Ecological regulation method for reservoirs during fish spawning season based on water temperature prediction

Abstract

This invention discloses a reservoir ecological scheduling method integrating water temperature prediction during fish spawning season, comprising: acquiring biological characteristic data of target fish species, determining the suitable water temperature range for spawning and parameters for determining the spawning window; constructing a multi-objective reservoir scheduling model including power generation benefits, flood control risks, and ecological scheduling objectives; generating multiple sets of candidate scheduling schemes using a multi-objective optimization algorithm; extracting physical information enhancement feature sets for candidate scheduling schemes during iterative solution, and inputting them into a pre-constructed downstream river water temperature prediction model to obtain a predicted water temperature sequence; dynamically determining the spawning window and calculating the ecological scheduling objectives based on the predicted water temperature sequence, the suitable water temperature range for spawning, and the parameters for determining the spawning window; and outputting the optimal scheduling scheme by combining the power generation benefit objective and the flood control risk objective. This invention improves the efficiency and reliability of downstream water temperature prediction and ecological objective evaluation in long-term multi-objective scheduling.

Description

Technical Field

[0001] This invention relates to the field of watershed water resources and water environment regulation technology, and in particular to a method for ecological scheduling of reservoirs during fish spawning season that integrates water temperature prediction. Background Technology

[0002] The construction and operation of large-scale water conservancy projects have profoundly altered the natural hydrological rhythms and thermodynamic states of rivers. After the impoundment of deep-water reservoirs, vertical temperature stratification easily forms, leading to changes in the annual distribution characteristics and peak phase of downstream water temperature. As an environmental factor controlling the reproductive rhythms of aquatic organisms, changes in water temperature's spatiotemporal heterogeneity can disrupt the life cycle succession of sensitive downstream aquatic organisms. Therefore, introducing a water temperature response mechanism into the multi-objective scheduling framework of reservoirs to effectively regulate the thermodynamic environment of downstream river channels has engineering and technical value for maintaining the habitat integrity and community stability of river systems.

[0003] Currently, research on reservoir scheduling for water temperature regulation mainly relies on coupled hydrodynamic and water quality numerical models to simulate the thermal evolution of the flow field, combined with conventional ecological baseflow or fixed habitat curves for evaluation and optimization. These numerical physics models suffer from bottlenecks when faced with long-term, multi-scenario scheduling simulations, including high computational costs and difficulty in integrating with high-frequency iterative multi-objective optimization frameworks. Furthermore, some schemes incorporating conventional surrogate models are prone to prediction distortion when faced with extreme hydrological and meteorological changes, and existing suitability assessment paradigms largely depend on static rules based on calendar cycles, failing to adequately characterize the differentiated physiological tolerance mechanisms of organisms under complex environmental stresses.

[0004] In summary, existing technologies struggle to effectively balance computational efficiency in spatiotemporal water temperature extrapolation with generalization accuracy under extreme conditions when dealing with large-scale scheduling optimization, and they lack dynamic adaptability in evaluating the response of aquatic ecological environment elements. Therefore, it is necessary to research an ecological scheduling method that can take into account both complex environmental boundary characterization and high-frequency optimization computational performance, thereby improving the decision-making reliability of large-scale water conservancy projects under multi-objective coordination. Summary of the Invention

[0005] Purpose of the invention: To provide a reservoir ecological scheduling method that integrates water temperature prediction for fish spawning season, in order to solve the problems of large calculation workload for water temperature extrapolation in existing reservoir ecological scheduling, mismatch between fixed calendar spawning season evaluation and actual reproductive rhythm, and difficulty in scheduling schemes to simultaneously take into account the water temperature requirements for power generation, flood control and fish spawning.

[0006] Technical solution: A reservoir ecological management method integrating water temperature prediction during fish spawning season, comprising:

[0007] Biological characteristic data of the target fish species were acquired to determine the suitable spawning water temperature range and spawning window parameters. A multi-objective reservoir scheduling model was constructed, including power generation benefit objectives, flood control risk objectives, and ecological scheduling objectives. A multi-objective optimization algorithm was used to iteratively solve the multi-objective reservoir scheduling model, generating multiple sets of candidate scheduling schemes. During the iterative solution process, a set of physical information enhancement features was extracted for each set of candidate scheduling schemes, and the set of physical information enhancement features was input into a pre-constructed downstream river water temperature prediction model to output the predicted water temperature sequence corresponding to the candidate scheduling scheme. Based on the predicted water temperature sequence, the suitable spawning water temperature range, and the spawning window parameters, the spawning window corresponding to the candidate scheduling scheme was determined, and the ecological scheduling objective of the candidate scheduling scheme was calculated. Combining the power generation benefit objective, flood control risk objective, and ecological scheduling objective, the optimal scheduling scheme was output from the multiple sets of candidate scheduling schemes.

[0008] Optionally, the parameters for determining the spawning window include the biological zero degree of gonadal development, the accumulated temperature threshold for gonadal maturation, and the number of days of spawning. This involves acquiring biological characteristic data of the target fish species, determining the suitable spawning water temperature range and the parameters for determining the spawning window, including: acquiring fish sample data collected under different temperature and current conditions, and constructing a biological characteristic database using body weight, gonadal weight, gonadal index, and histological section data; identifying water temperature-sensitive target fish species based on the biological characteristic database, and determining the suitable spawning water temperature range, the biological zero degree of gonadal development, the accumulated temperature threshold for gonadal maturation, and the number of days of spawning for the target fish species.

[0009] Optionally, acquiring biological characteristic data of the target fish and determining the suitable spawning water temperature range, biological zero degree for gonadal development, accumulated temperature threshold for gonadal maturation, and duration of spawning for the target fish includes: acquiring fish sample data collected under different temperature and current conditions, constructing a biological characteristic database using body weight, gonadal weight, gonadal index, and histological section data; identifying water temperature-sensitive target fish based on the biological characteristic database, and determining the suitable spawning water temperature range, biological zero degree for gonadal development, accumulated temperature threshold for gonadal maturation, and duration of spawning for the target fish.

[0010] Optionally, the physical information enhancement feature set includes at least the following feature categories: thermal stratification intensity features for characterizing the vertical heat distribution gradient inside the reservoir; hydraulic retention features for characterizing the water exchange rate and heating time inside the reservoir; and adaptive cumulative heat input features for characterizing the cumulative impact of previous meteorological conditions on the current outflow water temperature.

[0011] Optionally, the thermal stratification intensity feature, hydraulic retention feature, and adaptive cumulative heat input feature are constructed through the following steps: acquiring pre-collected meteorological driving data, daily inflow, and daily reservoir capacity of the target watershed, and extracting daily water level and daily outflow from candidate scheduling schemes; calculating the positive temperature deviation between the daily temperature and the preset annual average temperature in the meteorological driving data; multiplying the positive temperature deviation by the daily water level, and using the ratio of the product to the total flow as the thermal stratification intensity feature, where the total flow is the sum of the daily inflow and the daily outflow; calculating the ratio of the daily reservoir capacity to the daily outflow as the hydraulic retention feature; determining the historical impact days based on the numerical value of the hydraulic retention feature, constructing a dynamic accumulation window using the historical impact days, and accumulating the normalized historical shortwave radiation and historical temperature within the dynamic accumulation window to obtain the adaptive cumulative heat input feature.

[0012] Optionally, during the iterative solution process, a set of enhanced physical information features is extracted for each set of candidate scheduling schemes, and the set of enhanced physical information features is input into a pre-constructed downstream channel water temperature prediction model to output the predicted water temperature sequence corresponding to the candidate scheduling scheme. This includes: extracting the daily water level sequence and daily outflow sequence of the candidate scheduling scheme during the complete scheduling period, and obtaining the pre-stored daily meteorological driving sequence; for each day during the complete scheduling period, combining the water level, outflow, meteorological driving elements, and historical lag states of the preceding period to calculate and generate the set of enhanced physical information features for that day; concatenating the set of enhanced physical information features for each day during the complete scheduling period into an input feature matrix according to the time sequence; and inputting the input feature matrix into the downstream channel water temperature prediction model for a batch inference to obtain the predicted water temperature sequence covering the entire complete scheduling period.

[0013] Optionally, based on the predicted water temperature sequence, the suitable water temperature range for spawning, and the parameters for determining the spawning window, the spawning window corresponding to the candidate scheduling scheme is determined, and the ecological scheduling objective of the candidate scheduling scheme is calculated, including: calculating the cumulative effective accumulated temperature sequence on the time axis based on the predicted water temperature sequence and the zero degree of gonadal development biology; locking the start time by double comparing the cumulative effective accumulated temperature with the gonadal maturation accumulated temperature threshold, and the predicted water temperature of the day with the lower limit of the suitable water temperature for spawning, and calculating the end time by extending the number of days of spawning duration; extracting water temperature segments from the predicted water temperature sequence within the spawning window; calculating the degree of deviation of the water temperature value within the water temperature segment from the suitable water temperature range for spawning, and calculating the suitability score based on the degree of deviation, and using the cumulative suitability score within the spawning window as the ecological scheduling objective.

[0014] Optionally, in calculating the deviation of the water temperature value within a water temperature segment from the suitable spawning water temperature range, and in calculating the suitability score based on the deviation, an asymmetric water temperature suitability function is specifically applied. This includes: determining whether the predicted water temperature for the day within the spawning window is below the lower limit of the suitable spawning water temperature, within the suitable spawning water temperature range, or above the upper limit of the suitable spawning water temperature; when the predicted water temperature for the day is below the lower limit of the suitable spawning water temperature, a preset left deviation penalty parameter is used to calculate the low-temperature side suitability score; when the predicted water temperature for the day is within the suitable spawning water temperature range, the suitability score for the day is determined as the benchmark value for the full daily suitability score; when the predicted water temperature for the day is above the upper limit of the suitable spawning water temperature, a preset right deviation penalty parameter is used to calculate the high-temperature side suitability score; wherein, the values ​​of the left deviation penalty parameter and the right deviation penalty parameter are not equal.

[0015] Optionally, after outputting the optimal scheduling scheme from multiple candidate scheduling schemes, the process also includes an evaluation step on the improvement of spawning of target fish downstream of the dam: obtaining boundary data for different typical hydrological year scenarios pre-set; inputting the water level trajectory and outflow trajectory in the optimal scheduling scheme into the downstream river water temperature prediction model in combination with the boundary data to simulate and generate an optimized downstream water temperature sequence; statistically analyzing the duration and cumulative frequency of the optimized downstream water temperature sequence satisfying the suitable water temperature range for spawning; comparing the duration and cumulative frequency with historical baseline data when ecological scheduling was not implemented, calculating the improvement difference of the target fish spawning window period, and outputting the ecological improvement evaluation results.

