Algal bloom early warning and prevention method based on precursor pattern recognition

Abstract

The application discloses an algae-source lake flooding early warning and prevention and control method based on precursor pattern recognition, and aims at solving the problems of low accuracy of existing lake flooding early warning, difficulty in quantifying prediction uncertainty, and lack of closed-loop adaptive update of early warning and prevention and control. The application constructs training samples by collecting historical monitoring data of target water area, trains a precursor pattern recognition model containing a physically constrained neural dynamics model and a multi-task output network, performs time alignment and space matching on an observation data sequence online to extract features, uses ensemble Kalman filtering to iteratively assimilate state vectors to obtain assimilated state sequences and their uncertainties, drives the physically constrained neural dynamics model to generate a predicted state sequence, and then performs multi-task reasoning to output a lake flooding risk level and a predicted outbreak time window warning result. Furthermore, the warning result is converted into a prevention and control instruction, and assimilation correction and incremental training update are performed in combination with effect monitoring data, so that the technical effects of high-precision and rolling optimization of lake flooding early warning and whole-process linkage prevention and control are achieved.

Description

Technical Field

[0001] This invention relates to the field of environmental protection and water environment management, and in particular to a method for early warning and prevention of algal blooms in lacustrine environments based on precursor pattern recognition. Background Technology

[0002] Algal blooms typically occur in the shoreline or bays of eutrophic shallow lakes. Their formation generally involves the accumulation and decomposition of large amounts of cyanobacteria on the surface, leading to a decrease in dissolved oxygen. The sediment-water interface gradually transforms into a reducing environment, and reduction products such as iron and sulfur from the sediment enter the overlying water, causing the water to turn black and develop a foul odor. To address this problem, existing technologies have established a certain early warning and control system. Early warning primarily employs on-site monitoring and analysis, with monitoring indicators typically including meteorological conditions, surface algal accumulation, redox potential at the sediment-water interface, and ferrous ion and sulfide concentrations in the overlying water. Risk classification and outbreak assessment are based on empirical thresholds or criteria. For prevention and control, measures such as cyanobacteria interception and removal, the deployment of containment and capture facilities, aeration and oxygenation regulation, dredging and removal of contaminated sediment, and restoration of aquatic plants are implemented to reduce the probability of blooms or mitigate their impact. Simultaneously, some studies are attempting to introduce dynamic models or machine learning methods for prediction to improve the automation and foresight of early warning systems.

[0003] However, existing technologies still have the following shortcomings: First, early warning systems often rely on single indicators or static thresholds, making it difficult to reflect the continuous evolution of lake flooding under the coupling of multiple indicators, resulting in low accuracy in predicting outbreak time windows. Second, monitoring data often suffers from problems such as time asynchrony, spatial mismatch, noise, and missing data, lacking mechanisms for assimilating and correcting key states and quantifying uncertainties, making it difficult to assess the reliability of early warning results. Third, early warning and control implementation often lack closed-loop feedback, making it difficult to use control effect data to update prediction models and optimize strategies, hindering the achievement of rolling iterative, full-process coordinated control.

[0004] Therefore, a method for early warning and prevention of algal blooms in lakes that can overcome the shortcomings of the existing technologies is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for early warning and control of algal blooms based on precursor pattern recognition. Addressing the shortcomings of existing technologies that rely on static thresholds or single models, resulting in insufficient characterization of the bloom evolution process, noise and missing data in monitoring data making it difficult to quantify early warning uncertainty, and a lack of feedback and rolling optimization in early warning and control, this invention proposes a precursor pattern recognition model based on historical monitoring data. This model constructs training samples and trains a precursor pattern recognition model incorporating a physically constrained neurodynamic model and a multi-task output network. During online monitoring, the observed data is temporally aligned and spatially matched to extract features. An ensemble Kalman filter is used to iteratively assimilate the state vector to obtain an assimilated state sequence and its uncertainty. This drives the physically constrained neurodynamic model to generate a predicted state sequence and performs multi-task inference to output the bloom risk level and expected outbreak time window. Simultaneously, the early warning results are transformed into control instructions, and the model is updated through assimilation correction and incremental training using effect monitoring data. This closed-loop technical solution improves the accuracy and interpretability of early warnings, enhances credibility by outputting uncertainty, and achieves adaptive optimization of early warning and full-process coordinated control.

[0006] This invention provides a method for early warning and prevention of algal blooms in lacustrine environments based on precursor pattern recognition, comprising:

[0007] S1. Collect historical monitoring data of the target water area and form a training sample set according to a preset time slice. Each training sample includes observation data of precursor indicators and corresponding lake flooding labels. S2. Train a precursor pattern recognition model using the training sample set, including a physically constrained neural dynamics model and a multi-task output network. S3. Conduct online monitoring of the target water area to obtain observation data sequences. S4. Perform time alignment and spatial matching on the observation data sequences, and calculate feature sequences, including the numerical characteristics and rate of change characteristics of each precursor indicator. S5. Establish a state vector based on the feature sequences, and use ensemble Kalman filtering to iteratively calibrate the state vector. S6. Obtain the assimilation state sequence and its uncertainty; S7. Use the assimilation state sequence and its uncertainty to drive the physical constraint neurodynamic model in the precursor pattern recognition model to generate the predicted state sequence; S8. Input the predicted state sequence into the precursor pattern recognition model for multi-task reasoning to generate the early warning result; S9. Convert the early warning result into prevention and control instructions and execute the whole-process linkage prevention and control to obtain effect monitoring data; S0. Use the effect monitoring data to update the assimilation state sequence and use the effect monitoring data to update the parameters of the precursor pattern recognition model to generate the updated early warning result for the next round of whole-process linkage prevention and control execution.

[0008] Optionally, S1 includes:

[0009] Set preset time slots and collect historical monitoring data for each preset time slot in the target water area to form a training sample set;

[0010] Each training sample includes the meteorological conditions of the target water area, the amount of algae accumulation in the surface water, the redox potential of the mud-water interface, the ferrous ion concentration in the bottom and overlying water, and the sulfide concentration in the bottom and overlying water within the same preset time slice, and a lake flooding label is configured for the preset time slice.

[0011] The lake flooding label includes a lake flooding risk level label and an expected outbreak time window label. The lake flooding risk level label and the expected outbreak time window label are determined based on whether algal-originating lake flooding occurs within a preset prediction time domain after the preset time slice and the timing of algal-originating lake flooding.

