Intelligent ecological scheduling rehearsal method for inland river basin integrating scheduling process and ecological process

Abstract

The invention discloses an inland river basin intelligent ecological scheduling rehearsal method fusing a scheduling process and an ecological process. The inland river basin intelligent ecological scheduling rehearsal method comprises the steps of multi-source data acquisition and digital twinborn construction; carrying out reservoir intelligent scheduling reinforcement learning modeling; ecological process lag response modeling is carried out; spatial diffusion modeling of ecological influence; performing cross attention guided ecological response interpolation; and carrying out rehearsal and visual display on the ecological scheduling scheme. According to the method, a hydrological-ecological response modeling mechanism is introduced, and a time-space response relationship between scheduling behaviors such as water level and water volume and ecological indexes such as vegetation indexes and habitat indexes is combined, so that lagging response characteristics of an ecological process to the scheduling behaviors can be quantitatively described, and the defect that a traditional scheduling model is insufficient in ecological expression capability is overcome. A reinforcement learning algorithm is utilized to fuse multi-source data for state perception and strategy iteration, and the scheduling strategy can be dynamically adjusted according to the current hydrological situation and ecological feedback result of the watershed. Compared with static rule type scheduling, the regulation and control efficiency and ecological adaptability are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of smart water conservancy and ecological protection, and particularly relates to a method for smart ecological scheduling and pre-play of inland river basin by integrating scheduling process and ecological process. BACKGROUND

[0002] In recent years, with the increasing intensity of human activities and the deepening of climate change trends, the water resources regulation of inland river basin is facing multiple challenges such as imbalance of ecosystem structure, deterioration of water environment and degradation of biodiversity. The traditional method of basin reservoir scheduling mainly aims at flood control, water supply or power generation, ignoring the sensitive response of ecological process to hydrological regulation, and it is difficult to achieve the coordination and unity of water resources development and utilization and ecological protection.

[0003] Especially in inland river basin, the ecological system has significant nonlinearity, hysteresis and spatial heterogeneity to hydrological process. The ecological effects caused by scheduling behavior often show time-lag response, cumulative effect and regional difference, and traditional hydrological models are difficult to accurately depict the long-term, spatial diffusion effect of scheduling behavior on ecological system. In addition, the current scheduling system is mainly driven by static rules, lacking self-adaptive optimization ability, and difficult to meet the complex and changing ecological regulation requirements.

[0004] The existing scheduling method mainly relies on "hydrological-hydrodynamic model" (such as HEC-RAS, MIKE series), whose core logic is based on the physical simulation of hydrological cycle, that is, the runoff process is determined by scheduling scheme, and the final scheduling decision is obtained by river evolution calculation and water quantity evaluation. The basis of this mode is physical parameters such as water level and flow, but it does not establish a comprehensive evaluation system of ecological health, and it is difficult to reflect the ecological response process, for example, only short-term hydrological changes are evaluated, and long-term ecological processes such as vegetation restoration and species migration are not related. Secondly, the response of ecological process to hydrological change has significant hysteresis, but the existing method is difficult to quantify. For example, the static threshold method cannot reflect the seasonal difference, and the empirical regression model uses NDVI-runoff linear regression, but it cannot capture the multi-time-lag nonlinear response and cannot reflect the spatial heterogeneity of ecological evolution. At the same time, the hydrological model and the ecological model are run independently, and the ecological module is only used for post-processing evaluation, so the scheduling scheme cannot be optimized in real time. Therefore, a more perfect technical method is needed to quantify the ecological hysteresis response and directly map the optimized scheduling scheme to the multi-scale ecological spatio-temporal response process. SUMMARY

[0005] In view of the defects of the prior art, the present application proposes a method for smart ecological scheduling and pre-play by integrating scheduling process and ecological process, which quantifies the spatio-temporal influence of water resources scheduling on ecology by constructing an intelligent scheduling rule extraction model and a spatio-temporal coupling neural network model framework; and uses ecological response computer vision feedback to dynamically render the ecological evolution effect of the scheduling scheme.

[0006] An inland river basin intelligent ecological scheduling pre-play method fusing scheduling process and ecological process, comprising the following steps:

[0007] 1) Multi-source data acquisition and digital twin construction: Obtain and integrate multi-source heterogeneous data including reservoir scheduling data, hydrological and meteorological data, ecological monitoring data, remote sensing images, unmanned aerial vehicle measurement data and geographic information data, and construct an inland river basin digital twin system with virtual-real synchronization and real-time response capability, providing data support and spatial mapping foundation for scheduling and ecological process coupling modeling;

[0008] 2) Reservoir intelligent scheduling reinforcement learning modeling: Based on historical scheduling records, target function setting and water-ecological coupling index, a scheduling decision model based on reinforcement learning is constructed, and a knowledge augmented network KAN is used to extract interpretable rule strategy, optimizing multi-objective indexes of reservoir capacity utilization rate, ecological water demand satisfaction rate and downstream hydrological connectivity;

[0009] 3) Ecological process lag response modeling: Taking ecological response variables including NDVI, vegetation coverage and biomass index as modeling objects, a hybrid model constructed by bidirectional gate recurrent unit BiGRU and one-dimensional convolutional neural network TCN is used for modeling, and a time series attention mechanism is introduced to identify the lag relationship between ecological response and scheduling behavior, and output the predicted value sequence of ecological variables;

