adapted to complex water system morphologies such as river bifurcation, artificial canal systems, and wetland connectivity, greatly simplifying the workload of modeling complex watersheds.
[0034] S32. Integrate real-time inflow, regional precipitation, evapotranspiration, reservoir discharge, river level, and species abundance time-series monitoring data, and normalize and splice them to form a node attribute tensor. Multi-source monitoring data are input using a unified tensor normalization method, enabling simultaneous coupled simulation of multiple indicators, including hydrological, water quality, and biological data.
[0035] S33. A simulation network is built by stacking time-series graph convolutional units with GRU gated recursive units. The node state update formula is as follows: In the formula, Let be the node state vector at time t. For the set of adjacent nodes, For time-series edge weights, The ReLU activation function is used, and U and B are trainable network parameters. Iteratively, layer by layer, it outputs predicted sequences for water level, dissolved oxygen, and benthic organism density for the next 6 months. Each temporal graph convolutional unit is configured with an independent trainable neighborhood feature transformation weight matrix to complete spatial feature aggregation. The temporal graph convolution fused with GRU simultaneously captures both spatial water system transmission patterns and temporal hydrological evolution patterns. The independent neighborhood weight matrix distinguishes the intensity of upstream and downstream hydraulic transmission. The prediction results simultaneously include multi-dimensional ecological indicators of hydrology, water quality, and biology, providing complete pre-constraint data for scheduling optimization.
[0036] S34. Construct a causal directed acyclic graph, where the variable nodes include reservoir scheduling behavior, meteorological driving inputs, watershed hydrological response, and ecological evaluation indicators; perform counterfactual reasoning based on do-intervention sub-calculus, and calculate the average causal effect (ACE) of scheduling operations on the target ecological indicator using the following formula: The average causal effect quantifies the net impact of a single scheduling operation on ecological indicators, removes irrelevant interference variables such as rainfall and topography, and accurately identifies the optimal control flow range.
[0037] This system integrates a high-precision spatiotemporal deep learning segmentation network, a FAHP-DEA coupled dynamic water demand model, a time-series graph neural network causal simulation, and incremental transfer learning dynamic updating—a complete set of refined techniques—to fully connect the entire chain of watershed ecological space identification, dynamic ecological demand quantitative calculation, watershed multi-process spatiotemporal coupled simulation, and causal effect quantification of scheduling behavior. Based on multi-source remote sensing four-dimensional tensors and a time-series embedded U-Net network, it achieves full-domain, multi-temporal, high-precision ecological functional zoning. Incremental transfer learning continuously updates the zoning results, ensuring long-term accuracy of spatial base data. Based on a coupled model of fuzzy hierarchical analysis and data envelopment analysis, it achieves dynamic adaptive calculation of ecological water demand by region, season, and multiple objectives, objectively balancing ecological protection and domestic water needs. Based on a dynamic topological graph neural network, it synchronously simulates the temporal evolution of hydrology, water quality, and biology in the upper and lower reaches of the watershed, and combines causal inference to quantify the true contribution of each type of scheduling measure to ecological indicators. The entire module can provide high-precision, dynamic, multi-dimensional, and quantifiable ecological constraint inputs for backend scheduling optimization. It can accurately predict the evolution trend of downstream wetlands, fish habitats, and water quality under long-term scheduling operation. It can formulate differentiated ecological water use baselines for dry seasons, breeding seasons, and extreme drought periods, thereby improving the scientific nature, refinement, and long-term reliability of ecological scheduling schemes.
[0038] Example 3 Based on Example 1, step S4 includes the following steps: S41. The flow rate released in each time period and the opening parameters of each gate within the complete scheduling cycle are encoded into decision chromosomes. The NSGA-II multi-objective genetic algorithm is used to perform population iteration, outputting a Pareto optimal scheduling candidate solution set under multiple dimensions of ecology and economy. During the execution of the NSGA-II algorithm, population diversity is maintained through fast non-dominated sorting and congestion distance calculation. The flow rate time sequence is encoded into chromosomes to realize the global traversal search of continuous scheduling schemes. The NSGA-II fast non-dominated sorting generates multi-objective equilibrium candidate schemes, and the congestion distance ensures that the solution set covers multiple control tendencies such as ecological priority and economic priority, providing managers with multiple optional scheduling schemes.
