Lake and reservoir thermal stratification law identification method based on layered monitoring device

Abstract

The present application relates to the technical field of water environment monitoring, and discloses a lake and reservoir thermal stratification law identification method based on a layered monitoring device, which comprises processing original monitoring data to generate a space-time continuous data matrix, extracting thermocline boundary coordinates, oxycline boundary coordinates and anoxic water thickness parameters; inputting a multi-source meteorological and hydrological data set and a time series characteristic data set into a double-flow long short-term memory neural network based on a cross attention mechanism, introducing a joint loss function to optimize parameters, and outputting predicted jump layer boundaries and anoxic thickness; calculating a water body relative thermal resistance sequence based on the thermocline boundary coordinates to divide the evolution stage, and outputting a real thermal stratification evolution stage identification; calculating a water quality early warning index sequence, and outputting a bottom aeration instruction and a safe water intake depth. The present application combines meteorological driving and internal physical constraints to accurately identify thermal stratification laws, and provides objective control parameters for water quality early warning and water intake scheduling.

Description

Technical Field

[0001] This invention relates to the field of water environment monitoring technology, specifically to a method for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device. Background Technology

[0002] The thermal stratification formed by lakes and reservoirs hinders vertical material exchange in the water, leading to the depletion of dissolved oxygen at the bottom and causing water quality deterioration problems such as anaerobic phosphorus release from bottom sediments. Existing monitoring methods mostly rely on discrete sampling, which makes it difficult to reflect the continuous dynamic characteristics of the vertical space in deep water areas. When spatially fitting the observation data, due to the lack of constraints from physical boundary conditions, the fitting curve will produce numerical oscillations at the boundaries of the surface and bottom layers of the water body, resulting in errors in the boundary positioning of the thermocline and the oxygen stratum.

[0003] In terms of predicting the evolution of thermal stratification, existing models struggle to spatiotemporally couple external multi-source heterogeneous meteorological data with internal water quality stratification characteristics. Conventional deep learning models, as purely data-driven methods, lack the constraints of internal physical topological relationships within water bodies, and their predictive outputs may violate natural laws, such as the anomaly of thermoclines crossing oxygen strata.

[0004] Meanwhile, surface water is subject to short-term meteorological events such as wind, waves and rainfall, resulting in high-frequency jump data. Existing methods lack effective state smoothing processing, which can lead to misjudgments of long-term thermal stratification evolution stages. In addition, existing identification systems usually only output abstract water quality assessment indicators and cannot combine the stratification spatial boundaries to directly transform the prediction results into specific physical control parameters that guide the safe water intake gate scheduling of reservoirs and the deployment of aeration equipment. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device. This method solves the problems of large positioning errors of the spatial boundary of the stratification, abnormal results due to the lack of meteorological coupling and physical constraints in evolution prediction, misjudgments caused by short-term meteorological interference in stage determination, and the inability to output quantitative parameters that can directly guide the scheduling of reservoir projects.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device, the method comprising the following steps: Acquire raw monitoring data, multi-source meteorological and hydrological datasets, and time-series feature datasets; The raw monitoring data is processed to generate a spatiotemporal continuous data matrix; Based on the spatiotemporal continuous data matrix, extract the upper boundary coordinates and lower boundary coordinates of the thermocline. Based on the spatiotemporal continuous data matrix, the effective oxygenation layer, the coordinates of the upper and lower boundaries of the oxygenation layer, the lower boundary of the bottom anoxic water body, the upper boundary of the bottom anoxic water body, and the thickness parameter of the anoxic water body are extracted. The multi-source meteorological and hydrological dataset and the time-series feature dataset are input into a dual-flow long short-term memory neural network based on cross-attention mechanism. A joint loss function is introduced for parameter optimization, and the predicted lower boundary coordinates of the thermocline, the predicted lower boundary coordinates of the oxocline, and the predicted thickness of the anoxic water body are output. Based on the coordinates of the upper boundary of the thermocline and the coordinates of the lower boundary of the thermocline, a continuous sequence of relative thermal resistance of the water body is calculated to divide the thermal stratification evolution stages and output the true thermal stratification evolution stage identifiers. Calculate the water quality deterioration early warning index sequence and output the lake and reservoir thermal stratification pattern identification results, which include the start-up command of the bottom aeration equipment and the safe water intake depth parameter.

[0007] This invention utilizes a cubic spline interpolation algorithm to construct a spatiotemporally continuous data matrix, realizing the conversion of discrete observations into a spatially continuous physical field. Based on the first-order partial derivative of the continuous depth function, it extracts the boundary coordinates of the thermocline and the oxygen stratification layer, as well as the thickness of the bottom anoxic water body. It combines a dual-flow long short-term memory neural network to process heterogeneous meteorological and hydrological data and water quality stratification characteristics. It uses a cross-attention mechanism to establish a nonlinear mapping relationship between meteorological driving forces and the evolution of the internal habitat of the water body. This invention divides the thermal stratification evolution cycle by calculating the relative thermal resistance of the water body, and finally outputs water quality early warning sequences and scheduling instructions, providing quantitative data parameters that integrate meteorological, physical, and habitat characteristics.

[0008] Furthermore, the step of processing the raw monitoring data to generate a spatiotemporally continuous data matrix includes: The original monitoring data is subjected to anomaly removal and time synchronization processing, and vertical spatial continuity fitting is performed using a cubic spline interpolation algorithm along the depth axis. A sliding window mid-range filtering algorithm is used to remove abnormal jump data points. Combined with the water temperature change rate limit threshold and the dissolved oxygen change rate limit threshold, the change rate limit judgment is implemented and the excessive noise point filtering operation is performed. Extract the interpolation base point sequence and combine it with the observation value sequence to construct a discrete spatial coordinate node set, and configure the natural boundary condition that the second derivative value is equal to zero at the water body surface boundary node and bottom boundary node; By combining the continuity equations of each node in the discrete spatial coordinate node set with the natural boundary conditions, a tridiagonal linear equation system is constructed. The undetermined coefficients of each piecewise cubic polynomial are obtained by using the chasing method to reconstruct the high-resolution profile continuity function. To address the discontinuity of each profile in the time dimension, a linear interpolation algorithm is used along the time axis to perform time continuity processing, thereby obtaining a continuous observation sequence with equal time steps. By using the nodes of the continuous observation sequence with equal time steps as the row dimension and the discrete depth nodes extracted by the high-resolution profile continuous function at fixed intervals as the column dimension, a two-dimensional spatiotemporal continuous data matrix is ​​constructed.

[0009] This invention avoids data oscillations in the fitted curve at physical boundary positions by configuring natural boundary conditions and solving a tridiagonal linear equation system, ensuring the mathematical convergence and physical continuity of the high-resolution profile continuity function across the entire water depth profile. Simultaneously, it introduces time-axis linear interpolation and matrix row and column dimension reconstruction operations to achieve seamless coupling between temporal evolution attributes and spatial depth attributes, eliminating the data fragmentation problem caused by sensor sampling time differences, and providing a complete spatiotemporal continuous data basis covering the entire habitat cycle.

[0010] Furthermore, the steps for extracting the upper boundary coordinates and lower boundary coordinates of the thermocline based on the spatiotemporal continuous data matrix include: The first-order partial derivative of the water temperature continuous function in the spatiotemporal continuous data matrix with respect to depth is used to calculate the vertical gradient of water temperature. The system traverses the entire water body depth space to extract independent depth intervals where the absolute value of the vertical temperature gradient is continuously greater than or equal to a preset temperature gradient threshold. It then removes depth intervals whose actual spatial span is less than the effective thermocline thickness threshold and outputs the depth intervals that pass the screening as candidate thermoclines. When there are two or more independent candidate thermoclines at the same sampling time, the candidate thermocline with the largest absolute value of the vertical gradient of water temperature along the depth integral result, or the smallest distance between the lower boundary and the lower boundary of the main thermocline at the previous effective sampling time, is selected as the main thermocline. The starting shallow water depth coordinates of the candidate thermocline on the vertical water depth axis are used as the upper boundary coordinates of the thermocline, and the ending deep water depth coordinates of the candidate thermocline on the vertical water depth axis are used as the lower boundary coordinates of the thermocline.

[0011] This invention extracts the range of continuous change of vertical gradient of water body by calculating the first-order partial derivative of space, introduces spatial span screening and continuity matching conditions, eliminates unstructured data fluctuations caused by local thermal disturbances, and improves the accuracy of thermocline spatial boundary coordinate positioning.

[0012] Furthermore, the steps for extracting the effective oxocele, the coordinates of the upper and lower boundaries of the oxocele, the lower and upper boundaries of the anoxic water body, and the thickness parameter of the anoxic water body based on the spatiotemporal continuous data matrix include: The first-order partial derivative of the dissolved oxygen continuity function in the spatiotemporal continuous data matrix with respect to depth is used to calculate the vertical gradient of dissolved oxygen. The effective oxygen choke layer is defined as all independent depth intervals whose absolute value of the dissolved oxygen vertical gradient is continuously greater than or equal to the dissolved oxygen gradient threshold and whose actual spatial span is greater than or equal to the effective oxygen choke layer thickness threshold. The starting shallow water depth coordinates of the effective oxygenation layer on the vertical water depth axis are used as the upper boundary coordinates of the oxygenation layer, and the ending deep water depth coordinates of the effective oxygenation layer on the vertical water depth axis are used as the lower boundary coordinates of the oxygenation layer. The dissolved oxygen continuity function is traversed, and the hypoxic intervals where the dissolved oxygen measurement value is continuously less than or equal to the hypoxia determination threshold and is continuously connected to the bottom boundary of the water body are extracted as the bottom hypoxic water body. The lower boundary and upper boundary of the bottom hypoxic water body are extracted, and the depth difference between the lower boundary and the upper boundary of the bottom hypoxic water body is calculated and used as the thickness parameter of the hypoxic water body.

[0013] This invention quantifies the spatial location of the physical mixing and confined interface in aquatic ecosystems, accurately locates the boundary of the anaerobic phosphorus release risk zone in bottom sediments, and provides quantitative input indicators for assessing the spatial evolution of anoxic habitats.