[0010] 4) Spatial diffusion modeling of ecological impact: Based on ecological monitoring points, remote sensing grid units and their hydrological and ecological correlation, a dynamic graph structure is constructed, and a dynamic graph convolutional neural network DGCN is used to model the spatial diffusion process of ecological response, depicting the spatial propagation path and intensity of ecological process in the basin, and realizing the dynamic identification of ecological response propagation path driven by hydrology;

[0011] 5) Cross-attention guided ecological response interpolation: Using the obtained ecological response lag data points, a Kriging interpolation method with spatio-temporal cross-attention mechanism is used to reconstruct the ecological response lag field in the ecological monitoring blank area, realizing the continuous spatial expression of ecological impact state;

[0012] 6) Visualization expression and response feedback of ecological scheduling pre-play: The ecological response results driven by scheduling strategy are input into the digital twin visualization engine, and the ecological regulation impact is visualized by using the ecological adjustable factor driven module and the interpretable ecological response analysis module, including time series response curve, ecological sensitive area identification map and driving contribution heat map, and supporting multi-scheme interactive comparison and feedback correction of scheduling strategy.

[0013] The digital twin platform has dynamic interaction capability with the ecological scheduling model, realizing "scheduling-response-feedback" closed-loop simulation and pre-play.

[0014] Further, the scheduling decision model of step 2) adopts a two-stage training strategy, the first stage is based on historical scheduling records for policy initialization, and the second stage realizes policy update through online simulation feedback.

[0015] Further, the KAN rule extraction network of step 2) realizes the rule-based expression of the relationship between scheduling behavior and ecological indicators by embedding a sparse interpretable layer in the reinforcement learning process.

[0016] Further, the attention mechanism in the ecological lag modeling of step 3) is used to identify the key influence window of hydrological scheduling behavior on ecological indicators at different lag periods.

[0017] Further, the ecological propagation model constructed by the DGCN of step 4) is based on the river network topology within the basin, and considers the dynamic adjustment of the graph structure weight considering seasonal changes and human disturbance.

[0018] Further, the cross-attention mechanism of step 5) fuses the time series features of scheduling behavior and the graph structure features of ecological response, realizing joint expression of space-time features.

[0019] Further, the ecological response lag gradient field of step 5) is generated by a multi-source heterogeneous data-driven Kriging interpolation method, and is used for ecological fragile area identification and ecological risk warning.

[0020] The beneficial effects of the present application are:

[0021] The present application introduces a hydrological-ecological response modeling mechanism, combines the spatio-temporal response relationship between scheduling behavior such as water level and water quantity and ecological indicators such as vegetation index and habitat index, and can quantitatively depict the lag response characteristics of ecological process to scheduling behavior, making up for the defects of insufficient ecological expression ability of traditional scheduling models.

[0022] The present application uses reinforcement learning algorithm to fuse multi-source data for state perception and policy iteration, which can dynamically adjust the scheduling strategy according to the current hydrological situation and ecological feedback results of the basin, has self-adaptive and autonomous optimization ability, and significantly improves the regulation efficiency and ecological adaptability compared with static rule type scheduling.

[0023] The present application uses dynamic graph neural network to mine the time series dependence relationship and spatial propagation characteristics between ecological elements, solves the technical difficulty that the existing method is difficult to express the coupling relationship of multi-factor and multi-scale ecological process, and improves the accuracy and resolution of ecological response prediction.

[0024] The application introduces a cross-attention mechanism to strengthen the correlation information extraction capability between multi-source data, which can effectively extract deep interaction features between scheduling behavior, hydrological driving and ecological response, improve the sensitivity and interpretability of the model to the trend of key variables, and thus realize precise early warning and early intervention of ecological risk. Through the inversion and interpolation analysis of the ecological lag field, the Kriging interpolation is used to draw the lag time contour map to quantify the spatial gradient, and the law of "lagging with spatial distance attenuation" is explicitly expressed, which can fully show the change process of the ecological response caused by scheduling behavior in time and space. Combined with the ecological effect visualization interface, the scientific and intuitive auxiliary decision-making basis is provided for the managers, and the transparency and operability of ecological scheduling are improved. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 : The overall flowchart of the embodiment of the application;

[0026] Figure 2 : The hydrological-ecological response spatio-temporal coupling neural network architecture provided by the embodiment of the application. DETAILED DESCRIPTION

[0027] The technical solutions of the application will be described clearly and completely below with reference to the drawings.

[0028] As shown in Figure 1 , the method comprises the following steps:

[0029] Step 1: Obtain the water resources scheduling information of the inland river basin, including the historical operation data of reservoirs, the measured hydrological and meteorological data of local stations, ecological monitoring data, remote sensing inversion data (such as precipitation, evaporation, soil moisture, vegetation index), demand-side water consumption data and economic and social data of the middle and lower reaches, and other multi-source information, and construct the inland river water resources scheduling information dataset.

[0030] The hydrological and meteorological data mainly include annual monthly runoff data and precipitation and evaporation data, which are obtained from local hydrological monitoring stations or the Hydrology Yearbook. The runoff data can also be obtained from the Information Center of the Ministry of Water Resources, the management bureau of each basin (such as the Hydrology Bureau of the Heihe Water Conservancy Commission), and the hydrological database of the Gansu Provincial Water Resources Department. The precipitation and evaporation data and meteorological data such as air temperature can also be obtained from the China Meteorological Administration.