[0039] S42. Construct an Actor-Critic architecture deep reinforcement learning network, adding an entropy regularization constraint term to the network; use reservoir water level, real-time inflow, ecological satisfaction, and risk warning signals as a concatenated state vector. A continuous action space is constructed using flow regulation parameters within a continuous interval; the Bellman error loss calculation formula for the Critic network is as follows: Actor Network Policy Gradient Update Formula: The Actor-Critic architecture adapts to the continuous flow regulation action space of reservoirs, and entropy regularization prevents the strategy from converging to a local optimum too early. The dual loss function is used for collaborative training, and the scheduling strategy is continuously iterated based on the real-time ecological and hydrological feedback of the watershed, which has the ability to quickly adapt and adjust to extreme hydrological scenarios.
[0040] S43. Store all candidate scheduling schemes output by the genetic algorithm into the reinforcement learning experience replay pool, and iteratively update the model parameters of the Actor policy network and the Critic evaluation network based on the policy gradient method. The global solution set of the genetic algorithm expands the reinforcement learning experience pool, significantly reducing the number of trial and error iterations for the agent, and balancing the efficiency of searching for the global optimal solution with the ability to dynamically adjust in real time.
[0041] S44. The probability distribution of extreme hydrological events, water quality safety thresholds, total urban and rural water supply limits, and flood control water levels are all incorporated as hard constraints into the optimization process. Dispatch strategies that do not meet these constraints are eliminated, and the optimal compliant dispatch plan is output. Multiple hard constraints preemptively intercept illegal dispatch plans, simultaneously incorporating the risk probabilities of drought and flood extreme events. During the optimization process, risks such as ecological damage, water shortages, and exceeding flood control limits are proactively avoided, achieving optimal multi-objective regulation under safety constraints.
[0042] Step S4: Construct the multi-objective comprehensive reward function. The expression for the comprehensive reward function is as follows: The reward function simultaneously quantifies three types of indicators: ecological benefits, socio-economic benefits, and risk losses. The dynamic weights are adapted to different seasonal control emphases. The risk penalty item actively reduces the reward for high-risk scheduling strategies, guiding the agent to prioritize the generation of safe and eco-friendly outflow processes.
[0043] Therefore, a collaborative optimization architecture combining the NSGA-II multi-objective genetic algorithm and Actor-Critic deep reinforcement learning is adopted to achieve bidirectional complementarity between global Pareto solution set search and real-time dynamic adaptive regulation. The genetic algorithm generates multi-dimensional balanced scheduling candidate schemes in batches to expand the reinforcement learning experience pool and shorten the training convergence cycle of the agent. The reinforcement learning continuously iterates the scheduling strategy based on real-time watershed hydrological, ecological, and risk monitoring data to quickly respond to hydrological changes and extreme weather events. The multi-objective weights are dynamically balanced based on a comprehensive reward function that includes ecological benefits, socio-economic benefits, and risk losses. At the same time, flood control level, total water supply, water quality threshold, and extreme event probability are all set as hard constraints for pre-filtering. This ensures the priority of ecological indicators in the scheduling optimization process and can output multiple feasible scheduling schemes that take into account ecological protection, power generation and water supply benefits, and flood control and drought resistance safety. It is suitable for long-term annual and quarterly scheduling planning and also supports real-time dynamic flow regulation at daily and hourly scales, comprehensively improving the intelligence level, risk prevention and control capabilities, and multi-objective collaborative balance capabilities of water conservancy and hydropower project scheduling decisions.
[0044] Example 4 In conjunction with the above embodiments, the specific steps include: Step 1: Intelligent Identification and Zoning of Ecologically Sensitive Areas First, by utilizing multi-source remote sensing data (such as high-resolution satellite imagery and UAV aerial photography) and image segmentation algorithms in machine learning (such as U-Net variants that integrate spatiotemporal feature mechanisms), ecologically sensitive areas in the downstream watershed are automatically extracted and partitioned. This method can intelligently identify key ecological protection and restoration areas based on historical and real-time environmental data (including vegetation indices, water distribution, and biodiversity indicators), and achieve dynamic updates, providing a precise spatial basis for subsequent scheduling and assessment.