[0012] The occurrence of algal blooms is determined based on at least one of the following preset black and odorous criteria: either the on-site water black and odorous phenomenon record or the concentration of ferrous ions in the bottom and overlying water and the concentration of sulfides in the bottom and overlying water meet the record.

[0013] Optionally, S2 includes:

[0014] The training objective function of the precursor pattern recognition model is constructed using the training sample set, and the trained precursor pattern recognition model is obtained by iteratively updating the model parameters.

[0015] In the precursor pattern recognition model, the physical constraint neurodynamic model establishes at least the state evolution relationship of the dissolved oxygen state, the redox state of the mud-water interface, the ferrous ion state of the bottom layer and the overlying water, the sulfide state of the bottom layer and the overlying water, and the state of the surface algae accumulation, which is used to generate state prediction results corresponding to the training samples.

[0016] The multi-task output network outputs the lake flooding risk level prediction result and the expected outbreak time window prediction result corresponding to the training sample based on the state prediction result.

[0017] The training objective function includes a data fitting term and a physical constraint term. The data fitting term is used to characterize the error between the predicted lake flood risk level and the lake flood risk level label in the training sample set, as well as the error between the predicted outbreak time window and the predicted outbreak time window label in the training sample set. The physical constraint term is used to characterize the degree of deviation of the state prediction result from the state evolution relationship of the neural ordinary differential equation.

[0018] The model parameters of the physically constrained neurodynamic model and the multi-task output network are jointly trained by minimizing the training objective function, and the trained aura pattern recognition model is output.

[0019] Optionally, S3 includes:

[0020] Monitoring points are set up in areas prone to lacustrine flooding in the target waters, and observation data of precursor indicators corresponding to the same timestamp are acquired at preset time intervals during the early warning and prevention operation period, and the data are collected in chronological order to form an observation data sequence.

[0021] The meteorological conditions of the target water area include the temperature, wind direction, and wind speed near the water surface;

[0022] The amount of algae accumulation in the surface water was obtained by collecting cyanobacteria from the surface water using a plankton net, filtering out the water, and then calculating the amount of algae accumulation per unit area. The pore size of the plankton net was less than 64 μm.

[0023] The oxidation-reduction potential of the mud-water interface was obtained by measuring within a 2 cm range above and below the mud-water interface using a high-precision, high-sensitivity microelectrode system. The vertical resolution of the mud-water interface by the high-precision, high-sensitivity microelectrode system is greater than 200 μm.

[0024] The concentrations of ferrous ions and sulfides in the overlying water were obtained by collecting samples of the overlying water within 20 cm above the mud-water interface and performing chemical analysis. The concentration of ferrous ions in the overlying water was analyzed using the o-phenanthroline spectrophotometric method or the phenanthroline spectrophotometric method, and the concentration of sulfides in the overlying water was analyzed using the methylene blue method.

[0025] Optionally, S4 includes:

[0026] A unified time axis is set for the observation data sequence, and the meteorological conditions of the target water area, the amount of algae in the surface water, the redox potential of the mud-water interface, the ferrous ion concentration in the bottom and overlying water, and the sulfide concentration in the bottom and overlying water are mapped to the unified time axis to complete the time alignment.

[0027] When the sampling times of various precursor indicators are inconsistent, interpolation or resampling is used to convert each precursor indicator into data consistent with the unified time axis.

[0028] When multiple monitoring points are set up in the target water area, each precursor indicator is limited to the data corresponding to the same monitoring point, or the data corresponding to multiple monitoring points are aggregated according to a preset spatial range to complete spatial matching.

[0029] After completing time alignment and spatial matching, a feature sequence is constructed for each precursor indicator. The feature sequence includes numerical features and rate of change features, wherein the rate of change feature is obtained by dividing the numerical difference between adjacent time points by the time interval between adjacent time points.

[0030] Optionally, S5 includes:

[0031] Based on the feature sequence, an observation vector is constructed at each time point. The observation vector includes numerical features of the redox potential of the mud-water interface, the ferrous ion concentration in the bottom and overlying water, the sulfide concentration in the bottom and overlying water, and the amount of algae accumulation in the surface water.

[0032] For each time point, initialize a set of state vectors containing multiple set members. Each set member contains dissolved oxygen state, mud-water interface redox state, bottom and overlying water ferrous ion state, bottom and overlying water sulfide state, and surface algae accumulation state. The mud-water interface redox state is characterized by the mud-water interface redox potential, the bottom and overlying water ferrous ion state is characterized by the bottom and overlying water ferrous ion concentration, the bottom and overlying water sulfide state is characterized by the bottom and overlying water sulfide concentration, and the surface algae accumulation state is characterized by the amount of algae in the surface water.

[0033] For adjacent time points, the state transition relationship is used to predict each set member to obtain a predicted state vector set, and the predicted state vector set is mapped to a predicted observation vector set through the observation operator;

[0034] The Kalman gain is calculated based on the difference between the observed vector and the predicted observed vector set, and the predicted state vector set is corrected using the Kalman gain to obtain the assimilated state vector set.

[0035] The assimilation state sequence is determined by the set mean of the assimilation state vector set, and the uncertainty of the assimilation state sequence is determined by the set covariance of the assimilation state vector set. The assimilation state sequence and its uncertainty are then output.

[0036] Optionally, S6 includes:

[0037] The current state of the assimilation state sequence is taken as the initial state, and the meteorological conditions of the target water area and the amount of algae accumulation in the surface water corresponding to the current time in the feature sequence are taken as external driving quantities and input into the physical constraint neural dynamics model in the trained precursor pattern recognition model, so as to obtain the predicted state sequence by iteratively solving in the preset prediction time domain according to the preset prediction step size.

[0038] The predicted state sequence includes the redox potential of the mud-water interface, the ferrous ion concentration in the overlying water, and the sulfide concentration in the overlying water at each predicted time within the preset prediction time domain. The predicted state sequence also provides the prediction uncertainty corresponding to the assimilation state sequence and its uncertainty.

[0039] Optionally, the S7 includes:

[0040] The predicted state sequence is input into the multi-task output network of the trained precursor pattern recognition model to obtain the lake flooding risk level prediction result and the expected outbreak time window prediction result corresponding to the predicted state sequence.

[0041] The prediction results of the lake flood risk level include the probability of occurrence corresponding to each risk level, and the probability of occurrence is mapped to the lake flood risk level according to the preset risk level classification rules.