[0016] Optionally, the pre-built downstream river temperature prediction model is obtained through the following offline training steps: acquiring historical hydrological and meteorological sequences of the target watershed to construct a training sample set; constructing an initial machine learning model and defining the hyperparameter space to be optimized in the initial machine learning model; using a Bayesian optimization algorithm to optimize parameters in the hyperparameter space, and constructing a probabilistic surrogate model representing the mapping relationship between hyperparameters and model prediction performance based on historical evaluation results; iteratively selecting hyperparameter combinations to evaluate the performance of the initial machine learning model based on the probabilistic surrogate model and the acquisition function used to balance exploration and utilization; stopping the iteration when the preset training convergence condition is met, extracting the optimal hyperparameter combination at this time and fixing it into the initial machine learning model to obtain the pre-built downstream river temperature prediction model.

[0017] Optionally, the initial machine learning model is selected from at least one of the following: gradient boosting tree model, Gaussian process regression model, random forest model, and long short-term memory network model.

[0018] Beneficial effects: This invention improves the efficiency and reliability of downstream water temperature prediction and ecological target evaluation in long-term multi-objective scheduling. By extracting daily water level, daily outflow, inflow, reservoir capacity, and meteorological driving data from candidate scheduling schemes to construct a set of physical information-enhanced features, and inputting the complete scheduling period feature matrix into the downstream river water temperature prediction model for batch inference, the time consumed by water temperature extrapolation during high-frequency iterations of multi-objective evolutionary algorithms can be significantly reduced. By introducing the zero degree of gonadal development biology, the gonadal maturation accumulated temperature threshold, and the number of days of spawning, the spawning window can be dynamically locked according to the predicted water temperature sequence, avoiding time mismatch caused by fixed calendar evaluation. By using an asymmetric water temperature suitability function to distinguish between low-temperature and high-temperature deviation penalties, the ecological scheduling targets are more consistent with the physiological response differences of target fish to water temperature stress from different directions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the steps involved in constructing a scheduling model that incorporates multi-dimensional objectives and a water temperature prediction model for real-time simulation in this embodiment of the application.

[0020] Figure 2 This is a flowchart illustrating the steps for outputting the predicted water temperature sequence corresponding to the candidate scheduling scheme in this embodiment of the application.

[0021] Figure 3 This is a flowchart illustrating the steps involved in calculating the ecological scheduling target in an embodiment of this application.

[0022] Figure 4 This is a water temperature suitability curve for the asymmetric spawning of the target fish species in this embodiment of the application.

[0023] Figure 5 This is a flowchart illustrating the steps involved in offline training of the pre-built downstream river water temperature prediction model in this application embodiment. Detailed Implementation

[0024] Example 1, such as Figure 1 As shown, this embodiment details the process of regulating the water temperature of fish breeding habitats in the downstream river channel by combining a scheduling model incorporating multi-dimensional objectives with real-time simulations of a water temperature prediction model. In an optional implementation, the reservoir ecological scheduling method during the fish spawning season, which integrates water temperature prediction, specifically includes:

[0025] Step 101: Obtain biological characteristic data of the target fish, determine the suitable water temperature range for spawning of the target fish, and determine the parameters for the spawning window period.

[0026] The target fish species are those that exhibit physiological responses to changes in river water temperature and whose spawning activities are affected by temperature fluctuations, such as native fish species that lay adhesive or drifting eggs. Biological characteristic data can specifically include fish species composition, individual numbers, and biomass obtained through traditional fishing surveys, as well as data obtained through laboratory analysis such as body weight, gonadal weight, gonadal index, and distribution of gonadal developmental stages.

[0027] Specifically, in this embodiment, technicians set up representative survey sections in the reservoir area upstream of the dam and the river section downstream of the dam, collected samples using tools such as electrofishing and gillnets, and observed the maturity of oocytes or spermatogonia in conjunction with gonadal tissue sections. Based on this, statistical analysis was used to determine the water temperature range corresponding to the start of large-scale spawning in fish, thus obtaining the suitable spawning temperature range; and the biological zero point for gonadal development, the accumulated temperature threshold for gonadal maturation, and the duration of spawning were used as parameters to determine the spawning window. Specifically, the biological zero point for gonadal development was determined by the temperature corresponding to the start of effective gonadal development; the accumulated effective temperature threshold required for gonadal maturation was determined through years of monitoring or indoor temperature control experiments; and the duration of spawning was determined based on the measured spawning duration of the target fish species. For example, for target fish species downstream of a power station, the suitable spawning temperature range could be set to 18.0℃~22℃.

[0028] Step 102: Construct a multi-objective reservoir scheduling model, which includes power generation benefit objective, flood control risk objective, and ecological scheduling objective.

[0029] In this embodiment, the construction of a multi-objective reservoir scheduling model aims to coordinate engineering operational benefits with ecological protection needs. Specifically, the power generation benefit objective is calculated by determining the total power generation energy E of the power station during the scheduling period. total The formula for calculating the power generation energy in a single time period is as follows: E(t) = ρ * g * Qgen,t * Hnet,t * φ * Δt / (3.6 * 10^- ... 9 );

[0030] Where E(t) is the energy generated during time period t; ρ is the density of water, taken as 1000 kg / m³. 3 g is the acceleration due to gravity, with a value of 9.8 m / s². 2 Qgen,t represents the effective power generation flow rate in time period t; Hnet,t represents the effective head in time period t; φ represents the power generation efficiency, for example, a value of 0.84; Δt represents the calculation step size.

[0031] Accordingly, flood control risk targets are determined through the flood control risk index J. flood The calculation is performed using the following formula:

[0032] J flood (t)=Joutflow(t)+J level(t);

[0033] Wherein, Joutflow(t) represents the risk item of excessive outflow. When the outflow qout,t exceeds the preset flood warning flow Qwarn, a penalty value is calculated based on the excess amount: Joutflow(t) = max(qout,t - Qwarn,0); otherwise, it is 0. level (t) represents the risk of reservoir water level exceeding the limit. When the reservoir water level Ht exceeds the flood control water level H flood At that time, the penalty value is calculated based on the square of the water level deviation: J level (t)=α×max(Ht-H flood ,0) 2 Otherwise, take 0; where α is the risk dimension balance coefficient, used to adjust the water level exceeding the limit penalty to align with the power generation benefit target E. total For values ​​of the same magnitude, the specific values ​​can be determined based on the actual reservoir capacity characteristics and scheduling practices, through dimensional analysis or parameter verification methods.

[0034] In addition, the multi-objective reservoir scheduling model also includes preset constraints for verifying the physical feasibility of candidate scheduling schemes. These include: first, water balance constraints, where the reservoir capacity at a later time equals the reservoir capacity at the previous time plus the inflow and subtracting the outflow; second, water level change rate constraints, specifically limiting the water level rise rate to ≤2.17 m / day and the fall rate to ≤1.09 m / day, to avoid water level fluctuations affecting dam safety or the ecological environment; third, power station output constraints, for example, a power station's guaranteed output is set at 170 MW, and its installed capacity is set at 800 MW, with the actual output needing to fall within the aforementioned range; and fourth, outflow constraints, ensuring that the outflow is not less than the minimum ecological base flow, for example, 88 cubic meters per second.

[0035] In some embodiments, the preset constraints include: water balance constraints based on the relationship between inflow, outflow and reservoir capacity changes; upper and lower limits of reservoir capacity and water level based on flood control requirements and dead water level settings; water level change rate constraints to limit the rate of rise and fall of water level during reservoir temperature regulation operations; upper and lower limits of outflow based on downstream ecological base flow and downstream flood control safety; and power station output constraints based on the guaranteed output and installed capacity of the hydropower station.

[0036] Step 103: Use a multi-objective optimization algorithm to iteratively solve the multi-objective reservoir scheduling model and generate multiple sets of candidate scheduling schemes.

[0037] Specifically, this embodiment employs the NSGA-II non-dominated sorting genetic algorithm with an optimal individual retention strategy for iterative optimization. The system randomly initializes the population within the feasible region of the decision variables, with each individual representing a set of water level and flow combinations covering 365 days of the year. During each generation of evolution, the algorithm performs a fast non-dominated sorting of multiple candidate scheduling schemes within the current population based on power generation efficiency, flood control risk, and ecological scheduling objectives. To maintain solution diversity, the system also calculates the crowding distance between individuals. Offspring candidate scheduling schemes are generated through selection, crossover, and mutation operations. During this process, the algorithm filters individuals according to preset constraints; individuals that violate hard constraints are assigned low fitness values ​​or are eliminated.

[0038] In some embodiments, generating multiple sets of candidate scheduling schemes includes: randomly initializing an initial population within the feasible region of the decision variables of a multi-objective reservoir scheduling model; performing iterative evolution of the initial population using a non-dominated sorting genetic algorithm with an optimal individual retention strategy; in each generation of evolution, performing fast non-dominated sorting and congestion distance calculation on multiple sets of candidate scheduling schemes within the current population based on power generation benefit objectives, flood control risk objectives, and ecological scheduling objectives; generating offspring candidate scheduling schemes based on the sorting and calculation results through selection, crossover, and mutation operations; stopping iteration when the objective function value meets a preset convergence condition, extracting a Pareto optimal solution set composed of non-dominated solutions, and selecting the optimal scheduling scheme from the Pareto optimal solution set according to decision preferences.

[0039] Step 104: During the iterative solution process, extract the set of physical information enhancement features for each set of candidate scheduling schemes, input the set of physical information enhancement features into the pre-built downstream river water temperature prediction model, and output the predicted water temperature sequence corresponding to the candidate scheduling scheme.

[0040] Optionally, when evaluating the ecological performance of each candidate scheduling scheme, the system dynamically extracts the corresponding time-series features. Specifically, the physical information-enhanced feature set not only includes water level and flow rate from the decision variables, but also integrates external meteorological driving data and constructed variables reflecting thermodynamic characteristics. These features are organized into an input matrix of predetermined dimensions and input into the downstream channel water temperature prediction model. The prediction model performs a nonlinear mapping on the input data and outputs daily downstream cross-sectional water temperature data during the scheduling period, forming a predicted water temperature sequence.

[0041] Step 105: Based on the predicted water temperature sequence, suitable water temperature range for spawning, and parameters determined by the spawning window period, determine the spawning window period corresponding to the candidate scheduling scheme, calculate the ecological scheduling objective of the candidate scheduling scheme, and output the optimal scheduling scheme from multiple candidate scheduling schemes by combining the power generation benefit objective, flood control risk objective, and ecological scheduling objective.

[0042] In this embodiment, the system first calculates the cumulative effective accumulated temperature based on the predicted water temperature sequence and the zero degree of gonadal development biology. The moment when the cumulative effective accumulated temperature first reaches the gonadal maturation accumulated temperature threshold and the predicted water temperature on that day reaches the lower limit of the suitable spawning temperature is determined as the starting point of the spawning window. The ending point of the spawning window is then determined based on the number of days of spawning. Subsequently, the system calculates the water temperature suitability score only within this spawning window. The algorithm stops iterating when it reaches a preset convergence condition, such as when the Pareto front no longer moves significantly in 100 consecutive generations of evolution. Finally, from the generated Pareto optimal solution set, the scheme with the best overall performance in terms of power generation efficiency, flood control risk, and ecological scheduling objectives is selected based on decision preferences, and its water level trajectory and outflow trajectory during the scheduling period are output.

