Reservoir thermocline feature forecasting method considering climate and hydrological condition changes

Abstract

The invention provides a reservoir thermocline feature forecasting method considering climate and hydrological condition changes, belongs to the technical field of hydrological simulation and hydraulic engineering, and is suitable for a reservoir area without main branch convergence. The method comprises the following steps: collecting and preprocessing a multi-source data set; constructing a drainage basin runoff yield model and a facade two-dimensional water temperature model; main influence factors of the vertical water temperature are screened, and a data set is constructed; on the basis of a production and convergence model, a reservoir water balance equation model and a Copula model, 27 working conditions are designed by considering climate and hydrological elements and a reservoir dispatching coupling relation, and an expansion data set is generated by utilizing a facade two-dimensional water temperature model; an LSTM-BNN neural network model is established to carry out vertical water temperature probability forecasting; the model performance is evaluated and the occurrence time, location, thickness and strength of the quantified thermocline are identified. According to the method, effective expansion of the data set is realized through physical and reasonable working condition combination, meanwhile, redundancy of the training set is avoided, and the problem that the generalization ability of the model is insufficient under a short-series monitoring data scene is solved.

Description

Technical Field

[0001] This invention relates to the field of hydrological simulation and hydraulic engineering technology, and in particular to a method for predicting the characteristics of reservoir thermoclines that takes into account changes in climate and hydrological conditions. Background Technology

[0002] Water temperature is a key factor affecting aquatic ecosystems, aquatic biogeochemical cycles, and the efficiency of water resource utilization. The water temperature structure of large deep reservoirs is influenced by both natural meteorological factors and human activities, making accurate forecasting of vertical water temperature and thermocline characteristics crucial for ecological management.

[0003] Existing reservoir water temperature forecasting methods are divided into mechanistic models and data-driven models. Mechanistic models have clear physical mechanisms but high computational costs, making them difficult to meet real-time scheduling requirements. Data-driven models have high computational efficiency but are limited by short series of monitoring data, resulting in insufficient model generalization ability. Existing data augmentation methods mostly involve simple combinations of input elements, with a large number of operating conditions, some of which are detached from physical reality, leading to redundancy in the training set. Furthermore, existing forecasts mostly output deterministic water temperature values, failing to provide information on uncertainties and key characteristic parameters of the thermocline, thus failing to meet the needs of ecological scheduling decisions. Summary of the Invention

[0004] This invention provides a method for forecasting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions, addressing the following technical problems in existing technologies: Under short-series monitoring data scenarios, forecasting models struggle to capture the water temperature response patterns under the coupled effects of different climate and hydrological conditions and reservoir scheduling, resulting in insufficient generalization ability; when expanding data, the design of operating condition combinations is unreasonable, leading to problems such as being detached from physical reality, having a large number of data sets, and redundant training sets; and the forecast results are disconnected from management needs, lacking vertical water temperature probability forecast information and key characteristic parameters of the thermocline.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting the characteristics of the thermocline in reservoirs that takes into account changes in climate and hydrological conditions, applicable to reservoir areas without major tributaries, is described below: S1. Collect physical mechanism model datasets WD, including topographic dataset WDdx (reservoir cross-section data, watershed topographic raster data), hydrological element measured dataset WDsw (daily inflow, daily inflow temperature, watershed daily evapotranspiration data, watershed daily precipitation data), meteorological element measured dataset WDqx (daily air temperature, dew point temperature, wind speed, wind direction, cloud cover, solar radiation), water environment element measured dataset WDhj (water transparency), reservoir actual scheduling dataset WDtd (number of stacked gate layers, stacked gate top elevation, daily outflow, daily dam inlet water level), and dam inlet vertical water temperature measured dataset WDcxsw. Data sources include meteorological bureaus, integrated scheduling and operation management platforms, hydrological yearbooks, geospatial data clouds, ERA5 reanalysis data, etc.