[0042] The predicted outbreak time window is determined by the predicted remaining outbreak time output by the multi-task output network, or by the predicted moment when the lake flood risk level first reaches the medium risk level in the predicted state sequence.

[0043] Optionally, S8 includes:

[0044] The prevention and control level is determined based on the early warning results, and prevention and control instructions are generated and issued in accordance with the prevention and control level to implement the whole-process linkage prevention and control.

[0045] When the lake flood risk level characterized by the early warning result is low risk, the air-floating physical enclosure is kept in a non-outflow state, and the intermittent solar aeration device is operated.

[0046] When the lake flooding risk level indicated by the warning results reaches medium risk or above, cyanobacteria interception, capture, and emergency removal in the lakeshore zone will be implemented. This includes: controlling the air-floating physical enclosure to be in an outflow state within 1000 m of the shore for interception and capture; setting up cyanobacteria capture trenches outside the air-floating physical enclosures; when the algal sludge thickness in the capture trenches reaches more than 50% of the trench's thickness, using a sludge removal device to remove the algal sludge; and removing accumulated cyanobacteria in the surface water to control the algal accumulation in the surface water to 0.5 kg / m³. 2 the following;

[0047] When the lake flooding risk level indicated by the warning results reaches medium risk or above, aeration and oxygenation regulation is implemented. This regulation includes: starting high-intensity aeration on top of the operation of the intermittent solar-powered aeration device, ensuring an aeration flow rate greater than 0.18 m³ / s. 3 air / (min m 3 (water), and the emergency elimination standard is defined as the reduction of both the ferrous ion concentration and the sulfide concentration in the bottom and overlying water to below the preset low-risk threshold.

[0048] When the content of volatile acid sulfides in the sediment is found to be higher than 100 mg / kg and the content of ferrous iron in the sediment is higher than 5000 mg / kg, the polluted sediment should be removed. The dredging depth should be controlled to be greater than both the depth of volatile acid sulfides and the depth of ferrous iron pollution. After the removal of the polluted sediment is completed, a multi-series aquatic plant system should be constructed.

[0049] The execution process of the whole-process linkage prevention and control is recorded as effect monitoring data. The effect monitoring data includes meteorological conditions of the target water area, algae accumulation in the surface water, redox potential of the mud-water interface, ferrous ion concentration in the bottom and overlying water, sulfide concentration in the bottom and overlying water, as well as the enclosure status, floating mud removal amount and aeration flow rate corresponding to the whole-process linkage prevention and control.

[0050] Optionally, S9 includes:

[0051] The effect monitoring data is time-aligned and spatially matched according to a preset time interval to generate an updated feature sequence. The updated feature sequence includes the numerical characteristics and rate of change characteristics of the meteorological conditions of the target water area, the amount of algae accumulation in the surface water, the redox potential of the mud-water interface, the concentration of ferrous ions in the bottom and overlying water, and the concentration of sulfides in the bottom and overlying water.

[0052] An updated observation vector is constructed based on the updated feature sequence, and an ensemble Kalman filter is used to correct the assimilation state sequence to obtain the updated assimilation state sequence and its uncertainty.

[0053] Based on the effect monitoring data, the updated lake flooding labels corresponding to the preset time slices are determined, and the updated feature sequences and the updated lake flooding labels are aggregated into updated training samples.

[0054] Based on the updated training samples, incremental training is performed on the trained precursor pattern recognition model to minimize both the data fitting term and the physical constraint term in the training objective function, thus obtaining the updated precursor pattern recognition model.

[0055] The updated early warning results are generated using the updated precursor pattern recognition model, and the updated early warning results are used in the next round of full-process joint prevention and control execution.

[0056] The beneficial effects of this invention are:

[0057] 1. By performing time alignment and spatial matching on online monitoring data, and using ensemble Kalman filtering to iteratively assimilate and correct key state vectors, assimilated state sequences and their uncertainties can be obtained even in the presence of noise, missing measurements, and multi-source asynchronous sampling. This improves the accuracy of state estimation and provides a reliable quantitative basis for early warning.

[0058] 2. By introducing a precursor pattern recognition model that includes a physical constraint neurodynamic model and a multi-task output network, the physical constraint term is used to suppress unreasonable evolution and characterize the temporal changes of lake flood-related states, thereby achieving joint prediction of lake flood risk level and expected outbreak time window, and improving the foresight and accuracy of early warning.

[0059] 3. By converting early warning results into prevention and control instructions and recording the effect monitoring data of the entire process of coordinated prevention and control, the assimilation state sequence is further updated and the model is incrementally trained, forming a closed-loop adaptive optimization mechanism for early warning and prevention and control. This reduces the drift risk of the model caused by changes in the environment and enhances the prevention and control effect of long-term stable operation. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1 This is a flowchart of an early warning and control method for algal blooms based on precursor pattern recognition proposed in this invention;

[0062] Figure 2 Timing of algal blooms at different wind speeds;

[0063] Figure 3 The timing of lacustrine flooding varies depending on the amount of algae present.

[0064] Figure 4 It refers to the changes in the concentrations of ferrous iron and sulfides after the redox potential decreases;

[0065] Figure 5 It is to maintain the effect of clearing accumulated algae on the pre-control of algal blooms;

[0066] Figure 6 It demonstrates the effectiveness of high-intensity aeration in the emergency elimination of algal blooms in lakes.

[0067] Figure 7 This demonstrates the effectiveness of environmentally friendly dredging in controlling algal blooms in lakes. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figure 1 A method for early warning and prevention of algal blooms in lacustrine environments based on precursor pattern recognition, comprising:

[0070] S1. Collect historical monitoring data of the target water area and form a training sample set according to a preset time slice. Each training sample includes observation data of precursor indicators and corresponding lake flooding labels. S2. Train a precursor pattern recognition model using the training sample set, including a physically constrained neural dynamics model and a multi-task output network. S3. Conduct online monitoring of the target water area to obtain observation data sequences. S4. Perform time alignment and spatial matching on the observation data sequences, and calculate feature sequences, including the numerical characteristics and rate of change characteristics of each precursor indicator. S5. Establish a state vector based on the feature sequences, and use ensemble Kalman filtering to iteratively calibrate the state vector. S6. Obtain the assimilation state sequence and its uncertainty; S7. Use the assimilation state sequence and its uncertainty to drive the physical constraint neurodynamic model in the precursor pattern recognition model to generate the predicted state sequence; S8. Input the predicted state sequence into the precursor pattern recognition model for multi-task reasoning to generate the early warning result; S9. Convert the early warning result into prevention and control instructions and execute the whole-process linkage prevention and control to obtain effect monitoring data; S0. Use the effect monitoring data to update the assimilation state sequence and use the effect monitoring data to update the parameters of the precursor pattern recognition model to generate the updated early warning result for the next round of whole-process linkage prevention and control execution.