[0043] Example 2: This example details how multi-dimensional biological monitoring and experimental analysis are used to determine the target species for ecological regulation and the corresponding reproductive water temperature parameters. In one optional implementation, biological characteristic data of the target fish species are acquired, and the suitable spawning water temperature range, the biological zero point for gonadal development, the accumulated temperature threshold for gonadal maturation, and the duration of spawning are determined, specifically including:

[0044] Fish sample data under different temperature and current conditions were collected, and a biological feature library characterizing the reproductive and developmental process was constructed using body weight, gonad weight, gonad index, and histological section data. Target fish species significantly affected by water temperature and habitat fluctuations were identified, and the differences in water temperature requirements during the critical period of gonadal development and the spawning triggering period were analyzed. The water temperature range that meets the requirements for normal gonadal development and can induce spawning behavior was taken as the suitable water temperature range for spawning of target fish species, and the biological zero degree for gonadal development, the accumulated temperature threshold for gonadal maturation, and the number of days of spawning were further determined.

[0045] Specifically, in this embodiment, technicians conduct fish resource sampling in the downstream river section significantly affected by the low-temperature water released from the reservoir, and in the upstream reference river section less affected by the reservoir. The sampling frequency is dynamically adjusted according to the fish breeding season, for example, increasing the sampling frequency during March to July each year. After obtaining fish samples through methods such as electrofishing and gillnetting, morphological parameters such as total length, body length, and body weight of the individuals are measured on-site or in the laboratory. Furthermore, the gonadal organs of the fish are obtained through dissection and weighed; the specific formula for calculating the gonadal index is as follows:

[0046] GSI=(m g / m b )*100;

[0047] Among them, GSI is the gonadal index, used to quantitatively describe the proportion of total energy expenditure in fish gonadal development; m g For the gonadal quality of fish samples; m b The body mass of the fish sample.

[0048] Based on this, technicians fixed, dehydrated, embedded, sectioned, and stained the collected gonadal tissue to prepare histological sections. The developmental stages of oocytes or spermatogonia were observed under a microscope, for example, gonadal development was classified into stages 1-6. Biological characteristic data consisted of a calculated monthly mean sequence of gonadal indices, a sequence of gonadal quality changes, and the proportion of individuals at each developmental stage. By comparing the developmental progress of the downstream river section with a reference section at the same time point, if gonadal development of a certain species was found to be stagnant, degenerate, or delayed in the low-temperature environment downstream of the dam, the species was determined to be highly sensitive to water temperature changes and identified as the target fish species. Combining the measured water temperature records at the time of spawning behavior of the target fish, the water temperature range corresponding to the transition from gonadal development stage 4 to stage 5 or 6 was extracted, ultimately determining the suitable water temperature range for spawning; the biological zero degree of gonadal development was determined based on the low-temperature boundary at which gonadal development clearly begins; the accumulated temperature threshold for gonadal maturation was determined based on the correspondence between the daily effective water temperature accumulation and the gonadal maturation node; and the number of days of spawning duration was determined based on continuous spawning records.

[0049] Furthermore, in some optional implementations, in addition to applying gonadal indices and histological section analysis, blood steroid hormone levels can be introduced as a supplementary biological characteristic data. Specifically, the system can collect serum samples from the target fish and measure the levels of estradiol or testosterone. By analyzing the fluctuation patterns of hormone levels with water temperature changes, the adaptability range of the target fish to temperature fluctuations can be further verified at the molecular endocrine level, improving the accuracy of determining the suitable spawning water temperature range.

[0050] Furthermore, the selection logic for target fish species can be weighted and adjusted for rivers with different hydrological characteristics. For example, if a species exhibits high intolerance to water temperature fluctuations during gonadal development, it can be prioritized as the primary optimization target of the scheduling scheme. Through this multi-level identification and threshold extraction mechanism, the suitable spawning water temperature range and the spawning window determination parameters, composed of the biological zero degree of gonadal development, the accumulated temperature threshold for gonadal maturation, and the duration of spawning, can be provided for subsequent multi-objective scheduling models.

[0051] Example 3: This example details the mathematical analysis of the reservoir thermodynamic mechanism, transforming conventional raw hydrological and meteorological variables into surrogate indicators with physical meaning, thereby improving the generalization ability of the prediction model under complex operating conditions.

[0052] Optionally, the physical information enhancement feature set includes at least the following feature categories: thermal stratification intensity features for characterizing the vertical heat distribution gradient inside the reservoir; hydraulic retention features for characterizing the water exchange rate and heating time inside the reservoir; and adaptive cumulative heat input features for characterizing the cumulative impact of previous meteorological conditions on the current outflow water temperature.

[0053] In constructing surrogate prediction models, the input features sometimes consist only of raw time-series data such as water level, flow rate, temperature, and radiation. This requires the model to learn the physical coupling relationships between these variables through a black-box network. This embodiment introduces prior physical knowledge to construct the aforementioned three types of features, thereby reducing the model's dependence on the amount of training data and improving prediction accuracy.

[0054] Specifically, the thermal stratification intensity feature aims to calculate the potential vertical temperature gradient where the reservoir surface receives solar radiation heating while the bottom water temperature remains low; the hydraulic retention feature measures the timescale of water mass residence within the reservoir, which determines the time integral effect of heating of the water mass; and the adaptive cumulative heat input feature overcomes the shortcomings of the traditional fixed sliding window, dynamically adjusting the backtracking length of previous meteorological factors according to real-time hydrological conditions. In an optional implementation, the thermal stratification intensity feature, hydraulic retention feature, and adaptive cumulative heat input feature are constructed through the following steps:

[0055] Optionally, the system can acquire pre-collected meteorological driving data, daily inflow and daily reservoir capacity for the target watershed, and extract daily water level and daily outflow from candidate scheduling schemes.

[0056] In this embodiment, the meteorological driving data may specifically include air temperature, solar shortwave radiation, wind speed, and relative humidity. This type of data typically originates from meteorological observation stations within the watershed or from reanalysis datasets. In each evaluation of the multi-objective optimization iteration, for each predetermined individual, i.e., candidate scheduling scheme, the system parses the daily water level and daily outflow rate set by the scheme, and combines this with the externally input daily inflow rate and daily reservoir capacity to provide basic data support for the subsequent calculation of physical characteristics.

[0057] Optionally, the positive temperature deviation between the current day's temperature and the preset annual average temperature in the meteorological driving data is calculated; the positive temperature deviation is multiplied by the current day's water level, and the ratio of the product to the total flow is used as the thermal stratification intensity feature, where the total flow is the sum of the current day's inflow and the current day's outflow.

[0058] Specifically, the formula for calculating the intensity characteristics of thermal stratification is as follows:

[0059] SI t =((T air,t -T air,year ) + *H t ) / (q out,t +q in,t +ε);

[0060] Among them, SI t For the thermal stratification intensity characteristics on day t, T air,t Let T be the temperature on day t.air,year The preset average annual temperature, (T) air,t -T air,year ) + Indicates when T air,t Greater than T air,year Take the difference between the two values ​​when T air,t Less than or equal to T air,year When H is 0, t Let q be the water level on day t. out,t q represents the daily discharge flow on day t. in,t Let t be the daily inbound flow rate on day t, and ε be a preset small positive number to prevent division by zero anomalies.

[0061] Specifically, the numerator characterizes the driving force behind thermal stratification. The reservoir surface absorbs additional heat only when the temperature is above the annual average, hence a positive calculation is used. Higher temperatures and deeper water levels result in more pronounced vertical thermal stratification. The denominator characterizes the hybrid forces that disrupt thermal stratification; larger inflow and outflow rates lead to more intense water mixing and a greater weakening of thermal stratification. By introducing a small constant ε, for example, setting ε=1.0, the abnormal surge in eigenvalues ​​caused by the denominator approaching zero during the dry season and when the power station is completely shut down can be prevented. For example, under summer water storage conditions with higher temperatures, deeper water levels, and moderate outflow rates, SI t Taking a larger value, the model can identify that a significant thermal stratification has formed on the reservoir surface; however, during the high-flow discharge period of flood season, SI t The temperature approaches zero, reflecting the disruption of thermal stratification and the near-uniform mixing of the outflow water temperature, which is consistent with the actual physical process.

[0062] Optionally, the ratio of the reservoir capacity to the outflow on that day can be calculated as a hydraulic retention characteristic, and the calculation formula is as follows:

[0063] RT t =V t / ((q out,t +ε)*86400);

[0064] Among them, RT t V represents the hydraulic retention characteristics on day t. t Let q be the daily storage capacity on day t. out,t Let t represent the daily outflow rate on day t, 86400 represent the number of seconds per day, and ε represent the zero constant for flood control. This characteristic is used to calculate the average residence time of the water mass from inflow to outflow. RT t The larger the value, the more likely the reservoir is in a state of water storage or the discharge is small. The water in the reservoir has more time to absorb solar shortwave radiation, the less the outflow water temperature is affected by the weather changes of the day, and the stronger the lag effect.

[0065] Optionally, the number of historical impact days is determined based on the numerical value of the hydraulic retention characteristics, a dynamic accumulation window is constructed using the number of historical impact days, and within the dynamic accumulation window, the normalized historical shortwave radiation and historical temperature extracted from the meteorological driving data are accumulated to obtain the adaptive cumulative heat input characteristics.

[0066] Specifically, the construction process of the adaptive cumulative heat input feature is as follows: The hydraulic retention feature RT... t Round down to the nearest integer as the initial cumulative number of days. To prevent the window from extending indefinitely and introducing noise, set an upper limit threshold, such as 60 days. Take the smaller of these two values ​​as the final historical impact days n. t Within a defined dynamic accumulation window [tn] t The accumulation is performed within the range of [,t]. The accumulation calculation formula can be expressed as follows: add the ratio of the historical shortwave radiation of each day to its maximum value within the year, and the ratio of the historical temperature of each day to its highest temperature within the year, and then sum the sums of all dates within the window to obtain the adaptive cumulative heat input characteristics.

[0067] Normalization mitigates the differences in dimensions and numerical magnitudes between shortwave radiation and air temperature. This type of adaptive window design can dynamically adjust the backtracking length according to water flow: during the high-water season, the residence time is short, n t The automatic shrinkage reflects that the water temperature is mainly controlled by recent meteorological conditions; during the dry season, the retention time is long, n t With automatic extension, the model is able to capture the cumulative effect of heat over a longer time span.