[0006] S2. Preprocess the time series and spatial scatter plot subsets in the WD: use the optimal information binary segmentation method, variable bandwidth kernel density estimation method, and visual inspection to detect outliers; use linear interpolation to fill missing values ​​and correct outliers; align the time periods of all time series and unify them to a daily time scale to form an effective WD.

[0007] S3. Based on the watershed topographic raster data in WDsw and WDdx, a three-source Xin'anjiang model is constructed. This model has a decentralized structure and includes four levels: evapotranspiration calculation (three-layer evapotranspiration model), runoff calculation (runoff at full capacity), water source division calculation (free reservoir structure to divide surface runoff, interflow, and groundwater runoff), and confluence calculation (watershed confluence is calculated using the unit hydrograph or linear reservoir method, and river channel confluence is calculated using the Muskingan method and lag algorithm).

[0008] S4. Based on WDdx, WDsw, WDqx and WDtd, construct a two-dimensional water temperature model for the CE-QUAL-W2 facade. Its basic equations include the fluid continuity equation, x and z momentum equations, state equation, water level fluctuation equation, and component migration equation. They are: ① Fluid continuity equation: ; ②Momentum equation in the x-direction: ; ③ Momentum equation in the z-direction: ; ④ Water level fluctuation equation: ; ⑤ Component migration equation: ; Where U represents the longitudinal average velocity, W represents the vertical average velocity, B is the width of the water body, P is the atmospheric pressure, and the shear stresses in the x and z directions are respectively expressed as... , ρ is density; η is water depth; the diffusion coefficients of temperature or components in the longitudinal and vertical directions are represented by Dx and Dz, respectively; Φ is the average concentration of the component. It is the flow rate of the component per unit volume. It is the source and sink of the river on a horizontal average; α is the river channel tilt angle.

[0009] The operating platform is CE-QUAL-W2V4.1.

[0010] S5. Analyze the correlation between hydrological elements, meteorological elements, water environment elements, reservoir scheduling factors and vertical water temperature. Combine the degree of interference of elements on target elements in the neural network, data quality, data acquisition difficulty, physical reasons for water temperature stratification in the reservoir area, and physical mechanism of heat conduction in the reservoir water body to screen the main influencing factors of vertical water temperature. The screened elements include shortwave radiation, air temperature, inflow rate, inflow water temperature, outflow rate, and the height of the stacked beam gate.

[0011] S6. Using the selected main influencing factors as input elements and the pre-processed vertical water temperature in front of the dam as output elements, construct the data-driven model dataset SD1.

[0012] S7. Divide SD1 into training set SD1-1, validation set SD1-2, and test set SD1-3 according to the year. The division ratio can be adjusted according to the amount of data, for example, 3:1:1.

[0013] S8. Based on the hydrological yearbook precipitation statistics of nearby hydrological stations in the upstream basin of the reservoir area, the formula P=(m a) / (n+1 2a) (a=0, n is the number of years of precipitation data) Perform annual precipitation frequency analysis, take the extreme precipitation years with a frequency of more than 95% to calculate the daily precipitation average sequence JS1, take the extreme precipitation years with a frequency of less than 5% to calculate the daily precipitation average sequence JS2, and calculate the daily precipitation average sequence JS0 for many years of precipitation.

[0014] S9. Using the Xin'anjiang model with three water sources constructed in step S3, simulate the scenarios with precipitation of JS0, JS1, and JS2 respectively, and output the runoff sequences JL0, JL1, and JL2, which are respectively used as the reservoir inflow I0, I1, and I2 under different precipitation conditions.

[0015] S10. Based on the dynamic characteristics of reservoir scheduling and the water balance equation Ot=(St+1) St) / 86400+It (Ot is the average outflow during the period, St is the initial water storage during the period, St+1 is the final water storage during the period, It is the average inflow during the period), combined with the scheduling rules derived from historical data (reflected in the monthly target storage capacity), calculate the outflow sequences O0, O1, and O2 corresponding to I0, I1, and I2, and the water storage capacity is derived from the water level-storage capacity relationship curve.