[0014] Furthermore, the steps of inputting the multi-source meteorological and hydrological dataset and the time-series feature dataset into a dual-flow long short-term memory neural network based on a cross-attention mechanism, introducing a joint loss function for parameter optimization, and outputting the predicted coordinates of the lower boundary of the thermocline, the predicted coordinates of the lower boundary of the oxocline, and the predicted thickness of the anoxic water body include: The input tensor corresponding to the time-series feature dataset is obtained. The water quality feature extraction branch uses a stacked double-layer long short-term memory network to receive the water quality hierarchical feature subset in the input tensor and outputs the water quality hidden state sequence. Extract a subset of the multi-source meteorological and hydrological dataset from the multi-source meteorological and hydrological dataset. The meteorological feature extraction branch uses a single-layer long short-term memory network to receive the subset of the multi-source meteorological and hydrological dataset and outputs a meteorological hidden state sequence. In the cross-coupling attention layer, the water quality hidden state sequence is used as the query vector, and the meteorological hidden state sequence is used as the key vector and value vector to calculate the cross-attention weight matrix. After matrix multiplication, the water temperature and dissolved oxygen co-coupling feature vector of the fusion meteorological driving mechanism is obtained. The water temperature and dissolved oxygen co-coupling feature vector is input into the decoding regression network composed of three fully connected layers, and the predicted lower boundary coordinates of the thermocline, the predicted lower boundary coordinates of the anoxic layer, and the predicted thickness of the anoxic water body are output.

[0015] This invention extracts water quality stratification features and external meteorological driving features separately through a dual-stream independent branch processing architecture. It uses the water quality hidden state sequence as a query vector to query the meteorological conditions that cause the water quality state, thereby extracting external meteorological driving features that have a causal mapping relationship with water quality fluctuations and improving the extraction accuracy of multidimensional data features.

[0016] Furthermore, the steps for parameter optimization by introducing a joint loss function include: Obtain the mean squared error between the prediction evaluation tensor output by the model forward propagation and the true label tensor aligned with the supervised learning sample set. Subtract the predicted lower boundary coordinates of the thermocline from the predicted lower boundary coordinates of the oxycline, and then perform a maximum value operation with zero to obtain the maximum value. The maximum value is multiplied by the physical penalty weight coefficient of the constraint boundary conditions, and then added to the mean square error to obtain the joint loss function. The network weight parameters are updated using the backpropagation algorithm through the joint loss function.

[0017] This invention superimposes physical constraints on the mean square error evaluation of the network, and applies convergence constraints to address the spatial inversion anomaly where the thermocline crosses the oxycline in the predicted output, ensuring that the output values ​​of the neural network conform to physical laws.

[0018] Furthermore, the step of calculating a continuous sequence of relative thermal resistance of the water body based on the coordinates of the upper boundary and the lower boundary of the thermocline to delineate the stages of thermal stratification evolution includes: The absolute water temperature measurement parameters at the upper and lower boundary coordinates of the thermocline are converted into absolute density parameters of shallow and deep water layers. For each discrete time node, the first difference between the absolute density parameters of deep and shallow water layers is calculated, and the second difference between the maximum pure water density reference constant and the standard pure water density benchmark constant is calculated. The first difference is divided by the second difference to obtain the relative thermal resistance of the water body corresponding to each discrete time node. The relative thermal resistance of the water body corresponding to each discrete time node is arranged in the order of time occurrence to generate a continuous sequence of relative thermal resistance of the water body that characterizes the degree of mixing difficulty between the upper and lower water layers. Discrete time nodes in the continuous sequence of relative thermal resistance of the water body that are less than the resistance threshold of the formation period are marked as the mixing period stage, and discrete time nodes in the continuous sequence of relative thermal resistance of the water body that are greater than or equal to the resistance threshold of the stable period are marked as the stable period stage. For discrete time nodes in the continuous sequence of relative thermal resistance of the water body that are greater than or equal to the resistance threshold during the formation period and less than the resistance threshold during the stable period, the time first derivative of the lower boundary coordinate of the thermocline is calculated. Discrete time nodes with the time first derivative greater than zero are marked as the decay stage, and discrete time nodes with the time first derivative less than or equal to zero are marked as the formation stage.

[0019] This invention converts water temperature parameters into water density parameters and calculates the relative thermal resistance of the water body. Combined with the continuous time evolution trend of the lower boundary coordinates of the thermocline, it provides a quantitative division method that reflects the entire cycle of thermal evolution.

[0020] Furthermore, the step of outputting the true thermal stratification evolution stage identifier includes: A sliding window mode filtering algorithm is introduced to implement state smoothing. The aforementioned marked stages are used as state labels for various evolution stages. Within a set anti-interference sliding window, the frequency of occurrence of various evolution stage state labels is counted, and the evolution stage state label with the highest frequency is extracted as the corrected evolution stage of the current central time node. The reconstructed smooth stratified state evolution sequence, after removing non-periodic interference from short-term meteorological extreme events, is output. The reconstructed smooth stratified state evolution sequence is then spliced ​​with the predicted lower boundary coordinates of the thermocline and the predicted lower boundary coordinates of the oxygencline according to their time occurrence sequence to output the true thermal stratification evolution stage identifier.

[0021] This invention performs smoothing processing based on statistical frequencies on continuous time series, filters out short-term high-frequency misjudgment of abrupt changes caused by wind and wave disturbances and heavy rainfall meteorological events, and outputs hierarchical evolution stage identifiers that reflect long-term stable patterns.

[0022] Furthermore, the step of calculating the water quality deterioration early warning index sequence includes: Obtain the ratio of the predicted hypoxic water thickness to the maximum water depth, and multiply the ratio by the hypoxic thickness configuration weight attenuation coefficient. Calculate the effective oxygen jump layer strength parameter, obtain the ratio of the effective oxygen jump layer strength parameter to the preset maximum reference extreme value of oxygen jump layer strength, and multiply the ratio by the layered barrier configuration weight attenuation coefficient. The product of the ratio of the predicted hypoxic water body thickness to the maximum water body depth and the hypoxic thickness configuration weight attenuation coefficient, and the product of the ratio of the effective oxygenation layer strength parameter to the preset maximum baseline extreme value of oxygenation layer strength and the layer barrier configuration weight attenuation coefficient, are added to obtain the continuous water quality deterioration early warning index to calculate the water quality deterioration early warning index sequence on the continuous time axis.

[0023] This invention comprehensively considers the barrier effect of the expansion ratio of the hypoxic physical space and the strength of the oxygen stratum to construct a continuous evaluation index that reflects the degree of water deterioration in deep water areas.

[0024] Furthermore, the step of outputting the lake / reservoir thermal stratification pattern identification results, which include the bottom aeration device start-up command and the safe water intake depth parameter, includes: When the water quality deterioration early warning index in the water quality deterioration early warning index sequence is continuously greater than or equal to the early warning index alarm threshold, a bottom aeration device start command is generated, and the spatial midpoint depth between the lower boundary of the bottom hypoxic water body and the upper boundary of the bottom hypoxic water body is extracted as the deployment depth parameter of the bottom aeration device. Extract the coordinates of the upper boundary of the thermocline at the current time point and shift the coordinates of the upper boundary of the thermocline upward by a set value as a safe water intake depth parameter; If the safe water intake depth parameter after upward translation is less than the physical coordinate of zero water surface depth, then the safe water intake depth parameter is corrected to the depth physical coordinate limit parameter.

[0025] This invention transforms water body boundary values ​​into scheduling instructions to guide engineering operations, avoids the risk of water overflow in the hypoxic phosphorus release zone at the bottom, and provides data constraint parameters for adjusting the height of the water intake gate and the depth of the aeration and oxygenation device deployment in the reservoir.

[0026] This invention provides a method for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device. It has the following beneficial effects: 1. This invention generates a spatiotemporally continuous data matrix through a cubic spline interpolation algorithm and the configuration of natural boundary conditions. Based on the first-order partial derivatives of the continuous depth function of water temperature and dissolved oxygen, it extracts the spatial boundary coordinates of the thermocline and the oxygen stratum. The method of this invention converts discrete measurement data into a continuous vertical physical field, avoiding numerical oscillations at the boundary. At the same time, it combines spatial span screening conditions to extract independent depth intervals, eliminating unstructured data interference caused by local thermal disturbances or measurement noise, thereby improving the accuracy of locating the thermal stratification structure and the boundary of the hypoxic habitat at the bottom layer of the lake or reservoir.

[0027] 2. This invention employs a dual-stream long short-term memory neural network based on a cross-attention mechanism to extract multi-source meteorological features and water quality stratification features, and introduces a joint loss function with superimposed physical penalty weight coefficients for parameter optimization. The cross-attention mechanism is used to calculate the weight matrix of the meteorological hidden state and the water quality hidden state, establishing a nonlinear mapping relationship between external meteorological forcing events and the evolution of aquatic habitats. At the same time, the joint loss function imposes a forced convergence constraint on the spatial inversion anomaly of the thermocline crossing the oxygen cascade in the prediction output, so that the prediction results of the neural network strictly conform to the natural evolution law of water bodies, improving the reliability of the prediction model.

[0028] 3. This invention divides the thermal stratification evolution stages based on the continuous sequence of relative thermal resistance of the water body and the time derivative of the lower boundary coordinates of the thermocline. It uses a sliding window mode filtering algorithm to filter out non-periodic high-frequency jump characteristics caused by short-term meteorological events, and combines the water quality deterioration early warning index sequence to generate parameters for the deployment depth of bottom aeration equipment and safe water intake depth. This invention combines the physical damping properties of the water body with the statistical distribution frequency to ensure the stability of the thermal stratification cycle determination, and directly converts the early warning indicators into specific engineering control parameters, avoiding the risk of water body deterioration caused by the upwelling of the bottom anoxic phosphorus release zone. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the system timing interaction of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram illustrating the specific implementation process of acquiring the original monitoring data according to the present invention; Figure 4 This is a schematic diagram illustrating a specific implementation process for generating a spatiotemporally continuous data matrix according to the present invention. Figure 5 This is a schematic diagram illustrating the principle of quantization extraction of the characteristic parameters of the thermocline in this invention. Figure 6 This is a schematic diagram illustrating the principle of extracting features of the oxoclimate and anoxic water bodies according to the present invention. Figure 7 This is a schematic diagram illustrating the physical constraint mechanism of the dynamic coupling prediction model of the present invention. Figure 8 This is a schematic diagram illustrating the specific implementation process of determining the thermal stratification evolution stage of the present invention. Figure 9 A schematic diagram illustrating the specific implementation process of generating comprehensive early warning and control commands according to the present invention; Figure 10 This is a comparison chart of the prediction errors of the lower boundary of the thermocline layer for different models of the present invention at each stage of the layered evolution. Detailed Implementation

[0030] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] See attached document Figure 1This invention provides a system for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device. The system includes: a data preprocessing module, a feature extraction module, a coupling analysis module, a stage determination module, and a result output module.