[0068] Furthermore, in some optional implementations, the physical information enhancement feature set may also include inflow water temperature lag features. When the watershed has an upstream water temperature online monitoring station, the lag time parameter between inflow and outflow water temperatures can be extracted through correlation analysis, and the delayed inflow water temperature sequence can be incorporated into the input feature matrix, further improving the model's accuracy in describing the upstream and downstream heat transfer process.

[0069] Example 4: This example details how, during the iterative solution process, a set of enhanced physical information features is extracted for each group of candidate scheduling schemes. This enhanced physical information feature set is then input into a pre-constructed downstream river water temperature prediction model, outputting the predicted water temperature sequence corresponding to the candidate scheduling scheme, such as... Figure 2 As shown, it specifically includes:

[0070] Step 401: Extract the daily water level sequence and daily outflow sequence of the candidate scheduling scheme during the complete scheduling period, and obtain the pre-stored daily meteorological driving sequence.

[0071] In this embodiment, for each individual in the non-dominated sorting genetic algorithm population, i.e., a candidate scheduling scheme, the scheme essentially contains a sequence of decision variables covering the entire computation cycle. In this embodiment, the complete scheduling period is in days, covering the entire computation cycle set by the scheduling task. The system parses the daily water level sequence and daily outflow sequence with 365 consecutive time steps from the scheme. Simultaneously, the system reads a fixed daily meteorological driving sequence. The daily meteorological driving sequence specifically includes the daily average temperature and daily average shortwave radiation. Since all candidate schemes in the population face the same external meteorological boundary conditions, this driving sequence is kept constant in memory during a single optimization task to reduce the time overhead caused by repeated data read / write operations.

[0072] Step 402: For each day within the complete scheduling period, combine the water level, outflow, meteorological driving factors, and historical lag states of the preceding period read from the daily water level sequence, daily outflow sequence, and daily meteorological driving sequence to calculate and generate the physical information enhancement feature set for that day.

[0073] Specifically, for a predetermined date with time index t, the system reads the water level, outflow, temperature, and shortwave radiation for that day. To capture the time inertia effect of hydrological transport and heat response, the system extracts the historical lag states of preceding time periods.

[0074] For example, the system extracts the water level and flow rate of the previous day (t-1), and calculates the moving average of the outflow and the moving average of the temperature over the previous 7 days. Furthermore, to implicitly encode the seasonal cyclical distribution pattern, a day number parameter is introduced, with a value range of 1 to 365. Based on this, the system calculates the thermal stratification intensity characteristics, hydraulic retention characteristics, and adaptive cumulative heat input characteristics for the current day. The system concatenates and combines the variable data from these various dimensions into a one-dimensional feature vector, serving as a set of physical information enhancement features.

[0075] Step 403: Concatenate the daily physical information enhancement feature sets within the complete scheduling period into an input feature matrix according to the time sequence.

[0076] After generating the daily one-dimensional feature vector, the system stacks these feature vectors as independent data rows in an increasing time sequence from 1 to 365. This operation combines the vector sequence, originally discretely distributed across the cyclic time steps, into a two-dimensional floating-point array, i.e., the input feature matrix. For example, if a single one-dimensional feature vector contains 12 specific features, the final constructed matrix will have 365 rows and 12 columns.

[0077] Step 404: Input the input feature matrix into the downstream river water temperature prediction model for a batch inference to obtain a predicted water temperature sequence covering the entire scheduling period.

[0078] Optionally, conventional sequence prediction, due to its time-step iterative calling pattern, can lead to a significant increase in total computation time in multi-objective optimization scenarios with population sizes reaching hundreds and iterations exceeding tens of thousands, failing to meet the timeliness requirements of engineering scheduling. In this embodiment, the system treats the assembled input feature matrix as a whole as a computational input batch, passing it through in a single pass to the downstream river water temperature prediction model already loaded into the runtime environment. The downstream river water temperature prediction model executes internal parallel operation instructions such as matrix multiplication, outputting a result array containing 365 predicted water temperature values, i.e., the predicted water temperature sequence.

[0079] Furthermore, in a specific algorithm engineering configuration scenario, the population size of the multi-objective optimization algorithm is set to 400, and the maximum number of evaluations is set to 50,000. The system takes approximately 0.5 milliseconds to execute a batch inference operation containing 365 samples. Calculations show that the total global time for 50,000 model batch inference calls remains within 25 seconds. This data flow mechanism ensures that the inference time of the proxy model is lower than the computational overhead of the genetic operation itself, achieving a closed loop for ecological assessment target calculation while maintaining the accuracy of water temperature prediction.

[0080] Example 5: This example details the calculation of the ecological scheduling objective in a multi-objective reservoir scheduling model. Based on the predicted water temperature sequence, the suitable spawning water temperature range, and the parameters for determining the spawning window, the spawning window period corresponding to the candidate scheduling scheme is determined, and the ecological scheduling objective is calculated. Figure 3 As shown, it specifically includes:

[0081] Step 501: Calculate the cumulative effective temperature sequence on the time axis based on the predicted water temperature sequence and the preset gonadal development biology zero degree.

[0082] Correspondingly, in natural ecosystems, fish reproductive rhythms are driven by the cumulative effect of water temperature, rather than being solely governed by a fixed calendar time. In this embodiment, the biological zero degree of gonadal development refers to the baseline starting temperature in the calculation of effective accumulated temperature for the target fish, hereinafter referred to as biological zero degree, which can be determined based on indoor controlled experiments of the target fish. In this embodiment, the date on which the gonads of the target fish begin to develop is taken as the starting point t0 of accumulated temperature. According to the reproductive biology characteristics of fish, gonadal development begins when the water temperature steadily rises above the biological zero degree of gonadal development in spring; for example, t0 is taken as March 1st each year. The effective accumulated temperature is defined as:

[0083] ;

[0084] in, The predicted water temperature (°C) of the downstream river channel is output by the downstream river channel water temperature prediction model on day t. The biological zero point (°C) for gonadal development in the target fish is the temperature below which gonadal development is halted. The value is 1 day, which is the time step.

[0085] The system uses the start date of the scheduling period as the zero point of time and extracts the predicted water temperature values ​​for each day. For each day, the system determines whether the predicted water temperature is greater than the gonadal development biological zero degree. If so, the difference between the water temperature and the gonadal development biological zero degree is calculated as the effective accumulated temperature for that day; otherwise, the effective accumulated temperature for that day is set to 0. The system sequentially accumulates the effective accumulated temperatures of each day along the time axis, generating a cumulative effective accumulated temperature sequence of the same length as the number of scheduling days.

[0086] Step 502: The start time is locked by comparing the accumulated effective temperature with the gonadal maturation temperature threshold, the predicted water temperature of the day with the lower limit of the suitable water temperature for spawning, and the end time is calculated by extending the preset number of spawning days, so as to dynamically determine the spawning window period corresponding to the candidate scheduling scheme.

[0087] Preferably, to address the misalignment between the observation period and the actual breeding period in years of abnormal water temperature caused by the use of static calendar months in traditional ecological scheduling evaluations, this embodiment dynamically determines the spawning window period corresponding to each candidate scheduling scheme based on the gonadal maturation accumulated temperature threshold and the number of days of spawning duration in the spawning window determination parameters. The system pre-stores reproductive parameters obtained through ecological measurements, among which the core parameters include the accumulated temperature threshold required to trigger spawning (i.e., gonadal maturation accumulated temperature) and the typical number of days of spawning duration for the fish species. These parameters can be obtained through years of field monitoring and gonadal histological analysis. Furthermore, when determining the window, the system no longer searches for two independent accumulated temperature thresholds, but instead performs a time-series retrieval of the cumulative effective accumulated temperature sequence based on the water temperature evolution curve generated by the candidate scheduling scheme itself: the moment when both the accumulated temperature and water temperature conditions are met simultaneously for the first time is taken as the start time of the spawning window, and the end time is calculated by adding a fixed number of days of spawning duration to the start time. Each scheme adaptively determines the ecological evaluation period based on its water temperature process, rather than using a fixed calendar interval.

[0088] The specific definition is as follows: When the effective accumulated temperature first reaches the gonadal maturation accumulated temperature threshold... At this time, the fish's gonads have matured and are physiologically ready to spawn. The spawning window is defined as follows:

[0089] ;

[0090] in, The accumulated temperature threshold (°C·d) required for the gonadal development and maturation of fish. The lower limit of the suitable water temperature for spawning (°C). The significance of setting the above dual conditions is that: if the accumulated temperature reaches the standard but the water temperature on the day is lower than the critical value, although the fish have the physiological ability to spawn, they lack the triggering conditions, and the spawning behavior will not actually occur.

[0091] The end time of the spawning window is defined as:

[0092] ;

[0093] in, This represents the number of days the spawning activity lasts, determined based on the reproductive characteristics of fish.

[0094] Step 503: Extract water temperature segments from the predicted water temperature sequence during the spawning window.

[0095] Accordingly, after obtaining the start and end times, the system performs an array slicing operation on the predicted water temperature sequence covering the entire year, extracting a subsequence specifically corresponding to the actual spawning cycle of fish, thus obtaining the water temperature segment. Subsequent suitability evaluation will only be performed within this water temperature segment, eliminating the interference of non-breeding season water temperature fluctuations on the objective function evaluation.

[0096] Step 504: Calculate the degree of deviation of the water temperature value within the water temperature segment from the suitable water temperature range for spawning, and calculate the suitability score based on the degree of deviation. Use the cumulative suitability score within the spawning window as the ecological scheduling target.

[0097] In this embodiment, the ecological assessment focuses on calculating the supporting capacity of water temperature conditions for the reproduction of the target fish species. The system reads water temperature values ​​within a daily temperature segment and calculates the absolute or relative difference between these values ​​and the preset optimal spawning temperature range. The greater the deviation, the less suitable the thermodynamic environment is for egg cell hatching and survival on that day, resulting in a lower suitability score. The system sums the suitability scores for each day within the spawning window and outputs the cumulative value as the ecological scheduling objective used by the multi-objective optimization algorithm during iterative evolution.

[0098] In one optional implementation, the deviation of the water temperature value within a water temperature segment from the suitable spawning water temperature range is calculated, and a suitability score is calculated based on the deviation. Specifically, an asymmetric water temperature suitability function is applied for processing, including:

[0099] Optionally, it can be determined whether the predicted water temperature on the day during the spawning window is lower than the lower limit of the suitable spawning water temperature, within the suitable spawning water temperature range, or higher than the upper limit of the suitable spawning water temperature.

[0100] Because fish live in complex aquatic environments, their physiological functions exhibit different tolerances to upper and lower temperature limits. To reflect this biological specificity, the system introduces directional judgment logic before calculating the daily fitness score. The system uses numerical comparison calculations to detect the relationship between the water temperature value corresponding to the time step of the day and the lower and upper limits of the suitable spawning water temperature, distinguishing whether the water temperature environment on that day is deviating from the low-temperature side, within the suitable range, or deviating from the high-temperature side.