[0031] The data is preprocessed, mainly including linear interpolation of missing values, and the formula is as follows:

[0032]

[0033] Where (x1, y1) and (x2, y2) are known data points before and after the missing point.

[0034] In addition to linear interpolation, the 3σ rule is used to remove outliers and align the time to UTC.

[0035] Ecological monitoring data can be obtained from global water resources and irrigation data provided by the United Nations Food and Agriculture Organization (FAO) AQUASTAT, or from free and open source satellite remote sensing data platforms such as NASA Earthdata. The Earth Data dataset is built on satellite observations and ground measurements from NASA, and can provide MODIS vegetation index, evapotranspiration data, SMAP soil moisture data, and TRMM / GPM precipitation data through multi-source data fusion technology. The present application can also use

[0036] The normalized difference vegetation index (NDVI) with a resolution of 30 meters is calculated using the red band and near-infrared band data of Landsat 8 / 9 (OLI / TIRS) remote sensing satellites in the Landsat series:

[0037]

[0038] where NIR is the near-infrared band and Red is the red band.

[0039] The 30m resolution vegetation coverage data is then calculated using the pixel bisection model:

[0040]

[0041] where NDVI soil is the bare soil NDVI value, which is the value of the 5% cumulative frequency of NDVI in the study area or the average NDVI value of the non-vegetation coverage area, with a typical range of 0.05-0.20. NDVI veg is the pure vegetation NDVI value, which is the value of the 95% cumulative frequency of NDVI in the study area or the average NDVI value of the dense vegetation area, with a typical range of 0.70-0.95.

[0042] Agricultural water can be calculated based on the 100m resolution evapotranspiration and productivity data in the FAO global agricultural water data:

[0043] Agricultural water consumption = ∑(ET x A)

[0044] where ET is the actual evapotranspiration of the farmland pixel and A is the pixel area.

[0045] Agricultural water consumption = ∑(ET x A)

[0046] Industrial and domestic water can also be obtained from the GDP, population, and industry water intensity data provided by the World Bank database (i.e. the aforementioned economic and social data).

[0047] To realize the multi-source perception, mechanism expression, and visual presentation of the "hydrology-ecology" co-evolution process in inland river basins, this step is based on an integrated space-ground-terrestrial perception network, combining digital elevation model (DEM), multispectral remote sensing images, and in-situ ecological monitoring point data to build a full-factor digital twin platform for the basin.

[0048] 1. Terrain loading and modeling

[0049] Use THREE.TerrainLoader or custom Loader to read DEM data in GeoTIFF format and generate height map textures. Create terrain meshes using THREE.PlaneGeometry or THREE.BufferGeometry, set vertex Z coordinates based on elevation values, and apply grayscale gradient materials. Load satellite images using THREE.TextureLoader as the map attribute of the terrain mesh. Use THREE.MeshLambertMaterial or THREE.MeshStandardMaterial for realistic rendering.

[0050] 2. Three-dimensional modeling of river channels and embedding of engineering facilities

[0051] Parse GeoJSON river data and generate three-dimensional rivers using THREE.Line or THREE.TubeGeometry. Implement river smoothing using THREE.CurvePath, and dynamically adjust line width based on flow size. Pre-fabricate models such as dams and gates using tools like Blender, and export them in glTF format. Load and position them at actual coordinates using GLTFLoader, add click events to display engineering parameters, and complete the modeling of the river channel.

[0052] Step 2: Intelligent reservoir scheduling reinforcement learning modeling

[0053] 1. Multi-objective collaborative optimization problem modeling

[0054] To address the multi-objective collaborative needs of economic and social water supply safety, ecological water demand guarantee, and compliance with scheduling constraints in reservoir scheduling, an intelligent scheduling simulation environment is constructed that integrates complex constraint conditions. This environment is based on mathematical modeling, accurately depicting the dynamic evolution characteristics, multi-objective coupling mechanisms, and operational constraint conditions of the reservoir system, providing a digital experimental platform for intelligent optimization methods such as reinforcement learning.

[0055] In constructing the reinforcement learning model environment, the reservoir dispatch characteristic parameters including the output coefficient of the hydropower station, the water head loss, the flood control storage capacity, the upper and lower limits of water level, and the like need to be accessed, and the typical characteristic curves such as the water level-storage capacity curve and the flow-tail water level curve need to be accessed. For each dispatch period, the selectable water level range, the outflow interval and the corresponding power generation capacity need to be defined, and the outflow of the past N periods is introduced as the state history. The action space is set as the outflow Q out (t) of the current period

[0056] (1) Flood control safety reward function:

[0057]

[0058] Wherein, Q out (t) is the outflow of the reservoir at period t, Q safe (t) is the safe discharge of the river channel determined based on historical flood frequency analysis.

[0059] (2) Ecological guarantee reward function:

[0060]

[0061] Wherein, is the lower limit of the ecological base flow determined based on the needs of wetland maintenance and fish survival, and the application adopts the 90% average flow calculation method in the dry season, that is, according to the frequency result of the long series data, the monthly average flow with a water frequency of 90% in a year is selected as the ecological flow.