[0032] The data preprocessing module receives discrete time-series data collected by external devices. The discrete time-series data includes water temperature and dissolved oxygen sequences at different depths obtained at a set sampling frequency. The data preprocessing module performs abnormal mutation removal and time synchronization operations on the discrete time-series data. The data preprocessing module calls a spatial interpolation algorithm to perform vertical continuous fitting of the discrete time-series data along the depth axis. The data preprocessing module generates a spatiotemporal continuous data matrix with depth as the continuous independent variable.

[0033] The feature extraction module is connected to the data preprocessing module. The feature extraction module receives a spatiotemporal continuous data matrix. The feature extraction module uses the water temperature function and dissolved oxygen function in the spatiotemporal continuous data matrix to obtain the spatial first-order partial derivatives to obtain the vertical gradient of water temperature and the vertical gradient of dissolved oxygen. Based on the set gradient threshold and concentration threshold, the feature extraction module extracts candidate thermocline boundary, thermocline intensity, anoxic layer boundary and anoxic layer parameters from the spatiotemporal continuous data matrix.

[0034] The coupling analysis module is connected to the feature extraction module. The coupling analysis module receives parameters of the thermocline boundary, the oxygen-cline boundary, and the anoxic layer. The coupling analysis module uses a sliding time window algorithm to extract multi-source meteorological and hydrological features and water quality stratification features to construct a supervised learning sample set. The coupling analysis module constructs a dual-flow long short-term memory neural network based on the cross-attention mechanism as a dynamic coupling prediction model to calculate the predicted coordinates of the lower boundary of the thermocline, the lower boundary of the oxygen-cline, and the thickness of the anoxic water body in the future time period.

[0035] The stage determination module is connected to the data preprocessing module and the coupling analysis module respectively. The stage determination module receives the spatiotemporal continuous data matrix, the thermocline boundary depth parameter and the predicted spatial boundary evolution feature vector. The stage determination module calculates the relative thermal resistance of the water body based on the absolute density of the upper and lower water bodies, and divides the initial stage of thermal stratification evolution by combining the lower boundary coordinates of the thermocline. The stage determination module introduces a sliding window mode filtering algorithm to perform state smoothing processing to distinguish between short-term meteorological disturbances and the actual evolution trend. The stage determination module outputs the water body thermal stratification life cycle state identifier based on the matching results.

[0036] The results output module is connected to the stage determination module. The results output module receives the water body thermal stratification life cycle status identifier, the final thermocline lower boundary and oxygen consumption warning time. The results output module generates a visualization chart that includes the stratification evolution process, dynamic warning and vertical spatial distribution range. The results output module renders and outputs the final thermal stratification identification result to the terminal.

[0037] See attached document Figure 2 This invention provides a method for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device. The method includes the following steps: S100, the data preprocessing module receives raw monitoring data continuously collected by the vertical stratified monitoring device. The raw monitoring data includes the collection time, the discrete water depth location of the probe, the water temperature at each depth, and the dissolved oxygen concentration at each depth. At the same time, it synchronously accesses external meteorological and hydrological information to obtain a multi-source meteorological and hydrological dataset, and extracts time series features based on the raw monitoring data to obtain a time series feature dataset.

[0038] S200, the data preprocessing module performs anomaly removal and time synchronization processing on the raw monitoring data, and uses a cubic spline interpolation algorithm along the depth axis to perform vertical spatial continuity fitting, constructing a high-resolution profile continuity function with depth as the continuous independent variable, and generating a spatiotemporal continuous data matrix.

[0039] S300, the feature extraction module calculates the first-order partial derivative of the water temperature continuous function in the spatiotemporal continuous data matrix with respect to depth to calculate the water temperature vertical gradient, extracts continuous depth intervals with gradients greater than the preset temperature gradient threshold as candidate thermoclines, and extracts the upper boundary, initial lower boundary and intensity parameters of the candidate thermoclines.

[0040] S400, the feature extraction module calculates the first-order partial derivative of the dissolved oxygen continuous function in the spatiotemporal continuous data matrix with respect to depth to calculate the dissolved oxygen vertical gradient, identifies the depth interval where the gradient is greater than the preset dissolved oxygen gradient threshold as the oxygenation layer, and simultaneously determines the water layer space where the absolute concentration is lower than the preset concentration threshold as the hypoxic layer or anoxic layer.

[0041] The S500 coupling analysis module establishes a spatiotemporal synergy and dynamic coupling prediction model for water temperature and dissolved oxygen. It extracts multi-source meteorological and hydrological datasets and time-series feature datasets to construct a sample set. It uses a dual-flow long short-term memory neural network based on the cross-attention mechanism to implement multimodal feature fusion and regression decoding. It also introduces a joint loss function that includes environmental fluid dynamics physical constraints to optimize model parameters and outputs the spatial boundary absolute depth evolution state and anoxic water body thickness parameters within a specific time step in the future.

[0042] The S600 stage determination module combines the relative thermal resistance of the water body and the coordinates of the lower boundary of the thermocline to divide the thermal stratification evolution stages. It also introduces a sliding window mode filtering algorithm to perform state smoothing processing. Based on the statistical distribution probability on the continuous time series, it filters out the false short-term disturbances in water temperature caused by extreme weather and outputs the true thermal stratification evolution stage identifier.

[0043] S700, the results output module renders and outputs the identification results of lake and reservoir thermal stratification patterns. The identification results include the time series diagram of the evolution of the thermocline's start and end depth based on continuous spatial reconstruction and reverse boundary correction, the oxygen consumption early warning time data after baseline drift correction, the evolution process markers of each stage after filtering out short-term false disturbances, and the vertical distribution results of water temperature, dissolved oxygen and anoxic water bodies in time and depth profiles.

[0044] See attached document Figure 3 The specific implementation of step S100 can be handled through the following sub-steps: S101. A vertical monitoring device is deployed at a representative central point or deep water area of ​​the target lake or reservoir. The vertical monitoring device includes a buoy, a control mechanism, a steel cable, a counterweight, a temperature probe, and a dissolved oxygen probe. The upper end of the steel cable is connected to the buoy, and the lower end of the steel cable is connected to the counterweight. Under the tension of the counterweight, the steel cable remains vertically suspended in the water. The temperature probe and dissolved oxygen probe are deployed on the steel cable at a preset interval along the water depth direction. In the actual deployment of the temperature probe and dissolved oxygen probe, a non-uniform probe deployment interval is adopted. Surface monitoring nodes and bottom monitoring nodes are deployed at a depth of 0.5m below the water surface and a height of 0.5m above the bottom of the water, respectively.

[0045] This implementation uses the real-time water surface position as the zero point of depth and the vertical downward direction as the positive direction of the depth coordinate. The actual monitoring depth of each probe is obtained by a pressure-type depth sensor set at the corresponding probe position, or by conversion based on the real-time water level, the installation position of the probe on the steel cable, and the measurement results of the steel cable inclination angle, in order to correct the nominal installation depth error caused by water level changes or steel cable skew.

[0046] For the deployment of detection nodes in the middle of the water body, when historical continuous water temperature profile reference data is lacking, an empirical model is used to calculate and predict the depth of the temperature-variant layer center and the depth of the mixing layer. Within a 2.0m temperature-variant interval extending above and below the predicted temperature-variant layer center depth, a 0.5m spacing is used for denser deployment. Outside this temperature-variant interval, in other water layer sections, a spacing of 1.0m to 2.0m is used. The water area determines the wind power transmission path length and the water surface heat absorption area. The empirical model calculation mechanism for predicting the temperature-variant layer center depth, and the relevant quantitative calculation formulas, are as follows: ; In the formula, To predict the absolute depth of the thermosphere center; The water area parameter of the target lake or reservoir.

[0047] The empirical model formula for calculating the depth of the hybrid layer is constructed as follows: ; In the formula, This represents the absolute depth of the hybrid layer. The water area parameter of the target lake or reservoir.

[0048] When the predicted temperature change layer center depth, mixing layer depth, or temperature change range calculated based on the empirical model exceeds the actual water depth range of the target point, the corresponding depth is corrected to the effective monitoring range consisting of 0.5m below the water surface to 0.5m above the bottom of the water. When the correction positions of two or more probes overlap, one of the overlapping probes is retained, and the remaining probes are adjusted to the adjacent effective water layer at intervals of not less than 0.5m.

[0049] S102, the control mechanism inside the vertical monitoring device drives temperature probes and dissolved oxygen probes distributed at different water depths to synchronously collect environmental parameters according to a pre-set sampling frequency. The pre-set sampling frequency is configured to a time interval of 10 minutes to 60 minutes. The control mechanism acquires discrete observation values ​​at each depth. The control mechanism binds the collected physical parameters with the current sampling time and the spatial location of the probe and constructs a multi-dimensional single observation data tuple. The single observation data tuple includes water depth location parameters, collection time parameters, water temperature parameters at each depth, and dissolved oxygen parameters at each depth. The control mechanism synchronously collects and packages auxiliary water quality parameter fields such as pH, conductivity, redox potential, turbidity, total nitrogen, total phosphorus, and chlorophyll, and integrates the aforementioned fields to construct a complete water quality monitoring dataset.

[0050] The auxiliary water quality parameters are optional input fields. They are only used for result verification and auxiliary display data when not involved in subsequent thermocline identification or prediction model calculations. They are not necessary inputs for determining the thermocline, oxocline, and anoxic layer.

[0051] S103, the data preprocessing module listens to the remote communication port to receive the sequence of observation data tuples sent by the control agency through the wireless network. The data preprocessing module performs cyclic redundancy check and frame header and tail matching parsing on the arriving data frames in the memory buffer area. The data preprocessing module performs operations such as discarding damaged data frames that fail the cyclic redundancy check and issuing retransmission instructions to the control agency. It extracts each observation data tuple and appends each observation data tuple to the time series discrete data table.