[0071] In this specific embodiment, S1 includes:

[0072] A historical monitoring database for the target water area was established, and fixed monitoring points in areas prone to lacustrine flooding were selected as target monitoring points. Monitoring records from these target monitoring points over a continuous 365 days were used as the source of historical monitoring data. All monitoring records carried a unified Beijing time timestamp and included meteorological conditions of the target water area, algal accumulation in the surface water, redox potential at the mud-water interface, ferrous ion concentration in the bottom and overlying water, and sulfide concentration in the bottom and overlying water. The meteorological conditions of the target water area were determined by temperature, wind direction, and wind speed output from automatic weather stations near the water surface. The algal accumulation in the surface water was measured on-site per unit area. Record the redox potential at the mud-water interface in mV, and record the ferrous ion concentration and sulfide concentration in the overlying water in mV, respectively. The sampling depth is limited to 0 to 20 cm above the mud-water interface;

[0073] Historical monitoring data is divided into preset time slices, with preset time slice lengths. Take 1 hour, starting from the time slice start time. and the end time Definition of the first The system divides data into time slices and aggregates multiple records of the same indicator within each time slice to obtain the precursor indicator observation data for that time slice. The aggregation rules are as follows: temperature is calculated using the arithmetic mean, wind speed is calculated using the arithmetic mean, and wind direction is calculated using... to After converting to unit vectors, the average was calculated, and then the angle was inversely calculated. The arithmetic mean of algal bloom in the surface water, the arithmetic mean of the redox potential at the mud-water interface, the arithmetic mean of the ferrous ion concentration in the bottom and overlying water, and the arithmetic mean of the sulfide concentration in the bottom and overlying water were all calculated. Quality control was performed on the missing data. The quality control rule was that if the proportion of missing data for any indicator exceeded a certain threshold in any time slice... Then discard the samples corresponding to that time slice, with the missing test rate not exceeding [percentage missing]. Then, the index is linearly interpolated between adjacent time slices according to the timestamp to ensure that each training sample has temperature, wind direction, wind speed, surface water algae accumulation, mud-water interface redox potential, bottom and overlying water ferrous ion concentration and bottom and overlying water sulfide concentration in the same time slice.

[0074] After constructing the precursor indicator observation data for the training samples, lake flooding labels were configured for each time slice, and these labels were defined as lake flooding risk level labels and expected outbreak time window labels. The criteria for determining the occurrence of black and odorous lake flooding were defined at each historical timestamp. The above is defined as a formula:

[0075] ;

[0076] in Indicates time Indicator of whether algal blooms have occurred. This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true, and otherwise takes the value _____. Indicates time The corresponding on-site records of black and odorous water bodies are binary-coded based on whether or not the water is black and odorous. Indicates time The concentration of ferrous ions in the underlying water layer. Indicates time Sulfide concentration in the underlying water. The black and odor threshold representing the concentration of ferrous ions is taken as follows: The black and odorous threshold representing the sulfide concentration is taken as follows: ,symbol Represents a logical OR operation;

[0077] Preset prediction time domain Take 72 hours, and the interval after the end of each time slice. Internal hourly scanning of historical timestamps to find matching The earliest moment is recorded as the time of lake flooding corresponding to that time slice, and the expected outbreak time window label is determined accordingly. The expected outbreak time window label is divided into four categories, corresponding to "no outbreak", "0 to 24 h", "24 to 48 h", and "48 to 72 h", respectively. "No outbreak" means that the entire time slice is not affected. There are no conditions that satisfy the condition within the range. At any given moment, the three types of time windows indicate that the difference between the time of lake flooding and the time slice ends falls within the corresponding interval;

[0078] The risk level labels for lacustrine flooding are divided into three categories: low risk, medium risk, and high risk, and are determined by the same scan results. Low risk indicates that the expected outbreak time window is labeled "no outbreak", medium risk indicates that the expected outbreak time window is labeled "48 to 72 hours", and high risk indicates that the expected outbreak time window is labeled "0 to 24 hours" or "24 to 48 hours".

[0079] The aggregated precursor indicator observation data within each time slice are mapped one-to-one with the corresponding lake flood risk level label and expected outbreak time window label, and written into the training sample set to obtain the training dataset.

[0080] In this specific embodiment, S2 includes:

[0081] A precursor pattern recognition model is constructed and trained based on a training sample set. The precursor pattern recognition model consists of a physically constrained neurodynamic model and a multi-task output network.

[0082] The physically constrained neurodynamic model uses neural ordinary differential equations to model the evolution of the state vector, which is denoted as:

[0083] ;

[0084] in Indicates time, It indicates the dissolved oxygen state of the overlying water and is characterized by dissolved oxygen concentration, using historical dissolved oxygen sensor records as monitoring data. It represents the redox state of the mud-water interface and is characterized by the redox potential of the mud-water interface. It indicates the state of ferrous ions in the overlying water of the bottom layer and is characterized by the concentration of ferrous ions in the overlying water of the bottom layer. This indicates the state of sulfides in the overlying water and is characterized by the concentration of sulfides in the overlying water. It indicates the state of surface algae accumulation and is characterized by the amount of algae accumulation in the surface water.