[0062] (3) Water supply guarantee reward function:

[0063]

[0064] Wherein, D total (t) is the total water demand for agricultural irrigation and domestic water at period t, Q supply (t) is the actual water supply.

[0065] (4) Hydropower generation reward function:

[0066]

[0067] Wherein, Q max is the maximum design flow of the hydropower station, H max is the maximum design water head of the hydropower station, Q max ·H max

[0068] is the theoretical maximum power generation potential, Q out(t) is the actual discharge of period t, η is the turbine power generation efficiency, and ΔH is the water head difference. Meanwhile, divide by the designed power generation to normalize.

[0069] The comprehensive reward function is:

[0070] Q norm = ω1·F 防洪 + ω2·F 生态 + ω3·F 供水 + ω4·F 发电

[0071] where ω i is the weight coefficient of each target, which can be adjusted according to actual needs.

[0072] State transition and constraint processing:

[0073] The water balance equation is:

[0074] V(t+1) = V(t) + I(t) - Q out (t) - S(t) - E(t)

[0075] where V(t) is the reservoir storage at period t, I(t) is the inflow, S(t) is the abandoned water, and E(t) is the evaporation and seepage loss.

[0076] If the reservoir capacity is out of bounds or the ecological water supply is severely insufficient for consecutive periods, the current simulation is terminated.

[0077] 2. Reinforcement learning scheduling model solution based on policy gradient

[0078] The scheduling policy is optimized using the reinforcement learning method based on policy gradient. The input of the policy network includes the initial water level of each reservoir at the current period, the inflow, and the current time index. The intermediate layer structure is based on a double-layer fully connected network, with ReLU as the activation function, and a normalization layer is inserted before activation to improve convergence speed and training stability.

[0079] After the construction of the scheduling environment and the agent, the policy network interacts with the simulation environment to continuously sample data and update the policy network until the policy converges, and finally the optimal scheduling trajectory that meets the multi-objective requirements is obtained, i.e., the reservoir discharge sequence of each consecutive period.

[0080] 3. Key factor identification and design of interpretable agent

[0081] To improve the model transparency and the explainability of the scheduling strategy, a factor importance analysis model based on random forest is constructed. The input driving factors include monthly average temperature, evapotranspiration, precipitation, upstream inflow, reservoir water level, inflow, discharge, irrigation water demand, ecological water demand, and industrial water, a total of 10 variables. According to the importance ranking, the top 40% variables are selected as the key factors of water storage scheduling.

[0082] Subsequently, the selected key factors and the corresponding scheduling decisions are taken as input samples, and an interpretable neural network (KAN) model is introduced for strategy rule extraction. KAN is different from the traditional MLP structure, which uses a learnable spline function (such as B-spline) instead of a fixed activation function, combines a sparse network structure (narrow hidden layer + learnable grid) to realize parameter compression, and at the same time retains the function visualization capability, so that the learned scheduling rules have a clear mathematical analytical expression, thereby improving the explainability of the model.

[0083] The present application integrates policy gradient optimization algorithm, machine learning feature selection method and KAN model with explainability, constructs a "simulation-learning-analysis" three-in-one intelligent scheduling closed-loop system, balances model performance and result explainability, and significantly improves the intelligentization and decision support capability of reservoir multi-objective scheduling problem.

[0084] Step 3: Ecological process lag response modeling

[0085] This step is based on the reservoir optimization scheduling decision and the environmental covariate, constructs a multi-dimensional driving factor matrix, and integrates time series modeling and explainability mechanism to deeply mine the long-term and short-term lag response relationship of the ecological system to hydrological changes. By introducing the attention enhancement architecture of time convolution network (TCN) and bidirectional gate recurrent unit (BiGRU), the dynamic learning and mechanism quantization of ecological response are realized, and the SHAP value and heat map visualization framework are combined to improve the explainability and transparency of the model.

[0086] (1) Multi-dimensional spatio-temporal driving factor matrix construction

[0087] Firstly, the reservoir optimization scheduling output (such as reservoir discharge sequence) and environmental covariates (such as climate, hydrology, soil, vegetation, etc.) obtained in step 2 are integrated to construct a driving factor input matrix containing time series features. The main operation is as follows:

[0088] Ecological variable selection and covariate construction:

[0089] According to the inland river basin ecological vulnerability evaluation system, typical ecological indicators (such as normalized difference vegetation index NDVI, oasis area, etc.) are selected as the ecological state reference. The Spearman rank correlation coefficient method is used to analyze the correlation between the candidate environmental variables and the ecological indicators, and the variables with significant correlation are reserved as the covariate input. The calculation formula is:

[0090] Variable standardization processing:

[0091]

[0092] In the formula, X is the original data, μ represents the mean of the data set, and σ represents the standard deviation of the data set

[0093] Constructing a three-dimensional tensor:

[0094] The three-dimensional tensor input structure is constructed, and the format is [B, T, F], that is, "batch size x time step x feature number". Wherein, the time step T should be greater than or equal to the maximum lag period of the ecological response, so that the model can capture the complete time sequence dependence.

[0095] (2) Attention enhanced TCN-BiGRU network modeling

[0096] Considering that the ecological response has both long-term accumulation effect of hydrological conditions and short-term driving effect of scheduling behavior, the application proposes an attention enhanced network architecture combining TCN and BiGRU (TCN-BiGRU-Attention) for identifying multi-scale time features of the ecological response. The network structure of the TCN-BiGRU-Attention model is as shown in Figure 2 .

