A deep reservoir ecological regulation method, system and device based on a data-driven model of a coupled physical mechanism

Abstract

A deep reservoir ecological scheduling method, system and device based on a data-driven model of a coupled physical mechanism belong to the technical field of water resources management and environmental protection. First, a physical water temperature model is constructed based on measured data, diversified water temperature change scenarios are generated, and a deep learning model with physical mechanism constraints is constructed. Second, sensitivity analysis is carried out to identify key influencing factors driving water temperature change, a reservoir optimization scheduling model is constructed, and a deep learning model with physical mechanism constraints is coupled to derive a set of scheduling rules under the premise of meeting water temperature targets. Finally, multi-index optimization analysis is carried out. A deep reservoir ecological scheduling system and electronic device are also provided to implement the above deep reservoir ecological scheduling method. The present application can scientifically derive a set of power generation scheduling rules for deep reservoirs, filter out the optimal scheduling scheme with both ecological and economic benefits by introducing a multi-index optimization method, and achieve the overall balance of ecological demand and power generation benefit.

Description

Technical Field

[0001] This invention belongs to the field of water resources management and environmental protection technology, and relates to a method, system and equipment for ecological scheduling of the Shenzhen-Dalian Reservoir based on a data-driven model with coupled physical mechanisms. Specifically, it relates to a method, system and equipment for ecological scheduling of the Shenzhen-Dalian Reservoir based on a deep learning model constrained by physical mechanisms. Background Technology

[0002] In the operation and management of the Shenzhen-Dalian Reservoir Group, water temperature regulation is crucial for maintaining the health of the downstream ecosystem. The impoundment and release methods of the Shenzhen-Dalian Reservoir significantly alter the downstream water temperature process, thereby threatening the reproduction and propagation of fish downstream of the dam. To reconcile the conflict between hydropower generation and the downstream ecological water temperature requirements, accurately simulating the spatiotemporal dynamic distribution of water temperature in the Shenzhen-Dalian Reservoir Group has become a key technical challenge.

[0003] Physically driven hydrodynamic-temperature models (such as three-dimensional numerical models) can accurately depict the temperature distribution of water bodies in time and space, and have strong physical explanatory power, especially in the formation and evolution of vertical thermoclines and the combined effects of meteorological and hydrological conditions. For example, He et al. [HEW, LIANJ, DUH, et al. Source tracking and temperature prediction of discharged water in a reservoir based on a 3-D hydro-thermal-tracer model [J]. Journal of Hydro-environment Research, 2018, 20: 9-21] established a three-dimensional water temperature model for the Sanbanxi Reservoir, analyzed the spatiotemporal distribution of water temperature within the reservoir, and realized multi-level water temperature prediction; Jiang et al. [JIANGB, WANGF, NIG. Heating Impact of a Tropical Reservoir Downstream Water Temperature: A Case Study of the Jinghong Dayue Lancang River [J]. Water, 2018, 10(7): 951] established a Delft 3D model for the Jinghong Reservoir, quantified and analyzed the impact of reservoir storage on downstream water temperature. However, these models are computationally intensive and take a long time to run. They need to be called frequently in simulation optimization across multiple scenarios and time periods, which significantly reduces computational efficiency and limits their application in optimization scheduling and real-time decision-making.

[0004] To overcome the aforementioned challenges, deep learning methods, particularly Long Short-Term Memory (LSTM) networks, have been introduced into water temperature prediction. LSTM has a significant advantage in capturing time-series features and can effectively represent the temporal changes in water temperature. For example, Wang et al. [QIU R, WANG Y, RHOADS B, et al. River water temperature forecasting using a deep learning method[J]. Journal of Hydrology, 2021, 595:126016] proposed an LSTM-based method for predicting surface water temperature in large deep reservoirs; Qiu et al. [READ JS, JIAX, WILLARD J, et al. Process-Guided Deep Learning Predictions of Lake Water Temperature[J]. Water Resources Research, 2019, 55(11):9173-9190] explored the potential of LSTM in predicting daily river water temperature and quantifying the temporal changes in thermal conditions caused by climate change and dam construction. Domestic and international scholars have not only made theoretical breakthroughs in the field of water temperature prediction but have also proposed feasible solutions in production practice. For example, Chinese invention patent (application number 202110321843.1) provides a method for predicting ocean surface water temperature. It performs mode decomposition (EMD) on the acquired ocean surface water temperature time series data to obtain intrinsic mode (IMF) components and residuals. Based on the IMF components, residuals, and an LSTM deep learning network, predictions are made and synthesized to obtain the final predicted ocean surface water temperature. Chinese invention patent (application number 202411908372.4) provides a system for rapid short-term prediction of water temperature in large deep reservoirs. It uses principal component analysis to screen and reconstruct the input dataset, trains an LSTM-Transformer network model using the released water temperature data, and optimizes the hyperparameters of the trained model using a Bayesian optimization method to achieve rapid short-term water temperature prediction. While existing technologies can predict surface water temperature or released water temperature, under the complex spatial structure of large and deep reservoirs, relying solely on LSTM cannot account for the spatial distribution details of water temperature. More importantly, due to the lack of inherent constraints from hydrodynamic-thermal processes, purely data-driven predictions may produce results that violate physical laws, such as minimum temperatures below 4°C or discontinuities or reversals in the water temperature-density profile, thereby affecting the accuracy of downstream water temperature predictions and the reliability of ecological scheduling schemes.

[0005] Therefore, it is urgent to introduce physical mechanisms as constraints into deep learning models so that the prediction process follows the basic physical laws of water temperature evolution and leverages the advantages of deep learning models in extracting time patterns. This would enable high-precision spatiotemporal characterization of the water temperature released from the deep reservoir and couple it with the optimization model to support ecological scheduling for the water temperature needs of power generation and downstream fish farming. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a data-driven model based on coupled physical mechanisms for the ecological scheduling of the Shenda Reservoir, including a method, system, and equipment. This invention comprehensively utilizes physical water temperature simulation, deep learning proxy modeling, and optimized scheduling methods to deduce the ecological scheduling strategy for the Shenda Reservoir. It can quickly predict the downstream water temperature and, under the constraint of maximizing the ecological water temperature requirements of the three production areas, deduce the optimal power generation scheduling rules for the Shenda Reservoir. This approach balances the benefits of hydropower generation with the water temperature control requirements of key fish habitats, thereby optimizing the comprehensive management of watershed water resource utilization and ecological environmental protection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A data-driven model based on coupled physical mechanisms proposes an ecological scheduling method for large and deep reservoirs. This method can quickly derive scheduling rules for large and deep reservoirs that balance power generation and downstream fish habitat water temperature. The method includes the following steps:

[0009] The first step, based on measured data from the study area, is to construct a physical water temperature model, and then input data and simulate water temperature, as follows:

[0010] Step 1.1: Collect measured data of the study area, including meteorological data, hydrological data, topographic data, and water temperature monitoring data.

[0011] The meteorological data includes air temperature, solar radiation, wind speed, wind direction, and relative humidity; the hydrological data includes rainfall, runoff, and reservoir operation processes within the basin; the topographic data includes digital elevation model (DEM), catchment area, water flow relationship, and flow velocity distribution; and the water temperature monitoring data includes surface water temperature and vertical temperature profiles at different times and locations.

[0012] Step 1.2: After collecting the measured data in Step 1.1, a preliminary physical water temperature model is constructed and model parameters are set, including water flow velocity. diffusion coefficient Vertical water temperature gradient index Specifically:

[0013] water flow velocity Calculated based on the Navier-Stokes equations; diffusion coefficient Determined based on empirical formulas, measured data, or model calibration; Vertical water temperature gradient index Used to characterize the thermal stratification process.

[0014] Step 1.3: Based on the physical water temperature model initially constructed in Step 1.2, the physical water temperature model is sequentially divided into horizontal grids, discretized into grids, and layered vertically, as follows:

[0015] Step 1.3.1: Select the horizontal grid type, using an unstructured grid, and choose either a triangular or quadrilateral grid to complete the description of the boundary features of the lake, river, or reservoir.

[0016] Step 1.3.2: Perform horizontal mesh generation and set the time step, and set the spatial step. Set time step The mesh is discretized using the finite difference method or the finite volume method, with a time step of [missing information]. The Courant condition must be met:

[0017] (1)

[0018] in, This represents the maximum value of the water flow velocity. By satisfying this condition, numerical oscillations during heat transport can be avoided, ensuring the stability of the physical water temperature model.

[0019] Step 1.3.3 involves vertical stratification, that is, dividing the water into several sections along the depth direction. Layers of equal thickness or layers of uniform thickness, with each layer having a uniform internal temperature distribution. Lakes use layers of equal thickness, while reservoirs use layers of equal thickness. Layered.

[0020] Step 1.4: After completing Step 1.3, set the boundary conditions and initial conditions for the physical water temperature model. Specifically:

[0021] First, the boundary conditions of the physical water temperature model are set: the physical water temperature model needs to define the hydrodynamic boundary and the heat transfer boundary of the external boundary. The hydrodynamic boundary: flow rate or water level boundary conditions are set at the inlet and outlet. The inlet can be input with upstream water flow rate, and the outlet can be set with downstream controlled water level or free outflow conditions. The heat transfer boundary: atmospheric-water interface heat exchange conditions are set at the water surface, including solar shortwave radiation, longwave radiation, latent heat flux, and sensible heat flux. These fluxes are usually driven by meteorological data. Simultaneously, for the bottom boundary of the reservoir, the heat transfer effect of the bottom sediment must also be considered.