[0052] The time-series discrete data table uses discrete time points as the horizontal axis and discrete probe deployment depth as the vertical axis. The cells in the time-series discrete data table record the water temperature and dissolved oxygen measurement values ​​under the corresponding spatiotemporal coordinates. The data preprocessing module continuously accumulates observation data tuples to form raw monitoring data, and extracts and integrates the raw monitoring data according to the time series dimension to construct a time-series feature dataset.

[0053] The data preprocessing module also receives temperature, wind speed and solar radiation data provided by meteorological acquisition units set on the buoy or external meteorological data interfaces, and receives water level, inflow and outflow data provided by water level gauges, flow meters or reservoir scheduling systems. All meteorological and hydrological data are bound to the acquisition timestamp and stored together with the original monitoring data according to a unified time base to form a multi-source meteorological and hydrological dataset required for subsequent coupled analysis.

[0054] See attached document Figure 4 The specific implementation of step S200 can be handled through the following sub-steps: S201, the data preprocessing module extracts the raw discrete time-series data, performs time synchronization processing, sets a sliding window and uses a sliding window mid-range filtering algorithm to remove abnormal jump data points. The sliding window contains at least 3 consecutive sampling data points. When the sampling interval is 10 min to 20 min, the window length is set to 30 min to 60 min. When the sampling interval is greater than 20 min, the window length is adjusted according to the sampling interval to cover at least 3 consecutive sampling times. The water temperature change rate limit threshold is set to 0.5℃ / 10 min to 2.0℃ / 10 min and the dissolved oxygen change rate limit threshold is set to 1.0 mg / L / 10 min to 3.0 mg / L / 10 min. Change rate limit judgment is implemented and the operation of filtering out excessive noise points is performed. The data preprocessing module converts the difference between adjacent observation values ​​into a change rate in 10 min units according to the actual time interval between two adjacent observation data, and then compares it with the corresponding change rate limit threshold.

[0055] The outlier data points that are removed are first marked as missing values. When the duration of consecutive missing values ​​between adjacent valid observations is no more than 3 standard sampling intervals, linear interpolation is used to fill in the missing values. When the duration of consecutive missing values ​​is more than 3 standard sampling intervals, straddle interpolation is not performed, and the corresponding time interval is marked as invalid data interval. The data in the invalid data interval is only used for missing status display and does not participate in layer extraction, model training, or control command generation.

[0056] S202, based on a standardized discrete profile sequence, uses a data preprocessing module to extract the actual deployment depth values ​​of each probe at a specific sampling time as an interpolation base point sequence. Simultaneously, it extracts the corresponding depth water temperature measurement value sequence or dissolved oxygen measurement value sequence as an observation value sequence. The depth values ​​of each interpolation base point are combined with the observation values ​​to construct a discrete spatial coordinate node set. When constructing the cubic spline interpolation function, a natural boundary condition is configured where the second derivative value at the water surface boundary node and the bottom boundary node is equal to zero. This ensures that the curvature of the fitting curve at the water surface and bottom boundary is zero, avoiding bending oscillations at the boundary positions without actual measurement basis. It establishes the smooth convergence transition constraint conditions of the reconstructed high-resolution spatial surface at the upper and lower physical boundaries, providing mathematical completeness input parameters for the fitting curve across the entire water depth profile space.

[0057] S203, the data preprocessing module constructs cubic polynomials along the depth axis within the depth interval formed by two adjacent interpolation base points. Based on the principle of cubic spline interpolation, each piecewise cubic polynomial is constrained to satisfy the continuity conditions of function values, first derivatives, and second derivatives at its internal nodes. A tridiagonal linear equation system is constructed by simultaneously solving the continuity equations at each node and the boundary conditions. This tridiagonal linear equation system possesses a strict diagonal dominance property under natural boundary conditions to ensure that the catch-up method solution process does not produce division-to-zero anomalies or error amplification. The catch-up method is used to solve the tridiagonal linear equation system to obtain the undetermined coefficients of each piecewise cubic polynomial, resulting in the vertically continuous fitting function expression at a specific sampling time. The calculation formula for the vertically continuous fitting function expression of the high-resolution profile continuity function at a specific sampling time is constructed as follows: ; In the formula, It is a piecewise cubic polynomial continuous function within a specific depth interval; The independent variable is the continuous depth. The depth coordinates of a specific interpolation base point; For the zero-order term of a cubic polynomial within a specific depth interval, there are undetermined constant coefficients. These are the undetermined coefficients of the linear term in a cubic polynomial within a specific depth interval; These are the undetermined coefficients of the quadratic term in a cubic polynomial within a specific depth interval; These are the undetermined coefficients of the cubic term in a cubic polynomial within a specific depth interval; A positive integer index number representing a sequence of a specific depth interval.

[0058] S204, the data preprocessing module traverses all sampling times on the standard time axis and repeatedly executes the cubic polynomial interpolation solution process to construct a high-resolution profile continuous function with depth as the continuous independent variable. This transforms the originally discrete distribution of measured values ​​into a high-resolution continuous numerical distribution with a grid spacing of 0.01m to 0.1m, and avoids the problem of missing the true gradient peak characteristics of the water body due to the discrete spatial gaps. To address the discontinuity of each profile in the time dimension, a linear interpolation algorithm is used along the time axis to achieve temporal continuity, obtaining continuous observation sequences with equal time steps. The nodes of these continuous observation sequences with equal time steps are used as the row dimension, and the discrete depth nodes extracted by the high-resolution profile continuity function at fixed intervals are used as the column dimension, constructing a two-dimensional spatiotemporal continuous data matrix. In this matrix structure, each row represents the vertical profile state at different depths under the same discrete time node, and each column represents the dynamic evolution sequence that continuously changes over time at the same specific depth, thus achieving two-dimensional coupling between the time axis and the spatial axis. The data preprocessing module integrates the high-resolution profile continuity functions at all times to generate a complete spatiotemporal continuous data matrix containing both water temperature and dissolved oxygen spatiotemporal continuous data matrices. This generated spatiotemporal continuous data matrix serves as a unified data base, used as input to the feature extraction module to accurately extract the boundary coordinate parameters of the thermocline, oxygen stratum, and the underlying anoxic water body.

[0059] The data preprocessing module performs a range check on the interpolation results within each adjacent probe interval. When the cubic spline interpolation result exceeds the numerical range formed by the adjacent measured values ​​and exceeds the preset tolerance, the corresponding interval is refitted using the shape-preserving piecewise cubic interpolation method, or the interpolation result is limited to the range formed by the adjacent measured values ​​and the preset tolerance. The data preprocessing module synchronously records the measured, interpolated, or invalid data quality identifiers corresponding to each grid point.

[0060] See attached document Figure 5 The specific implementation of step S300 can be handled through the following sub-steps: S301, the feature extraction module retrieves the water temperature continuous function from the spatiotemporal continuous data matrix. For a specific sampling time, it calculates the spatial first-order partial derivative of the water temperature continuous function with respect to the depth variable, obtaining the water temperature vertical gradient continuous function, which characterizes the degree of temperature change in the vertical direction of the water body. The absolute value of the water temperature vertical gradient continuous function is taken as the water temperature vertical gradient intensity to eliminate the influence of positive or negative signs caused by the depth coordinate direction and the increase or decrease of water temperature with depth. The calculation formula for obtaining the water temperature vertical gradient continuous function by calculating the spatial first-order partial derivative of the water temperature continuous function is constructed as follows: ; In the formula, It is a continuous function of the vertical gradient of water temperature; It is a spatially continuous function of water temperature with depth as the independent variable; For continuous depth, the independent variable is .

[0061] S302, the feature extraction module performs candidate interval screening based on the water temperature vertical gradient continuity function. It sets a preset temperature gradient threshold of 0.2℃ / m to 0.5℃ / m, traverses the entire water depth space, and extracts all independent depth intervals whose absolute value of the water temperature vertical gradient continuity function is continuously greater than or equal to the preset temperature gradient threshold. The feature extraction module simultaneously sets an effective thermocline thickness threshold of 0.5m to 1.0m, calculates the actual spatial span of each independent depth interval, and removes depth intervals whose actual spatial span is less than the effective thermocline thickness threshold. The combined screening operation filters out non-structural temperature fluctuation intervals caused by local microscale thermal disturbances in the water body, and outputs the depth intervals that pass the screening as candidate thermoclines.

[0062] When the interval between two adjacent candidate thermoclines is less than the preset merging distance, the two candidate thermoclines are merged into one candidate thermocline. The preset merging distance is 0.2m to 0.5m. When there are more than two independent candidate thermoclines at the same sampling time, the candidate thermocline with the largest absolute value of the vertical water temperature gradient along the depth is selected as the main thermocline. When the difference between the integral results of two candidate thermoclines is less than a preset ratio, the candidate thermocline with the smallest distance between the lower boundary coordinates of the candidate thermocline and the lower boundary coordinates of the main thermocline at the previous effective sampling time is selected as the main thermocline. The preset ratio is set to a range of 5% to 10%, and the subsequent thermocline boundary coordinates and thermocline intensity refer to the parameters corresponding to the main thermocline.

[0063] S303, the feature extraction module performs a quantization extraction operation on the spatial boundary parameters of each extracted candidate thermocline. It locates the starting shallow water depth coordinates of the candidate thermocline on the vertical axis of water depth as the upper boundary coordinates of the thermocline, and locates the ending deep water depth coordinates of the candidate thermocline on the vertical axis of water depth as the lower boundary coordinates of the thermocline. The feature extraction module extracts the water temperature measurement values ​​at the depths where the upper boundary coordinates of the thermocline are located as the upper boundary temperature parameter, and extracts the water temperature measurement values ​​at the depths where the lower boundary coordinates of the thermocline are located as the lower boundary temperature parameter.

[0064] S304, the feature extraction module calculates the intensity parameter of the thermocline based on the extracted upper and lower bound temperature parameters, as well as the combination of the upper and lower boundary coordinates of the thermocline. The thermocline intensity parameter is defined as the average temperature gradient inside the candidate thermocline. The quantization formula for calculating the thermocline intensity parameter based on the boundary extraction results is constructed as follows: ; In the formula, These are the strength parameters of the thermocline. The absolute temperature parameter of the upper boundary of the thermocline at the depth where the upper boundary coordinates are located; The absolute temperature parameter of the lower boundary of the thermocline at the depth where the lower boundary coordinates are located; This represents the absolute depth value of the lower boundary coordinates of the thermocline. This represents the absolute depth value of the upper boundary coordinates of the thermocline.