[0101] Optionally, if the predicted water temperature is lower than the lower limit of the suitable water temperature for spawning, the low-temperature side suitability score is calculated using a preset left deviation penalty parameter.

[0102] Specifically, when a low-temperature deviation is detected, the system executes the following calculation formula: S low =S base -a L *(T lower -T pred );

[0103] Among them, S low For the calculated low-temperature suitability score, S base For the pre-configured daily fitness score benchmark, a L The preset left deviation penalty parameter was determined by fitting the tolerance curve of the target fish through an indoor temperature-controlled experiment. lower To meet the lower limit of the suitable water temperature for spawning, T pred The system predicts the water temperature for the day. Through the above calculations, the system applies a linear deduction to water temperatures below a threshold.

[0104] Optionally, if the predicted water temperature for the day is within the suitable water temperature range for spawning, the suitability score for that day is determined as the baseline value for the full suitability score for that day.

[0105] Specifically, when the system determines through numerical comparison that Tlower ≤ Tpred(t) ≤ Tupper, it indicates that the thermodynamic environment of that day meets the hatching and reproductive development requirements of the target fish eggs, and there is no cold or heat stress. At this time, the system does not invoke the penalty parameter and executes the algorithm: Sopt = Sbase; where Sopt is the calculated optimal state score. Through this logic, the system provides sufficient full-score incentives for the downstream river water temperature conditions within the optimal temperature range, thereby guiding the scheduling model to ensure that the downstream river is within the optimal temperature range as much as possible during optimization.

[0106] Optionally, if the predicted water temperature for the day is higher than the upper limit of the suitable water temperature for spawning, the high-temperature side suitability score is calculated using a preset right-side deviation penalty parameter.

[0107] Accordingly, when a deviation from high temperature is determined, the system executes the following calculation formula: Shigh=Sbase-aR*(Tpred-Tupper); where Shigh is the calculated high-temperature side suitability score, Sbase is the pre-configured daily suitability full score baseline value, aR is the preset right deviation penalty parameter, which is determined by fitting the tolerance curve of the target fish through indoor temperature control experiments, Tpred is the predicted water temperature for the day, and Tupper is the upper limit of the suitable water temperature for spawning.

[0108] Optionally, the values ​​of the left deviation penalty parameter and the right deviation penalty parameter are not equal, and heterogeneous penalties are applied to water temperature deviations in different directions. The daily suitability score is determined based on the low temperature side suitability score, the full score within the interval, or the high temperature side suitability score.

[0109] In this embodiment, technicians extracted the tolerance extreme boundary of the target fish through indoor controlled experiments and transcriptome analysis. Based on this, the left deviation penalty parameter and the right deviation penalty parameter were set to be numerically asymmetrical.

[0110] For example, assuming the suitable spawning temperature range for the target fish is 18.0~22℃, and the maximum daily suitability score is 10.0 points. The current predicted water temperature is 16.0℃, below the lower limit of the suitable spawning temperature, with a temperature deviation of 2.0℃. The system calls the preset left deviation penalty parameter, with a value of 1.5. Substituting this into the formula, the low-temperature suitability score is calculated to be 10.0 - 1.5 * 2.0 = 7.0 points. In another scenario, the current predicted water temperature is 24.0℃, above the upper limit of the suitable spawning temperature, with the same temperature deviation of 2.0℃. Since research shows that this target fish responds more severely to high-temperature stress, the system calls the preset right deviation penalty parameter, with a value of 2.5. Substituting this into the formula, the high-temperature suitability score is calculated to be 10.0 - 2.5 * 2.0 = 5.0 points. It can be seen that for cases where the absolute deviation is 2.0℃, the system implemented a heterogeneity penalty based on biological mechanisms. The final high-temperature suitability score was significantly lower than the low-temperature score, forcing the optimization scheduling model to more proactively avoid potential high-temperature runoff risks during the evolutionary process. Conversely, if the current predicted water temperature is 20.0℃, within the suitable range of 18.0~22.0℃, the system determines there is no temperature deviation and assigns a perfect suitability score of 10.0 for that day. The specific water temperature suitability for asymmetric spawning of the target fish species is as follows: Figure 4 As shown in the figure. Point a is the low-temperature penalty point, corresponding to a water temperature of 16℃ and a suitability score of 7; point b is the high-temperature severe penalty point, corresponding to a water temperature of 24℃ and a suitability score of 5.

[0111] Furthermore, in some alternative implementations, for specific fish species exhibiting strong nonlinear biological responses, the linear function used to calculate the low-temperature and high-temperature fitness scores can be replaced with an exponential penalty function. That is, the penalty term is implemented by calculating the exponential power of the deviation, causing the penalty to increase rapidly as the water temperature approaches the lethal limit, further strengthening the model's ecological constraint performance.

[0112] Example 6: This example details the offline pre-training mechanism of the prediction model, optional feature contribution verification methods, and the post-event effect evaluation process of ecological scheduling based on scenario replay. In an optional implementation, such as... Figure 5 As shown, the pre-built downstream river temperature prediction model is obtained through the following offline training steps:

[0113] Optionally, historical hydrological and meteorological sequences of the target watershed can be obtained to construct a training sample set.

[0114] In this embodiment, the historical hydrological and meteorological sequence refers to long-term measured data obtained in advance from the watershed hydrological database or meteorological observation station since the dam's construction. This type of data specifically covers daily inflow, outflow, dam-front water level, measured water temperature, air temperature, and solar shortwave radiation, among other physical quantities. The system performs missing value imputation and outlier removal on the acquired data, and then divides it into non-overlapping datasets according to chronological order or random sampling. For example, the system divides the historical hydrological and meteorological sequence into a training set for model fitting, a validation set for hyperparameter optimization, and a test set for final generalization error evaluation, respectively, at proportions of 70%, 15%, and 15%.

[0115] Optionally, an initial machine learning model is constructed, and the hyperparameter space to be optimized for the initial machine learning model is defined.

[0116] Furthermore, due to the nonlinearity and time lag inherent in reservoir thermodynamic processes, traditional empirical formulas for water temperature regression are insufficient to achieve the desired fitting accuracy. Therefore, the system initializes an untrained computing network, i.e., the initial machine learning model, during the offline phase. To control model complexity and prevent overfitting, technicians need to configure basic parameter boundaries for this model. The system defines a multi-dimensional numerical interval matrix in memory as the hyperparameter space to be optimized. The training objective function of the initial machine learning model uses the mean squared error (MSE) loss function, fitting the model parameters by minimizing the squared mean of the difference between the predicted and measured water temperatures. For the Long Short-Term Memory (LSTM) network model, the Backpropagation Theorem (BPTT) algorithm combined with the Adam adaptive moment estimator optimizer is used for parameter updates. Technicians can adaptively adjust training hyperparameters such as the learning rate according to the actual data scale and convergence requirements. For example, for tree-structured algorithms, this space may include a tree depth range limited to 3–10 and a learning rate range limited to 0.01–0.3; for time series networks, this space may include a preset range such as a hidden layer node number limit of 32–256.

[0117] Optionally, the initial machine learning model is selected from at least one of the following: gradient boosting tree model, Gaussian process regression model, random forest model, and long short-term memory network model.

[0118] To address the technical challenges of complex water temperature prediction, this embodiment provides four specific parallel alternative computational architectures. Specifically, when using a gradient boosting tree model, the system continuously fits the residuals of previous predictions by sequentially constructing multiple decision trees, capturing nonlinear mapping patterns. When using a Gaussian process regression model, the system utilizes a radial basis function kernel to measure the covariance relationship between input feature vectors, outputting not only the mean water temperature prediction but also the prediction variance to characterize uncertainty. When using a random forest model, the system constructs multiple independent decision trees through sampling with replacement and takes their average value, thereby reducing prediction variance. When using a long short-term memory network model, the system utilizes internal forget gates and input gate state units to autonomously retain or discard historical temporal features in the sequence, effectively handling long-term hydrological lag effects. In actual deployment, the system can use the model with the highest cross-validation score as the final operating architecture.

[0119] Optionally, a Bayesian optimization algorithm is used to optimize parameters in the hyperparameter space, and a probabilistic surrogate model representing the mapping relationship between hyperparameters and model prediction performance is constructed based on the initial random sampling evaluation results and the accumulated historical evaluation results.

[0120] Correspondingly, traditional parameter optimization methods such as grid search incur significant computational overhead when dealing with high-dimensional parameter spaces. To improve offline construction efficiency, the system introduces a Bayesian optimization algorithm. This algorithm treats the hyperparameter evaluation process as an optimization of a black-box objective function. After each evaluation, the system records the current hyperparameter combination and its root mean square error on the validation set. Based on historical evaluation results, the system typically uses a Gaussian process to construct a probabilistic surrogate model. This surrogate model can predict the expected error and its confidence interval at any point in the hyperparameter space, constructing a smooth mapping surface.

[0121] Optionally, based on the probabilistic surrogate model and a pre-defined acquisition function for balancing exploration and exploitation, the hyperparameter combination is iteratively selected to evaluate the performance of the initial machine learning model, and the performance evaluation results are fed back as new historical evaluation results to update the posterior distribution of the probabilistic surrogate model.

[0122] Accordingly, in the decision-making stage of each iteration of optimization, the system calculates the value of the acquisition function. Specifically, the acquisition function can employ a confidence upper bound strategy or an expected improvement strategy. The system determines the next set of hyperparameter coordinates to be tested by maximizing the value of the acquisition function. This calculation process balances the development process utilizing known low-error regions with the trial-and-error process exploring high-uncertainty unknown regions. After selecting the parameters, the system configures them into the initial machine learning model for retraining and validation set error evaluation, and feeds back the newly generated error results to update the posterior distribution of the probabilistic surrogate model.

[0123] Optionally, when the preset training convergence condition is reached, the iteration stops, and the optimal hyperparameter combination at this time is extracted from the historical evaluation results and solidified into the initial machine learning model to obtain the pre-built downstream river water temperature prediction model.

[0124] Furthermore, the system monitors the termination marker of the iteration process in real time. Preset training convergence conditions can be set such that the algorithm reaches its maximum number of evaluations, for example, 100, or that the decrease in validation set error is less than 0.001 over 10 consecutive iterations. Once these conditions are met, the optimization loop terminates. The system extracts the set of hyperparameter configurations with the smallest corresponding error from historical records, burns them into the network structure as fixed parameters, and finally outputs a pre-built downstream river water temperature prediction model that can be called in real time.

[0125] Optionally, to verify the rationality of the physical information-enhanced feature set, the Shapley value attribution method can be used to decompose the prediction results of the pre-built downstream river water temperature prediction model into the independent contributions of each feature in the training sample set. This step serves as a model interpretation and feature verification step and does not limit the training method of the downstream river water temperature prediction model. Optionally, machine learning models often have black-box properties, making it difficult to intuitively explain the degree of influence of each input feature on the prediction output. To verify the effectiveness of the physical features, the system performs feature attribution calculations offline. The Shapley value attribution method originates from cooperative game theory. Its mechanism is to regard the prediction model as a cooperative game process involving multiple features, and to fairly distribute the difference between the predicted water temperature output of the model and the baseline expected value as the total payoff to each input feature component participating in the calculation.