[0022] After setting the boundary conditions, the initial conditions for the water temperature physical model are set: at the start of the simulation, the initial water temperature field for the entire study area is given. The initial water temperature is obtained through the following methods: interpolation using measured water temperature data (such as vertical temperature profiles and surface water temperature monitoring values) to obtain the initial distribution; if complete observation data is lacking, the regional average temperature or the historical average temperature for the same period can be used as the initial field.

[0023] At this point, the initial settings for the physical water temperature model have been completed.

[0024] Step 1.5, based on the physical water temperature model set up in steps 1.2 to 1.4, simulates the spatiotemporal variation of water temperature within the study area by solving the water heat transport equation. Specifically:

[0025] Step 1.5.1, the water temperature simulation process adopts... The turbulence model was closed, and calculations were performed using the ADI algorithm based on the finite volume method. The ADI algorithm discretizes the Navier-Stokes equations to obtain a numerical solution for the flow field. Then, calculations were performed based on the heat conduction equation using the finite element method. The heat conduction equation was discretized, and the finite element method was used to numerically solve it, yielding a numerical solution for the water temperature.

[0026] The heat conduction equation used in the water temperature simulation is as follows:

[0027] (2)

[0028] (3)

[0029] in, As an intermediate variable, Density of water; For reference density; , , They represent , , The velocity component in the direction, where Indicates the direction of water flow. Indicates the horizontal and vertical direction of water flow. Indicates the direction perpendicular to the water flow; Represented as time; The height of the free surface above the reference plane (z=0); This refers to the water depth below the reference surface; This refers to water temperature; This refers to the horizontal eddy diffusion coefficient; This refers to the vertical eddy diffusion coefficient; The Lamé coefficient is the coefficient in the direction of water flow. The Lamé coefficient is the ratio of horizontal to vertical water flow direction.

[0030] Step 1.5.2, based on the numerical results of Step 1.5.1, performs dynamic simulation updates: During the operation of the physical water temperature model, the water temperature field is updated at each time step. By calculating the convection and diffusion processes of heat within each grid cell, and superimposing the effects of surface meteorological factors, bottom sediment heat transfer, and other heat source terms, the vertical distribution of water temperature and the overall spatiotemporal variation trend can be gradually obtained.

[0031] Step 1.5.3: Finally, output the simulation results: The simulation results include the temperature distribution of the surface and deep layers of the water, the evolution of the thermal stratification structure, and the temporal changes in water temperature at key observation points. Through visualization methods such as two-dimensional profiles, three-dimensional temperature fields, and time series curves, the spatial distribution and seasonal variation characteristics of water temperature in the study area can be intuitively presented, and the simulation results data can be obtained.

[0032] Step 1.6: To ensure the physical water temperature model accurately reflects the actual water temperature changes, the simulation results from Step 1.5.3 are compared with the measured data using evaluation indicators to complete the calibration and verification of the physical water temperature model. Specifically:

[0033] The evaluation metrics include Nash curve efficiency (NSE), percentage bias (PBIAS), and average relative bias (ARD%), and the evaluation criteria are as follows:

[0034] When NSE > 0.85, PBIAS < 30%, and ARD < 25% simultaneously, the physical water temperature model is considered to have good performance and can reliably predict water temperature in practical applications. The physical water temperature model is then validated. Validation of the physical water temperature model involves applying the calibrated model to different time periods and regions to verify its accuracy under different scenarios. During validation, new measured data is compared with the simulation results output by the physical water temperature model to ensure that the model has high predictive power and reliability.

[0035] If the conditions are not met simultaneously, return to step 1.2, adjust the parameters of the physical water temperature model, and then repeat the subsequent steps until the evaluation index conditions are met simultaneously.

[0036] The second step involves using the physical water temperature model built in the first step to construct and run simulation scenarios based on collected measured data, generating diverse water temperature change scenarios as training datasets for the deep learning model. Specifically:

[0037] Step 2.1: Determine the uncertainty of the scenario based on the measured data:

[0038] To build rich training data for deep learning, based on the physical water temperature model constructed in the first step, measured data covering multiple years were used to simultaneously reflect input uncertainty and situation uncertainty.

[0039] The simulation scenarios include: long-term meteorological series: hourly / daily sequences of air temperature, shortwave / longwave radiation, wind speed / direction, relative humidity, cloud cover, precipitation, etc. Long-term hydrological series: continuous sequences of inflow / outflow processes, water level, flow velocity, evaporation, inflow water temperature, etc.; operational scheduling scenarios (different release rules, different water replenishment strategies) can be added. External heat flux and optical scenarios: surface atmosphere-water interface heat exchange terms (sensible heat / latent heat / net radiation), bottom sediment heat transfer, inflow heat flux, nearshore shading, light attenuation coefficient (affected by turbidity / suspended sediment), etc.; scenario-based perturbations can be applied to key parameters (e.g., decreased transparency, increased solar radiation). The key objective is to calculate the water temperature field using different year and scenario combinations. The spatiotemporal evolution (including thermal stratification) provides diverse and generalizable training samples for deep learning models.

[0040] Step 2.2: Then, using the physical water temperature model completed in Step 1, simulate and output time series data and spatial series data:

[0041] In each long-series scenario, the physical water temperature model is driven by the multi-dimensional long-series features composed of meteorological, hydrological, and heat flux data to solve the water temperature transport equation and output temperature simulation results data under temporal and spatial variations.

[0042] The time-series data includes: continuous sequences of water temperature at key locations (buoy points, inlets / outlets, representative cross-sections) over time, as well as temporal changes in diagnostic parameters such as thermal stratification intensity and thermocline depth. The spatial sequence data includes: temperature distribution in different grid cells / depth layers, two-dimensional horizontal distribution and vertical profile at typical moments, and the evolution of the thermal stratification structure (temperature and thickness of the surface / thermocline / bottom layers).

[0043] The input data and output are matched according to the year, disturbance type, and parameter set to form paired result data of input feature and target variable.

[0044] Step 2.3: Organize and divide the output data from Step 2.2 to form a training dataset:

[0045] Step 2.3.1: To ensure the breadth and representativeness of the deep learning model training, the dataset is divided into three parts:

[0046] Training set: Used to train deep learning models, it contains a large amount of data and covers a variety of different simulation scenarios;

[0047] Validation set: Used to validate the model's performance during training and to help tune parameters;

[0048] Test set: Used to evaluate the model's prediction accuracy and ensure that the model has good generalization ability.

[0049] The data ratios of the training set, validation set, and test set are 60%, 20%, and 20%, respectively.

[0050] Step 2.3.2, Normalization and Standardization: To eliminate the dimensional differences between different input features, the simulation results of the physical water temperature model are preprocessed after the training dataset is partitioned, including normalization and standardization. This ensures that the input data is on a similar scale, which helps improve the training efficiency of subsequent deep learning models. The normalization formula is:

[0051] (4)

[0052] In the formula, and The minimum and maximum values ​​of the input data.

[0053] The third step, based on the training dataset obtained in the second step, is to construct a deep learning model constrained by physical mechanisms. Specifically:

[0054] Step 3.1: After obtaining the training dataset, the process moves to the deep learning model construction phase. The following are the detailed steps for constructing the deep learning model:

[0055] Step 3.1.1, Design the deep learning model architecture:

[0056] Input layer design: The input layer of a deep learning model receives multiple long-sequence features, including: meteorological factors: air temperature, solar radiation, wind speed, wind direction, relative humidity; hydrological features: inflow rate, lake / reservoir water level, water flow velocity, inflow water temperature; external heat flux and optical parameters: air-water interface heat exchange terms (sensible heat, latent heat, net radiation), substrate heat transfer, etc. The number of these input feature dimensions corresponds to the number of neurons.

[0057] Hidden layer design: ReLU (RectifiedLinearUnit) is used as the activation function to alleviate the gradient vanishing problem and improve the convergence speed.

[0058] Output layer design: The output layer of the deep learning model consists of multiple neurons used to predict water temperature-related indicators. Based on the research objective, it is divided into: single-point prediction: outputting the water temperature change over time at a representative location; profile prediction: outputting the temperature distribution across the entire vertical temperature profile; indicator prediction: outputting the intensity of the thermocline, the depth of the thermocline, or the surface-to-bottom temperature difference. ).

[0059] Step 3.1.2, design the loss function for the physical mechanism constraining the neural network in the deep learning model:

[0060] The most crucial aspect of deep learning model architecture is determining the loss function design: incorporating physical knowledge as a loss function into the training of the neural network. This invention uses density as a bridge, designing key physical relationships (mathematical equations) between water temperature, density, and depth for reservoir water temperature simulation as constraints on the loss function. These physical relationships include ① the water temperature-density relationship and ② the density-depth relationship, as detailed below:

[0061] ① Water temperature-density relationship;

[0062] water temperature and density The following physical equations are known to be nonlinearly related to each other:

[0063] (5)

[0064] in, Indicates water temperature. Indicates density;

[0065] ②Density-depth relationship;

[0066] In the vertical structure of water bodies such as reservoirs, the density of water generally increases monotonically with depth; that is, the density of water is greater closer to the bottom. Assuming there are two water bodies, 1 and 2, in the vertical direction, this physical law can be described by the following inequality:

[0067] (6)

[0068] in, Indicates depth and time The water density at that location; Indicates depth and time The water density at that location; Indicates the depth of water body 1; This indicates the depth of water body 2.