[0085] External driving quantity is denoted as:

[0086] ;

[0087] in Indicates the temperature near the water surface. Indicates wind speed. Indicates the wind direction azimuth angle in radians;

[0088] Let the function on the right-hand side of the constant differential equation be denoted as... ,in The parameters to be trained in the physically constrained neurodynamic model are represented by the weights and biases of a multilayer perceptron. The multilayer perceptron is configured with a 9-dimensional input dimension and consists of... and The sequence is obtained by splicing together three hidden layers with 64 neurons in each layer, and the activation function is... The output dimension is 5-dimensional and is used to output the state derivative;

[0089] For each training sample The observations corresponding to the end of its time slice are constructed as the initial state. and in the preset prediction time domain Inner fixed step size The predicted state sequence is obtained by iterative solution using a fourth-order Runge-Kutta numerical integrator. ,in Indicates training samples In the The prediction state vector corresponding to each prediction step;

[0090] The multi-task output network takes the predicted state sequence as input and outputs the predicted lake flooding risk level and the predicted outbreak time window. The multi-task output network consists of a single-layer gated recurrent unit (GRU) and two task output heads. The GRU has a 5-dimensional input dimension and a time step of [missing information]. The hidden state dimension of GRU is set to 32, and the hidden state of the last time step is used as the shared representation vector. The output head of the lake flood risk level is a two-layer fully connected network with dimensions as follows: In the final layer, Softmax is used to obtain the probability vectors for the three risk levels. ,in This index represents low-risk, medium-risk, and high-risk categories. The output header, representing the expected outbreak time window, is a two-layer fully connected network with dimensions as follows: In the final layer, Softmax is used to obtain the probability vectors for the four types of time windows. ,in Indicates no outbreak, 0 to to Up to 72 h category index;

[0091] Before training, all inputs are standardized. The mean and standard deviation used for standardization are obtained by statistical analysis of the training sample set in one dimension and are kept consistent during the training and inference phases.

[0092] Construct the training objective function and apply it to... With multi-task output network parameters For joint training, the training objective function is written as:

[0093] ;

[0094] in Represents the training objective function. This represents the total number of training samples. Indicates training samples The one-hot encoded component of the lake flood risk level label, Indicates training samples The one-hot encoded component of the expected outbreak time window label, Represents the natural logarithm function. Represents the weight of the physical constraint term and takes Indicates training samples In the The external driving force corresponding to each prediction step The L2 norm is used to characterize the predicted state sequence pairs derived from the functions of the constant differential equation. The degree of deviation from the defined state evolution relationship;

[0095] Using the Adam optimizer and For iterative updates, Adam's learning rate is set to... The first-order moment coefficients are set to 0.9, the second-order moment coefficients are set to 0.999, and the numerical stability term is set to... The batch size was set to 128, the number of training rounds was set to 300, and the gradient norm was pruned with a threshold of 1.0 to ensure numerical stability, resulting in the trained precursor pattern recognition model.

[0096] In this specific embodiment, S3 includes:

[0097] Three fixed monitoring points were set up in the lake flood-prone area of ​​the target waters and labeled as follows: and ,in and All are located within 1000 m of the shoreline and each monitoring point is equipped with the same set of online monitoring and sampling units. The early warning and prevention operation period is set from 00:00 on June 1st to 24:00 on October 31st every year.

[0098] Set the preset time interval as ,in This indicates the data acquisition time interval for online monitoring, with the start time of the runtime segment as the reference. , will the The unified timestamp for this monitoring is: ,in Indicates the first The Beijing time timestamp corresponding to this monitoring. This represents the monitoring sequence number and is a non-negative integer;

[0099] At each timestamp Automatic weather stations near the water surface simultaneously collect meteorological conditions of the target water area and generate temperature data. Wind speed With wind direction ,in This indicates the temperature near the water surface, expressed in degrees Celsius. This indicates the wind speed near the water surface, expressed in meters per second. It indicates the azimuth angle of the wind direction near the water surface, and the unit is degrees, with due north as zero degrees and clockwise.

[0100] At each timestamp The surface water area at each monitoring point is [area value missing]. A floating collection frame was used to collect surface-accumulated cyanobacteria within the collection frame area using a planktonic net with an aperture of less than 64 micrometers. The collected material was filtered through a sieve to remove free water, and after a constant drip filtration time of ten minutes, it was weighed using an electronic balance with a range of 2000 g and a graduation of 0.1 g to obtain the wet weight. The amount of algae accumulated in the surface water was then calculated by dividing the wet weight by the area of ​​the collection frame. ,in This represents the amount of algae accumulated per unit area, with units of kg / m². 2 ;

[0101] At each timestamp The redox potential of the mud-water interface is obtained by vertically profiling within a 2 cm range above and below the mud-water interface using a high-precision, high-sensitivity microelectrode system. The potential value at the mud-water interface location in the profile is then taken as the redox potential of the mud-water interface. The vertical resolution of the microelectrode system is greater than 200 μm and The unit is millivolt;

[0102] At each timestamp A bottom overlying water sample was collected 10 cm above the mud-water interface using a fixed-depth water sampler, with the sampling depth limited to within 20 cm above the mud-water interface. The water samples were aliquoted into nitrogen-filled, sealed brown sample bottles and stored at 4°C in the dark. Chemical analysis was then performed within two hours, including the determination of the ferrous ion concentration in the bottom overlying water. The concentration of sulfide in the overlying water was determined by phenanthroxazine spectrophotometry, with units of mg / L. The methylene blue method was used for determination, and the unit is mg / L;

[0103] The same monitoring point at the same timestamp The obtained observation data of each precursor index are used to form an observation vector:

[0104] ;

[0105] in Represents timestamp The corresponding precursor indicator observation data vectors are generated, and the observation vectors of each monitoring point are collected in chronological order to form an observation data sequence, which is then uploaded to the early warning and prevention server.

[0106] In this specific embodiment, S4 includes:

[0107] A unified time axis is established for the observed data sequence, and feature sequences are generated after time alignment and spatial matching are performed.

[0108] The unified timeline is denoted as ,in Indicates the first A unified timestamp and satisfying , Indicates the start time of the early warning and prevention operation period. Indicates the preset time interval and takes Indicates the total number of unified timestamps within the runtime period;

[0109] Spatial matching uses a fixed strategy to limit the monitoring points to And only use monitoring points The corresponding data, used as online observation data for the target water area, comes from monitoring points. and The data is not included in the subsequent processing in step S4;

[0110] Time alignment for monitoring points At any sampling time The obtained temperature Wind speed ,wind direction Algae accumulation in surface water Redox potential at the mud-water interface ferrous ion concentration in the overlying water at the bottom layer and the concentration of sulfides in the overlying water Perform a unified mapping, setting the allowed time deviation threshold to... When a certain indicator is at a timestamp Sampling time nearby satisfy At that time, the sampled value is directly used as the indicator. The alignment value at the location is denoted as and When there is no satisfying When sampling values, the index is in The index is obtained by performing linear interpolation on the two most recent valid records before and after. The alignment value at the specified location is used, and the time interval between the two valid records must not exceed 6 hours. If the time interval exceeds 6 hours, the indicator is considered to be in a state of readiness. Detect missing values ​​and discard the feature vector corresponding to the timestamp to avoid introducing long-term missing values ​​into subsequent inference;