[0045] In the practical operation of intelligent identification and zoning of ecologically sensitive areas, the entire process needs to fully integrate remote sensing information processing, spatiotemporal data mining, and deep learning algorithms to achieve detailed identification of the complex geographical and ecological spaces of the downstream watershed. At the data level, after spatial positioning and registration, high-resolution satellite remote sensing imagery and UAV ground pseudo-color imagery are further integrated with multi-dimensional information such as surface reflectance, multi-temporal vegetation indices (e.g., NDVI, EVI), water body indices (e.g., MNDWI), and surface temperature to construct a three-dimensional tensor. Where H and W represent the region height and width, respectively, D is the multi-source feature dimension, and T represents the time dimension. This tensor serves as the input for subsequent algorithms, merging upstream historical data and real-time monitoring data to ensure the timeliness and richness of the data. Simultaneously, biodiversity data from field survey sampling points are used as supervisory labels for the training set, enabling multimodal training and evaluation.
[0046] The algorithm backbone adopts a variant of the spatiotemporal embedding U-Net. Traditional U-Net has a typical encoder-decoder structure, i.e., downsampling to capture large-scale spatial features and upsampling to recover segmentation boundaries. This scheme introduces a "spatial-temporal attention mechanism" and a "multi-scale feature fusion module." Specifically, a spatiotemporal attention layer is inserted into the skip connection path of the standard U-Net. Let the output feature of the encoder's \(l\) layer be denoted as... Based on this, spatiotemporal self-attention weights are constructed. Defined as: here, Indicates the location ,time eigenvectors, The inner product operation measures the similarity of the same region at different times. By fusing cross-temporal information, the model can dynamically focus on spatial units with drastic ecological fluctuations and volatile states during the monitoring period, thereby improving the accuracy of segmenting environmentally sensitive areas. Multi-scale feature fusion, through cascading backbone feature maps at different resolutions, achieves coordinated control over ecological details and the overall pattern.
[0047] The loss function employs multi-branch supervision, with one branch using boundary-aware Dice loss to facilitate accurate identification of small patches and ecologically sensitive points in boundary areas. The formula is as follows: in and These are the predicted value and the actual label for pixel i, respectively. To prevent small constants with zero denominators, another branch uses joint cross-entropy loss to focus on overall classification accuracy. During training, dynamic weight allocation is integrated to ensure good generalization results for both small ecological anomalies and main regions.
[0048] The model inference stage implements end-to-end processing: inputting multi-source remote sensing data from different time slices of the watershed, the model outputs the ecological sensitivity probability of each pixel (i.e., the probability distribution of a pixel belonging to a high-sensitivity area). Then, a spatial adaptive thresholding method based on Gaussian kernel density estimation is used to partition the probability map into three categories: "core protection area," "transition buffer zone," and "general area." Spatial partitioning can also incorporate graph convolution post-processing based on the topological relationships of the watershed river network, topography, and ecological corridors to avoid the appearance of unreasonable small patches or fragmented areas.
[0049] Each time the latest imagery or field ecological monitoring data is obtained, the automatic zoning model can invoke an incremental transfer learning strategy to fine-tune historical model parameters, accelerating adaptation to new ecological scenarios and ensuring that the distribution of ecologically sensitive areas can be dynamically adjusted according to seasonal and climatic events. Furthermore, all processing workflows can interface with cloud platforms or geographic information service systems, enabling automatic model deployment and rapid visualization of zoning results, providing a high-resolution, dynamically available spatial foundation for subsequent scheduling assessments and ecological optimization.
[0050] Step 2: Dynamic Modeling of Multi-Objective Ecological Demands Based on ecologically sensitive areas, and considering the needs of species habitat, wetland function, and ecosystem services, a multi-objective dynamic ecological demand model is constructed. Fuzzy comprehensive evaluation and hierarchical analysis (FAHP-DEA ensemble model) are introduced, with weights assigned to different ecological functions, and ecological demand is dynamically adjusted based on real-time hydrological information such as rainfall, runoff, and seasonal variations. This model can integrate ecosystem vulnerability, seasonal demand, and emergency response, achieving the quantification and dynamic adjustment of multi-objective ecological demands.
[0051] The realization of dynamic modeling of multi-objective ecological demands relies on the results of ecologically sensitive zoning. It represents the diverse hydrological demands of different functional zones within the region (fish spawning zones, wetland and waterbird habitats, water purification zones, and farmland irrigation areas) in a formalized, dynamic, and multi-objective interactive manner. Each type of ecological unit often has unique and time-varying suitable ranges for watershed water quantity, water quality, flow velocity, and connectivity. To scientifically express and measure this complex relationship, a fuzzy comprehensive evaluation (FCE) combined with the analytical hierarchy process (AHP) is introduced to form a hierarchical system of ecological demands based on multiple criteria, weights, and scenarios. This system is coupled with data envelopment analysis (DEA) to solve and dynamically balance resource allocation conflicts between different ecological objectives.