[0069] When depth Less than At that time, that is Compare As you get closer to the surface, the density of the water should not be greater than the density at deeper depths. This reflects the physical property that density increases with depth.

[0070] However, when predicting temperature and then converting it to density using a deep learning model, the model may violate monotonicity constraints. For example, due to prediction errors, the deep learning model might give unphysical results for certain depth pairs, such as shallow water having a higher density than deep water. To incorporate physical laws into the training of deep learning models, a loss function based on physical constraints needs to be constructed to penalize these violations.

[0071] Therefore, we should consider Depth value and An unlabeled dataset of input features on a regular grid with a time step. For any pair of consecutive depth values... and Above, the deep learning model is computed at time step. The differences in density estimates are shown below:

[0072] (7)

[0073] in, This indicates a violation of the rules regarding depth. and time The physical equations; This indicates the depth and time The density value; This indicates the depth and time The density value.

[0074] This can be evaluated as The non-zero occurrence of . Therefore, this invention can treat the average of all physical violations for each consecutive depth pair and time step as a physics-based loss function:

[0075] (8)

[0076] in, This represents the cumulative loss under physical constraints; Indicates the water temperature of a specific body of water; Indicates the number of neurons over time; Indicates the number of deep neurons;

[0077] Based on physical loss, a deep learning model constrained by physical mechanisms was realized. In particular, in the special problem of water temperature stratification modeling in reservoirs, by introducing physical mechanism-based constraints into the loss function, the predicted water temperature results are kept consistent with key physical variables such as density and depth. This ensures that the output of the deep learning model conforms to the basic laws of thermodynamics and fluid mechanics, and obtains a physically reasonable solution.

[0078] Step 3.1.3: Train the deep learning model based on the already constructed model:

[0079] Training dataset: Using the training set from step 2.3, the input features are meteorological data, hydrological data, and water temperature input data, the target variable is the water temperature change process (such as surface water temperature, vertical water temperature, etc.), and the loss function is the mean squared error (MSE). The difference between the predicted value and the true value of the deep learning model is calculated.

[0080] Model performance optimization: The Adam optimizer is employed. This optimization algorithm combines adaptive learning rate and momentum methods to improve the efficiency and convergence speed of deep learning model training. The Adam algorithm updates the parameters of each layer according to the following update rules:

[0081] (9)

[0082] in , They represent the times as and The learning value, For learning rate, and For estimates of the first and second moments, To prevent division by zero of constants.

[0083] Regularization: To avoid overfitting, regularization techniques, also known as weight decay, are used. This involves adding a regularization term to the loss function to control model complexity.

[0084] (10)

[0085] in This represents the loss of the entire function. Indicates the average error. The regularization coefficient is . These are the weight parameters in the model.

[0086] Training strategy: An early stopping strategy is employed to prevent overfitting. Training is stopped early when performance on the validation set no longer improves. This strategy effectively saves training time while avoiding overfitting to the training data.

[0087] Step 3.2, Validation and Evaluation of Deep Learning Model After Training:

[0088] To ensure the accuracy and generalization ability of the deep learning model, cross-validation is used to ensure its stability and reliability. Specifically:

[0089] Cross-validation: This method uses K-fold cross-validation to evaluate deep learning models. The data is divided into K subsets, and one subset is used as the validation set each time, while the remaining subsets are used for training. Finally, the average performance of the deep learning model is evaluated. This method effectively reduces the random errors introduced by data partitioning.

[0090] The evaluation metrics primarily use root mean square error (RMSE) and non-mean square error (NSE). A deep learning model is considered to have a good fit and can meet the requirements of subsequent simulations when both RMSE and NSE are less than 1 and greater than 0.85 simultaneously. If both conditions are not met, the process returns to step two, increasing the amount of training data required for the deep learning model.

[0091] The fourth step involves conducting sensitivity analysis based on the deep learning model built in the third step. This systematically identifies the key influencing factors driving changes in the temperature of the water released from the Shenda Reservoir, as follows:

[0092] First, the SHAP value is calculated using the Shapley additive interpretation method based on game theory to calculate the marginal contribution of the input features. Then, statistical analysis of the SHAP values ​​is performed to obtain the contribution distribution of each input feature, generating a SHAP summary plot to demonstrate the degree of influence of each input feature on the deep learning model's prediction of the released water temperature. Finally, based on the SHAP results, key sensitive factors are identified, and input features with higher SHAP values ​​are selected as key sensitive factors to guide subsequent reservoir ecological scheduling decisions.

[0093] The fifth step involves constructing a reservoir optimization scheduling model based on the key sensitive factors from the fourth step, and coupling it with the deep learning model from the third step. This process, while satisfying the water temperature target, derives a suitable set of scheduling rules. Specifically:

[0094] Step 5.1, first, define the framework for the optimized scheduling model:

[0095] To improve optimization efficiency and avoid the high computational overhead caused by repeated running of the physical water temperature model, this invention introduces a pre-built deep learning model into the outflow water temperature prediction to replace the original physical water temperature model and quickly simulate the water temperature process under different scheduling schemes. The prediction results are directly coupled with the optimization scheduling model, thereby efficiently calculating statistical indicators related to outflow water temperature in each iteration.

[0096] Step 5.2: After setting the framework in Step 5.1, set the optimization objectives and determine the decision variables:

[0097] Determine the power generation target: Maximize the multi-year average power generation of the Shenda Reservoir as the power generation efficiency optimization target, and provide an economic evaluation basis for subsequent dispatching schemes.

[0098] (11)

[0099] in, This represents the total amount of electricity generated. Indicates the total number of scheduling periods; Indicates that the reservoir is at time Hydropower generation; Indicates the time step;

[0100] Determine the target for downstream water temperature: The ecological optimization target is the rate at which the downstream water temperature meets the suitable temperature range for downstream fish habitats. This rate is defined as the proportion of time periods during which the downstream water temperature is within the suitable temperature range out of the total number of scheduling periods. This target quantitatively reflects the effectiveness of the scheduling scheme in ensuring ecological water temperature; a higher target is better.

[0101] (12)

[0102] in, Downflow water temperature guarantee rate (%) Indicates the total number of scheduling periods; Indicates time as The temperature of the discharged water; , These represent the lower and upper limits of water temperature suitable for fish habitats, respectively; a higher guarantee rate indicates that the scheduling plan is more conducive to ecological water temperature protection.

[0103] Decision variables: In order to ensure that the predetermined ecological and power generation goals can be achieved under different water inflow scenarios, it is necessary to optimize the reservoir power generation scheduling rules to make them applicable in the long term.

[0104] Step 5.3: After completing Step 5.2, set constraints for the reservoir optimization scheduling model:

[0105] Optimal reservoir scheduling generally follows constraints such as water balance, reservoir capacity, reservoir discharge, power generation flow, reservoir station output, scheduling rule capacity parameter constraints, and scheduling diagram output coefficient constraints. Because this invention needs to take into account the downstream river water temperature and ecological needs, it adds a constraint on the guarantee rate of downstream fish habitat suitable water temperature.

[0106] (1) Water balance constraints;

[0107] (13)

[0108] In the formula, These represent the reservoir at time . and The water storage capacity; , These respectively represent the reservoir at Real-time inbound and outbound flow; Indicates the time step.

[0109] (2) Reservoir capacity constraints;

[0110] (14)

[0111] In the formula Indicates that the reservoir is The amount of water stored at any given time; , These represent the dead storage capacity and the water storage volume corresponding to the normal storage level of the reservoir, respectively.

[0112] (3) Reservoir discharge constraints;

[0113] (15)

[0114] In the formula, Indicates that the reservoir is Real-time outbound flow; Indicates that the reservoir is Power generation flow rate at any given moment; Indicates that the reservoir is The flow rate in the downstream section at any given time.

[0115] (4) Power generation flow constraints;

[0116] (16)

[0117] In the formula, Indicates that the reservoir is Power generation flow rate at any given moment; This indicates the flow rate at the maximum flow capacity of the reservoir.

[0118] (5) Output constraints of hydropower stations;

[0119] (17)

[0120] In the formula, Indicates that the reservoir is Hydropower generation at any given time; This indicates the installed capacity of the reservoir.

[0121] (6) Scheduling rule capacity parameter constraints;

[0122] (18)

[0123] In the formula, , These represent the dead storage capacity and the water storage volume corresponding to the normal storage level of the reservoir, respectively. They represent the reservoir's number Monthly power limit line, increased power line and increase output line The corresponding water storage capacity.

[0124] (7) Output coefficient constraints in scheduling diagram

[0125] (19)

[0126] In the formula, These represent limiting the output line and increasing the output line, respectively. and increase output line The corresponding output coefficient; This indicates the installed capacity of the reservoir; This indicates the guaranteed output of the reservoir.

[0127] Step 5.4, finally, the solution algorithm for the reservoir optimization scheduling model is set:

[0128] A heuristic algorithm is employed to solve the reservoir optimization scheduling model, analyzing a multi-objective, nonlinear, large-scale reservoir optimization scheduling problem while balancing computational efficiency and solution quality. Through multi-objective optimization scheduling solutions, a set of scheduling rules that simultaneously meet the requirements of the discharge water temperature guarantee rate is obtained, providing scientific support for reservoir operation and management and laying the foundation for multi-index optimization.