[0016] S11. Based on ERA5 reanalysis data from 1991 to 2021, three diurnal temperature sequences were constructed: the baseline climatological temperature sequence QW0 (T_mean), the warm phase temperature sequence QW1 (T_mean+σ), and the cold phase temperature sequence QW2 (T_mean+σ). σ).

[0017] S12. Construct a conditional probability framework based on Sklar's theorem. That is, according to Sklar's theorem, the joint distribution function FQW of air temperature (SW) and water temperature (SWI), SWI(q,i), can be decomposed as follows: ; in, and These are the marginal cumulative distribution functions of air temperature and water temperature, respectively. The original data is mapped to a uniform distribution space of [0,1] through probability integral transformation. For Copula functions, the parameters precisely characterize the nonlinear, asymmetric dependency structure between two variables; The Copula joint distribution model is used to derive the inflow water temperature sequences SW0, SW1, and SW2 corresponding to QW0, QW1, and QW2. First, the optimal marginal probability distributions are fitted to the air and water temperatures respectively. Then, the Copula function parameters are estimated, and finally, the conditional distribution and median predicted value of the reservoir water temperature are derived. That is, the inflow water temperature is derived using the Copula model. , , Conditional distribution function: ; Based on this conditional distribution, the results can be calculated for this specific temperature scenario. Below is the inflow water temperature sequence. The conditional median prediction value is used to generate inflow water temperature prediction sequences under warm phase (QW1, SWI1), baseline climatological state (QW0, SWI0) and cold phase (QW2, SWI2) scenarios, respectively.

[0018] S13. Based on the number of stacked beam doors and the top elevation information of the stacked beam doors in WDtd, determine the three common stacked beam door scheduling scenarios: DLM0, DLM1, and DLM2.

[0019] S14. Construct simulation conditions, which are combinations of [JS0-JS1-JS2 (I0-I1-I2, O0-O1-O2)], [QW0-QW1-QW2 (SW0-SW1-SW2)], and [DLM0-DLM1-DLM2], for a total of 27 conditions.

[0020] S15. Using the CE-QUAL-W2 model built in step S4, run 27 different operating conditions and collect the vertical water temperature data in front of the dam from the results of each operating condition. After summarizing, organize the data into a dataset SD2 that is consistent with the format of SD1-1, and use it as the second training set.

[0021] S16. Establish an LSTM-BNN neural network model, and optimize the model parameters using SD1-1 and SD2 as training sets and SD1-2 as validation sets. By analyzing the time-delay correlation between the surface water temperature in front of the dam and the daily air temperature data and the inflow water temperature data (using the sliding Pearson correlation coefficient, partial correlation coefficient, mutual information and Granger causality test), determine the input time step parameters based on the maximum time delay.

[0022] S17. Using SD1-3 as the input parameters of the test set, the optimized model is used to obtain the vertical water temperature forecast results.

[0023] S18. The performance of probabilistic forecasts is evaluated using the Predictive Interval Coverage (PICP), the Predictive Interval Average Width (PINAW), and the Continuous Ranking Probability Score (CRPS), while the performance of deterministic forecasts is evaluated using the Mean Absolute Error (MAE).

[0024] The first method is the prediction interval coverage rate, which characterizes the probability that the target value falls within the prediction interval. It is a key indicator for evaluating the reliability of interval prediction. The closer it is to 1, the better the model performance. The formula is as follows: ; ; Where N is the total number of samples participating in the evaluation. A binary variable, used to represent Is it within the prediction range? Among them.

[0025] The second method is prediction interval normalized average width (PINAW), which introduces the average bandwidth of the interval to reflect the clarity of the prediction, avoiding the prediction interval from being too wide and losing its reference value in the pursuit of reliability alone. The formula is as follows: .