[0052] The mathematical foundation for this modeling process is built from the bottom up. Within the AHP framework, the objective layer is "maximizing regional ecosystem services throughout the year," while the criterion layer is divided into species habitat (water level / flow velocity indicators), wetland function (area / water holding time indicators), ecosystem services (water quality / carbon sink, buffering capacity), and human livelihoods (water resources / irrigation), etc. The scenario layer consists of specific flow, time, and quality control indicators for each sensitive area. An expert scoring method is used to establish a pairwise judgment matrix to assess the importance of each need: for any criterion... Construct a judgment matrix element And satisfying the consistency condition, the criterion weights are obtained by normalizing the principal eigenvectors. ,satisfy Against the backdrop of dynamic changes in parameters such as daily rainfall, inflow, water price, and pollution load, fuzzy membership functions correlate ecological demand indicators with actual hydrological conditions. Let the ecological demand threshold be an interval... The current hydrological status is as follows: Define the triangular membership function. It reflects the degree to which the current state meets the target requirements. This function is widely used for ecological variables such as suitable water levels for different species and optimal inundation depths for wetlands.
[0053] FAHP dynamically integrates membership functions with the hierarchical weights of AHP. For each type of ecological unit, it periodically recalculates the weights of each objective and the satisfaction of criteria based on real-time downstream hydrological data (rainfall, irrigation water, groundwater recharge, etc.) and external disturbances (flood season, dry season, pollution events, etc.). The overall multi-objective ecological demand vector can be represented by the following formula: Assuming m regions and n ecological function objectives, the ecological demand at time t is... Then, for each target, the comprehensive membership degree is: in This represents the contribution weight of region i to objective j. These are the relevant state variables for that area. Membership function evaluates the degree of satisfaction.
[0054] The model's dynamism is reflected in the time-varying feedback of weights and thresholds. Specifically, as seasonal variables change (monthly inflow, temperature variations, and plant and animal breeding seasons), the AHP-FAHP weights are updated via a dual-track approach, drawing on both expert knowledge and big data training. On one hand, historical experience is considered; on the other hand, real-time monitoring data is used to correct weight fluctuations through Markov Chain Monte Carlo (MCMC) sampling. In the event of ecological emergencies (sudden pollution, drought, extreme floods), a disturbance response mechanism is introduced to automatically prioritize sensitive ecological targets, achieving adaptive adjustments for extreme scenarios.
[0055] To balance multiple objectives and resource allocation conflicts, Data Envelopment Analysis (DEA) is introduced to evaluate the effectiveness of different scheduling schemes. DEA constructs decision units (DMUs) to measure the relative efficiency of ecological satisfaction across different watersheds or scheduling schemes. Inputs are defined as adjustable resources (flow rate, clean water replenishment), and outputs are defined as the satisfaction of each ecological objective. The DEA model takes the following form: in , The output and input variables are weighted separately, and the model solves for the optimal ratio to determine the relative efficiency of the current ecological goal, thus providing a quantitative basis for dynamic resource reallocation and priority scheduling. To enhance the robustness of the model, the entropy weight method is used to fine-tune the DEA boundary scheme, and fuzzy DEA is used for special scenarios to adapt to the uncertainty of input and output.
[0056] Step 3: Watershed Process Simulation and Causal Reasoning Analysis Based on a clear understanding of ecological needs, a watershed process intelligent simulation algorithm based on graph neural networks (GNNs) is introduced to predict the spatiotemporal evolution of hydrodynamics, water quality, and biological responses in various ecological units of the downstream watershed under multiple inputs (inflow, rainfall, and human scheduling behavior). Simultaneously, by combining causal reasoning networks, the main contribution of scheduling measures to changes in key ecological indicators is explored, clarifying the causal link between scheduling and ecological response, and providing strong information for the scientific assessment of scheduling efficiency.
[0057] The core task of watershed process simulation and causal reasoning analysis is to realistically reproduce the spatiotemporal evolution of hydrodynamics, water quality, and ecosystems within a watershed under complex natural and anthropogenic disturbances, and further reveal the main causes of ecological effects after the intervention of different scheduling strategies. Traditional physical models usually rely on a large number of parameter calibrations and prior assumptions, while this method adopts a watershed intelligent simulation architecture based on graph neural networks (GNNs), which is naturally adapted to the network structure characteristics of watersheds and the complex interactions of multiple nodes and multiple sides.