[0129] The sixth step, based on the set of scheduling rules obtained in the fifth step, further combines the characteristics of downstream water temperature changes, power generation operation characteristics, and seasonal electricity price factors to conduct a multi-indicator optimization analysis. From the perspective of comprehensive ecological and economic benefits, the optimal reservoir scheduling scheme is selected. Specifically:

[0130] Step 6.1, determine the preferred indicators:

[0131] This invention determines the duration of unsuitable water temperature and the depth of damage as water temperature indicators, and determines the power generation efficiency based on utilization hours and seasonal electricity prices as power generation efficiency indicators. The specific formula is as follows:

[0132] Unsuitable water temperature duration This reflects the duration for which the discharged water temperature continuously deviates from the suitable range. The calculation formula is as follows:

[0133] (20)

[0134] in The value range is [0,1], and the closer the value is to 1, the closer the water temperature is to the suitable condition. The duration of continuous deviation of water temperature predicted by the model; This is the target reference value, which is the maximum allowable continuous deviation time in an ecologically acceptable manner.

[0135] Depth of damage This reflects the maximum deviation between the discharged water temperature and the suitable range. The calculation formula is:

[0136] (twenty one)

[0137] in To minimize the impact of water temperature, the value range is [0,1]. The closer the value is to 1, the closer the water temperature is to the suitable range. The predicted deviation of water temperature; The target reference value is the maximum acceptable allowable temperature deviation.

[0138] Hours of use The calculation formula is used to examine how close the unit's operating time is to the theoretical full-capacity utilization hours.

[0139] (twenty two)

[0140] in The utilization hours index has a value range of [0,1]. The closer the value is to 1, the higher the unit utilization efficiency. This represents the actual utilization hours of the generating units under the candidate scheduling scheme. The target utilization hours can usually be set as the midpoint between the upper and lower limits; These represent the theoretically acceptable maximum and minimum number of utilization hours, respectively.

[0141] Power generation efficiency of seasonal electricity prices The objective function is to maximize the power generation benefits of the cascade hydropower stations during the scheduling period. The calculation formula is as follows:

[0142] (twenty three)

[0143] in As an indicator of the efficiency of electricity generation under seasonal electricity prices; Indicates the time step; Indicates the first Hydropower station Year Real-time electricity pricing; Indicates the total number of scheduling periods; Indicates the number of hydroelectric power stations; Indicates the total number of years; Indicates the first Hydropower station output coefficient; Indicates the first Hydropower station Year Real-time power generation flow; Indicates the first Hydropower station Year Water level at all times.

[0144] Step 6.2, Multi-index Optimization Method:

[0145] Based on the set of scheduling rules obtained in step 5, the scheduling process corresponding to each rule is simulated and calculated. Based on this, the water temperature index and power generation benefit index involved in step 6 are calculated. Finally, a weighted objective function is used for evaluation, which measures the performance of each scheduling rule in maintaining the suitability of the downstream water temperature and evaluates its power generation output level under different seasonal electricity price conditions, thereby achieving a comprehensive comparison and optimization of water temperature control and power generation benefits.

[0146] The objective function is: to calculate a weighted comprehensive score based on the water temperature index and power generation efficiency index of all scheduling schemes.

[0147] (twenty four)

[0148] in, The weights are respectively the duration of unsuitable water temperature, depth of damage, number of utilization hours, and the power generation benefits considering seasonal electricity prices; These represent the scores for each indicator, with scores ranging from [0,1]. The duration of unsuitable water temperature is an important factor. To disrupt the depth index, To utilize the hourly index, This is an indicator of the efficiency of electricity generation under seasonal electricity prices.

[0149] By comparing the comprehensive scores of all candidate schemes, the scheme with the highest comprehensive score is selected as the optimal ecological scheduling scheme.

[0150] An ecological scheduling system for the deep reservoir based on a data-driven model with coupled physical mechanisms includes the following modules:

[0151] The physical water temperature model construction module establishes a physical water temperature model based on measured data from the study area. This module covers data input, model parameter setting, mesh generation and vertical layering, boundary and initial condition settings, water temperature simulation, and result output. It supports dynamic model updates and visualization of simulation results, and performs calibration and verification by comparing the model with measured data to ensure that the model accurately reflects the water temperature change process, providing reliable support for subsequent work.

[0152] The module for generating the training dataset outputs time-series and spatial-series temperature results by running a physical water temperature model under different scenarios. The resulting data is then organized and divided into training, validation, and test sets, and preprocessed using normalization and standardization methods to obtain a high-quality dataset suitable for training deep learning models.

[0153] The deep learning model building module completes the construction of a deep learning model constrained by physical mechanisms: model architecture design, setting input layers, hidden layers, and output layers to achieve single-point prediction, profile prediction, and index prediction; the relationship between water temperature and density, and the relationship between density and depth are introduced into the loss function constraint to ensure that the prediction results conform to physical laws; mean square error, Adam optimizer, regularization, and early stopping strategies are used to improve model performance; finally, the deep learning model is verified and evaluated through K-fold cross-validation and indicators such as root mean square error and Nash efficiency coefficient to ensure the model's accuracy and generalization ability.

[0154] The deep learning model training module: Based on the obtained training dataset, it establishes and trains a deep learning model, determines feature importance through SHAP analysis, and ensures the prediction accuracy and interpretability of the deep learning model.

[0155] The reservoir optimization scheduling module: constructs a reservoir optimization scheduling model by coupling a deep learning model, and derives a suitable set of scheduling rules under the premise of meeting the water temperature target.

[0156] The multi-indicator optimization output optimal solution module comprehensively considers the characteristics of downstream water temperature changes, power generation operation characteristics, and seasonal electricity price factors to select the ecological dispatching scheme of the Shenzhen-Da Reservoir that best meets the overall needs, providing a strong guarantee for the sustainable management of water resources in the Shenzhen-Da Reservoir.

[0157] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described deep reservoir ecological scheduling method based on a data-driven model with coupled physical mechanisms.

[0158] The beneficial effects of this invention are as follows:

[0159] (1) This invention achieves high-fidelity and fast prediction of outflow water temperature by embedding physical mechanisms into the loss function to constrain the training of deep learning models. This method effectively solves the efficiency bottleneck caused by the complexity of traditional physical water temperature models and the long running time, while overcoming the shortcomings of pure deep learning models that are prone to prediction inaccuracies due to the lack of mechanistic constraints on the water temperature change process.

[0160] (2) This invention uses a reservoir optimization scheduling model coupled with a deep learning model to scientifically and accurately derive the set of power generation scheduling rules for deep and large reservoirs. By introducing a multi-index optimization method, it comprehensively evaluates water temperature improvement indicators (including the duration of unsuitable water temperature and the depth of damage) and power generation benefit indicators (including utilization hours and seasonal electricity price benefits) to achieve a balance between ecological needs and power generation benefits, and finally selects the optimal scheduling scheme that combines ecological benefits and economic benefits.

[0161] (3) This invention provides a set of efficient and intelligent decision support tools for the coordinated optimization of water temperature and power generation in the Shenzhen-Dalian Reservoir. It not only improves the scientific and refined level of management, but also promotes the protection of the water ecological environment and the sustainable use of water energy resources. Attached Figure Description

[0162] Figure 1 A flowchart of the ecological scheduling method for the Shenzhen-Dalian Reservoir based on a deep learning model constrained by physical mechanisms, provided by this invention;

[0163] Figure 2 A schematic diagram of the grid division of the study area;

[0164] Figure 3 A two-dimensional profile of water temperature changes in the study area;

[0165] Figure 4 A distribution chart of SHAP values ​​for deep learning models (five main factors). Detailed Implementation

[0166] The technical solutions will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0167] This embodiment provides a method, system, and equipment for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms. It can take into account both the benefits of hydropower generation and the need for water temperature control in key fish habitats, and scientifically and efficiently derive ecological scheduling strategies for deep and large reservoirs, providing decision support for regional sustainable development and ecological protection.

[0168] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below, taking a deep reservoir in the Lancang River as an example, in conjunction with the accompanying drawings and specific embodiments.

[0169] See Figure 1 A data-driven model based on coupled physical mechanisms for ecological scheduling of large and deep reservoirs includes:

[0170] The first step, based on measured data from the study area, is to construct a physical water temperature model, and then input data and simulate water temperature, as follows:

[0171] Step 1.1: Collect measured data of the research area of ​​the deep reservoir. The measured data includes meteorological data, hydrological data, topographic data, and water temperature monitoring data, with a time series from 2014 to 2018.

[0172] The meteorological data includes air temperature, solar radiation, wind speed, wind direction, and relative humidity; the hydrological data includes rainfall, runoff, and reservoir operation processes within the basin; the topographic data includes digital elevation model (DEM), catchment area, water flow relationship, and flow velocity distribution; and the water temperature monitoring data includes surface water temperature and vertical temperature profiles at different times and locations.

[0173] Step 1.2: After collecting the measured data, a preliminary physical water temperature model is constructed using Delft3D software, and model parameters are set, including water flow velocity. diffusion coefficient Vertical water temperature gradient index .

[0174] water flow velocity The diffusion coefficient is calculated to be 0.06 m / s based on the Navier-Stokes equations; Based on empirical formulas and measured data, the value was determined to be 0.0063; Vertical water temperature gradient index. The value is set at 50.

[0175] Step 1.3: Based on the physical water temperature model initially constructed in Step 1.2, the physical water temperature model is sequentially divided into horizontal grids, discretized into grids, and layered vertically, as follows:

[0176] Step 1.3.1 involves selecting the horizontal mesh type, using an unstructured mesh, and choosing a triangular mesh to describe the boundary characteristics of the deep reservoir. Figure 2 As shown.