[0065] See attached document Figure 6 The specific implementation of step S400 can be handled through the following sub-steps: S401, the feature extraction module retrieves the dissolved oxygen continuity function from the spatiotemporal continuous data matrix. For a specific sampling time, it calculates the spatial first-order partial derivative of the dissolved oxygen continuity function with respect to the depth variable, obtaining the dissolved oxygen vertical gradient continuity function, which characterizes the severity of dissolved oxygen concentration decay in the vertical direction of the water body. The severity of dissolved oxygen concentration decay in the vertical direction of the water body directly maps to the objective physical state of the interface between biological respiration oxygen consumption and physical mixing constraints in the aquatic ecosystem. The calculation formula for obtaining the dissolved oxygen vertical gradient continuity function by calculating the spatial first-order partial derivative of the dissolved oxygen continuity function is constructed as follows: ; In the formula, It is a continuous function of the vertical gradient of dissolved oxygen; It is a spatially continuous function of dissolved oxygen with depth as the independent variable; For continuous depth, the independent variable is .

[0066] S402, the feature extraction module performs oxygen stratum space interval screening operation based on the dissolved oxygen vertical gradient continuity function, sets the dissolved oxygen gradient threshold to the range of 0.5 mg / L / m to 2.0 mg / L / m, traverses the entire water depth space and extracts all independent depth intervals whose absolute value of the dissolved oxygen vertical gradient continuity function is continuously greater than or equal to the dissolved oxygen gradient threshold.

[0067] The feature extraction module synchronously sets the effective oxygen cascade thickness threshold to a range of 0.5m to 1.0m. The feature extraction module calculates the actual spatial span of each independent depth interval and removes depth intervals with an actual spatial span smaller than the effective oxygen cascade thickness threshold. Through combined screening operations, it avoids misjudging non-structural dissolved oxygen fluctuation intervals caused by sensor hardware noise or local aggregation photosynthesis of phytoplankton as the main body of the cascade. If no independent depth interval that meets the combined screening conditions is extracted within the entire water depth range, it is determined that there is no oxygen cascade structure at the current sampling time and the oxygen cascade boundary quantization extraction operation at the current sampling time is terminated. The feature extraction module outputs the depth intervals that pass the screening as the effective oxygen cascade.

[0068] When there are more than two effective oxygen chokes at the same sampling time, the effective oxygen choke located near or below the main thermocline and with the highest continuity with the oxygen choke boundary at the previous effective sampling time is selected as the main oxygen choke. When there are still more than two effective oxygen chokes that meet the conditions, the effective oxygen choke with the largest absolute value of the dissolved oxygen vertical gradient along the depth is selected as the main oxygen choke. The subsequent oxygen choke boundary coordinates and oxygen choke intensities all refer to the parameters corresponding to the main oxygen choke.

[0069] S403, the feature extraction module extracts the boundary and intensity parameters of each effective oxygenation layer, locates the starting shallow water end depth coordinates of the effective oxygenation layer on the vertical axis of water depth as the upper boundary coordinates of the oxygenation layer, and locates the ending deep water end depth coordinates of the effective oxygenation layer on the vertical axis of water depth as the lower boundary coordinates of the oxygenation layer.

[0070] The feature extraction module extracts the dissolved oxygen measurement values ​​at the depth of the upper boundary coordinates of the oxoclimax as the upper boundary dissolved oxygen parameter, and extracts the dissolved oxygen measurement values ​​at the depth of the lower boundary coordinates of the oxoclimax as the lower boundary dissolved oxygen parameter. Based on the extracted upper and lower boundary dissolved oxygen parameters, as well as the combination of the upper and lower boundary coordinates of the oxoclimax, the feature extraction module calculates the effective oxoclimax intensity parameter. The effective oxoclimax intensity parameter quantitatively maps the gradient constraint physical potential energy of the redox environment isolation in the vertical space of the water body. The quantitative mapping formula for the effective oxoclimax intensity parameter established based on the absolute boundary parameter is constructed as follows: ; In the formula, The effective oxygen choke intensity parameter; The upper limit of dissolved oxygen is the absolute parameter at the depth where the upper boundary coordinates of the oxygen climax are located. The lower limit of dissolved oxygen absolute parameter at the depth where the lower boundary coordinates of the oxygen climax are located; This represents the absolute depth value of the lower boundary coordinates of the oxygen-climate layer; This represents the absolute depth value of the upper boundary coordinates of the oxygen climax layer.

[0071] S404, the feature extraction module performs hypoxic water body feature extraction operation based on the dissolved oxygen continuous function, and sets hypoxia judgment threshold and hypoxia judgment threshold respectively, wherein the hypoxia judgment threshold is less than the hypoxia judgment threshold; in a specific embodiment, the hypoxia judgment threshold is set to 2.0 mg / L and the hypoxia judgment threshold is set to 1.0 mg / L. When the dissolved oxygen concentration of the water body is lower than the hypoxia judgment threshold, it indicates that the bottom water body has entered an extremely restricted hypoxic evolution stage. This habitat state will trigger phosphorus release by anaerobic microorganisms in the bottom sediment.

[0072] The feature extraction module iterates through the dissolved oxygen continuity function and extracts water depth intervals where the dissolved oxygen measurement value is consistently less than or equal to the anoxic threshold. If the dissolved oxygen continuity function measurement value is greater than the anoxic threshold in all water depth spaces, or if there is no anoxic interval connected to the bottom boundary, the thickness parameter of the anoxic water body is recorded as zero. Otherwise, the feature extraction module extracts each independent water depth interval where the dissolved oxygen measurement value is consistently less than or equal to the anoxic threshold, and identifies the anoxic interval that is continuously connected to the bottom boundary as the bottom anoxic water body.

[0073] Extract the lower and upper boundaries of the bottom hypoxic water body. Specifically, determine the top depth coordinates of the bottom hypoxic water body as the upper boundary of the bottom hypoxic water body, and determine the maximum depth coordinates of the bottom bed as the lower boundary of the bottom hypoxic water body.

[0074] For independent anoxic zones in the middle layer that are not connected to the bottom of the water body, the upper and lower boundaries of the independent anoxic zones in the middle layer are recorded separately, but the thickness of the anoxic water body at the bottom layer is not included. The feature extraction module calculates the depth difference between the lower boundary and the upper boundary of the anoxic water body at the bottom layer and uses the depth difference as the thickness parameter of the anoxic water body. The feature extraction module archives and stores the oxygen stratification boundary parameters, effective oxygen stratification intensity parameters, and anoxic water body thickness parameters at each sampling time according to the time series, and outputs a dissolved oxygen stratification time series feature dataset.

[0075] See attached document Figure 7 The specific implementation of step S500 can be handled through the following sub-steps: S501, the coupling analysis module extracts the time-series feature dataset and the multi-source meteorological and hydrological dataset and performs timestamp alignment. The multi-source meteorological and hydrological dataset includes continuous temperature, wind speed, and solar radiation sequences. To meet the training requirements of the dynamic coupling prediction model, the coupling analysis module performs the creation of the model training sample dataset. Specifically, the coupling analysis module uses a sliding time window algorithm to slice the aligned multi-source meteorological and hydrological dataset and the time-series feature dataset to construct a supervised learning sample set. The time window length is set to 24h to 72h and the sliding time step is set to 1h. The time-series feature dataset and the multi-source meteorological and hydrological dataset within the sliding time window are concatenated and constructed as the input tensor.

[0076] The time dimension of the input tensor is defined as the number of historical time steps covered by the sliding time window, and the feature dimension is equal to the total number of channels of the multi-source input parameters. The coupling analysis module extracts the water stratification features of the future specific predicted time step adjacent to the sliding time window as the true label tensor. By mapping and pairing the input tensor with the true label tensor, supervised learning sample data pairs for model training are formed. The true label tensor covers the true lower boundary coordinates of the thermocline and the true lower boundary coordinates of the oxocline at a specific future time, as well as the true thickness parameter of the anoxic water body.

[0077] S502, the specific construction operation for the dynamic coupling prediction model includes: the coupling analysis module constructs a dual-stream long short-term memory neural network based on the cross-attention mechanism as the dynamic coupling prediction model. The dynamic coupling prediction model is internally divided into water quality feature extraction branch, meteorological feature extraction branch and cross-coupling attention layer, and establishes the network computing architecture of time series feature extraction and cross-modal feature fusion.

[0078] The water quality feature extraction branch uses a stacked two-layer long short-term memory network to receive a subset of water quality hierarchical features from the input tensor and outputs a water quality hidden state sequence. The feature dimension of the water quality hidden state sequence is configured to be 64 to 128. The meteorological feature extraction branch uses a single-layer long short-term memory network to receive a subset of multi-source meteorological and hydrological datasets and outputs a meteorological hidden state sequence with the same feature dimension configuration of 64 to 128. The coupling analysis module synchronously inputs the water quality hidden state sequence and the meteorological hidden state sequence into the cross-coupling attention layer. The cross-coupling attention layer is based on the physical causal relationship between meteorological and water quality changes. It uses the water quality hidden state sequence as the query vector and the meteorological hidden state sequence as the key vector and value vector to calculate the cross-attention weight matrix. The specific cross-attention calculation formula is as follows: ; In the formula, The query vector matrix for transforming the hidden state sequence of water quality; The key vector matrix for transforming the meteorological hidden state sequence; This is a value vector matrix transformed from a sequence of meteorological hidden states; express The transpose of the matrix; Indicates the dimension of the key vector; A scaling factor used to prevent gradient vanishing due to excessively large inner products.

[0079] After matrix multiplication, the water quality status is matched and addressed using meteorological conditions to obtain the water temperature and dissolved oxygen co-coupled feature vectors that integrate the meteorological driving mechanism. The cross-coupled attention layer inputs the water temperature and dissolved oxygen co-coupled feature vectors into the decoding regression network composed of three fully connected layers.