[0097] TCN layer extracts long-term dependence features:

[0098] The original tensor input is first subjected to a time convolution network (TCN) to extract long-term evolution patterns across time scales. The TCN output is a three-dimensional tensor in the form of [B, F, T], which needs to be further adjusted in format to match the input requirements of the BiGRU.

[0099] Channel reconstruction and format adaptation:

[0100] A 1x1 convolution layer is added at the end of the TCN to realize output channel transformation and dimension rearrangement by adjusting the number of convolution kernels, so that the output format is converted to [B, T, F], which can be directly used as the input of the BiGRU module.

[0101] BiGRU layer extracts short-term dynamic features:

[0102] After the input is subjected to a bidirectional gated recurrent unit (BiGRU), a bidirectional hidden state sequence is output:

[0103] H={h1,h2,…,h t}

[0104] The sequence fuses information of the previous and subsequent time steps, and can fully reflect the sensitive response of the system to instantaneous disturbances.

[0105] Attention mechanism fusion and output generation:

[0106] Introduce attention mechanism, calculate the weight contribution of each time step in modeling target, weight coefficient α t The calculation formula is:

[0107] α t = softmax(u T tanh(W·h t +b))

[0108] The attention weighted result is fused into the context vector ccc, and the final prediction result is generated through the fully connected layer.

[0109] Model calibration and performance optimization:

[0110] The particle swarm optimization algorithm (PSO) is used to search and optimize the network hyperparameters (such as hidden dimension, learning rate, convolution kernel width, etc.) globally, to improve the model fitting accuracy and generalization ability.

[0111] (3) Explainability analysis of ecological response mechanism

[0112] To enhance the mechanism transparency and decision explainability of the model, the SHAP analysis method is introduced to quantify and evaluate the contribution of each input variable and its lag time step:

[0113] The SHAP value is used to calculate the marginal effect of each variable on the model output at different time steps;

[0114] The heat map is used to visualize the matrix and show the response strength and lag distribution of ecological response to each factor;

[0115] The key variables and key time windows are located to reveal the causal chain and response rule between hydrology and ecological system.

[0116] Through the above methods, an ecological response analysis framework is formed, which integrates multi-source factor fusion, dynamic response learning and mechanism visualization, significantly improves the model's ability to identify ecological lag effects and its physical explainability, and provides strong technical support for subsequent dispatching strategy evaluation and ecological regulation.

[0117] Step 4: Spatial diffusion modeling of ecological impact

[0118] To further reveal the spatial differentiation of ecological response lag in the inland river basin of Northwest China, a spatial heterogeneous graph convolutional neural network (DGCN) architecture is designed in this step, which integrates dynamic graph structure perception ability. This architecture combines the dynamic node characteristics of hydro-ecosystem, inputs the multi-scale time series features extracted in step 3 as node attributes, and adjusts the connection relationship between nodes in real time through dynamic adjacency matrix update mechanism, to model the spatial diffusion characteristics of ecological response and identify regional heterogeneity.

[0119] I. Graph structure input and data definition

[0120] The input of the graph neural network (GNN) model consists of three parts:

[0121] 1. Graph topology (nodes + edges)

[0122] Nodes represent observation locations in space where ecological response needs to be calculated.

[0123] Edges represent the connection relationship between nodes, which can be determined by geographical distance, hydrological connectivity or ecological connection.

[0124] 2. Node feature vector

[0125] Each node contains the following attribute information:

[0126] The multi-scale time series embedding vector output by the TCN-BiGRU-Attention model in step 3;

[0127] Static hydrological variables, such as annual average precipitation, groundwater depth, and runoff;

[0128] Geospatial information, such as latitude, longitude, and elevation.

[0129] 3. Edge feature vector

[0130] Characterizes the strength and type of connection between two nodes, including:

[0131] Natural water flow direction of the river;

[0132] Permeability of groundwater channel;

[0133] Semantic connectivity based on ecological similarity, etc.

[0134] II. Edge weight calculation and adjacency matrix construction

[0135] Calculate the distance between nodes based on geographical location, using the spherical distance formula:

[0136]

[0137] Where d ij is the spherical distance between node i and node j; r is the radius of the earth; αi , α j are the longitude of i, j points, respectively i , β j are the latitude of i, j points, respectively.

[0138] The weight elements of the adjacency matrix W are defined based on the above distance as follows:

[0139]

[0140] where σ 2 is a scale parameter to control the degree of weight decay; ε is a sparsification threshold to filter out weak connections.

[0141] III. Dynamic graph construction and real-time updating mechanism

[0142] Unlike static graph structure, the present application adopts a graph structure that can be dynamically updated with changes in ecological state (dynamic graph), which can automatically adjust the topological relationship of nodes and edges according to changes in basin ecology and hydrological conditions. Typical trigger conditions for node adjustment include:

[0143] (1) The NDVI value of a certain location is less than the set threshold for 12 consecutive months, which is considered as an ecological degradation area;

[0144] (2) The connection mode of the river system has changed (such as diversion or drying up);

[0145] (3) A new observation point appears or an existing observation point is deleted;

[0146] (4) New water conservancy projects and other human interventions cause changes in system structure.