[0026] The third method is the continuous ranked probability score (CRPS), which represents the weighted average of the cumulative distribution differences between forecasts and actual observations, characterizing the probabilistic forecasting performance of the model. It is a standard method for measuring the overall effectiveness of interval forecasts; the smaller the value, the higher the accuracy of the interval forecast. The formula is as follows: ; in, For the nth observation, For each forecast member in the cumulative distribution function obtained for the nth forecast The corresponding cumulative probability; This is a step function. The closer CRPS is to 0, the more accurate the forecast.

[0027] Deterministic forecast performance was achieved using mean absolute error (MAE).

[0028] The fourth method is the Mean Absolute Error (MAE), which is used to evaluate the performance of deterministic forecasts. The formula is as follows: ; in, For the nth simulated value, For the nth observation, the closer the MAE is to 0, the higher the forecast accuracy.

[0029] S19. Determine the water depth at the bottom sill of the intake by combining the water level in front of the dam and the elevation of the bottom sill of the intake; determine the time of appearance of the thermocline by the first occurrence of the three-day moving average of the daily average temperature difference between the surface water temperature and the water temperature at the bottom sill of the intake when it exceeds 1℃ for three consecutive days; determine the location of the thermocline by the water depth above and below the water layer with a vertical temperature gradient greater than 0.2℃ / m; determine the thickness of the thermocline by the difference in water depth between the upper and lower parts; determine the strength of the thermocline by the vertical temperature gradient.

[0030] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention analyzes the coupled changes of climate and hydrological elements by coupling the runoff generation and runoff model, the water balance equation, and the Copula model. It designs 27 physically reasonable combinations of operating conditions, effectively expanding the dataset while controlling the number of operating conditions. This solves the problem of insufficient model generalization ability under short-series monitoring data scenarios, and avoids training set redundancy.

[0031] 2. An LSTM-BNN neural network model was established to achieve probabilistic prediction of vertical water temperature. At the same time, the occurrence time, location, thickness and intensity of the thermocline were accurately identified and quantified, filling the gap in existing methods that could not simultaneously provide uncertainty information and complete features of the thermocline.

[0032] 3. The method has high computational efficiency and reliable forecast accuracy, and can provide direct and usable decision support for reservoir ecological scheduling and precise water temperature management, which helps to mitigate the impact of low-temperature water discharge and maintain the health of downstream ecosystems. Attached Figure Description

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a technical flowchart of an embodiment of the present invention; Figure 2 This is the model dataset structure of an embodiment of the present invention; Figure 3 This is a structural diagram of the Xin'anjiang River water source model according to Embodiment 3 of the present invention; Figure 4 This is a diagram of the running interface of the CE-QUAL-W2 model according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the physical mechanism of heat transfer and mixing in reservoir water according to an embodiment of the present invention; Figure 6 This is a partial example diagram of the output features of the data-driven model dataset SD1 in an embodiment of the present invention; Figure 7 This is a partial example diagram of the input features of the data-driven model dataset SD1 in an embodiment of the present invention; Figure 8 This is an example diagram illustrating the design and comparison of optimized working condition combinations for the data-physical dual-drive model based on the front-to-back coupling fusion strategy according to an embodiment of the present invention. Figure 9 This is an example diagram illustrating the time-lag correlation between regional air temperature and surface water temperature based on kernel density estimation and the sliding Pearson correlation coefficient method, as described in this embodiment of the invention. Figure 10 This is a comparison chart of deterministic forecast results and actual measurements in an embodiment of the present invention; Figure 11 This is a graph showing the probability forecast results of water temperature at a depth of 30m for a 28-day forecast period according to an embodiment of the present invention. Figure 12 This is an example diagram showing the calculation results of the thermocline feature in an embodiment of the present invention. Detailed Implementation

[0034] The specific embodiments of the present invention will be further described in detail with reference to the accompanying drawings.

[0035] This embodiment takes a reservoir without major tributaries as the research object, and the technical process is as follows: Figure 1 As shown.