[0177] Step 1.3.2: Perform horizontal mesh generation and set the time step, and set the spatial step. Set the time step to 500m. For a time step of 2 minutes, the mesh is discretized using either the finite difference method or the finite volume method, with a time step of [missing information]. The calculations show that the Courant condition is satisfied.

[0178] Step 1.3.3 involves vertical stratification, dividing the water into 20 layers along the depth direction. Each layer has a uniform internal temperature distribution. The layer design is suitable for the hydrodynamic conditions of deep and large reservoirs.

[0179] Step 1.4: After completing Step 1.3, set the boundary conditions and initial conditions for the physical water temperature model. Specifically:

[0180] First, the boundary conditions of the physical water temperature model are set: the physical water temperature model needs to define the hydrodynamic boundary and the heat transfer boundary of the external boundary. The hydrodynamic boundary: flow rate or water level boundary conditions are set at the inlet and outlet. The inlet can be input with upstream water flow rate, and the outlet can be set with downstream controlled water level or free outflow conditions. The heat transfer boundary: atmospheric-water interface heat exchange conditions are set at the water surface, including solar shortwave radiation, longwave radiation, latent heat flux, and sensible heat flux. These fluxes are usually driven by meteorological data. Simultaneously, for the bottom boundary of the reservoir, the heat transfer effect of the bottom sediment must also be considered.

[0181] After setting the boundary conditions, the initial conditions for the water temperature physical model are set: at the start of the simulation, the initial water temperature field for the entire study area is given. The initial water temperature can be obtained through the following methods: interpolation using measured water temperature data (such as vertical temperature profiles and surface water temperature monitoring values) to obtain the initial distribution; for locations lacking observation data, the regional average temperature is used as the initial field.

[0182] At this point, the initial settings for the physical water temperature model have been completed.

[0183] Step 1.5, based on the physical water temperature model set up in steps 1.2 to 1.4, simulates the spatiotemporal variation of water temperature within the study area by solving the water heat transport equation. Specifically:

[0184] Step 1.5.1, the water temperature simulation process adopts... The turbulence model was closed, and calculations were performed using the ADI algorithm based on the finite volume method. The ADI algorithm discretizes the Navier-Stokes equations to obtain a numerical solution for the flow field. Then, calculations were performed based on the heat conduction equations of the finite element method. The heat conduction equations were discretized, and numerical solutions were obtained using the finite element method to obtain a numerical solution for the water temperature. The heat conduction equations used for water temperature simulation are shown in equations (2) and (3).

[0185] Step 1.5.2, based on the numerical results of Step 1.5.1, performs dynamic simulation updates: During the operation of the physical water temperature model, the water temperature field is updated at each time step. By calculating the convection and diffusion processes of heat within each grid cell, and superimposing the effects of surface meteorological factors, bottom sediment heat transfer, and other heat source terms, the vertical distribution of water temperature and the overall spatiotemporal variation trend can be gradually obtained.

[0186] Step 1.5.3, finally, output the simulation results: The simulation results include the temperature distribution of the surface and deep water layers, the evolution of the thermal stratification structure, and the temporal changes in water temperature at key observation points. Through visualization methods such as two-dimensional profiles, three-dimensional temperature fields, and time-series curves, the spatial distribution and seasonal variation characteristics of water temperature within the study area can be intuitively presented, yielding simulation result data, such as... Figure 3 As shown.

[0187] Step 1.6: To ensure the physical water temperature model accurately reflects the actual water temperature changes, the simulation results from Step 1.5.3 are compared with the measured data using evaluation indicators to complete the calibration and verification of the physical water temperature model. Specifically:

[0188] The evaluation metrics include Nash curve efficiency (NSE), percentage bias (PBIAS), and average relative bias (ARD%). The evaluation criteria are as follows:

[0189] When NSE>0.85, PBIAS<30%, and ARD<25% are simultaneously satisfied, the physical water temperature model is considered to have good performance and can reliably predict water temperature in practical applications, thus validating the physical water temperature model.

[0190] If the conditions are not met simultaneously, return to step 1.2, adjust the parameters of the physical water temperature model, and then repeat the subsequent steps until the evaluation index conditions are met simultaneously.

[0191] After repeated adjustments, the NSE of the water temperature simulation results was 0.92, PBIAS was 20%, and ARD was 16%, which met the evaluation index requirements and can be considered to be satisfied simultaneously.

[0192] The second step involves using the physical water temperature model built in the first step to construct and run simulation scenarios based on collected measured data, generating diverse water temperature change scenarios as training datasets for the deep learning model. Specifically:

[0193] Step 2.1: Determine the uncertainty of the scenario based on the measured data:

[0194] To build rich training data for deep learning, based on the physical water temperature model constructed in the first step, measured data covering multiple years were used to simultaneously reflect input uncertainty and situation uncertainty.

[0195] The simulation scenarios include: long-term meteorological series: hourly / daily sequences of air temperature, shortwave / longwave radiation, wind speed / direction, relative humidity, cloud cover, precipitation, etc. Long-term hydrological series: continuous sequences of inflow / outflow processes, water level, flow velocity, evaporation, inflow water temperature, etc.; operational scheduling scenarios (different release rules, different water replenishment strategies) can be added. External heat flux and optical scenarios: surface atmosphere-water interface heat exchange terms (sensible heat / latent heat / net radiation), bottom sediment heat transfer, inflow heat flux, nearshore shading, light attenuation coefficient (affected by turbidity / suspended sediment), etc.; scenario-based perturbations can be applied to key parameters (e.g., decreased transparency, increased solar radiation). The key objective is to calculate the water temperature field using different year and scenario combinations. The spatiotemporal evolution (including thermal stratification) provides diverse and generalizable training samples for deep learning models.

[0196] Step 2.2: Then, using the physical water temperature model completed in Step 1, simulate and output time series data and spatial series data:

[0197] In each long-series scenario, the physical water temperature model is driven by the multi-dimensional long-series features composed of meteorological, hydrological, and heat flux data to solve the water temperature transport equation and output temperature simulation results data under temporal and spatial variations.

[0198] The time-series data includes: continuous water temperature sequences at key locations (buoy points, inlets / outlets, representative cross-sections) over time, as well as temporal variations in diagnostic parameters such as thermal stratification intensity and thermocline depth. The spatial sequence data includes: temperature distribution across different grid cells / depth layers, two-dimensional horizontal distribution and vertical profiles at typical times, and the evolution of the thermal stratification structure (temperature and thickness of the surface / thermocline / bottom layers). Input and output data are mapped according to year, perturbation type, and parameter set to form paired result data of input features and target variables, totaling 3285 pairs of simulation result data, used for subsequent deep learning model training.

[0199] Step 2.3: Organize and divide the output data from Step 2.2 to form a training dataset:

[0200] Step 2.3.1: To ensure the breadth and representativeness of the deep learning model training, the dataset is divided into three parts:

[0201] Training set: Used to train deep learning models, it contains a large amount of data and covers a variety of different simulation scenarios;

[0202] Validation set: Used to validate the model's performance during training and to help tune parameters;

[0203] Test set: Used to evaluate the model's prediction accuracy and ensure that the model has good generalization ability.

[0204] The data ratios of the training set, validation set, and test set are 60%, 20%, and 20%, respectively, which are 1971 pairs, 657 pairs, and 657 pairs.

[0205] Step 2.3.2, Normalization and Standardization: To eliminate the dimensional differences between different input features, the simulation results of the physical water temperature model are preprocessed after the training dataset is partitioned, including normalization and standardization. This ensures that the input data is on a similar scale, which helps improve the training efficiency of subsequent deep learning models.

[0206] The third step, based on the training dataset obtained in the second step, is to construct a physical mechanism-constrained deep learning model (deep learning model). Specifically:

[0207] Step 3.1: After obtaining the training dataset, the process moves to the deep learning model construction phase. The following are the detailed steps for constructing the deep learning model:

[0208] Step 3.1.1, Design the deep learning model architecture:

[0209] Input layer design: The input layer of a deep learning model receives multiple long-sequence features, including: meteorological factors: air temperature, solar radiation, wind speed, wind direction, relative humidity; hydrological features: inflow rate, lake / reservoir water level, water flow velocity, inflow water temperature; external heat flux and optical parameters: air-water interface heat exchange terms (sensible heat, latent heat, net radiation), substrate heat transfer, etc. The number of these input feature dimensions corresponds to the number of neurons.

[0210] Hidden layer design: ReLU is used as the activation function to alleviate the gradient vanishing problem and improve the convergence speed.

[0211] Output layer design: The output layer of the deep learning model consists of multiple neurons used to predict water temperature-related indicators. Based on the research objective, it is divided into: single-point prediction: outputting the water temperature change over time at a representative location; profile prediction: outputting the temperature distribution across the entire vertical temperature profile; indicator prediction: outputting the intensity of the thermocline, the depth of the thermocline, or the surface-to-bottom temperature difference. ).

[0212] Step 3.1.2, design the loss function for the physical mechanism constraining the neural network in the deep learning model:

[0213] The most crucial aspect of deep learning model architecture is determining the loss function design: incorporating physical knowledge as a loss function into the training of the neural network. This invention uses density as a bridge, designing key physical relationships (mathematical equations) between water temperature, density, and depth for reservoir water temperature simulation as constraints on the loss function. These physical relationships include ① the water temperature-density relationship and ② the density-depth relationship, as detailed below:

[0214] ① Water temperature-density relationship;

[0215] The temperature Y and density ρ of water are known to be nonlinearly related according to equation (5):

[0216] ②Density-depth relationship;

[0217] In the vertical structure of water bodies such as reservoirs, the density of water usually increases monotonically with depth, that is, the closer to the bottom, the greater the density of water. This physical law is described by the inequality in equation (6).