[0111] After aligning all indicators, the wind direction will be... Convert to wind direction characteristics and ,in The wind direction azimuth is expressed in radians and proportionally converted from the wind direction in degrees to eliminate the difference between 0 and 1. Discontinuity at the point;

[0112] Then for each timestamp Constructing feature vectors ,in It includes numerical features and rate of change features. The numerical features are composed of... , as well as The composition and rate of change characteristics are calculated separately for each of the above numerical characteristics using a unified formula:

[0113] ;

[0114] in Representation of features timestamp The rate of change characteristics at that location Representation of features timestamp Numerical characteristics at that location Representation of features Previous timestamp Numerical characteristics at that location Represents the time interval between adjacent unified timestamps and takes Represents a unified timestamp sequence number and satisfies ,feature Limited to or One of them, and The rate of change feature at any given time is set to zero to ensure consistency of feature dimensions;

[0115] All timestamps The feature sequence is obtained by arranging them in chronological order. .

[0116] In this specific embodiment, S5 includes:

[0117] Based on feature sequences At each timestamp Construct observation vectors and perform ensemble Kalman filtering assimilation to obtain the assimilated state sequence and its uncertainties;

[0118] Among them, unified timestamp With preset time interval and The observation vector is denoted as ,in Represents timestamp The observed vector at that location, This indicates the numerical characteristics of the redox potential at the mud-water interface. This indicates the numerical characteristics of the ferrous ion concentration in the overlying water at the bottom layer. This indicates the numerical characteristics of sulfide concentration in the overlying water. This indicates the numerical characteristics of algal accumulation in surface water.

[0119] The state vector used for assimilation is denoted as ,in Represents timestamp The state vector at that point, It represents the dissolved oxygen state of the overlying water at the bottom layer and is characterized by dissolved oxygen concentration. It also participates in assimilation as a latent state that cannot be directly observed.

[0120] The initial collection size is set to ,in Indicates the number of members in the set, and Constructing the initial state vector set at each time step ,in Indicates the first The initial state vector of each set member. Represents the index of a set member, ranging from 1 to... The initial state and Taken from the observation vector The initial state Take a fixed value The five state components are then subjected to zero-mean Gaussian perturbations to form the ensemble discreteness, and the standard deviations of the perturbations are set to 1. and ;

[0121] For adjacent time points A prediction step is performed for each set member, using the right-hand side function of the neural ordinary differential equation of a physically constrained neurodynamic model. As a state transition relation and using fixed model parameters When predicting, the first The previous assimilation state of each set member. With external driving force Common input to ,in Indicates the first The collection members at the timestamp The assimilation state vector at that point, Represents timestamp External driving force at the location, Indicates the numerical characteristics of temperature. This indicates the numerical characteristics of wind speed. Indicates wind direction azimuth in radians;

[0122] To ensure consistency with step S2, during the call Former and The training set's dimension-wise mean vector saved in step S2 With the one-dimensional standard deviation vector Standardize and The normalized state derivative of the output is multiplied dimension by dimension. After restoring to the state derivative of the physical quantity, a fourth-order Runge-Kutta integrator is used with a step size of... The predicted state is obtained by completing one step of integration within the inner circle. ,in Indicates the first The collection members at the timestamp The predicted state vector at the location;

[0123] Process noise is superimposed on the predicted state to characterize the model error; the process noise covariance matrix is ​​denoted as... And take a diagonal matrix, with the five diagonal elements corresponding to... and The variance is set as and ;

[0124] Then, the predicted state vector set is mapped to the predicted observation vector set using the observation operator, denoted as [the operator is missing here]. It is a constant matrix, and its function is to extract sequentially from the state vector. and Four components with the observation vector Maintain consistency;

[0125] The observation noise covariance matrix is ​​denoted as And take a diagonal matrix, with the four diagonal elements corresponding to... and The variance is set as and ;

[0126] In the update step, the sample covariance is calculated based on the predicted state vector set and the predicted observation vector set, and the Kalman gain is obtained accordingly. And an assimilation correction is performed on each set member using a perturbation-based ensemble update method, the calculation relationship of which is as follows:

[0127] ;

[0128] in Represents timestamp The Kalman gain matrix at that location, The sample cross-covariance matrix represents the set of predicted state vectors and the set of predicted observation vectors. The sample covariance matrix represents the set of predicted observation vectors. Represents the observation noise covariance matrix. This represents finding the inverse of a matrix. Indicates the first The collection members at the timestamp The assimilation state vector at that point, Indicates the first The collection members at the timestamp The predicted state vector at that location, Represents timestamp The observed vector at that location, Indicates the first The observed perturbation vectors corresponding to each set member are zero-mean Gaussian random vectors with covariance of... , Represents the observation operator matrix;

[0129] After completing the iterative assimilation of all timestamps, for each timestamp Assimilate the set of state vectors The set mean is used as the state estimate of the assimilated state sequence at that moment, and the sample covariance of the assimilated state vector set is used as the uncertainty at that moment. The assimilated state sequence and its uncertainty are output.

[0130] In this specific embodiment, S6 includes:

[0131] Assimilate the state sequence at the current time The set of assimilated state vectors at a given location is used as the initial condition for prediction, and its uncertainty is combined to generate the prediction state sequence and prediction uncertainty.

[0132] The current time This represents the current timestamp on the unified timeline and is consistent with step S4. Consistent, the number of members in the set is taken from the value set in step S5. The initial state vector set is denoted as ,in Indicates the first The members of the set at time... The assimilated state vector at that point and This indicates the dissolved oxygen state of the overlying water. Indicates the redox potential state at the mud-water interface. This indicates the concentration of ferrous ions in the overlying water at the bottom layer. This indicates the concentration of sulfides in the overlying water. This indicates the state of surface algal aggregation. Represents the index of a set member, ranging from 1 to... ;

[0133] Set the preset prediction time domain to be consistent with step S2. And set the preset prediction step size to be consistent with step S4. The number of prediction steps is denoted as ;

[0134] Take the external driving quantity as the feature sequence at the current time. The corresponding meteorological conditions and surface algal blooms in the target water area remain constant throughout the entire prediction time domain. The external driving force is denoted as:

[0135] ;

[0136] in This indicates the numerical characteristics of temperature near the water surface. This indicates the numerical characteristics of wind speed. The azimuth angle is expressed in radians, and the external driving force for algae accumulation in surface water is taken as... And it is given by the numerical features after alignment in step S4;

[0137] To allow surface algae accumulation to participate in prediction as an external driving force, the surface algae accumulation state component in the set member state vector is clamped to [the specified value] for each prediction step. And use them as input components of the physical constraint neurodynamic model, so that the dissolved oxygen state, the redox state of the mud-water interface, the ferrous ion state of the bottom and overlying water and the sulfide state of the bottom and overlying water can be forward evolved under the background of algal accumulation in the same surface water body.