[0058] The watershed is discretized into nodes according to spatial units (such as sub-watersheds, water patches, core wetlands, and river segment nodes). Rivers, water bodies, and pipelines connecting nodes are represented as directed edges of the graph. The whole is abstracted into a dynamic diagram. It not only depicts the spatial pattern but also reflects the ever-changing processes under hydrodynamic scheduling or extreme weather conditions. Each node records multi-source input characteristics: real-time inflow. Precipitation (P), evapotranspiration (ET), and dispatching flow The monitored water level (H), habitat species abundance (B), and other variables were normalized and merged into a node attribute tensor. Edges between nodes can represent dynamic processes such as flow transmission and pollutant migration, and weights are defined. It reflects the impact of factors such as flow velocity, flow direction, distance, and control thresholds on downstream areas.
[0059] The GNN simulation model uses temporal graph convolutional units as its core, combined with gated recurrent units (such as GRU / LSTM) to capture spatiotemporal dynamics. At each time step, the state propagation process at node \(i\) is given by the following update formula: in It is the potential state vector of node i. Represents all adjacent nodes. The feature transformation function (which can be fully connected or convolutional) is used. The activation function (ReLU) uses trainable parameters U and B. The input feature set is dynamically fused with neighborhood propagation, capturing the nonlinear feedback of hydrodynamic transport, pollution diffusion, and ecological response under complex spatial and scheduling operations. Multi-layer stacking enables downstream nodes to perceive upstream cascading effects and adapt to the dynamic resetting of the watershed scheduling diagram structure (such as adding artificial diversion points or merging points).
[0060] The GNN simulation output is a sequence of ecophysical and hydrological prediction values for each node, such as the water level within the next T steps. Dissolved oxygen concentration, benthic organism density, etc., are used to support subsequent evaluation and optimization.
[0061] To identify the primary causes of ecological responses to scheduling measures, it is necessary to move beyond correlation and engage in causal reasoning. This paper introduces a combination of a structured causal model (SCM) and a causal graph network (CGN). A causal directed acyclic graph (DAG) is constructed, where variable nodes include key scheduling behaviors (such as "daily water release from a gate" or "water replenishment process of a reservoir"), meteorological inputs (such as extreme rainfall), watershed responses (such as "number of days with acceptable flow velocity in a certain section"), and ecological indicators (such as fish spawning success rate and biodiversity index).
[0062] In SCM, each variable Consists of all its direct "parent nodes" function and external noise Expression, that is By jointly fitting all variables of the DAG, a complete system causal dependency structure is formed. CGN explicitly constructs causal migration paths for each node and scheduling decision point in the watershed. By inserting observations or setting "do-intervention" operations, it simulates the counterfactual distribution of "if a certain scheduling method is adopted, the ecological indicators will change accordingly." That is, it applies Do-Calculus to calculate the distribution of the target ecological indicators under human intervention (such as implementing a specific scheduling for a node). Where E represents the ecological response variable, and A represents the target scheduling operation. To quantify the contribution of the principal cause, the "Average Causal Effect" (ACE) is commonly used. \] The model infers the net effect of scheduling on key downstream indicators under measured and counterfactual scenarios.
[0063] Model training employs a multi-task loss aggregation approach: the main loss is the mean squared error (MSE) of predictions for all ecological and hydrological indicators, supplemented by edge weights and structure learning losses from causal structure inference. End-to-end gradient training is conducted on actual historical scheduling scenarios and simulation experiments. Batch historical data, ground-based automatic observations, and remote sensing change homotopy are fed into the model, and the model's resolution and accuracy are continuously improved through cross-validation and expert correction.
[0064] Step 4: Optimize the ecological scheduling plan Finally, an adaptive ecological scheduling optimization method integrating reinforcement learning and evolutionary algorithms (multi-objective genetic algorithm + deep reinforcement learning DRL) is proposed. Using the ecological demands and prediction results obtained in steps two and three as objectives and constraints, reservoir scheduling operation parameters and discharge flow processes are treated as action variables. Through continuous interactive learning with historical actual operating conditions, the method seeks a scheduling strategy that optimally satisfies the key ecological demands downstream. Simultaneously, this optimization incorporates risk warning and socio-economic constraints, achieving an optimal scheduling scheme outputting a multi-dimensional balance between ecology, society, and economy.