[0036] First, data collection and preprocessing were performed. The following datasets were collected: topographic dataset WDdx (including reservoir cross-sectional data and watershed topographic raster data); measured hydrological data WDsw (daily inflow, daily inflow temperature, daily watershed evapotranspiration and precipitation data); measured meteorological data WDqx (daily air temperature, dew point temperature, wind speed, wind direction, cloud cover, and solar radiation); measured water environment data WDhj (water transparency); actual reservoir operation dataset WDtd (number of gate layers used, gate top elevation, daily outflow, and daily upstream water level); and measured vertical water temperature upstream of the dam dataset WDcxsw. The data covers five hydrological years. The dataset structure is as follows: Figure 2 As shown, the dataset covers six subsets, comprehensively supporting model building and validation. 32 outliers were detected using the optimal information binary segmentation method, variable bandwidth kernel density estimation method, and visual inspection. 28 missing values ​​were filled and outliers were corrected through linear interpolation. All time series periods were aligned and unified to a daily scale, forming an effective data structure (WD).

[0037] A model of the Xin'anjiang River with three water sources was constructed based on WDsw and watershed topographic raster data. The model structure is as follows: Figure 3 As shown. Figure 3 In the three-source Xin'anjiang model, a four-level structure is used to simulate the hydrological processes of the watershed. The first level is evapotranspiration calculation, which adopts a three-layer evaporation model, considering the uneven vertical distribution of soil moisture content and the evaporation processes of the upper, lower, and deeper soil layers. The main parameters include the water storage capacity of the upper and lower soil layers (WUM, WLM) and the evapotranspiration diffusion coefficient of the deeper layer (C). The second level is runoff calculation, which uses a water storage capacity-area distribution curve to address the problem of uneven spatial distribution of soil moisture content based on the full-storage runoff mechanism. The key parameters include the average water storage capacity of the watershed (WM) and the water storage capacity curve index (B). The third level is water source delineation, which divides runoff into three sources: surface, subsurface, and groundwater through a free-water reservoir structure. The free-water storage capacity-area distribution curve is used to consider the uneven distribution of water storage capacity, and the parameters involved include free-water storage capacity (SM), curve index (EX), and subsurface runoff outflow coefficients (KI, KG). The fourth level is runoff calculation, including slope runoff and channel runoff: slope runoff is calculated using the linear reservoir method for the three water sources, with decline coefficients CS, CI, and CG, respectively; channel runoff is calculated using the Muskingen piecewise continuous algorithm, with parameters being the channel storage coefficient (KE) and the flow-to-gravity coefficient (XE). This structure comprehensively represents the entire process from vertical evapotranspiration, runoff generation, water source allocation to horizontal runoff from slopes and channels.

[0038] The parameters for each layer are set according to Table 1, including the water storage capacity of the upper soil layer WUM=20mm, the water storage capacity of the lower soil layer WLM=60mm, and the average water storage capacity of the watershed WM=120mm, etc.

[0039] Table 1 - Functions, Calculation Methods, and Corresponding Parameters of Each Hierarchical Structure of the Xin'anjiang Model

[0040] The CE-QUAL-W2 model is built based on WDdx, WDsw, WDqx, and WDtd. The running interface is as follows: Figure 4 As shown, the simulation period is set to 1 year, and the time step is 60 seconds.

[0041] Analyze the correlation between various influencing factors and vertical water temperature, and combine this with the physical mechanisms of heat transfer and mixing in the reservoir water, such as... Figure 5 As shown in Table 2, the relevant physical elements were selected as the main influencing factors, including inflow rate, inflow water temperature, air temperature, outflow rate, height of the stacked beam gate, and water transparency, and a data-driven model dataset SD1 was constructed.

[0042] Table 2 - Summary of Physical Elements Related to Heat Transfer and Mixing in Reservoir Water

[0043] Note: Elements marked with ★ are the input elements of the selected data-driven model.

[0044] Local examples of SD1 input features are as follows: Figure 7 As shown, a partial example of the output elements is as follows: Figure 6 As shown, they are divided into SD1-1, SD1-2, and SD1-3 in a 3:1:1 ratio.