[0218] However, when predicting temperature and then converting it to density using a deep learning model, the model may violate monotonicity constraints. For example, due to prediction errors, the deep learning model might give unphysical results for certain depth pairs, such as shallow water having a higher density than deep water. To incorporate physical laws into the training of deep learning models, a loss function based on physical constraints needs to be constructed to penalize these violations.

[0219] Therefore, we should consider Depth value and An unlabeled dataset of input features on a regular grid with a time step. For any pair of consecutive depth values... and Above, the deep learning model is computed at time step. The difference in density estimation on the surface is shown in Equation (7).

[0220] This can be considered a violation of the concept of depth. and time The physical equations. This can be evaluated as The non-zero occurrence of . Therefore, the present invention can regard the average of all physical violations for each consecutive depth pair and time step as a physics-based loss function, as shown in Equation (8).

[0221] Based on physical loss, a physical mechanism constraint deep learning model was implemented. In particular, in the special problem of reservoir water temperature stratification modeling, by introducing a physical mechanism-based constraint condition into the loss function, the predicted water temperature results are kept consistent with key physical variables such as density and depth. This ensures that the output of the deep learning model conforms to the basic laws of thermodynamics and fluid mechanics, and obtains a physically reasonable solution.

[0222] Step 3.1.3: Train the deep learning model based on the already constructed model:

[0223] Training dataset: Using the training set from step 2.3, the input features are meteorological data, hydrological data, and water temperature input data, the target variable is the water temperature change process (such as surface water temperature, vertical water temperature, etc.), and the loss function is the mean squared error (MSE). The difference between the predicted value and the true value of the deep learning model is calculated.

[0224] Model performance optimization: The Adam optimizer is adopted, which combines adaptive learning rate and momentum methods to improve the efficiency and convergence speed of deep learning model training. The Adam algorithm updates the parameters of each layer through the update rule of Equation (9).

[0225] Regularization: In order to avoid overfitting, regularization techniques are used, also known as weight decay, which control the model complexity by adding a regularization term to the loss function, as shown in Equation (10).

[0226] Training strategy: An early stopping strategy is used to prevent overfitting. Training is stopped early when performance on the validation set no longer improves. This strategy effectively saves training time while avoiding overfitting to the training data.

[0227] Step 3.2, Validation and Evaluation of Deep Learning Model After Training:

[0228] To ensure the accuracy and generalization ability of the deep learning model, cross-validation is used to ensure its stability and reliability. Specifically:

[0229] Cross-validation: This method uses K-fold cross-validation to evaluate deep learning models. The data is divided into K subsets, and one subset is used as the validation set each time, while the remaining subsets are used for training. Finally, the average performance of the deep learning model is evaluated. This method effectively reduces the random errors introduced by data partitioning.

[0230] Evaluation indicators: The following indicators are mainly used:

[0231] Root Mean Square Error (RMSE): Measures the difference between the predictions of a deep learning model and the actual values.

[0232] Nash efficiency coefficient (NSE): Used to evaluate the fit of a deep learning model. The closer the value is to 1, the better the performance of the deep learning model.

[0233] After the final model cross-validation, the RMSE was within 0.31 and the NSE was above 0.87, which can be considered as the deep learning model having a good fitting effect.

[0234] The fourth step involves conducting sensitivity analysis based on the deep learning model built in the third step. This systematically identifies the key influencing factors driving changes in the temperature of the water released from the Shenda Reservoir, as follows:

[0235] First, the SHAP value is calculated using the Shapley additive interpretation method based on game theory to calculate the marginal contribution of the input features. Then, statistical analysis is performed on the SHAP values ​​to obtain the contribution distribution of each input feature, generating a SHAP summary plot to determine the degree of influence of each input feature on the deep learning model's prediction of the released water temperature. Finally, based on the SHAP results, key sensitive factors are identified, and the input feature with the highest SHAP value, namely the water level process, is selected as the key sensitive factor to guide subsequent reservoir ecological scheduling decisions.

[0236] The fifth step involves constructing a reservoir optimization scheduling model based on the key sensitive factors from the fourth step, and coupling it with the deep learning model from the third step. This process, while satisfying the water temperature target, derives a suitable set of scheduling rules. Specifically:

[0237] Step 5.1, first, define the framework for the optimized scheduling model:

[0238] To improve optimization efficiency and avoid the high computational overhead caused by repeated running of the physical water temperature model, this invention introduces a pre-built deep learning model into the outflow water temperature prediction to replace the original physical water temperature model and quickly simulate the water temperature process under different scheduling schemes. The prediction results are directly coupled with the optimization scheduling model, thereby efficiently calculating statistical indicators related to outflow water temperature in each iteration.

[0239] Step 5.2: After setting the framework in Step 5.1, set the optimization objectives and determine the decision variables:

[0240] Determine the power generation target: Maximize the average annual power generation of the Shenda Reservoir as the power generation benefit optimization target, and provide an economic evaluation basis for subsequent dispatching schemes, as shown in Equation (11).

[0241] Determine the downstream water temperature target: The ecological optimization target is the guarantee rate that the downstream water temperature meets the suitable water temperature of the downstream fish habitat. The guarantee rate is defined as the proportion of the number of time periods when the downstream water temperature is within the suitable temperature range to the total number of scheduling time periods. This target can quantitatively reflect the effectiveness of the scheduling scheme in ensuring ecological water temperature. The higher the target, the better, as shown in Equation (12).

[0242] Decision variables: In order to ensure that the predetermined ecological and power generation goals can be achieved under different water inflow scenarios, it is necessary to optimize the reservoir power generation scheduling rules to make them applicable in the long term.

[0243] Step 5.3: After completing Step 5.2, set constraints for the reservoir optimization scheduling model:

[0244] Reservoir scheduling optimization generally follows water balance constraints, reservoir capacity constraints, reservoir discharge constraints, power generation flow constraints, reservoir output constraints, and scheduling rule parameter range constraints. Since this invention needs to take into account the downstream river water temperature and ecological needs, it adds a guarantee rate constraint that the downstream water temperature meets the suitable water temperature for downstream fish habitats. Specifically, these are Equations (13), (14), (15), (16), (17), (18), and (19).

[0245] Step 5.4, finally, the solution algorithm for the reservoir optimization scheduling model is set:

[0246] To efficiently solve the above optimization problem, this invention employs a non-dominated sorting genetic algorithm (NSGA-II) with an elitist strategy. Its parameters are comprehensively optimized as follows: the population size is set to 200 to fully maintain solution set diversity; a simulated binary crossover operator is used, with a crossover probability of 0.7 and a distribution exponent of 20, to promote the effective exchange of superior genes; for mutation, a multinomial mutation operator is used, with a mutation probability of 0.3 and a distribution exponent of 20, enhancing local search capabilities while maintaining population diversity; the algorithm adopts a tournament selection strategy (size 2) and uses a maximum evolutionary generation of 1000 as the termination condition, thereby ensuring the convergence and distribution of solutions with reasonable computational cost.

[0247] Through multi-objective optimization scheduling, a set of scheduling rules that simultaneously meet the requirements of the discharge water temperature guarantee rate were obtained, providing scientific support for reservoir operation and management, and laying a solid foundation for subsequent multi-index optimization.

[0248] The sixth step, based on the set of scheduling rules obtained in the fifth step, further combines the characteristics of downstream water temperature changes, power generation operation characteristics, and seasonal electricity price factors to conduct a multi-indicator optimization analysis. From the perspective of comprehensive ecological and economic benefits, the optimal reservoir scheduling scheme is selected. Specifically:

[0249] Step 6.1, determine the preferred indicators:

[0250] This invention determines the duration of unsuitable water temperature and the depth of damage as water temperature indicators, and determines the power generation benefits of utilization hours and seasonal electricity prices as power generation benefit indicators.

[0251] The water temperature index aims to ensure that the discharged water temperature is as close as possible to the suitable temperature range of fish habitats, thereby reducing the adverse impact on the aquatic ecosystem. It mainly includes two aspects: the duration of unsuitable water temperature and the depth of damage. The duration of unsuitable water temperature is used to characterize the longest period of continuous deviation of the discharged water temperature from the suitable range, reflecting the spatiotemporal continuity of adverse conditions, as shown in Equation (20). The depth of damage is used to measure the magnitude and severity of the deviation from the suitable temperature range, taking into account both extreme values ​​and average deviation levels, as shown in Equation (21). By comprehensively considering these two indicators, it is possible to effectively prevent long-term and large-scale abnormal deviations in the discharged water temperature, thereby maintaining the stability and suitability of the aquatic ecosystem.

[0252] The power generation benefit indicators aim to ensure that reservoir operation achieves a high economic return while taking into account ecological protection. They mainly include two aspects: utilization hours and power generation benefit considering seasonal electricity prices. Utilization hours are used to measure the ratio of the actual annual operating time of the unit to the full-load operating time corresponding to the installed capacity, reflecting the impact of reservoir scheduling on the unit utilization rate and the degree of capacity release, as shown in Equation (22). Seasonal electricity price power generation benefit combines the differences in electricity prices and the distribution of power generation in different seasons to quantify the economic benefit level of the scheduling strategy in which priority is given to power generation in high electricity price seasons and reasonable operation is arranged in low electricity price seasons, as shown in Equation (23). By comprehensively considering these two indicators, the power generation efficiency and benefit performance of the scheduling rules under different electricity prices and water inflow conditions throughout the year can be effectively evaluated.