[0126] Optionally, based on the independent contributions of each feature, a recursive path calculation or local linear approximation is used to solve the marginal contribution gain of each physical information enhancement feature in the physical information enhancement feature set under different feature subsets to the predicted water temperature.

[0127] As an alternative implementation, the computational complexity increases exponentially with dimensionality because calculating the Shapley value requires traversing all possible permutations of features. The system provides differentiated alternative computation strategies based on the selected model type. When the initial machine learning model uses a tree structure, such as a gradient boosting tree, the system extracts the node splitting information of the tree and parses the assigned weights of the features through top-down recursive path operations. When using other black-box models such as Long Short-Term Memory networks, the system employs a kernel interpreter based on local linear approximation. Specifically, the system randomly perturbs the original input features to generate neighboring samples and assigns weights using a kernel function with exponential decay characteristics. By solving the weighted least squares regression equation, the system estimates the predicted water temperature change caused by adding each physical feature to the feature subset, i.e., the marginal contribution gain.

[0128] Optionally, the rationality of the set of physical information enhancement features can be verified by calculating and evaluating the relative importance of each physical information enhancement feature in water temperature prediction based on the marginal contribution gain output.

[0129] As an optional implementation, the system takes the absolute value of the marginal contribution gain of all samples in the test set and calculates the average to obtain a global-scale importance score. The system then outputs all features in descending order of importance score. If the evaluation results show that the thermal stratification intensity feature and the adaptive cumulative thermal input feature rank highly in the ranking list, it objectively confirms from a data-driven perspective that these physically derived variables do indeed provide high-value information gain for the prediction network, verifying the correctness and rationality of the preceding feature engineering steps.

[0130] In one optional implementation, after outputting the optimal scheduling scheme from multiple candidate scheduling schemes, the method further includes an evaluation step on the improvement of spawning of target fish downstream of the dam:

[0131] Optionally, boundary data of different typical hydrological year scenarios can be obtained in advance, and the water level trajectory and outflow trajectory in the optimal scheduling scheme can be combined with the boundary data and input into the downstream river water temperature prediction model to simulate and generate the optimized downstream water temperature sequence.

[0132] As an optional implementation, after the multi-objective reservoir scheduling model outputs the optimal scheduling scheme, the system performs a scenario replay operation to verify its ecological regulation effect. The system imports pre-configured climate and inflow boundary data for high-water years, normal-water years, and low-water years. The system reads the discharge flow curve and water level change trajectory set by the optimal scheduling scheme, assembles these trajectory vectors into a time series, and feeds them in batches into the downstream river water temperature prediction model for forward inference, outputting the complete predicted temperature curve under the predetermined typical hydrological year conditions, i.e., the optimized downstream water temperature series.

[0133] Optionally, the statistically optimized downstream water temperature sequence satisfies the duration and cumulative frequency of the suitable water temperature range for spawning.

[0134] In this embodiment, the system scans the optimized downstream water temperature sequence along the time axis and compares the daily temperature values ​​with the suitable spawning water temperature range for the target fish. The system is equipped with a time accumulator, which increases the value when the water temperature data falls within a suitable threshold range, thereby calculating the number of consecutive days that meet the ecological requirements, i.e., the duration. At the same time, it records the total number of times the condition is triggered throughout the year, i.e., the cumulative frequency.

[0135] Optionally, the duration and cumulative frequency are compared with pre-stored historical baseline data when ecological scheduling was not performed to calculate the improvement difference of the target fish's spawning window and output the ecological improvement assessment results.

[0136] As an optional implementation, the system retrieves historical monitoring data from the reservoir during conventional power generation operations as a comparison benchmark. The system subtracts the corresponding values ​​from the historical benchmark data for the duration and cumulative frequency obtained from the above statistics. For example, by comparing, the number of days and frequency of improvement during the spawning window of the target fish species can be obtained, and the ecological improvement assessment results are output. The system summarizes the improvement differences for all typical hydrological years into a matrix report, which serves as the final ecological improvement assessment result, providing an objective and calculable technical basis for evaluating the ecological scheduling effectiveness of water conservancy projects.

[0137] Example 7: This example serves as a comprehensive application of the above-mentioned reservoir ecological scheduling method based on water temperature prediction during the fish spawning season in a typical engineering scenario. Furthermore, this example is not a separate method independent of the aforementioned examples, but rather, within the same scheduling framework, it sequentially executes the determination of target fish parameters, generation of candidate scheduling schemes, embedding of water temperature prediction, calculation of ecological targets, and post-scheduling evaluation. Specifically, it can be as follows:

[0138] (1) Obtain biological characteristic data of the target fish, determine the suitable water temperature range for spawning of the target fish and the parameters for determining the spawning window period;

[0139] (2) Using the water level sequence and outflow sequence during the scheduling period as decision variables, a multi-objective reservoir scheduling model including power generation benefit objective, flood control risk objective and ecological scheduling objective is constructed, and multiple sets of candidate scheduling schemes are generated;

[0140] (3) Based on historical hydrological and meteorological data and physical information since the dam was built, the water temperature prediction model of the downstream river was trained. In the iterative evaluation of the multi-objective scheduling model, the input feature matrix of the complete scheduling period corresponding to the candidate scheduling scheme was input into the model at once, and the predicted water temperature sequence of the downstream river corresponding to the candidate scheduling scheme was calculated.

[0141] (4) Based on the predicted water temperature sequence, suitable water temperature range for spawning and spawning window period, determine the parameters dynamically to determine the spawning window period of the target fish, and calculate the ecological scheduling target using the asymmetric water temperature suitability function within the spawning window period; after outputting the optimal scheduling scheme, evaluate the improvement of spawning of the target fish by combining the boundary data of different typical hydrological year scenarios.

[0142] Step (1) specifically includes the following steps:

[0143] (1.1) Using traditional fishing, cross-sectional surveys and comparison of historical data, we analyzed the species composition and quantity, diversity index, dominant species and invasive species in the upper and lower reaches of the dam, and conducted historical comparisons of fish resources.

[0144] (1.2) Based on historical data and field surveys, fish species with significant differences in distribution between the upper and lower reaches of the dam were selected as key fish species for attention.

[0145] (1.3) Using ecological methods that combine fish gonadal tissue sections, gonadal index (GSI), body weight and gonadal weight measurements, fish species sensitive to water temperature were screened, and the suitable water temperature range for spawning, the biological zero degree of gonadal development, the accumulated temperature threshold for gonadal maturation and the number of days of spawning were determined accordingly. Among them, the biological zero degree of gonadal development, the accumulated temperature threshold for gonadal maturation and the number of days of spawning together constitute the parameters for determining the spawning window period.

[0146] Furthermore, step (2) specifically includes the following steps:

[0147] (2.1) The objectives of power generation benefit, flood control risk, and ecological scheduling are used as the objective functions of a multi-objective reservoir scheduling model oriented towards the water temperature demand during fish spawning season. Pareto optimal solution sets are formed between different objectives through non-dominated sorting and crowding distance calculations. When the optimization algorithm needs to be unified into a minimization problem, the power generation benefit and ecological scheduling objectives are equivalently transformed by taking negative values, while the flood control risk objective remains the minimization objective. The definitions and expressions of the different objectives are as follows:

[0148] a) Maximizing power generation efficiency:

[0149] (1)

[0150] (2)

[0151] (3)

[0152] in: =1000kg / m3; =9.8m / s²; Q gen,t =Effective power generation flow, m³ / s; H net,t =Effective head, m; =Power generation efficiency; ∆t =Time step, seconds; E(t) =Energy generated in time period t, MWh, taking the negative sign transforms it into a minimum problem.

[0153] b) Minimize flood risk:

[0154] (4)

[0155] (5)

[0156] Among them: J flood (t) = Flood risk index.

[0157] c) Maximizing suitable water temperature during the spawning period:

[0158] Given the high overlap between the critical spawning period of the target fish species in the reservoir and the period of most severe low-temperature water discharge, and considering the evaluation bias inherent in traditional static water level control methods, a dynamic suitability evaluation mechanism based on real-time machine learning model is further introduced into the multi-objective scheduling model. During iterations, the algorithm calculates the cumulative suitability score of the target fish species within the spawning window based on the predicted downstream river temperature and an asymmetric water temperature suitability function, aiming to maximize this score. The calculation formula is as follows:

[0159] The model aims to increase the water temperature in the downstream channel during the spawning window:

[0160] (6)

[0161] In the formula, The objective function value for ecological scheduling; To calculate the time step set of the spawning window period dynamically determined by using the predicted sequence; The water temperature suitability score for day t is calculated using an asymmetric penalty function:

[0162] s(t) (7)

[0163] in: The predicted water temperature on day t is output in real time using a pre-built downstream river water temperature prediction model. and It is divided into the lower and upper limits of the suitable water temperature for spawning of the target fish species; This is the baseline value for the maximum daily suitability score. and These are preset left and right deviation penalty parameters, respectively, to apply heterogeneous penalties for water temperature deviations in different directions.

[0164] (2.2) To ensure that the reservoir does not exceed its physical limits such as water storage capacity and flood discharge capacity during the scheduling process, and to take into account ecological needs and downstream water level safety, the following constraints are specified to provide feasibility and operability for the practical application of the model.

[0165] a) Reservoir water balance constraints.

[0166] Taking into account reservoir capacity, inflow, and total discharge, the reservoir water balance constraint expression is as follows:

[0167] (8)

[0168] In the formula, For storage capacity, For inbound flow, This refers to the total discharge capacity (including discharge for power generation, flood control, etc.).

[0169] b) Reservoir capacity and water level constraints.

[0170] Considering the reservoir's capacity and water level constraints, its flood control capacity limit is:

[0171] V min ≤V t ≤V max (9)

[0172] Among them, V max To meet the upper limit of flood control reservoir capacity, V min This is the lower limit of dead storage capacity.

[0173] c) Water temperature regulation and control of water level change rate.

[0174] Considering the rate of water level change during reservoir temperature regulation, the range of variation for both the rate of rise and the rate of fall in water level is limited as follows:

[0175] (10)

[0176] The water level is limited to a rise rate of ≤2.17 m / d and a fall rate of ≤1.09 m / d.

[0177] d) Water level upper and lower limits constraints.

[0178] (11)

[0179] Among them, H lower (t) and H upper (t) represents the upper and lower limits of the water level, t=1,2,...,T.

[0180] e) Outbound flow constraints.

[0181] (12)

[0182] Where, q min To ensure the minimum ecological baseflow for downstream ecological security, q max To ensure safe flow for downstream flood control.