[0045] Based on 30 years of precipitation statistics from a nearby hydrological station upstream of the reservoir, the annual precipitation frequency was calculated using the formula P=m / (n+1). Three extremely rainy years (more than 95% frequency) and three extremely dry years (less than 5% frequency) were selected to calculate JS1 and JS2 respectively. Simultaneously, the 30-year average daily precipitation sequence JS0 was calculated. The inflow rates I0, I1, and I2 corresponding to JS0, JS1, and JS2 were simulated using the Xin'anjiang Three-Source Model, and the corresponding outflow rates O0, O1, and O2 were calculated using the water balance equation. Based on ERA5 reanalysis data from 1991 to 2021, QW0, QW1, and QW2 were constructed, and the corresponding SW0, SW1, and SW2 were derived using the Gaussian Copula model. Based on WDtd, three scheduling scenarios were determined: DLM0 (all beam gates closed), DLM1 (beam gates open at level a), and DLM2 (beam gates open at level a). The combined operating conditions were designed as follows: Figure 8 As shown, compared with the traditional simple combination, the number of working conditions is greatly reduced while ensuring physical rationality, and 27 working conditions are finally constructed.

[0046] The CE-QUAL-W2 model was used to simulate 27 operating conditions, and vertical water temperature data in front of the dam was collected to form SD2. An LSTM-BNN neural network model was established, and parameters were optimized using SD1-1 and SD2 as the training set and SD1-2 as the validation set. The time-delay correlation analysis process and core results are as follows: Figure 9 As shown, by comprehensively considering multiple methods such as the sliding Pearson correlation coefficient and partial correlation coefficient, and taking into account hardware performance, the optimal input time step was finally determined to be 50 days. The SD1-3 data was input into the model to obtain the forecast results. The forecast results agree well with the measured data, and some results are as follows: Figure 10 , Figure 11 As shown.

[0047] By combining the water level in front of the dam and the elevation of the intake sill, the timing of the thermocline's appearance, the water depth and thickness at the upper and lower boundaries of the thermocline, and the average temperature gradient (thermocline intensity) can be determined, providing technical support for the formulation of a stratified water intake scheme for the reservoir. Some results are shown below. Figure 12 As shown.

[0048] The working principle of this invention is as follows: by collecting and preprocessing multi-source data, a dual-drive framework of physical mechanism model and data-driven model is constructed; effective operating conditions are designed taking into account the coupling relationship between climate and hydrological elements and reservoir scheduling to avoid redundancy in the final dataset; the physical mechanism model is used to simulate water temperature response under different climate, hydrological and scheduling scenarios to generate an expanded dataset; an LSTM-BNN model is trained based on the expanded dataset to achieve probabilistic forecasting of vertical water temperature; finally, key features of the thermocline are extracted from the forecast results through preset judgment criteria to form a forecast result that takes into account both uncertainty analysis and practical decision-making needs.