[0253] Step 6.2, Multi-index Optimization Method:

[0254] Based on the set of scheduling rules obtained in the fifth step, the scheduling process corresponding to each rule is simulated and calculated. Based on this, the water temperature index and power generation benefit index involved in the sixth step are calculated. Finally, a weighted objective function is used for evaluation. This measures the performance of each scheduling rule in maintaining the suitability of the downstream water temperature and evaluates its power generation output level under different seasonal electricity price conditions. This achieves a comprehensive comparison and optimization of water temperature control and power generation benefits, as shown in Equation (24).

[0255] By comparing the comprehensive scores of all candidate schemes, the scheme with the highest comprehensive score is selected as the optimal ecological scheduling scheme.

[0256] This embodiment also provides a data-driven model based on coupled physical mechanisms for the ecological scheduling system of the deep reservoir, which includes the following modules:

[0257] The physical water temperature model construction module: This physical water temperature model construction module establishes a physical water temperature model based on measured data of the study area, covering data input, model parameter setting, mesh generation and vertical layering, boundary and initial condition setting, water temperature simulation and result output.

[0258] The module for generating the training dataset outputs time-series and spatial-series temperature results by running a physical water temperature model under different scenarios. The resulting data is then organized and divided into training, validation, and test sets, and preprocessed using normalization and standardization methods to obtain a high-quality dataset suitable for training deep learning models.

[0259] The deep learning model building module completes the construction of a deep learning model constrained by physical mechanisms: model architecture design, setting input layers, hidden layers, and output layers to achieve single-point prediction, profile prediction, and index prediction; the relationship between water temperature and density, and the relationship between density and depth are introduced into the loss function constraint to ensure that the prediction results conform to physical laws; mean square error, Adam optimizer, regularization, and early stopping strategies are used to improve model performance; finally, the deep learning model is verified and evaluated through K-fold cross-validation and indicators such as root mean square error and Nash efficiency coefficient to ensure the model's accuracy and generalization ability.

[0260] The deep learning model training module: Based on the obtained training dataset, it establishes and trains a deep learning model, determines feature importance through SHAP analysis, and ensures the prediction accuracy and interpretability of the deep learning model.

[0261] The reservoir optimization scheduling module: constructs a reservoir optimization scheduling model by coupling a deep learning model, and derives a suitable set of scheduling rules under the premise of meeting the water temperature target.

[0262] The multi-indicator optimization output optimal solution module comprehensively considers the characteristics of downstream water temperature changes, power generation operation characteristics, and seasonal electricity price factors to select the ecological dispatching scheme of the Shenzhen-Da Reservoir that best meets the overall needs, providing a strong guarantee for the sustainable management of water resources in the Shenzhen-Da Reservoir.

[0263] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described deep-water reservoir ecological scheduling method based on a data-driven model with coupled physical mechanisms.

[0264] The above embodiments are merely illustrative of the implementation methods of the present invention, but should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A data-driven model based on coupled physical mechanisms for ecological scheduling of large and deep reservoirs, characterized in that, The ecological regulation method for the deep reservoir includes the following steps: The first step is to construct a physical water temperature model based on measured data from the study area, and then input the data and simulate the water temperature. Step 1.1: Collect measured data of the study area, including meteorological data, hydrological data, topographic data, and water temperature monitoring data; Step 1.2: After collecting the measured data in Step 1.1, a preliminary physical water temperature model is constructed and model parameters are set, including water flow velocity. diffusion coefficient Vertical water temperature gradient index ; Step 1.3 involves sequentially performing horizontal mesh generation, mesh discretization, and vertical layering on the physical water temperature model initially constructed in Step 1.2; Step 1.4: Set the boundary conditions and initial conditions for the physical water temperature model; Step 1.5: By solving the heat transport equation of the water body, the spatiotemporal changes of water temperature in the study area are simulated to obtain simulation results data; Step 1.6: Compare the simulation results data from Step 1.5 with the measured data using evaluation indicators to complete the calibration and verification of the physical water temperature model; The second step involves using the physical water temperature model built in the first step to construct and run a simulation scenario based on the collected measured data, generating diverse water temperature change scenarios as a training dataset for the deep learning model. The third step, based on the training dataset obtained in the second step, is to construct a deep learning model constrained by physical mechanisms. Specifically: Step 3.1: Constructing and training the deep learning model; Step 3.2: Validate and evaluate the trained deep learning model using cross-validation. The fourth step involves conducting sensitivity analysis based on the deep learning model built in the third step to identify the key influencing factors driving changes in the temperature of the water discharged from the Shenda Reservoir. The fifth step involves constructing a reservoir optimization scheduling model by combining the key sensitive factors from the fourth step and coupling it with the deep learning model from the third step. Under the premise of meeting the water temperature target, a suitable set of scheduling rules is derived. The sixth step, based on the set of scheduling rules obtained in the fifth step, further combines the characteristics of downstream water temperature changes, power generation operation characteristics, and seasonal electricity price factors to conduct multi-index optimization analysis, selecting the optimal reservoir scheduling scheme from the perspective of comprehensive ecological and economic benefits; specifically: Step 6.1, determine the preferred indicators: The duration of unsuitable water temperature and the depth of damage are used as water temperature indicators, while the power generation efficiency is measured by utilization hours and seasonal electricity prices; the specific formulas are as follows: Unsuitable water temperature duration This reflects the duration for which the discharged water temperature continuously deviates from the suitable range. The calculation formula is as follows: (20) ； in The value range is [0,1], and the closer the value is to 1, the closer the water temperature is to the suitable condition. The duration of continuous deviation of water temperature predicted by the model; The target reference value represents the maximum allowable duration of continuous deviation from the ecologically acceptable range. Depth of damage This reflects the maximum deviation between the discharged water temperature and the suitable range. The calculation formula is: (21) ； in To minimize the impact of water temperature, the value range is [0,1]. The closer the value is to 1, the closer the water temperature is to the suitable range. The predicted deviation of water temperature; The target reference value represents the maximum acceptable temperature deviation. Hours of use The calculation formula is used to examine how close the unit's operating time is to the theoretical full-capacity utilization hours. (22) ； in The utilization hours index has a value range of [0,1]. The closer the value is to 1, the higher the unit utilization efficiency. This represents the actual utilization hours of the generating units under the candidate scheduling scheme. The target utilization hours can usually be set as the midpoint between the upper and lower limits; These represent the theoretically acceptable maximum and minimum number of utilization hours, respectively. Power generation efficiency of seasonal electricity prices The objective function is to maximize the power generation benefits of the cascade hydropower stations during the scheduling period. The calculation formula is as follows: (23) in As an indicator of the efficiency of electricity generation under seasonal electricity prices; Indicates the time step; Indicates the first Hydropower station Year Real-time electricity pricing; Indicates the total number of scheduling periods; Indicates the number of hydroelectric power stations; Indicates the total number of years; Indicates the first Hydropower station output coefficient; Indicates the first Hydropower station Year Real-time power generation flow; Indicates the first Hydropower station Year Water level at all times; Step 6.2, Multi-index optimization method.

2. The method for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms as described in claim 1, characterized in that, The first step is as follows: Step 1.3 is as follows: Step 1.3.1: Using an unstructured mesh, select a triangular mesh or a quadrilateral mesh to complete the description of the boundary features of lakes, rivers or reservoirs; Step 1.3.2: Perform horizontal mesh generation and set the time step; set the spatial step. Set time step The mesh is discretized using the finite difference method or the finite volume method, with a time step of [missing information]. Satisfy the Courant condition: (1) ； in, This represents the maximum value of the water flow velocity; Step 1.3.3: Perform vertical stratification, dividing the water into several sections along the depth direction. Layers of equal thickness or layers of uniform thickness, with each layer having a uniform internal temperature distribution. Lakes use layers of equal thickness, while reservoirs use layers of equal thickness. Layering; Step 1.4 is as follows: The boundary conditions of the physical water temperature model are set as follows: the hydrodynamic boundary and the heat transfer boundary of the external boundary are defined; the hydrodynamic boundary includes: setting flow rate or water level boundary conditions at the inlet and outlet, the inlet is input with upstream water flow rate, and the outlet is set with downstream control water level or free outflow condition; the heat transfer boundary is: setting air-water interface heat exchange conditions at the water surface. Initial conditions are set for the water temperature physical model: an initial water temperature field is given; Step 1.5 is as follows: Step 1.5.1, the water temperature simulation process adopts... The turbulence model was closed, and the flow field was calculated using the ADI algorithm based on the finite volume method to obtain the numerical solution of the flow field. The heat conduction equation was calculated based on the finite element method. The heat conduction equation was discretized, and the finite element method was used to solve it numerically to obtain the numerical result of the water temperature. Step 1.5.2, based on the numerical results of Step 1.5.1, perform simulation dynamic updates: during the operation of the physical water temperature model, the water temperature field is updated at each time step; by calculating the convection and diffusion process of heat in each grid cell, and superimposing the effects of surface meteorological effects, bottom sediment heat transfer and other heat source terms, the vertical distribution of water temperature and the overall spatiotemporal variation trend can be gradually obtained. Step 1.5.3, output the simulation results: The simulation results include the temperature distribution of the surface and deep layers of the water, the evolution of the thermal stratification structure, and the temporal changes of water temperature at key observation points, which are used as simulation result data; Step 1.6: Compare the simulation results data from Step 1.5.3 with the measured data using evaluation indicators to complete the calibration and verification of the physical water temperature model; The evaluation metrics include Nash curve efficiency (NSE), percentage bias (PBIAS), and average relative bias (ARD%). The evaluation criteria are as follows: When NSE>0.85, PBIAS<30%, and ARD<25% are simultaneously satisfied, the performance of the physical water temperature model meets the requirements, and the physical water temperature model is validated. If the conditions are not met simultaneously, return to step 1.2 and adjust the parameters of the physical water temperature model until the evaluation index conditions are met simultaneously.