[0080] S503, the coupling analysis module performs forward propagation based on the decoding regression network and outputs the prediction tensor. The number of neurons in the output layer of the decoding regression network is set to 3 to correspond to the output of various spatial feature parameters. The values ​​of each item in the prediction tensor are directly mapped to the absolute depth evolution state of the spatial boundary of the thermal stratification and dissolved oxygen dissipation layer in the water body within the next 24h to 168h. The prediction tensor includes the predicted lower boundary coordinates of the thermocline and the lower boundary coordinates of the oxygencline at future times, as well as the predicted thickness of the anoxic water body. The coupling analysis module establishes a joint loss function that combines the mean square error algorithm with the physical constraint penalty term.

[0081] The joint loss function calculation formula for the dynamically coupled prediction model, which performs iterative parameter optimization during backpropagation, is constructed as follows: ; In the formula, For the joint loss function; The prediction evaluation tensor output by the model's forward propagation; To supervise the learning of the true label tensor for the alignment and assignment of the sample set; The physical penalty weight coefficient for constraining boundary conditions; The output of the forward inference is the predicted depth coordinates of the lower boundary of the thermocline. These are the predicted depth coordinates of the lower boundary of the oxygen climax, output by the forward inference.

[0082] The joint loss function introduces a mathematical mechanism that constrains the mathematical knowledge of natural water body physics on the basis of conventional numerical prediction errors. It transforms the prior physical law of environmental fluid dynamics, that the lower boundary depth of the thermocline in the vertical spatial profile of the water body is shallower than or equal to the lower boundary depth of the oxygencline, into a numerical penalty calculation boundary for the parameter iteration of the dynamic coupling prediction model.

[0083] S504, the training method for the dynamic coupling prediction model, specifically includes the following model training steps: First, the coupling analysis module divides the supervised learning sample set into a training set with a data ratio of 0.8 and a validation set with a data ratio of 0.2 according to the time sequence; Second, the network structure parameters and weight parameters of the dynamic coupling prediction model are initialized. The coupling analysis module sets the initial learning rate to 0.001 and the batch processing number to 64, and selects the existing adaptive moment estimation optimization algorithm as the solver for network weight optimization. Subsequently, the training phase begins. In each training iteration, the input tensors in the training set are fed into the dynamic coupling prediction model in batches according to the set batch processing quantity to perform forward propagation calculation. The dynamic coupling prediction model outputs a prediction evaluation tensor. Further, the prediction evaluation tensor output by forward propagation and the corresponding true label tensor are combined and substituted into the joint loss function formula to calculate the joint loss value of the current batch. The error gradient of the joint loss function with respect to the weight nodes inside each network layer of the model is calculated through the backpropagation algorithm, and the network weight parameters are updated based on the calculated error gradient using the adaptive moment estimation optimization algorithm. After each training iteration, forward propagation inference is performed on the updated dynamic coupling prediction model using validation set data. The measured value of the joint loss function on the validation set is calculated and monitored to evaluate the model's generalization prediction ability. Finally, to prevent overfitting, an early stopping validation mechanism is activated, and the iterative training steps of the dynamic coupling prediction model are continuously executed until the measured value of the joint loss function on the validation set no longer decreases for 10 consecutive iterations. At this point, the training process ends and exits.

[0084] The coupling analysis module saves the converged dynamic coupling prediction model network weight parameters and deploys them in the online evaluation system. The online evaluation system receives real-time multi-source time-series feature data and outputs the prediction space boundary evolution feature vector through forward propagation inference operations.

[0085] The dynamic coupling prediction model outputs predictions of future water stratification boundaries and the thickness of hypoxic water bodies, providing a quantitative basis for early warning of hypoxia risks in the bottom layer.

[0086] See attached document Figure 8 The specific implementation of step S600 can be handled through the following sub-steps: S601, the stage determination module extracts the thermocline boundary depth parameter from the time-series feature dataset, and converts the absolute water temperature measurement parameters at the upper and lower boundary coordinates of the thermocline into absolute density parameters of the shallow and deep water layers based on the empirical polynomial of the water state equation; the stage determination module calculates a continuous sequence of relative thermal resistance of the water body, representing the ease of mixing between the upper and lower water layers, based on the extracted upper and lower water density parameters. The relative thermal resistance of the water body quantifies the proportion of mechanical work required for natural wind to disrupt the vertical density stratification of the water body. The formula for calculating the relative thermal resistance of the water body, representing the physical damping limitation effect of mixing between the upper and lower water layers, is constructed as follows: ; In the formula, The relative thermal resistance of the water body; The absolute density parameter of the deep water layer at the depth where the lower boundary coordinates of the thermocline are located; The absolute density parameter of the shallow water layer at the depth where the upper boundary coordinates of the thermocline are located; This is the reference constant for the maximum density of pure water at a water temperature of 4℃. It is the standard constant for the density of pure water at a water temperature of 5℃.

[0087] S602, the stage determination module performs an initial determination of the thermostratification evolution stage based on the continuous sequence of relative thermal resistance of the water body and the lower boundary coordinates of the thermocline. If the current discrete time node lacks valid thermocline boundary coordinate data, the stage determination module directly determines and marks the current discrete time node as the fully mixed stage. The stage determination module sets the resistance threshold for the formation stage to a range of 20 to 50 and the resistance threshold for the stable stage to a range of 80 to 150. The stage determination module marks discrete time nodes where the relative thermal resistance of the water body is less than the resistance threshold for the formation stage as the mixed stage. The phase determination module calculates the time first derivative of the lower boundary coordinates of the thermocline for discrete time nodes where the relative thermal resistance of the water body is greater than or equal to the resistance threshold of the formation period and less than the resistance threshold of the stable period. The phase determination module marks discrete time nodes with a time first derivative greater than zero as the decay phase and discrete time nodes with a time first derivative less than or equal to zero as the formation phase. The phase determination module marks discrete time nodes where the relative thermal resistance of the water body is greater than or equal to the resistance threshold of the stable period as the stable phase. The phase determination module combines the marking results of various discrete time nodes and outputs an initial stratification state sequence covering the complete time span.

[0088] S603, the stage determination module performs anti-interference filtering on the initial hierarchical state sequence. The stage determination module introduces a sliding window mode filtering algorithm to perform state smoothing. The stage determination module sets the anti-interference sliding window length to a range of 72h to 168h. Within the set anti-interference sliding window, the stage determination module counts the frequency of occurrence of various evolution stage state labels: The stage determination module extracts the evolution stage state label with the highest frequency within the anti-interference sliding window as the corrected evolution stage of the current central time node. The state smoothing operation filters out the high-frequency jump characteristic state caused by the violent fluctuation of shallow water temperature caused by natural wind and wave disturbances and heavy rainfall meteorological events based on the statistical distribution probability on the continuous time series. The anti-interference filtering operation outputs a reconstructed smooth layered state evolution sequence that removes the non-periodic interference results of short-term meteorological extreme events.

[0089] S604, the stage determination module combines the predicted lower boundary coordinates of the thermocline and the predicted lower boundary coordinates of the oxocline output by the dynamic coupling prediction model to perform a comprehensive output operation on the evolution trend. The stage determination module extracts the predicted lower boundary coordinates of the thermocline and the predicted lower boundary coordinates of the oxocline to quantify the stable state of water stratification at future moments. The stage determination module reconstructs the smooth stratification state evolution sequence and performs time axis alignment and splicing with the predicted lower boundary coordinates of the thermocline and the predicted lower boundary coordinates of the oxocline according to the time occurrence sequence to output the true thermal stratification evolution stage identifier.

[0090] Based on the real thermal stratification evolution stage identifiers, the stage determination module extracts the water body habitat evolution patterns within a complete time span and outputs a comprehensive evolution pattern map that includes the water body thermal stratification cycle determination results and bottom layer hypoxia risk warning indicators. The comprehensive evolution pattern map quantitatively maps the dynamic coupling evolution state of the water body's internal environment fluid mechanics and provides bottom layer objective data parameters to support the reservoir's stratified water intake scheduling actions and the start-stop control actions of aeration and oxygenation equipment.

[0091] See attached document Figure 9 The specific implementation of step S700 can be handled through the following sub-steps: S701, the result output module extracts the temporal feature dataset, the predicted spatial boundary evolution feature vector, and the smoothed hierarchical state evolution sequence. The result output module uses the hourly timestamp of the predicted spatial boundary evolution feature vector as the time axis reference to perform alignment and resampling operations on multimodal heterogeneous data. The result output module performs three-dimensional spatiotemporal gridding reconstruction operations on the aligned multimodal heterogeneous data. The result output module sets the time axis grid resolution to 1h and the water depth axis grid resolution to the range of 0.1m to 0.5m. The result output module maps discrete time nodes and spatial depth coordinates to the interior of a standard three-dimensional spatiotemporal data cube.

[0092] If there are missing data grids inside the standard three-dimensional spatiotemporal data cube due to packet loss in hardware sensor communication, the result output module calls a three-dimensional spatial interpolation algorithm based on inverse distance weighting to fill the missing data grids. The result output module sets the distance decay power exponent of the inverse distance weighting algorithm to 2 to quantify the nonlinear decay law of spatial distance on the correlation of physical and chemical parameters. The missing data grid filling mechanism avoids the rendering engine from triggering rendering interruption errors when encountering missing matrix elements. The result output module outputs a multidimensional habitat evolution data cube with continuous spatiotemporal domain distribution characteristics.

[0093] S702, the result output module performs continuous calculation of the water quality deterioration early warning index based on the multidimensional habitat evolution data cube. The water quality deterioration early warning index quantitatively maps the ecological stress effect of the superposition and fusion of the diffusion area of ​​the anaerobic environment in the bottom water body and the degree of physical isolation of the vertical habitat. The proportion of the thickness of the anoxic water body represents the spatial scale of phosphorus release from the bottom reducing sediment, and the intensity of the oxygen choke represents the physical barrier effect that blocks the exchange of substances between the upper and lower layers of the water body.

[0094] The formula for calculating the water quality deterioration early warning index, which reflects the cumulative effect of habitat degradation, is as follows: ; In the formula, It serves as a continuous early warning index for water quality deterioration; Assign a weighted attenuation coefficient to the hypoxia thickness; To predict the absolute value of the thickness of anoxic water bodies; The maximum depth of the water body; Configure a weighted attenuation coefficient for the stratified barrier; The effective oxygen choke intensity parameter; This is the preset maximum baseline value for the strength of the oxygen quench layer.