[0138] The physical constraint neurodynamic model uses the right-hand side function of the neuronormal differential equation. And fix its model parameters ,in It consists of a multilayer sensory mechanism with three hidden layers, each layer having 64 neurons and an activation function of . The input is the concatenation of the state vector and the external driving quantity; the output is the state derivative vector.

[0139] For each set member Call within each prediction step A fourth-order Runge-Kutta integrator is used with a step size of 1. Numerical integration is performed to obtain the predicted state vector. ,in Indicated by time Forward prediction from the starting point The first step after Each set member predicts the state vector and Take 0 to During the integration process, the standardization and destandardization of the input and output use the one-dimensional mean vector and one-dimensional standard deviation vector of the training set saved in step S2 to ensure consistency with the training distribution. After each integration step, the same process noise covariance matrix as in step S5 is superimposed. To propagate assimilation uncertainty to prediction uncertainty;

[0140] After completing the forward propagation of all set members, at each prediction step The set mean and set covariance of the predicted states are calculated to form the predicted state sequence and prediction uncertainty. The calculation relationship is as follows:

[0141] ,

[0142] ;

[0143] in Indicates the prediction step The mean vector of the predicted state set at the location, Indicates the prediction step The covariance matrix of the predicted state set at the given location is used as the prediction uncertainty. Indicates transpose;

[0144] Will The predicted state sequence is output, from which the predicted redox potential sequence of the mud-water interface is extracted. Predicted sequence of ferrous ion concentration in the overlying water. Predicted sequence of sulfide concentration in overlying water At the same time by Extraction of corresponding diagonal elements , and The prediction variance is used as the prediction uncertainty of the three indicators and is output together with the predicted state sequence for multi-task inference.

[0145] In this specific embodiment, S7 includes:

[0146] Predict the state sequence The multi-task output network of the trained precursor pattern recognition model is input to generate early warning results, whereby... Indicates the current time Forward prediction from the starting point The mean vector of the predicted state set for each step:

[0147] ;

[0148] This indicates the predicted dissolved oxygen value in the overlying water. This represents the predicted value of the redox potential at the mud-water interface. This indicates the predicted concentration of ferrous ions in the overlying water. This indicates the predicted concentration of sulfides in the overlying water. This represents the predicted value of algal bloom in surface water. This indicates the number of prediction steps, and is set to 72.

[0149] The structure of the multi-task output network is the same as in step S2, and its model parameters are fixed. Its input is a length of The state sequence and the time-by-time mean vector of the training set saved according to step S2 With the one-dimensional standard deviation vector After standardization, the data is fed into a single GRU layer to obtain a shared representation vector. The GRU hidden state dimension is 32, and the hidden state at the last time step is taken as the hidden state. The risk level output head outputs the lake flood risk level prediction results, and the time window output head outputs the expected outbreak time window prediction results.

[0150] The prediction results of the lake flood risk level are output in the form of three types of occurrence probabilities and denoted as follows: and ,in Indicates the probability of a low-risk event. Indicates the probability of medium risk occurring. This indicates the probability of high-risk occurrence, and maps the category with the highest probability of occurrence to the current moment according to a preset risk level classification rule. The risk level of lake flooding;

[0151] The predicted outbreak time window results are output in the form of four types of occurrence probabilities and denoted as follows: and ,in This indicates the probability of a time window during which an outbreak does not occur. This represents the probability within a time window of 0 to 24 hours. This represents the probability within a 24- to 48-hour time window. This represents the probability of a time window from 48 to 72 hours, and maps the time window category with the highest probability of occurrence to the current time. The expected outbreak window;

[0152] Meanwhile, in order to convert the time window results into outbreak remainder prediction values ​​that can be directly used for linkage control, outbreak remainder prediction values ​​are constructed according to the representative remainder times of the four types of time windows. And used to generate the start and end times of the expected outbreak time window, wherein the remaining time is respectively set as , and ,in This represents the remaining time for non-outbreak categories outside the preset prediction time domain. The midpoint of the 0-24 hour time window represents the remaining time. The midpoint of the 24- to 48-hour time window represents the remaining time. The midpoint of the 48- to 72-hour time window represents the remaining time. in Indicates the current time The predicted time remaining of the outbreak, Indicates the current time In the The predicted probability of an outbreak within a given time window. Indicates the first The class represents the remaining time corresponding to the expected outbreak time window. This indicates the time window category index, ranging from 1 to 4;

[0153] Lake flood risk levels and the probability vectors corresponding to each risk level. Expected outbreak time window, and probability vector of occurrence for each time window. and by Targeted outbreak time They are collectively packaged into early warning results and output for determining the prevention and control level and generating prevention and control instructions.

[0154] In this specific embodiment, S8 includes:

[0155] The early warning and prevention server at the current moment Receive early warning results and analyze them to obtain the lake flood risk level. With the corresponding occurrence probability vector ,in Represents the current timestamp on a unified timeline. This indicates the current level of lake flooding risk and specifies it as low, medium, or high risk. Indicates the probability of a low-risk event. Indicates the probability of medium risk occurring. Indicates the probability of a high risk occurring;

[0156] The server will Mapping to prevention and control levels and generating prevention and control instruction sets, which are then sent to the field programmable logic controller (PLC) to trigger full-process linkage prevention and control. Among these, when... When the risk level is low, the prevention and control level is set to Level 1 and a low-risk operation strategy is implemented. The low-risk operation strategy includes setting the enclosure state of the air-floating physical enclosure to a non-outflow state and keeping the upper edge of the enclosure below the water surface. To avoid impacting navigation and water exchange, the intermittent solar aeration device was activated and operated at a fixed rhythm, set to run for 10 minutes every 60 minutes, with the aeration device outputting an air volume of [missing information]. The air-floating physical enclosure consists of a float, a flexible enclosure skirt, and a lifting actuator, and the enclosure state is controlled by a two-position control quantity output by a PLC. The intermittent solar aeration device consists of a solar power supply module, an energy storage module, and a microporous aeration disc, and the start / stop and air volume are controlled by a PLC through relays and solenoid valves.