[0049] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method of forecasting characteristics of a thermocline of a reservoir taking into account changes in climatic hydrological conditions, characterized in that, Includes the following steps: S1. Collect the physical mechanism model dataset WD, which includes the topographic dataset WDdx, the measured hydrological element dataset WDsw, the measured meteorological element dataset WDqx, the measured water environment element dataset WDhj, the actual reservoir scheduling dataset WDtd, and the measured vertical water temperature dataset WDcxsw in front of the dam. S2. Preprocess the time series and spatial scatter plot subsets in the WD to form an effective WD; S3. Construct a watershed runoff generation and confluence model based on watershed topographic raster data in WDsw and WDdx; S4. Based on WDdx, WDsw, WDqx, and WDtd, construct a two-dimensional water temperature model of the reservoir elevation; S5. Analyze the correlation between hydrological elements, meteorological elements, water environment elements, reservoir scheduling factors and vertical water temperature, and determine the main influencing factors of vertical water temperature by combining multi-dimensional screening criteria. S6. Using the main influencing factors as input and the preprocessed vertical water temperature in front of the dam as output, construct the data-driven model dataset SD1; S7. Divide SD1 into training set SD1-1, validation set SD1-2, and test set SD1-3; S8. Calculate the daily precipitation sequences JS1, JS2, and JS0 for extreme precipitation years and multi-year average precipitation years based on precipitation statistics; S9. Simulate the inflow sequences I0, I1, and I2 corresponding to JS0, JS1, and JS2 using a watershed runoff generation and confluence model; S10. Based on the water balance equation and scheduling rules, calculate the outflow sequences O0, O1, and O2 corresponding to I0, I1, and I2. S11. Based on ERA5 reanalysis data, three diurnal temperature sequences QW0, QW1, and QW2 were constructed. S12. The inflow water temperature sequences SW0, SW1, and SW2 corresponding to QW0, QW1, and QW2 are derived using the Copula joint distribution model. S13. Based on WDtd, determine three stacked beam gate scheduling scenarios: DLM0, DLM1, and DLM2. S14. Construct a combined simulation condition of [JS0-JS1-JS2, QW0-QW1-QW2, DLM0-DLM1-DLM2]; S15. Run the above working conditions using the two-dimensional water temperature model of the facade, collect the vertical water temperature data in front of the dam, and organize it into dataset SD2; S16. Establish an LSTM-BNN neural network model, optimize parameters using SD1-1 and SD2 as training sets and SD1-2 as validation sets, and determine the input time step. S17. Input SD1-3 into the optimized model to obtain the forecast results; S18. Use preset indicators to evaluate model performance; S19. By combining the water level in front of the dam with the elevation of the bottom sill of the intake, identify and quantify the occurrence time, location, thickness and intensity of the thermocline.

2. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, The preprocessing in step S2 includes: outlier detection using the optimal information binary segmentation method, variable bandwidth kernel density estimation method, and visual inspection method; processing missing and outlier values ​​using linear interpolation; and aligning all time series periods and unifying the time scale.

3. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, The runoff generation and confluence model in step S3 is the Xin'anjiang model with three water sources. The model includes four levels: evapotranspiration calculation, runoff generation calculation, water source division calculation, and confluence calculation.

4. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, The vertical two-dimensional water temperature model in step S4 is the CE-QUAL-W2 model, whose basic equations include the fluid continuity equation, x and z momentum equations, state equation, water level fluctuation equation, and component migration equation.

5. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, The selection criteria for step S5 include the degree of interference of the elements on the target elements in the neural network, data quality, ease of data acquisition, physical reasons for the temperature stratification of the reservoir water, and physical mechanisms of heat conduction in the reservoir water.

6. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, In step S8, the frequency of extreme precipitation years is calculated using the formula P=(m a) / (n+1 2a), where a=0, n is the number of years of precipitation data, 95% or more frequency is an extremely rainy year, and less than 5% frequency is an extremely dry year.

7. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, In step S11, QW0 is the baseline climatological temperature sequence T_mean, QW1 is the warm phase temperature sequence T_mean+σ, and QW2 is the cold phase temperature sequence T_mean. σ, the climate baseline period for the ERA5 reanalysis data is 1991–2021.

8. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, The number of combined simulation conditions in step S14 is 3×3×3=27.

9. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, The preset indicators in step S18 include: prediction interval coverage (PICP), prediction interval average width (PINAW), and continuous ranking probability score (CRPS), which characterize the performance of probabilistic forecasts; and mean absolute error (MAE), which characterizes the performance of deterministic forecasts.

10. The method for predicting the characteristics of a reservoir thermocline considering changes in climate and hydrological conditions as described in claim 1, characterized in that, In step S19: the thermocline appears when the daily average temperature difference between the surface water and the bottom plate elevation of the intake water exceeds 1℃ for three consecutive days; the location of the thermocline is determined by the water depth above and below the water layer with a vertical temperature gradient greater than 0.2℃ / m; the thickness of the thermocline is the difference between the upper and lower water depths; and the intensity of the thermocline is the vertical temperature gradient.

Images