3. The method for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms as described in claim 2, characterized in that, The second step is specifically as follows: Step 2.

1. Determine the uncertainty of the scenario based on the measured data: Based on the physical water temperature model built in the first step, measured data covering multiple years were used. Step 2.

2. Using the physical water temperature model completed in Step 1, simulate and output time series data and spatial series data: The physical water temperature model is driven by multidimensional long-sequence features composed of meteorological, hydrological, and heat flux data, and outputs temperature simulation results data under temporal and spatial variations. Step 2.

3. Organize and divide the temperature simulation results data output in Step 2.2 to form a training dataset, which is divided into a training set, a validation set, and a test set. After dividing the training dataset, perform data preprocessing on the physical water temperature model simulation results, including normalization and standardization.

4. The method for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms as described in claim 3, characterized in that, In the third step mentioned above: Step 3.1 specifically involves: Step 3.1.1, Design the deep learning model architecture: Input layer design: The input layer of a deep learning model receives multiple multidimensional long sequence features; Hidden layer design: ReLU is used as the activation function; Output layer design: Composed of multiple neurons, used to predict water temperature-related indicators; categorized according to research objectives: Single-point prediction: Outputs the water temperature change over time at a representative location; Profile prediction: Outputs the temperature distribution across the entire vertical temperature profile; Indicator prediction: Outputs the intensity of the thermocline, the depth of the thermocline, or the surface-to-bottom temperature difference. Step 3.1.2, design the loss function for the physical mechanism constraining the neural network in the deep learning model: Using density as a bridge, the key physical relationships between water temperature, density, and depth in reservoir water temperature simulation are used as constraints for the loss function. These physical relationships include water temperature-density relationships and density-depth relationships, as detailed below: ① Water temperature-density relationship; (5) ； in, Indicates water temperature. Indicates density; ②Density-depth relationship; Assuming there are two bodies of water, 1 and 2, in the vertical direction, then: If d1 < d2(6); in, Indicates depth and time The water density at that location; Indicates depth and time The water density at that location; Indicates the depth of water body 1; Indicates the depth of water body 2; When depth Less than At that time, that is Compare When closer to the water surface, the corresponding water density should not be greater than the density of the deeper layers; To incorporate physical laws into the training of deep learning models, a loss function based on physical constraints is constructed; Depth value and An unlabeled dataset of input features on a regular grid with a time step; in any pair of consecutive depth values and Above, the deep learning model is computed at time step. The differences in density estimates are shown below: (7) ； in, This indicates a violation of the rules regarding depth. and time The physical equations; This indicates the depth and time The density value; This indicates the depth and time The density value; The average of all physical violations for each consecutive depth pair and time step is considered as a physics-based loss function: (8) ； in, This represents the cumulative loss under physical constraints; Indicates the water temperature of a specific body of water; Indicates the number of neurons over time; Indicates the number of deep neurons; A deep learning model constrained by physical mechanisms is implemented based on physical loss. Step 3.1.3: Train the deep learning model: Training dataset: Using the training set from step 2.3, the input features are meteorological data, hydrological data, and water temperature input data, the target variable is the water temperature change process, and the loss function is the mean squared error (MSE). The difference between the predicted value and the true value of the deep learning model is calculated. Model performance optimization: Adam optimizer was used; Regularization: Controlling the complexity of deep learning models by adding a regularization term to the loss function. (10) ； in, This represents the loss of the entire function. Indicates the average error. The regularization coefficient is . These are weight parameters; Training strategy: Employ early stopping to prevent overfitting; In step 3.2: Evaluation metrics are used, including root mean square error (RMSE) and Nash efficiency coefficient (NSE). When both RMSE and NSE are satisfied, the deep learning model is considered to have a good fit to meet the simulation requirements. If both conditions are not met, the process returns to step two and the amount of training data required for the deep learning model is increased.

5. The method for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms as described in claim 4, characterized in that, The fourth step is as follows: First, the SHAP value is calculated using the Shapley additive interpretation method based on game theory to calculate the marginal contribution of the input features. Then, statistical analysis of the SHAP value is performed to obtain the contribution distribution of each input feature, generating a SHAP summary diagram to show the degree of influence of each input feature on the deep learning model's prediction of the outflow water temperature. Finally, based on the SHAP results, key sensitive factors are identified, and input features with higher SHAP values ​​are selected as key sensitive factors to guide subsequent reservoir ecological scheduling decisions.

6. The method for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms as described in claim 5, characterized in that, The fifth step is as follows: Step 5.1, first, define the framework for the optimized scheduling model: The existing deep learning model is introduced into the prediction of outflow water temperature, and its prediction results are directly coupled with the optimization scheduling model to efficiently calculate the statistical indicators related to outflow water temperature in each iteration. Step 5.2: Set optimization objectives and determine decision variables: Determine the power generation target: Maximize the multi-year average power generation of the Shenda Reservoir as the power generation efficiency optimization target; (11) ； in, This represents the total amount of electricity generated. Indicates the total number of scheduling periods; Indicates that the reservoir is at time Hydropower generation; Indicates the time step; Determine the target for downstream water temperature: The ecological optimization target is the rate at which the downstream water temperature meets the suitable temperature range for downstream fish habitats. The guarantee rate is defined as the proportion of the number of time periods during which the downstream water temperature is within the suitable temperature range to the total number of scheduling time periods. (12) ； in, Downflow water temperature guarantee rate (%) Indicates the total number of scheduling periods; Indicates time as The temperature of the discharged water; , These represent the lower and upper limits of water temperature suitable for fish habitats, respectively. Decision variables: Optimize reservoir power generation scheduling rules; Step 5.3: Set constraints for the reservoir optimization scheduling model: The constraints include water balance constraints, reservoir capacity constraints, reservoir discharge constraints, power generation flow constraints, reservoir station output constraints, reservoir capacity parameter constraints in dispatching rules, and output coefficient constraints in dispatching diagrams. Step 5.4, finally, the solution algorithm for the reservoir optimization scheduling model is set: A heuristic algorithm was used to solve the reservoir optimization scheduling model, and a set of scheduling rules that simultaneously meet the requirements of the discharge water temperature guarantee rate were obtained.

7. The method for ecological scheduling of deep and large reservoirs based on a data-driven model with coupled physical mechanisms as described in claim 6, characterized in that, Step 6.2 of the sixth step specifically includes: Based on the set of scheduling rules obtained in step 5, the scheduling process corresponding to each rule is simulated and calculated. Based on this, the water temperature index and power generation benefit index involved in step 6 are calculated. Finally, a weighted objective function is used for evaluation to measure the performance of each scheduling rule in maintaining the suitability of the downstream water temperature, and to assess its power generation output level under different seasonal electricity price conditions, so as to achieve a comprehensive comparison and optimization of water temperature control and power generation benefits. The objective function is: to calculate a weighted comprehensive score based on the water temperature index and power generation efficiency index of all scheduling schemes. (24) ； in, , , , The weights are respectively the duration of unsuitable water temperature, depth of damage, number of utilization hours, and the power generation benefits considering seasonal electricity prices; , , , These represent the scores for each indicator, with scores ranging from [0,1]. The duration of unsuitable water temperature is an important factor. To disrupt the depth index, To utilize the hourly index, As an indicator of the efficiency of electricity generation under seasonal electricity prices; By comparing the comprehensive scores of all candidate schemes, the scheme with the highest comprehensive score is selected as the optimal ecological scheduling scheme.

8. A data-driven model based on coupled physical mechanisms for ecological scheduling of large and deep reservoirs, characterized in that, The ecological scheduling system for the Shenzhen-Dongming Reservoir, based on a data-driven model with coupled physical mechanisms, implements the ecological scheduling method for the Shenzhen-Dongming Reservoir as described in any one of claims 1-7. This includes a physical water temperature model construction module, a training dataset generation module, a deep learning model construction module, a deep learning model training module, a reservoir optimization scheduling module, and a multi-index optimization and optimal solution output module. Specifically: The physical water temperature model construction module establishes a physical water temperature model based on measured data from the study area. The module for generating training datasets: By running the physical water temperature model under different scenarios, it outputs temperature results in time series and spatial series; it forms training sets, validation sets and test sets, and preprocesses them using normalization and standardization methods to obtain datasets that can be used for training deep learning models; The deep learning model building module: completes the construction of deep learning models constrained by physical mechanisms; The deep learning model training module: Based on the obtained training dataset, it establishes and trains a deep learning model, and determines the importance of features through SHAP analysis; The reservoir optimization scheduling module: constructs a reservoir optimization scheduling model by coupling a deep learning model, and derives a suitable set of scheduling rules under the premise of meeting the water temperature target; The multi-indicator optimization output optimal solution module: comprehensively considers the characteristics of water temperature change, power generation operation characteristics and seasonal electricity price factors, and selects the ecological dispatching scheme of the Shenda Reservoir that best meets the overall needs.

9. An electronic device, characterized in that, The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned deep-water reservoir ecological scheduling method based on a data-driven model with coupled physical mechanisms.

Images