[0147] Edge features will also change with the change of nodes, and the dynamic updating rules of edge structure include:

[0148] Adjust the water flow connection mode based on the change of node hydrological properties;

[0149] Re-estimate the weight of the edge according to the ecological similarity index;

[0150] Re-screen effective connections according to proximity to ensure the representativeness of the graph structure.

[0151] IV. Real-time updating algorithm of adjacency matrix based on K-NN

[0152] To realize the adaptive evolution of dynamic graph structure, K-Nearest Neighbor algorithm is used to dynamically update the adjacency matrix, and the process is as follows:

[0153] (1) Distance matrix recalculation: calculate the Euclidean distance between all pairs of nodes to generate distance matrix D;

[0154] (2) K-Nearest Neighbor Screening: For each node i, select the nearest K neighbor nodes from the distance matrix D as its connected objects;

[0155] (3) Update Adjacency Matrix Elements: Adjust the adjacency matrix elements according to the K-Nearest Neighbor results, only keep the edges corresponding to the K-Nearest Neighbor relationship, and reweight according to the updated distance between nodes.

[0156] Through the above dynamic graph modeling method, the spatial evolution trend and structural change characteristics of ecological response can be accurately captured, providing support for regional-scale ecological security evaluation and regulation.

[0157] Step 5: Cross-Attention Guided Ecological Response Interpolation

[0158] Based on the completion of time and space dimension ecological response feature extraction, this step further introduces the cross-attention mechanism (Cross-Attention), realizes the deep collaborative modeling of spatio-temporal features, and generates an interactive feature tensor that integrates time dynamics and spatial heterogeneity. Based on this interactive tensor, the Kriging interpolation algorithm is used to construct the ecological response lag contour map, to explicitly depict the spatial distribution delay characteristics of the basin ecological feedback, and to reveal the nonlinear spatio-temporal coupling and lag law between hydrological driving and ecological response.

[0159] I. Spatio-temporal Cross-Attention Feature Fusion

[0160] This step establishes a dynamic interactive mapping mechanism between the extracted time features (output by GRU-Attention) and spatial features (output by graph neural network), generating a learnable spatio-temporal fusion representation. The specific process is as follows:

[0161] (1) Time feature input and Query generation:

[0162] Input the time feature vector T output by GRU-Attention into the linear mapping matrix WQ to generate the Query vector in the time domain:

[0163] Q = W Q T

[0164] (2) Spatial feature input and Key / Value generation:

[0165] Map the spatial feature matrix S output by the GNN module to the Key and Value spaces respectively to get:

[0166] K = W K S

[0167] V = W V S

[0168] (3) Cross-attention weight calculation:

[0169] The time feature Query is matched with the space feature Key by dot product, and the weight matrix is calculated by using the scaled dot product attention mechanism:

[0170]

[0171] In the formula, Q' represents repeating expansion of Q to a matrix consistent with the dimension of K, d k is the dimension of the Key vector.

[0172] (4) Interaction feature tensor output:

[0173] By fusing the attention weight and the Value tensor, the space response representation Ts dominated by time and the time response representation ST dominated by space are obtained respectively:

[0174] T S = A·V

[0175] S T = A T ·V

[0176] Finally, the interaction feature tensor that fuses the time lag and the spatial topology is constructed, providing a basis for subsequent spatial modeling.

[0177] II. Construction of time lag contour map based on Kriging interpolation

[0178] In order to explicitly depict the lag gradient field of ecological response in different regions, the Kriging interpolation method is introduced to perform spatial interpolation analysis on the interaction feature tensor, and the steps are as follows:

[0179] (1) Variance function modeling:

[0180] Firstly, based on the difference of ecological response lag between feature point pairs, the empirical variogram function is constructed:

[0181]

[0182] Where h is the spatial distance, N(h) is the number of sample pairs with distance h, τ i represents the ecological lag value of the i-th spatial node.

[0183] Fit the exponential model:

[0184] γ(h) = c0 + c(1 - e -h / a )

[0185] Where c0 is the base, c is the variance of the range, and a is the range distance.

[0186] (2) Kriging equation system solving

[0187] After obtaining the variogram model, the Kriging interpolation weight coefficient matrix is constructed to solve the lag value of the spatial prediction point. The interpolation results can be used to generate the ecological lag contour map, and further construct the ecological response delay field.

[0188] Through this method, the response time lag distribution characteristics and nonlinear spatial gradient law of the ecosystem to the hydrological regulation behavior at the basin scale can be systematically revealed, which has important application value in river ecological management, regulation strategy optimization and lag risk assessment.

[0189] Step 6: Visualization of ecological regulation pre-visualization and response feedback

[0190] Further introduce dynamic particle system (DPS) and ecological cellular automata model (ECA-M) to simulate the spatial and temporal dynamic process of runoff diffusion and ecological succession, generate the continuous gradient field of ecological response, and realize the whole process visualization modeling and expression of hydrological regulation and ecological feedback.

[0191] 1、Data access and particle emitter mechanism

[0192] JSON format is used for transmission, and time series runoff data is input. New data is received in real time through WebSocket, or REST API is polled every 5 minutes. Then the dynamic particle system is used to simulate the runoff diffusion process, so as to realize the visual real-time display of the water resources regulation process. First, use the position coordinates in riverSections to create particle emitters at key river sections, and then set the particle parameters: the particle size base is set to 0.1, the base color is set to blue (0x0066cc), and the maximum life cycle is set to 300 seconds.