[0095] The results output module sets the weighting coefficient for anoxic thickness to 0.6 and the weighting coefficient for stratification barrier to 0.4. The results output module sets the maximum reference value for the preset oxygen stratum intensity to 5.0 mg / L / m. The results output module calculates the water quality deterioration early warning index sequence on the continuous time axis and outputs a comprehensive evolution pattern map covering water temperature isolines, dissolved oxygen isolines, and the water quality deterioration early warning index sequence.

[0096] S703, the result output module generates the underlying command parameters for reservoir stratified water intake scheduling and aeration and oxygenation equipment start-stop control actions based on the comprehensive evolution law map. The result output module sets the early warning index alarm threshold to a range of 0.5 to 0.8. When the water quality deterioration early warning index is continuously greater than or equal to the early warning index alarm threshold, the result output module generates the underlying aeration equipment start command.

[0097] The output module extracts the coordinates of the upper boundary of the thermocline at the current time point and shifts the coordinates upward by 1.0m to 2.0m as the safe water intake depth parameter. If the safe water intake depth parameter after upward shift is less than the physical coordinates of the zero depth of the water surface, the output module corrects the safe water intake depth parameter to a depth parameter of 0.5m. The forced depth correction logic compensates for the solution anomalies caused by the extremely shallow water layer structure during the dry season, which leads to the theoretical calculation elevation deviating from the physical boundary of the water body. The output module extracts the spatial midpoint depth between the lower boundary and the upper boundary of the anoxic water body as the deployment depth parameter of the bottom aeration equipment.

[0098] S704, the result output module serializes and encapsulates the control strategy set, which includes the start-up command of the underlying aeration equipment, the safe water intake depth parameter, and the deployment depth parameter. The network distribution and deserialization of the control strategy set communication transmission can be implemented according to the existing IoT application layer communication protocol. The result output module pushes the comprehensive evolution law map and the control strategy set to the large screen display terminal and the reservoir electromechanical control system at a fixed time period.

[0099] The comprehensive output processing logic executed by the result output module reduces the dimensionality of microscopic spatiotemporal water body physicochemical profile data and maps it into a concrete physical space operation index for macroscopic reservoir operation and management. The water body thermal stratification evolution law and bottom hypoxia risk early warning index output by objective calculation methods provide a data source for the active intervention physical mechanism of deep physicochemical habitat deterioration events in lakes and reservoirs. The result output module synchronously writes the feature parameters and control instructions generated by full processing into the local time series database for persistent storage.

[0100] To aid in understanding the technical solution of this invention, a specific application example and test case of monitoring and early warning of thermal stratification and water quality evolution in deep-water reservoirs are provided below.

[0101] The implementation example was deployed in a test reservoir with an area of ​​45 square kilometers and a maximum water depth of 60 meters in the central water intake area. The execution time of the application example covered the complete cycle of thermal stratification evolution of the water body.

[0102] According to the system architecture configuration, vertical monitoring devices are deployed in the section from below the test water surface to the bottom. The data preprocessing module receives discrete time series data sent by the vertical monitoring devices, calls the sliding time window mid-range filtering algorithm to locate data nodes whose rate of change exceeds the limit and performs abnormal mutation removal operation. The data preprocessing module executes the cubic spline interpolation algorithm to generate a high-resolution profile continuous function and constructs a spatiotemporal continuous data matrix containing water temperature and dissolved oxygen.

[0103] The feature extraction module extracts the continuous functions of water temperature and dissolved oxygen within the spatiotemporal continuous data matrix, and calculates the spatial first-order partial derivatives of the independent variable with respect to depth. The feature extraction module traverses the continuous function of vertical gradient of water temperature and extracts spatially independent depth intervals with values ​​continuously greater than 0.3 degrees Celsius per meter as the upper and lower boundaries of the thermocline. The feature extraction module traverses the continuous function of vertical gradient of dissolved oxygen and extracts depth intervals with values ​​continuously greater than 1.0 milligrams per liter per meter as the boundary of the oxygenation layer. Simultaneously, a hypoxia threshold of 1.0 milligrams per liter is set, and the thickness parameter of the hypoxic water body is determined and output.

[0104] The coupling analysis module reconstructs meteorological and hydrological data and water quality time-series characteristics into input tensors, which are then input into a dual-flow long short-term memory neural network based on a cross-attention mechanism. The dynamic coupling prediction model outputs a predicted spatial boundary evolution feature vector containing the predicted coordinates of the lower boundary of the thermocline, the predicted coordinates of the lower boundary of the oxocline, and the predicted thickness of the anoxic water body through forward propagation inference. The stage determination module calculates the continuous sequence of relative thermal resistance of the water body based on the absolute density parameters of the shallow water layer and the absolute density parameters of the deep water layer, and divides the initial stage of thermal stratification evolution by combining the coordinates of the lower boundary of the thermocline.

[0105] The stage determination module applies a sliding window mode filtering algorithm to perform state smoothing and outputs the thermal stratification life cycle state identifier. The result output module calculates a continuous sequence of water quality deterioration early warning index based on a preset formula, generates start-up instructions for the bottom aeration equipment, and outputs safe water intake depth parameters and aeration equipment deployment depth parameters by combining the upper boundary coordinates of the thermocline.

[0106] The test cases were used to compare the predictive performance of the method of this invention, the traditional threshold empirical model, and the standard long short-term memory network model in a test reservoir. The test period was divided into the formation stage, the stable stage, and the decline stage according to the natural time series. The evaluation indicators were set as the root mean square error of the predicted depth of the lower boundary of the thermocline in the next 24 hours, the root mean square error of the predicted depth of the lower boundary of the oxocline, the F1 score of the state evolution stage, and the boundary physical violation rate. The boundary physical violation rate was defined as the percentage of test samples in which the predicted depth of the lower boundary of the thermocline was deeper than the predicted depth of the lower boundary of the oxocline.

[0107] The specific test results are shown in the table below.

[0108] Multi-stage comparative test data table of predictive performance and physical violation rate: According to the table above and Figure 10 The data shows that the prediction error and state determination accuracy of the method of the present invention are better than those of the traditional threshold empirical model and the standard long short-term memory network model in different stratification evolution test stages. The traditional threshold empirical model lacks a time-series mapping mechanism and environmental covariate fusion channel. In the decay stage when the thermal stratification structure undergoes drastic changes, the root mean square error of the prediction of the lower boundary of the thermocline reaches 3.15m, which cannot provide reliable spatial boundary coordinates for reservoir scheduling.

[0109] Furthermore, traditional threshold empirical models employ univariate threshold-triggered calculation rules and lack collaborative calculation channels for multi-source heterogeneous parameters. They cannot simultaneously output the relative spatial coordinates of the lower boundary of the thermocline and the lower boundary of the oxygencline, resulting in invalid data in the statistical results of traditional threshold empirical models in the dimension of boundary physical violation rate.

[0110] The standard long short-term memory network model has a time-series processing architecture, and the root mean square error of various deep prediction errors is reduced compared with the traditional threshold empirical model. However, the boundary physics violation rate of the standard long short-term memory network model rises to 17.4% during the decay phase, reflecting that conventional data-driven models will output boundary reversal prediction coordinates that violate the common sense of fluid mechanics when meteorological conditions fluctuate.

[0111] The coupling analysis module of the present invention constructs a dynamic coupling prediction model including a cross-coupled attention layer. It performs matrix multiplication on the hidden state sequences of the water quality feature extraction branch and the meteorological feature extraction branch to obtain the cooperative coupling feature vector. The root mean square error of the prediction of the lower boundary of the thermocline during the formation, stable and decay stages is maintained in the range of 0.35m to 0.54m.

[0112] To address the issue of boundary physical violations, this invention establishes a joint loss function to introduce a hydrodynamic constraint mechanism for the natural water body environment. An error penalty operator is applied to numerical nodes where the predicted lower boundary of the thermocline in the vertical spatial profile of the water body is deeper than the predicted lower boundary of the oxycline. The joint loss function constrains the iteration direction of the network weights of the dynamically coupled prediction model, ensuring that the boundary physical violation rate of this invention remains at 0% throughout the entire test period. This quantitatively proves that the forward propagation output of the model satisfies the physical logic of the natural spatial hierarchy.

[0113] The stage determination module introduces a sliding window mode filtering algorithm based on the calculation mechanism of relative thermal resistance of water bodies. It filters out the short-term temperature disturbance jump characteristics caused by local rainfall and wind waves based on the statistical distribution probability of the time series. The method of this invention maintains a state determination F1 score of 0.93 in the decay stage of unstable boundary morphology. Test data shows that the predicted spatial boundary evolution feature vector extracted by the method of this invention, combined with physical prior rules and multi-source feature fusion architecture, has numerical stability. The hierarchical evolution process and early warning command control parameters generated by the result output module conform to the objective laws of natural habitat change and meet the data input requirements for water quality monitoring and control equipment in water sources to perform precise physical intervention operations.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the thermal stratification patterns of lakes and reservoirs based on a stratified monitoring device, characterized in that, The method includes the following steps: Acquire raw monitoring data, multi-source meteorological and hydrological datasets, and time-series feature datasets; The raw monitoring data is processed to generate a spatiotemporal continuous data matrix; Based on the spatiotemporal continuous data matrix, extract the upper boundary coordinates and lower boundary coordinates of the thermocline. Based on the spatiotemporal continuous data matrix, the effective oxygenation layer, the coordinates of the upper and lower boundaries of the oxygenation layer, the lower boundary of the bottom anoxic water body, the upper boundary of the bottom anoxic water body, and the thickness parameter of the anoxic water body are extracted. The multi-source meteorological and hydrological dataset and the time-series feature dataset are input into a dual-flow long short-term memory neural network based on cross-attention mechanism. A joint loss function is introduced for parameter optimization, and the predicted lower boundary coordinates of the thermocline, the predicted lower boundary coordinates of the oxocline, and the predicted thickness of the anoxic water body are output. Based on the coordinates of the upper boundary of the thermocline and the coordinates of the lower boundary of the thermocline, a continuous sequence of relative thermal resistance of the water body is calculated to divide the thermal stratification evolution stages and output the true thermal stratification evolution stage identifiers. Calculate the water quality deterioration early warning index sequence and output the lake and reservoir thermal stratification pattern identification results, which include the start-up command of the bottom aeration equipment and the safe water intake depth parameter.

2. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for processing the raw monitoring data to generate a spatiotemporally continuous data matrix include: The original monitoring data is subjected to anomaly removal and time synchronization processing, and vertical spatial continuity fitting is performed using a cubic spline interpolation algorithm along the depth axis. A sliding window mid-range filtering algorithm is used to remove abnormal jump data points. Combined with the water temperature change rate limit threshold and the dissolved oxygen change rate limit threshold, the change rate limit judgment is implemented and the excessive noise point filtering operation is performed. Extract the interpolation base point sequence and combine it with the observation value sequence to construct a discrete spatial coordinate node set, and configure the natural boundary condition that the second derivative value is equal to zero at the water body surface boundary node and bottom boundary node; By combining the continuity equations of each node in the discrete spatial coordinate node set with the natural boundary conditions, a tridiagonal linear equation system is constructed. The undetermined coefficients of each piecewise cubic polynomial are obtained by using the chasing method to reconstruct the high-resolution profile continuity function. To address the discontinuity of each profile in the time dimension, a linear interpolation algorithm is used along the time axis to perform time continuity processing, thereby obtaining a continuous observation sequence with equal time steps. By using the nodes of the continuous observation sequence with equal time steps as the row dimension and the discrete depth nodes extracted by the high-resolution profile continuous function at fixed intervals as the column dimension, a two-dimensional spatiotemporal continuous data matrix is ​​constructed.

3. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for extracting the upper and lower boundary coordinates of the thermocline based on the spatiotemporal continuous data matrix include: The first-order partial derivative of the water temperature continuous function in the spatiotemporal continuous data matrix with respect to depth is used to calculate the vertical gradient of water temperature. The system traverses the entire water body depth space to extract independent depth intervals where the absolute value of the vertical temperature gradient is continuously greater than or equal to a preset temperature gradient threshold. It then removes depth intervals whose actual spatial span is less than the effective thermocline thickness threshold and outputs the depth intervals that pass the screening as candidate thermoclines. When there are two or more independent candidate thermoclines at the same sampling time, the candidate thermocline with the largest absolute value of the vertical gradient of water temperature along the depth integral result, or the smallest distance between the lower boundary and the lower boundary of the main thermocline at the previous effective sampling time, is selected as the main thermocline. The starting shallow water depth coordinates of the candidate thermocline on the vertical water depth axis are used as the upper boundary coordinates of the thermocline, and the ending deep water depth coordinates of the candidate thermocline on the vertical water depth axis are used as the lower boundary coordinates of the thermocline.

4. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for extracting the effective oxocele, the coordinates of the upper and lower boundaries of the oxocele, the lower and upper boundaries of the anoxic water body, and the thickness parameter of the anoxic water body based on the spatiotemporal continuous data matrix include: The first-order partial derivative of the dissolved oxygen continuity function in the spatiotemporal continuous data matrix with respect to depth is used to calculate the vertical gradient of dissolved oxygen. The effective oxygen choke layer is defined as all independent depth intervals whose absolute value of the dissolved oxygen vertical gradient is continuously greater than or equal to the dissolved oxygen gradient threshold and whose actual spatial span is greater than or equal to the effective oxygen choke layer thickness threshold. The starting shallow water depth coordinates of the effective oxygenation layer on the vertical water depth axis are used as the upper boundary coordinates of the oxygenation layer, and the ending deep water depth coordinates of the effective oxygenation layer on the vertical water depth axis are used as the lower boundary coordinates of the oxygenation layer. The dissolved oxygen continuity function is traversed, and the hypoxic intervals where the dissolved oxygen measurement value is continuously less than or equal to the hypoxia determination threshold and is continuously connected to the bottom boundary of the water body are extracted as the bottom hypoxic water body. The lower boundary and upper boundary of the bottom hypoxic water body are extracted, and the depth difference between the lower boundary and the upper boundary of the bottom hypoxic water body is calculated and used as the thickness parameter of the hypoxic water body.

5. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps of inputting the multi-source meteorological and hydrological dataset and the time-series feature dataset into a dual-flow long short-term memory neural network based on a cross-attention mechanism, introducing a joint loss function for parameter optimization, and outputting the predicted coordinates of the lower boundary of the thermocline, the predicted coordinates of the lower boundary of the oxocline, and the predicted thickness of the anoxic water body include: The input tensor corresponding to the time-series feature dataset is obtained. The water quality feature extraction branch uses a stacked double-layer long short-term memory network to receive the water quality hierarchical feature subset in the input tensor and outputs the water quality hidden state sequence. Extract a subset of the multi-source meteorological and hydrological dataset from the multi-source meteorological and hydrological dataset. The meteorological feature extraction branch uses a single-layer long short-term memory network to receive the subset of the multi-source meteorological and hydrological dataset and outputs a meteorological hidden state sequence. In the cross-coupling attention layer, the water quality hidden state sequence is used as the query vector, and the meteorological hidden state sequence is used as the key vector and value vector to calculate the cross-attention weight matrix. After matrix multiplication, the water temperature and dissolved oxygen co-coupling feature vector of the fusion meteorological driving mechanism is obtained. The water temperature and dissolved oxygen co-coupling feature vector is input into the decoding regression network composed of three fully connected layers, and the predicted lower boundary coordinates of the thermocline, the predicted lower boundary coordinates of the anoxic layer, and the predicted thickness of the anoxic water body are output.

6. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps of inputting the multi-source meteorological and hydrological dataset and the time-series feature dataset into a dual-flow long short-term memory neural network based on a cross-attention mechanism, introducing a joint loss function for parameter optimization, and outputting the predicted coordinates of the lower boundary of the thermocline, the predicted coordinates of the lower boundary of the oxocline, and the predicted thickness of the anoxic water body include: Obtain the mean squared error between the prediction evaluation tensor output by the model forward propagation and the true label tensor aligned with the supervised learning sample set. Subtract the predicted lower boundary coordinates of the thermocline from the predicted lower boundary coordinates of the oxycline, and then perform a maximum value operation with zero to obtain the maximum value. The maximum value is multiplied by the physical penalty weight coefficient of the constraint boundary conditions, and then added to the mean square error to obtain the joint loss function. The network weight parameters are updated using the backpropagation algorithm through the joint loss function.

7. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for calculating a continuous sequence of relative thermal resistance of the water body based on the coordinates of the upper and lower boundaries of the thermocline to delineate the stages of thermostratification evolution include: The absolute water temperature measurement parameters at the upper and lower boundary coordinates of the thermocline are converted into absolute density parameters of shallow and deep water layers. For each discrete time node, the first difference between the absolute density parameters of deep and shallow water layers is calculated, and the second difference between the maximum pure water density reference constant and the standard pure water density benchmark constant is calculated. The first difference is divided by the second difference to obtain the relative thermal resistance of the water body corresponding to each discrete time node. The relative thermal resistance of the water body corresponding to each discrete time node is arranged in the order of time occurrence to generate a continuous sequence of relative thermal resistance of the water body that characterizes the degree of mixing difficulty between the upper and lower water layers. Discrete time nodes in the continuous sequence of relative thermal resistance of the water body that are less than the resistance threshold of the formation period are marked as the mixing period stage, and discrete time nodes in the continuous sequence of relative thermal resistance of the water body that are greater than or equal to the resistance threshold of the stable period are marked as the stable period stage. For discrete time nodes in the continuous sequence of relative thermal resistance of the water body that are between the resistance threshold of the formation period and the resistance threshold of the stable period, the time first derivative of the lower boundary coordinate of the thermocline is calculated. Discrete time nodes with the time first derivative greater than zero are marked as the decay period stage, and discrete time nodes with the time first derivative less than or equal to zero are marked as the formation period stage.

8. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for outputting accurate thermal stratification evolution stage identifiers include: A sliding window mode filtering algorithm is introduced to implement state smoothing. The aforementioned marked stages are used as state labels for various evolution stages. Within a set anti-interference sliding window, the frequency of occurrence of various evolution stage state labels is counted, and the evolution stage state label with the highest frequency is extracted as the corrected evolution stage of the current central time node. The reconstructed smooth stratified state evolution sequence, after removing non-periodic interference from short-term meteorological extreme events, is output. The reconstructed smooth stratified state evolution sequence is then spliced ​​with the predicted lower boundary coordinates of the thermocline and the predicted lower boundary coordinates of the oxygencline according to their time occurrence sequence to output the true thermal stratification evolution stage identifier.

9. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for calculating the water quality deterioration early warning index sequence include: Obtain the ratio of the predicted hypoxic water thickness to the maximum water depth, and multiply the ratio by the hypoxic thickness configuration weight attenuation coefficient. Calculate the effective oxygen jump layer strength parameter, obtain the ratio of the effective oxygen jump layer strength parameter to the preset maximum reference extreme value of oxygen jump layer strength, and multiply the ratio by the layered barrier configuration weight attenuation coefficient. The product of the ratio of the predicted hypoxic water body thickness to the maximum water body depth and the hypoxic thickness configuration weight attenuation coefficient, and the product of the ratio of the effective oxygenation layer strength parameter to the preset maximum baseline extreme value of oxygenation layer strength and the layer barrier configuration weight attenuation coefficient, are added to obtain the continuous water quality deterioration early warning index to calculate the water quality deterioration early warning index sequence on the continuous time axis.

10. The method for identifying the thermal stratification pattern of lakes and reservoirs based on a stratified monitoring device according to claim 1, characterized in that, The steps for outputting the lake / reservoir thermal stratification pattern identification results, which include the bottom aeration equipment start-up command and the safe water intake depth parameter, are as follows: When the water quality deterioration early warning index in the water quality deterioration early warning index sequence is continuously greater than or equal to the early warning index alarm threshold, a bottom aeration device start command is generated, and the spatial midpoint depth between the lower boundary of the bottom hypoxic water body and the upper boundary of the bottom hypoxic water body is extracted as the deployment depth parameter of the bottom aeration device. Extract the coordinates of the upper boundary of the thermocline at the current time point and shift the coordinates of the upper boundary of the thermocline upward by a set value as a safe water intake depth parameter; If the safe water intake depth parameter after upward translation is less than the physical coordinate of zero water surface depth, then the safe water intake depth parameter is corrected to the depth physical coordinate limit parameter.

Images