[0065] Intelligent optimization of ecological scheduling aims to achieve an efficient balance of ecological, social, and economic objectives under multi-objective and multi-constraint scenarios, in dynamic environments, and under uncertain risk conditions. This process uses the aforementioned ecological demand model and watershed process simulation system as inputs, treating high-dimensional coherent control variables such as scheduling operation parameters and downstream flow processes as decision-making actions. An adaptive optimization framework is designed that integrates multi-objective genetic algorithms (such as NSGA-II or MOEA / D) with deep reinforcement learning (DRL, typically DDPG, SAC, etc.). This decoupled-cooperative strategy combines global search capabilities with the advantages of dynamic intelligent learning, adapting to the highly nonlinear and complex constraints of actual hydrological regulation.
[0066] The optimization problem is modeled as an interactive system of state-action-reward. The entire watershed scheduling state space... The state vector integrates reservoir water level, inflow, system water demand, current ecological indicators, and risk warning signals. Covering key hydrodynamic and ecological elements; space for action Includes operational parameters such as discharge flow rate, gate regulation, water replenishment, and flow restriction at various time periods, written as... Based on the ecological demand and forecast indicators from steps two and three, the objective function of intelligent optimization should maximize the satisfaction of ecological demand at multiple downstream points and ensure water storage safety, while also taking into account socio-economic water use and risk costs.
[0067] A composite multi-objective function is defined as follows: in, Describes the rate at which demand for the j-th type of ecosystem service / species / wetland is met. Characterizing socioeconomic indicators (such as agricultural and urban water supply efficiency). For dynamic weights, This is the risk loss function (such as the expected loss from scheduling failures, floods, and extreme ecological events). This is the risk penalty coefficient. The scheduling strategy to be optimized.
[0068] The multi-objective genetic algorithm first encodes the entire action sequence (such as a monthly / quarterly outflow series) into chromosomes. After initializing the population, genetic operations such as selection, crossover, and mutation are used to iteratively evolve and screen scheduling candidates that constitute a Pareto optimal front across multiple ecological, social, and risk objectives. For example, based on NSGA-II, the multi-objective fitness of individuals is evaluated, and fast non-dominated sorting and crowding distance are used to maintain population diversity. A simulator (GNN + causal inference) performs batch rolling evaluations of each candidate scheme to screen out the optimal or "satisfactory" scheme family among ecological needs, economic benefits, and regulatory safety.
[0069] Genetic algorithms can find the distribution of solutions through global search, but they often struggle to adapt precisely to the continuously changing hydrological and ecological inputs in complex dynamic regulation. Deep reinforcement learning is introduced to supplement the "rapid adaptation and autonomous learning" capabilities. With agent-environment feedback as the core, state transitions rely on the actual watershed hydrodynamic response and the output of an ecological simulator. The agent is based on a policy network. Select an action to receive an instant reward. The actor-critic network (DDPG) dynamically optimizes a strategy based on empirical replay and soft-target entropy regularization in a high-dimensional continuous action space. The critic network minimizes the Bellman error. The actor network improves overall ecological and social benefits through policy gradient ascent, and the policy update formula is: After generating the optimal initial scheduling solution in each round, the genetic algorithm can be used to expand the experience pool of the reinforcement learning agent and accelerate policy training. Conversely, the efficient policies obtained by the DRL agent in extreme situations can also be used to enrich the mutation / optimization mechanism of the genetic algorithm. The two work together dynamically and complement each other.
[0070] The optimized framework tightly integrates risk warnings and socio-economic constraints. When indicator thresholds trigger downstream ecological and social red lines (water levels, pollutant exceedances, drought, etc.), mandatory inclusion of these constraints is required. The filtering process dynamically adjusts the model's weights to correlate economic benefits with ecological dividends. Simultaneously, the risk loss function, using extreme event probability distribution predictions from the simulator, considers not only the mean target but also strengthens the control over extremely low-probability risks.
[0071] Those skilled in the art will understand that the above embodiments can be modified in form and detail in practical applications without departing from the spirit and scope of the invention.