[0193] 2、Flow and flow rate driven mechanism

[0194] The number of particles generated is determined by the current section flow Q t Driven:

[0195]

[0196] Where: Q b is the reference flow, which is 1 m 3 / s here, and ρ is the density coefficient, which is an adjustable parameter.

[0197] Then determine the motion rate of the generated particles according to the flow rate:

[0198]

[0199] Where: Q t : the real-time flow value of the current river section (from the input data), Qmin : Minimum flow supported by the system, Q max : Maximum flow supported by the system, v min : Minimum velocity of particles, 0.1 units / sec, v max : Maximum velocity of particles, 0.5 units / sec.

[0200] 3. Flow constraints and visual enhancements

[0201] Hydrodynamic constraints: The particle constraints of the present invention are divided into hydrodynamic and boundary constraints. The hydrodynamic constraints follow the pre-computed DEM gradient, moving along the river channel slope direction, while applying random perturbations to simulate turbulence:

[0202] v turb = 0.05 x (rand(v) - 0.5)

[0203] Boundary constraints refer to the constraints received by particles when they exceed the boundaries of the river channel: if in the floodplain area, the particle velocity is halved, and if it exceeds the maximum floodplain range, the life cycle is immediately ended.

[0204] Visual enhancement techniques: For different sizes of flow, a color change from blue to red is used. When the flow is greater than 100 cubic meters per second, the particle tail effect is added using THREE.BufferGeometry to generate trajectory lines, and the normal map dynamic intensity is enhanced, activating the water surface ripple Shader. At the same time, LOD control is used to simplify the page: when the view distance from the lens is far, the particle quantity is simplified by 50%; when the distance from the lens is medium, the particle quantity is 80%; when the distance from the lens is close, the particle quantity is 100% rendered.

[0205]

[0206] Finally, the mouse hover displays real-time data of runoff: current position, current flow, and cumulative water volume. A playback control bar is added, and the replenishment scheduling start time is marked as a key event on the time axis.

[0207] The Ecological Cellular Automata (ECA-M) model based on ecological indicators simulates the ecological succession process:

[0208] 1. Discrete modeling of ecological state

[0209] The Time Delay Coupled Neural Network (TD-CNN) model based on runoff data can obtain local NDVI values. The CA model driven by ecological indicators is constructed using NDVI data to simulate riverbank forest succession and wetland expansion. The continuous NDVI values need to be discretized into a limited number of ecological states for CA simulation. The present invention is divided into 4 states, and the threshold is determined by analyzing the distribution of historical NDVI data (such as a histogram), combined with visual interpretation of remote sensing images or known land cover maps. The threshold is determined as follows:

[0210] ​S1: Low vegetation / bare land / water (NDVI < 0.1)

[0211] S2: Grassland / sparse shrubs (0.1 <= NDVI < 0.3)

[0212] S3: Dense shrubs / young forest (0.3 <= NDVI < 0.7)

[0213] S4: Mature forest / dense wetland (NDVI >= 0.7)

[0214] 2. CA model modeling and succession mechanism

[0215] NDVI data is processed into raster data with the same spatial resolution and DEM alignment. According to the NDVI data of the first year, the threshold rule is applied to generate the initial ecological state raster map. The invention simulates how vegetation tends to higher NDVI natural succession over time, but is affected by the current state, neighborhood state and NDVI dynamic change, and may be degraded due to disturbance. Therefore, the natural succession trend is the dominant rule, which means that if the current state is greater than the state threshold, and the duration of the state is greater than or equal to the minimum duration of the next higher state, the state has a probability of transition, otherwise it maintains the status quo. At the same time, if there is degradation caused by disturbance such as drought, then the negative change needs to meet the threshold value Δdeg:

[0216] NDVI t-1 -NDVI t = Δdeg

[0217] The processed NDVI data is brought into the CA model, and the annual state map of the digital encoding grid is output.

[0218] 3. Three-dimensional ecological state rendering and gradient visualization

[0219] Each element in the state array is mapped to the corresponding position on the three-dimensional terrain, and Three.js needs to load the state data by year. According to the state material mapping:

[0220] S1: Water is represented by a blue translucent material; bare land is represented by a sandy / gray material.

[0221] S2: Grassland is represented by a green material

[0222] S3: Young forest is represented by a darker green material, and a sparse low tree model is placed.

[0223] S4: Mature forest is represented by a dark green material, and a dense tall tree model is placed.

[0224] For the cells whose state changes between adjacent years, a blending factor alpha is used to blend the state materials between the two time periods when rendering. For tree models, position and height can be interpolated during the transition years. In addition to using discrete state encoding, NDVI values can also be interpolated directly in the shader to interpolate colors more continuously, which can show vegetation changes more continuously.

[0225] The above technical process can be implemented by an integrated software platform deployed on a computer system supporting WebGL and high-performance graphics rendering. The system supports the following functions: data reading and synchronization; DEM and remote sensing raster processing; particle system simulation and LOD rendering; cellular automata rule analysis and execution; three-dimensional interactive interface and multi-dimensional visualization display.