[0183] f) Flood control constraints for water outflow.

[0184] (13)

[0185] Among them: Q warn =Flood control warning flow, Q crit =Flood control flow restriction.

[0186] (14)

[0187] Where H flood α is the flood control level of the reservoir, and α is the risk weighting coefficient.

[0188] g) Power plant output constraints.

[0189] The output of a hydropower station is generally between its guaranteed output and installed capacity. A certain power station has a guaranteed output of 170MW and an installed capacity of 800MW.

[0190] (15)

[0191] Where: P t,min This represents the lower limit of power generation output, i.e., the guaranteed output, measured in MW.

[0192] P t,max This represents the upper limit of power generation output, i.e., installed capacity, measured in MW.

[0193] h) Non-negativity constraint.

[0194] Water level, reservoir capacity, flow rate, water head, and water consumption are all taken as non-negative values.

[0195] (2.3) Based on the second-generation non-dominated genetic algorithm NSGA-II, the multi-objective reservoir scheduling model facing the water temperature demand of the downstream river is optimized and solved. When the maximum number of iterations is reached, the change amplitude of the Pareto front is less than the preset threshold for several consecutive generations, or the objective function value meets the preset convergence condition, the iteration stops, and the Pareto optimal solution set composed of non-dominated solutions is obtained. The optimal scheduling scheme is selected according to the decision preference, and the corresponding water level trajectory and outflow trajectory are output.

[0196] Furthermore, step (3) specifically includes the following steps:

[0197] (3.1) Collect data on water level, flow rate, water temperature, meteorological data, inflow, and reservoir capacity. Construct a set of physical information enhancement features that includes thermal stratification intensity characteristics, hydraulic retention characteristics, and adaptive cumulative heat input characteristics. Then, use various machine learning algorithms to establish a water temperature prediction model for the downstream channel. By learning complex nonlinear relationships from historical data, the machine learning model can capture the inherent laws of the hydrological system, thereby achieving more accurate predictions. The formulas used in different methods are as follows:

[0198] (3.1.1) XGBoost

[0199] In this embodiment, XGBoost is mainly used to characterize the nonlinear response relationship of water temperature to flow rate, water level, and meteorological variables. By systematically searching and optimizing key hyperparameters such as the number of trees, maximum depth, learning rate, and sampling ratio, the model achieves optimal prediction performance while ensuring generalization ability. Its objective function is shown in equations (16) and (17).

[0200] The objective function of XGBoost consists of two parts: a loss function and a regularization term, expressed as follows:

[0201] (16)

[0202] in, It is a loss function that measures the difference between the actual value yi and the predicted value. The error between; It is a regularization term that controls model complexity and is usually defined as:

[0203] (17)

[0204] Where T is the number of leaves in the tree, w j γ and λ are the weights of the leaf nodes, and γ and λ are regularization parameters. Through efficient splitting computation and pruning strategies, XGBoost can achieve fast training on large-scale datasets while maintaining high computational efficiency and robustness, especially suitable for scenarios with missing and outlier values.

[0205] (3.1.2) Gaussian process.

[0206] Gaussian process regression (GPR), as a typical Bayesian nonparametric method, provides uncertainty quantification information for prediction results by probabilistically modeling the function space. In water temperature prediction, this characteristic is of great significance for assessing the risk of scheduling decisions. This invention constructs a GPR model based on equations (18)-(23), achieving a balance between prediction accuracy and computational efficiency through the reasonable selection of the kernel function and its hyperparameters. In the GPR model, the objective function f(x) is assumed to be a Gaussian process:

[0207] (18)

[0208] Where m(x) = E[f(x)] is the mean function, It is the covariance function (kernel function).

[0209] In regression tasks, observations are typically represented as:

[0210] (19)

[0211] Where h(x) is the linear mapping of the input features, β is the regression coefficient, and ε is Gaussian noise.

[0212] The joint distribution (Gaussian distribution) takes the form of:

[0213] (20)

[0214] The covariance matrix K(X,X) is calculated using the kernel function. Using the training data X,y, the conditional distribution of the predicted points can be derived as follows:

[0215] (twenty one)

[0216] The predicted mean and variance are as follows:

[0217] (twenty two)

[0218] (twenty three)

[0219] in, Input the feature vector for the point to be predicted. To predict the mean, To predict variance, Let be the covariance vector between the point to be predicted and the training samples. The kernel function value is the value of the point to be predicted; in the above expression, " ", as a subscript, indicates the identifier of the point to be predicted.

[0220] (3.1.3) Random Forest.

[0221] Random forest models effectively reduce the risk of overfitting from a single model by constructing multiple independent decision trees and integrating their prediction results. Simultaneously, their built-in feature importance assessment mechanism can be used to analyze the relative contributions of different hydrological and meteorological factors to the downstream water temperature change, providing support for subsequent model interpretation. The prediction process is shown in equation (24). The prediction process of random forest regression can be expressed as:

[0222] (twenty four)

[0223] In the formula, Let represent the final prediction result, N be the number of decision trees, and fi(x) represent the prediction value of the i-th decision tree for the input x. By integrating the prediction results of multiple decision trees, random forests can effectively handle high-dimensional and complex nonlinear relationships while improving prediction performance.

[0224] (3.1.4) Long Short-Term Memory Network.

[0225] To address the significant temporal correlation and lag effect of water temperature downstream of the dam, a Long Short-Term Memory (LSTM) network is introduced for modeling. LSTM effectively captures the long-term dependence features in the water temperature sequence through gating mechanisms such as forget gate, input gate, and output gate. Its core calculation process is shown in equations (25)-(30). This model is particularly suitable for describing the dynamic evolution of water temperature at different time scales.

[0226] The core computational process of LSTM is as follows:

[0227] a) Forget Gate: The forget gate controls which information needs to be deleted from the memory unit. Given the input xt at the current time step and the hidden state ht-1 at the previous time step, the output of the forget gate is:

[0228] (25)

[0229] Among them, f t is the activation value of the gate, representing the proportion that needs to be forgotten; W f b is the weight matrix of the forget gate; f σ is the forget gate bias vector; σ is the Sigmoid activation function.

[0230] b) Input Gate: The input gate determines which information from the current input needs to be updated in the memory unit. Its calculation formula includes the input gate activation value and candidate memory values:

[0231] (26)

[0232] (27)

[0233] In the formula, i t The activation value of the input gate indicates the proportion of memory cells that need to be updated; Candidate memory values ​​represent potential memory updates; W i and W C b is the weight matrix for the input gate and candidate memory; i and b C is the bias vector for the input gate and candidate memory; tanh is the hyperbolic tangent function.

[0234] c) Update the memory unit: Combining the results of the forget gate and the input gate, update the state of the memory unit:

[0235] (28)

[0236] In the formula, C t and C t-1 This represents the memory cell states at the current time step and the previous time step.

[0237] d) Output gates: Output gates determine which information in the memory cells will be used to update the hidden state.

[0238] (29)

[0239] (30)

[0240] In the formula, h is the activation value of the output gate, representing the effect of the memory cell state on the hidden state; t The hidden state at the current time step, used to pass to the next time step or the output layer; W o b is the weight matrix of the output gate; o This is the bias vector for the output gate.

[0241] (3.2) Bayesian optimization is then used to compare and optimize the hyperparameters of various machine learning methods. The contribution of the physical information enhancement feature set to the water temperature prediction results can be verified by combining the Shapley value interpretation analysis, so as to select the optimal water temperature prediction model for the downstream river channel.

[0242] (3.2.1) Bayesian optimization.

[0243] The core idea of ​​Bayesian optimization is to establish a probabilistic model to approximate the true objective function, i.e., the mapping relationship between hyperparameters and model performance. Based on existing evaluation results, this probabilistic model can not only predict the performance of new hyperparameter combinations but also quantify the uncertainty of this prediction. Based on the prediction results and their uncertainty, the most promising hyperparameter points are selected for evaluation in each iteration, thereby gradually approaching the global optimum. This active learning strategy is more efficient than passive grid or random search. The key steps of Bayesian method for optimizing hyperparameters are as follows:

[0244] a) Constructing a probability model:

[0245] Based on the existing hyperparameter combinations and objective function values, a probabilistic model P(f(x)) for the objective function is established.

[0246] b) Select the acquisition function and determine the optimal point:

[0247] The acquisition function a(x) defines how to balance exploration and exploitation, with the goal of finding new optimal points:

[0248] (31)

[0249] Commonly used data acquisition functions include Expectation Improvement (EI) and Upper Confidence Bound (UCB):

[0250] Expected Improvement (EI):

[0251] (32)

[0252] Where f(xbest) is the current optimal value.

[0253] Confidence Upper Bound (UCB):

[0254] UCB(x)=μ(x)+k·σ(x)(33)

[0255] In the formula, μ(x) and σ(x) are the prediction mean and uncertainty, respectively, and k controls the trade-off between exploration and exploitation.

[0256] c) Evaluate the objective function and update the model:

[0257] Selected point The probability model is updated by evaluating the objective function f(x).

[0258] Bayesian optimization has high accuracy, requires no manual adjustment, and is highly versatile. It is especially suitable for scenarios with limited computing resources and complex objective functions, such as the optimization of learning rate, regularization parameters, and network structure in deep learning models.

[0259] (3.2.2) Interpretive analysis of Shapley value.

[0260] The core idea of ​​the SHAP method is to decompose the model's prediction results into the contributions of each feature, and to ensure that this decomposition is economically fair through the mathematical concept of Shapley value. Introduced into the field of machine learning, the Shapley value can be used to measure the contribution of a single feature to the model's prediction results.

[0261] Specifically, let the feature set be N and the target feature be i, the formula for calculating its Shapley value is:

[0262] (34)

[0263] In the formula, N is the set of all features; S is any subset of features that does not contain feature i. υ(S) is the number of features in set S; υ(S) is the contribution of subset S to the prediction result; The Shapley value represents the marginal gain of the prediction result after adding feature i to a subset S. This design ensures that the Shapley value satisfies several important properties: efficiency (the sum of the Shapley values ​​of all features equals the difference between the model output and the baseline value), symmetry (features with the same marginal contribution have the same Shapley value), dummy property (features with no contribution receive a zero contribution value), and additivity (the Shapley values ​​of multiple models can be directly added). These properties make the Shapley value a reasonable and fair indicator for measuring feature importance.

[0264] However, accurately calculating the Shapley value requires evaluating 2|N| distinct feature subsets, which becomes computationally infeasible when the feature dimension is high. To overcome this difficulty, the SHAP library provides optimized computation methods for different types of models. For tree-based models (such as XGBoost and Random Forest), SHAP TreeExplainer leverages the structural properties of trees, recursively calculating the contribution of each path to reduce computational complexity to polynomial order. For arbitrary black-box models, SHAPKernelExplainer employs a model-independent approach based on the idea of ​​local linear approximation, estimating the Shapley value using weighted least squares, where sample weights are determined by an exponential kernel function. Furthermore, SHAP supports Monte Carlo sampling to further accelerate computation.