Claims
1. A method for ecological scheduling assessment and optimization in downstream watersheds of water conservancy and hydropower projects, characterized in that, Includes the following steps: S1. Intelligent identification and zoning of ecologically sensitive areas: By integrating multi-source remote sensing datasets and image segmentation algorithms, the automatic extraction and spatial zoning of ecologically sensitive areas across the entire downstream watershed is completed. S2. Multi-objective dynamic modeling of ecological demand: Based on the zoning results of ecologically sensitive areas output in step S1, a multi-objective dynamic ecological demand model is constructed. The weights of various ecological functions are set by fuzzy comprehensive evaluation coupled hierarchical analysis method. The ecological demand of the whole region is iteratively updated based on real-time hydrological data of rainfall, runoff, and seasonal time series. S3. Watershed process simulation and causal reasoning analysis: Construct a watershed process simulation model based on graph neural network, and output the spatiotemporal evolution prediction sequence of hydrodynamic, water quality and aquatic biological response of each downstream ecological unit; build a structured causal model to identify the causal dependency links between scheduling intervention behavior and various ecological indicators. S4. Adaptive optimization of ecological scheduling scheme: The multi-objective ecological demand output in step S2 and the watershed ecological simulation prediction results output in step S3 are used as optimization objectives and constraint boundaries. The reservoir scheduling operation parameters and the time-period discharge flow process are set as decision action variables. The algorithm integrates reinforcement learning and evolutionary algorithms to iteratively solve the optimal scheduling strategy. The strategy solving stage incorporates risk warning thresholds and socio-economic water use constraints to complete the scheme selection.
2. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 1, characterized in that, Step S1 includes the following steps: S11. Simultaneously acquire high-resolution satellite remote sensing imagery and UAV aerial ground imagery, and uniformly perform spatial positioning and geometric registration on both types of imagery; integrate multiple environmental data such as surface reflectance, multi-temporal vegetation index, water index, and surface temperature to construct a multimodal spatiotemporal three-dimensional tensor group. ,in For regional height, For area width, For multi-source feature dimensions, For the time dimension; collect data from biodiversity survey sites in the watershed, and construct an image segmentation model to train a supervised label set; S12. A spatiotemporal embedding U-Net variant model is adopted as the image segmentation backbone network. The backbone network has a built-in spatiotemporal attention module and a multi-scale feature fusion module. The spatiotemporal attention module calculates the similarity weights of features across time phases. The calculation formula is: In the formula, For encoder number Layer, Time ,Location eigenvectors, This refers to the vector dot product operation. S13. Based on the Gaussian kernel density estimation coupled spatial adaptive threshold method, the pixel sensitivity probability map output by the model is divided into regions. The regions obtained are fixedly divided into three categories: core protection zone, transition buffer zone, and general zone. S14. Introduce graph convolution algorithm, and combine the topological relationship of watershed river network system, topography elevation, and ecological connectivity channels to perform post-processing optimization on the partition vector boundary to eliminate broken and fragmented patches.
3. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 2, characterized in that, In step S12, the spatiotemporal embedding U-Net variant model employs a dual-branch composite loss function, which is composed of a weighted superposition of a boundary-aware Dice loss branch and a pixel-level cross-entropy loss branch. The formula for calculating the boundary-aware Dice loss is as follows: In the formula, For the first One pixel predicted value, For the first The actual labeled value of each pixel. To prevent the denominator from being zero, a constant, This represents the total number of pixels in the image.
4. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 1, characterized in that, Step S2 includes the following steps: S21. Construct a hierarchical framework for hierarchical analysis, which is divided into three layers from top to bottom: target layer, criteria layer, and scheme layer. The criteria layer is fixed with four categories of criteria: species habitat, wetland function, ecosystem services, and human water use. An expert scoring method is used to construct a pairwise judgment matrix for the criterion layer. After consistency verification, the matrix is normalized and solved to obtain the static basic weights of each criterion. S22. Construct a triangular fuzzy membership function for each type of ecological unit. Binding suitable hydrological threshold ranges for each ecological indicator The piecewise expression for the triangular fuzzy membership function is: In the formula, These are real-time hydrological index observations. For the first The lower limit threshold of each ecological indicator For the first The upper limit threshold for each ecological indicator; time Next Ecological target demand The calculation formula is: In the formula, For the first The weight of each ecological goal criterion The total number of watershed zones, For the first Partition to the first Contribution weight to ecological goals for Time of the first Partition corresponding to the first The hydrological state variables of the target For the corresponding triangular fuzzy membership function; S23. Introduce a data envelopment analysis model, treating each watershed zone and each preset scheduling scheme as an independent decision-making unit, and calculate the relative input-output efficiency of ecological goal satisfaction for different units; the data envelopment analysis objective function is: In the formula, As the weight of ecological output indicators, As the weight of water resource input indicators, To achieve ecological goals, For adjustable water resource input, the basic weights are corrected using the entropy weight method, and under extreme hydrological scenarios, fuzzy data envelopment analysis is switched to complete adaptive parameter correction.
5. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 1, characterized in that, Step S3 includes the following steps: S31. Discretize the downstream watershed into graph nodes according to sub-river segments, wetland patches, and sub-watershed units. Use natural river channels, artificial water conveyance pipelines, and water body connectivity sections as directed edges between nodes to construct a watershed topology graph that is dynamically updated with hydrological time series. ; S32. Collect real-time inflow, regional precipitation, watershed evapotranspiration, reservoir discharge, river level, and regional species abundance data from multiple dimensions. After standardization and normalization, these data are spliced together to form attribute tensors specific to each node. ; S33. A watershed simulation network is constructed using time-series graph convolutional units superimposed with gated recursive units. GRU or LSTM are selected as the gated recursive units. The network node state update formula is: In the formula, for Time Node Latent state vector, For nodes The set of all adjacent nodes, for Time Node To node edge weights, For the characteristic transformation function, It is the ReLU activation function. The network can be trained parameters; the feature space interaction and temporal state update of the neighboring nodes are completed layer by layer, and the predicted sequence of water level, dissolved oxygen and benthic organism density for future multi-step length is output. S34. Construct a causal directed acyclic graph, where the variable nodes include reservoir scheduling behavior, meteorological driving inputs, watershed hydrological response, and ecological evaluation indicators; perform counterfactual reasoning based on do-interventional factors, and calculate the average causal effect (ACE) of scheduling operations on the target ecological indicator using the following formula: In the formula, As an ecological response indicator, For scheduling operations, a binary intervention variable, To estimate the causal interference, we quantify the contribution percentage of each variable.
6. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 1, characterized in that, Step S4 includes the following steps: S41. Encode the flow rate and gate opening parameters for each time period within the complete scheduling cycle into decision chromosomes, and use the NSGA-II multi-objective genetic algorithm to perform population iteration, outputting a Pareto optimal scheduling candidate solution set under ecological and economic dimensions. S42. Construct an Actor-Critic architecture deep reinforcement learning network, adding an entropy regularization constraint term to the network; use reservoir water level, real-time inflow, ecological satisfaction, and risk warning signals as a concatenated state vector. A continuous action space is constructed using flow regulation parameters within a continuous interval; the Bellman error loss calculation formula for the Critic network is: The gradient update formula for the Actor network policy is: In the formula, For the current value network, For soft target value networks, This is the total reward value for a single step. As a reward discount factor, For policy networks; S43. Store all candidate scheduling schemes output by the genetic algorithm into the reinforcement learning experience replay pool, and iteratively update the model parameters of the Actor policy network and the Critic evaluation network based on the policy gradient method. S44. The probability distribution of extreme hydrological events, water quality safety threshold, total urban and rural water supply red line, and flood control water level are all embedded as hard constraints into the optimization process. Dispatch strategies that do not meet the constraints are eliminated, and the compliant optimal dispatching scheme is output.
7. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 6, characterized in that, Step S4 involves constructing a multi-objective comprehensive reward function, the expression of which is: In the formula, For the first Ecological goals and benefits For the first Socioeconomic benefits, For dynamic weights, For extreme event risk loss function, As a risk penalty coefficient, The scheduling strategy needs to be optimized.
8. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 5, characterized in that, In step S33, each layer of the temporal graph convolutional unit is set with an independent trainable neighborhood feature transformation weight matrix. After the neighborhood node features are mapped by the weight matrix, they are aggregated to the central node to complete the spatial information transmission.
9. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 6, characterized in that, In step S41, the NSGA-II algorithm maintains population diversity through fast non-dominated sorting and crowding distance calculation.
10. The ecological scheduling assessment and optimization method for downstream basins of water conservancy and hydropower projects as described in claim 1, characterized in that, After the partitioning in step S1 is completed, the incremental transfer learning process is started. The spatiotemporal embedding U-Net variant model is fine-tuned based on the latest remote sensing images and field ecological monitoring sampling data. The model update cycle is synchronized with the collection and update frequency of remote sensing data and field monitoring data.