[0226] The three-dimensional dynamic visualization platform can meet the transition of spatial dimension from traditional two-dimensional static map to three-dimensional digital twin watershed, and the evolution of time dimension from discrete time point screenshot to dynamic time sequence deduction. Combined with active interactive optimization technology, it can dynamically display the processes of riparian vegetation succession and wetland expansion and shrinkage, realize interactive adjustment of scheme parameters and second-level viewing of ecological response, and solve the limitations of using GIS platform to render only the coverage range of river network at a single time point and unable to express the ecological gradual change process.

[0227] In specific implementation, the method proposed by the technical solution of the present application can be automatically run by a person skilled in the art using computer software technology, and the system device for realizing the method, such as a computer readable storage medium storing the corresponding computer program of the technical solution of the present application and a computer device including the running of the corresponding computer program, should also be within the protection scope of the present application.

[0228] It should be understood that the above description of the preferred embodiments is more detailed and should not be considered as limiting the scope of protection of the present application. Those skilled in the art can make substitutions or modifications without departing from the scope of protection claimed by the present application under the inspiration of the present application, and all fall within the protection scope of the present application. The scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for intelligent ecological scheduling simulation in inland river basins that integrates scheduling processes and ecological processes, characterized in that, Includes the following steps: 1) Multi-source data acquisition and digital twin construction: Acquire and integrate multi-source heterogeneous data, including reservoir scheduling data, hydrological and meteorological data, ecological monitoring data, remote sensing images, UAV measured data and geographic information data, and construct an inland river basin digital twin system with virtual and real synchronization and real-time response capabilities, providing data support and spatial mapping foundation for the coupled modeling of scheduling and ecological processes; 2) Reinforcement learning modeling for intelligent reservoir scheduling: Based on historical scheduling records, objective function settings and water-ecological coupling indicators, a scheduling decision model based on reinforcement learning is constructed, and the knowledge-enhanced network KAN is used to extract interpretable rule strategies to optimize multi-objective indicators such as reservoir capacity utilization, ecological water demand satisfaction rate and downstream hydrological connectivity. 3) Ecological process lag response modeling: Taking ecological response variables, including NDVI, vegetation cover and biomass indicators as modeling objects, a hybrid model constructed by bidirectional gated recurrent unit (BiGRU) and one-dimensional convolutional neural network (TCN) is used for modeling, and a temporal attention mechanism is introduced to identify the lag relationship between ecological response and scheduling behavior, and output the predicted value sequence of ecological variables. 4) Spatial diffusion modeling of ecological impacts: Based on ecological monitoring points, remote sensing grid units and their hydrological and ecological correlations, a dynamic graph structure is constructed. Dynamic graph convolutional neural network (DGCN) is used to model the spatial diffusion process of ecological response, characterize the spatial propagation path and intensity of ecological process in the watershed, and realize the dynamic identification of ecological response propagation path driven by hydrology. 5) Cross-attention guided ecological response interpolation: Using the acquired ecological response lag data points, the Kriging interpolation method with spatiotemporal cross-attention mechanism is used to reconstruct the ecological response lag field in the ecological monitoring blank area, so as to realize the continuous spatial expression of the ecological impact state. 6) Visualization and response feedback of ecological scheduling simulation: The ecological response results driven by the scheduling strategy are input into the digital twin visualization engine. The ecological adjustable factor driving module and the interpretable ecological response analysis module are used to visualize the impact of ecological regulation, including time-series response curves, ecological sensitive area identification maps and driving contribution heat maps. It also supports interactive comparison of multiple schemes and feedback correction of scheduling strategies.

2. The intelligent ecological scheduling simulation method for inland river basins that integrates scheduling processes and ecological processes according to claim 1, characterized in that, Step 2) The reinforcement learning scheduling decision model adopts a two-stage training strategy. The first stage initializes the policy based on historical scheduling records, and the second stage updates the policy through online simulation feedback.

3. The intelligent ecological scheduling simulation method for inland river basins that integrates scheduling processes and ecological processes according to claim 1, characterized in that, Step 2) The KAN rule extraction network, by embedding a sparse interpretable layer during reinforcement learning, achieves a rule-based expression of the relationship between scheduling behavior and ecological indicators.

4. The intelligent ecological scheduling simulation method for inland river basins integrating scheduling processes and ecological processes as described in claim 1, characterized in that, Step 3) describes an attention mechanism in ecological lag modeling used to identify key impact windows of hydrological scheduling behavior on ecological indicators at different lag periods.

5. The intelligent ecological scheduling simulation method for inland river basins that integrates scheduling processes and ecological processes according to claim 1, characterized in that, Step 4) The ecological propagation model constructed by DGCN is based on the river network topology within the watershed and considers the dynamic adjustment of graph structure weights due to seasonal changes and human interference.

6. The intelligent ecological scheduling simulation method for inland river basins integrating scheduling processes and ecological processes as described in claim 1, characterized in that, Step 5) describes a cross-attention mechanism that integrates the time-series features of scheduling behavior with the structural features of the ecological response map to achieve joint expression of spatiotemporal features.

7. The intelligent ecological scheduling simulation method for inland river basins integrating scheduling processes and ecological processes as described in claim 1, characterized in that, Step 5) The ecological response hysteresis gradient field is generated using a Kriging interpolation method driven by multi-source heterogeneous data and is used for the identification of ecologically vulnerable areas and early warning of ecological risks.

Images