[0265] (3.3) The optimal downstream channel water temperature prediction model selected through comparison is directly embedded into the multi-objective scheduling optimization framework via code. In each generation evolution evaluation stage of the NSGA-II algorithm, the water level and flow sequences of candidate individuals are extracted to construct a physical information enhancement feature matrix, which is then input into the model for batch inference. The predicted water temperature sequence is output in real time, and the ecological scheduling target is calculated accordingly. Thus, a multi-objective scheduling model for reservoirs that can respond quickly and drive the iterative evolution of scheduling strategies is obtained.

[0266] Finally, step (4) includes the following steps:

[0267] (4.1) After selecting the optimal scheduling scheme from the Pareto optimal solution set, typical hydrological year scenario boundary data such as meteorological and inflow data are obtained for high-water years, normal-water years, and low-water years. The water level trajectory and outflow trajectory in the optimal scheduling scheme are input together with the above boundary data into the downstream river water temperature prediction model for scenario playback evaluation, and the optimized downstream water temperature sequence is generated by simulation. If it is necessary to formulate scheduling schemes for different typical hydrological years in engineering applications, the multi-objective reservoir scheduling model can also be resolved under the boundary data of each typical hydrological year.

[0268] (4.2) Scan the optimized downstream water temperature sequence, and statistically analyze the duration and cumulative frequency of the temperature range suitable for spawning. Compare the duration and cumulative frequency with historical baseline data when ecological scheduling was not implemented to calculate the improvement difference in the spawning window period of the target fish species. Output the ecological improvement assessment results. The above improvement difference is used to evaluate the improvement effect of the optimal scheduling scheme on the spawning window period of the target fish species under typical hydrological year scenarios.

[0269] This invention constructs a downstream river water temperature prediction model based on historical hydrological, water temperature, and meteorological data, and embeds it into a multi-objective reservoir scheduling model. This allows for real-time calculation of downstream river water temperature changes under different scheduling schemes during the optimization process, avoiding the problems of large computational load and slow response speed of traditional numerical water temperature models, thus improving the efficiency of ecological scheduling decisions. Furthermore, through a multi-objective optimization method, it comprehensively considers power generation benefits, flood control safety, and ecological needs within the same scheduling framework, avoiding the systemic risks that may be caused by pursuing only power generation or a single ecological objective, and achieving a reasonable balance among multiple objectives in the scheduling results.

[0270] Correspondingly, by explicitly constructing physical information enhancement features such as thermal stratification intensity proxy, hydraulic residence time, and adaptive cumulative heat input, the complex reservoir hydrothermal physical mechanism is mathematically analyzed, reducing the model's dependence on training data and improving the generalization accuracy and robustness of the prediction model under complex boundary conditions.

[0271] Furthermore, based on ecological principles, the actual spawning time window of the target organism is dynamically locked by effectively accumulating temperature, which alleviates the evaluation cycle misalignment error caused by abnormal hydrological years. At the same time, an asymmetric heterogeneous penalty function is introduced to calculate the differentiated physiological tolerance limits of individual organisms to high and low temperature stresses, making the discharge scheme output by the optimized model more consistent with the actual ecological breeding needs of rivers and improving the decision reliability of large-scale water conservancy projects under multi-objective coordinated scheduling.

[0272] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A reservoir ecological management method integrating water temperature prediction during the fish spawning season, characterized in that, include: Obtain biological characteristic data of the target fish species, determine the suitable water temperature range for spawning and the parameters for determining the spawning window period; A multi-objective reservoir scheduling model is constructed, which includes power generation benefit objective, flood control risk objective, and ecological scheduling objective. A multi-objective optimization algorithm is used to iteratively solve the multi-objective reservoir scheduling model, generating multiple sets of candidate scheduling schemes; During the iterative solution process, a set of physical information enhancement features is extracted for each set of candidate scheduling schemes, and the set of physical information enhancement features is input into a pre-built downstream river water temperature prediction model to output the predicted water temperature sequence corresponding to the candidate scheduling scheme. Based on the predicted water temperature sequence, suitable spawning water temperature range, and spawning window period determination parameters, the spawning window period corresponding to the candidate scheduling scheme is determined, and the ecological scheduling target of the candidate scheduling scheme is calculated. Combining power generation efficiency targets, flood control risk targets, and ecological dispatch targets, the optimal dispatch scheme is output from multiple candidate dispatch schemes; The physical information enhancement feature set should contain at least the following feature categories: Thermal stratification intensity characteristics used to characterize the vertical heat distribution gradient within a reservoir; Hydraulic retention characteristics used to characterize the water exchange rate and heating time within a reservoir; An adaptive cumulative heat input feature used to characterize the cumulative impact of previous meteorological conditions on the current outflow water temperature; Thermal stratification intensity characteristics, hydraulic retention characteristics, and adaptive cumulative heat input characteristics are constructed through the following steps: Acquire pre-collected meteorological driving data, daily inflow and daily reservoir capacity for the target watershed, and extract daily water level and daily outflow from candidate scheduling schemes; Calculate the positive temperature deviation between the daily temperature and the preset annual average temperature in the meteorological driving data; multiply the positive temperature deviation by the daily water level, and use the ratio of the product to the total flow as the thermal stratification intensity characteristic, where the total flow is the sum of the daily inflow and the daily outflow; Calculate the ratio of the reservoir capacity to the outflow on the same day as a hydraulic retention characteristic; The historical impact days are determined numerically based on the hydraulic retention characteristics. A dynamic accumulation window is constructed using the historical impact days. Within the dynamic accumulation window, the normalized historical shortwave radiation and historical temperature are accumulated to obtain the adaptive cumulative heat input characteristics.

2. The method according to claim 1, characterized in that, The parameters for determining the spawning window include the zero degree of gonadal development biology, the accumulated temperature threshold for gonadal maturation, and the number of days of spawning. Obtain biological characteristic data of the target fish species, determine the suitable water temperature range for spawning and the parameters for determining the spawning window, including: Fish sample data collected under different temperature and current conditions were obtained, and a biological feature database was constructed using body weight, gonad weight, gonad index, and histological section data. Based on a biological feature database, target fish species sensitive to water temperature are identified, and the suitable water temperature range for spawning, the biological zero degree for gonadal development, the accumulated temperature threshold for gonadal maturation, and the number of days for spawning are determined.

3. The method according to claim 1, characterized in that, During the iterative solution process, a set of physical information enhancement features is extracted for each set of candidate scheduling schemes. This set of features is then input into a pre-built downstream river water temperature prediction model, outputting the predicted water temperature sequence corresponding to each candidate scheduling scheme, including: Extract the daily water level sequence and daily outflow sequence of the candidate scheduling scheme during the complete scheduling period, and obtain the pre-stored daily meteorological driving sequence; For each day within the complete scheduling period, the set of enhanced physical information features for that day is calculated by combining the water level, outflow, meteorological driving factors, and historical lag conditions of the preceding period. The daily physical information enhancement feature sets within the complete scheduling period are concatenated in time sequence to form the input feature matrix; The input feature matrix is ​​fed into the downstream river water temperature prediction model for a batch inference to obtain a predicted water temperature sequence covering the entire scheduling period.

4. The method according to claim 1, characterized in that, Based on the predicted water temperature sequence, suitable spawning water temperature range, and parameters for determining the spawning window, the spawning window corresponding to the candidate scheduling scheme is determined, and the ecological scheduling objectives of the candidate scheduling scheme are calculated, including: The cumulative effective accumulated temperature sequence on the time axis is calculated based on the predicted water temperature sequence and the zero degree of gonadal development biology. By comparing the accumulated effective temperature with the gonadal maturation temperature threshold, the predicted water temperature of the day with the lower limit of the suitable water temperature for spawning, the start time is locked, and the end time is calculated by extending the number of days of spawning, and the spawning window period corresponding to the candidate scheduling scheme is dynamically determined. Extract water temperature segments from the predicted water temperature sequence during the spawning window; The deviation of water temperature values ​​within a water temperature segment from the suitable spawning water temperature range is calculated, and a suitability score is calculated based on the deviation. The cumulative suitability score within the spawning window is used as the ecological scheduling target.

5. The method according to claim 4, characterized in that, The suitability score is calculated based on the degree of deviation, specifically using an asymmetric water temperature suitability function, including: Determine whether the predicted water temperature on the day of the spawning window is lower than the lower limit of the suitable spawning water temperature, within the suitable spawning water temperature range, or higher than the upper limit of the suitable spawning water temperature. When the predicted water temperature is lower than the lower limit of the suitable water temperature for spawning, the low temperature side suitability score is calculated using the preset left deviation penalty parameter; When the predicted water temperature for the day is within the suitable water temperature range for spawning, the suitability score for that day is determined as the baseline value for the full suitability score for that day. When the predicted water temperature is higher than the upper limit of the suitable water temperature for spawning, the high temperature side suitability score is calculated using the preset right deviation penalty parameter; The values ​​of the left deviation penalty parameter and the right deviation penalty parameter are not equal.

6. The method according to claim 1, characterized in that, After outputting the optimal scheduling scheme from multiple candidate scheduling schemes, the process also includes an evaluation step on the improvement of spawning of target fish downstream of the dam: Obtain boundary data for different typical hydrological year scenarios pre-set, and input the water level trajectory and outflow trajectory in the optimal scheduling scheme into the downstream river water temperature prediction model to simulate and generate the optimized downstream water temperature sequence. The statistically optimized downstream water temperature sequence satisfies the duration and cumulative frequency of the suitable water temperature range for spawning. The duration and cumulative frequency are compared with historical baseline data when ecological scheduling was not implemented to calculate the improvement difference of the target fish's spawning window period, and the ecological improvement assessment results are output.

7. The method according to claim 1, characterized in that, The pre-built downstream river water temperature prediction model was obtained through the following offline training steps: Obtain historical hydrological and meteorological sequences of the target watershed to construct a training sample set; Construct an initial machine learning model and define the hyperparameter space to be optimized for the initial machine learning model; A Bayesian optimization algorithm is used to optimize parameters in the hyperparameter space, and a probabilistic surrogate model representing the mapping relationship between hyperparameters and model prediction performance is constructed based on historical evaluation results. Based on the probabilistic surrogate model and the acquisition function used to balance exploration and exploitation, the performance of the initial machine learning model is evaluated by iteratively selecting hyperparameter combinations. When the preset training convergence condition is met, the iteration stops, the optimal hyperparameter combination at this time is extracted and fixed into the initial machine learning model, and the pre-constructed downstream river water temperature prediction model is obtained.

8. The method according to claim 7, characterized in that, The initial machine learning model is selected from at least one of the following: gradient boosting tree model, Gaussian process regression model, random forest model, and long short-term memory network model.

Images