[0157] when When the risk level reaches medium or high, the prevention and control level is set to Level II. While maintaining the operation of intermittent solar aeration devices, the interception, capture, and emergency removal of cyanobacteria in the lakeside zone, as well as aeration and oxygenation regulation, are implemented. The interception, capture, and emergency removal of cyanobacteria in the lakeside zone includes setting the air-floating physical enclosure to the effluent state, maintaining the upper edge of the enclosure 0.20 m above the water surface, limiting the enclosure's deployment range to within 1000 m of the shoreline, and ensuring the enclosure is closed to form an interception zone to capture surface-accumulated cyanobacteria. Simultaneously, cyanobacteria capture trenches are set along the enclosure direction outside the air-floating physical enclosure, with an effective depth of 1.0 m. The thickness of the algae sludge in the trench is measured using an ultrasonic level gauge installed on the trench sidewall, and a thickness of 50% of the effective depth of the trench is used as a trigger condition. When the trigger condition is met, the PLC activates a sludge removal device to pump and remove the algae sludge in the trench and transport the removed algae sludge to a temporary storage tank on shore. The pumping flow rate of the sludge removal device is set to... The process continues until the algal sludge thickness is reduced to below 0.30 m, and the cumulative suction volume during the removal process is converted into the amount of floating sludge removed. Record it, among which Indicates the timestamp The corresponding amount of floating sludge removed within the control cycle, in meters. 3 ;

[0158] While completing the removal of algae and sludge from the trenches, an emergency operation to remove surface blue-green algae was initiated, and a mechanical dredging vessel was used to operate along the inner channel of the enclosure to reduce the amount of algae accumulating in the surface water. ,in Indicates the amount of algae accumulation in surface water, with units of . and obtained through online monitoring Not higher than 0.5 kg / m 2 As a condition for achieving the standard for surface cyanobacteria removal;

[0159] Aeration and oxygenation control includes activating a high-intensity aeration system while maintaining the operation of the intermittent solar aeration device. The high-intensity aeration system consists of a shore-based blower, aeration main pipe, and a microporous aeration disc array deployed at the bottom layer. The blower speed is controlled by a frequency converter and PLC to achieve closed-loop regulation of air volume. The aeration flow rate is normalized to the unit water volume to obtain the aeration intensity per unit volume. And maintain it above the threshold, the relationship is defined as:

[0160] ;

[0161] in Represents timestamp Aeration intensity per unit volume and the unit is Represents timestamp The aeration flow rate output by the blower to the aeration main pipe, and the unit is... This indicates the water volume of the enclosure and control area during Level II prevention and control, with units of 1. It is obtained by multiplying the enclosed area of ​​the enclosure by the average water depth;

[0162] The concentration of ferrous ions in the underlying water With the concentration of sulfides in the overlying water Simultaneously, reducing the risk level to below the preset low-risk threshold is considered a condition for emergency elimination compliance. This indicates the concentration of ferrous ions in the overlying water at the bottom layer, with units of 1. This indicates the concentration of sulfides in the overlying water at the bottom layer, and the unit is . The preset low-risk threshold is set to and Furthermore, the PLC determines in real time whether the standards are met based on the online chemical analysis results;

[0163] When sediment monitoring results indicate that the content of volatile acid sulfides in the sediment exceeds 100 mg / kg and the content of ferrous iron in the sediment exceeds 5000 mg / kg, contaminated sediment removal is carried out. The content of volatile acid sulfides in the sediment is determined by collecting columnar sediment samples, stratifying them vertically at 5 cm intervals, and then using an acid stripping-absorption titration method. The content of ferrous iron in the sediment is determined by inductively coupled plasma atomic emission spectrometry (ICP-AES) after acid digestion of the same stratified sample. Contaminated sediment removal utilizes a dredging system with a dredging depth set greater than both the depth of volatile acid sulfide contamination and the depth of ferrous iron contamination, with an additional 0.10 m safety margin to ensure complete removal of the contaminated layer. After the contaminated sediment removal is completed, a multi-series aquatic plant system is constructed in the dredged area. This multi-series aquatic plant system consists of emergent plant strips, floating-leaved plant strips, and submerged plant strips, planted with reeds, water lilies, and Vallisneria natans, respectively, at a planting density of 4 plants per strip. strain With 16 plants And by using fences to spatially isolate each plant zone, stable survival can be ensured;

[0164] In the process of joint prevention and control, To record the periodic generation of effect monitoring data and upload it to the early warning and prevention server, the effect monitoring data is generated at each timestamp. At least include meteorological conditions of the target waters. and Algae accumulation in surface water Redox potential at the mud-water interface The concentration of ferrous ions in the water covering the bottom layer With the concentration of sulfides in the overlying water And the isolation status corresponding to the whole-process joint prevention and control. Floating mud removal volume With aeration flow rate ,in The enclosure state is represented by 0 for non-outflow state and 0 for outflow state. The unit is mV, thus completing the deterministic conversion and execution of early warning results into prevention and control instructions and forming an effect monitoring dataset.

[0165] In this specific embodiment, S9 includes:

[0166] The early warning and prevention server at the end of each control cycle Receive effect monitoring data and use Temporal alignment and spatial matching are performed at the temporal granularity to generate an updated feature sequence for updating, where Indicates the current unified timestamp, This indicates a preset time interval, and spatial matching is limited to using only monitoring points. The corresponding data is aligned using the same unified timeline as in step S4. and temperature Wind speed ,wind direction Algae accumulation in surface water Redox potential at the mud-water interface ferrous ion concentration in the overlying water at the bottom layer With the concentration of sulfides in the overlying water Perform interpolation to complete and remove missing measurements, wind direction Convert to and This, along with other indicators, constitutes numerical features, and the rate of change feature is further calculated to obtain an updated feature vector. ,in Represents timestamp The updated feature vector at the location contains both numerical and rate-of-change features, thus... Composition and updating of feature sequences;

[0167] Based on the updated feature sequence at each timestamp Construct and update the observation vector:

[0168] ;

[0169] in This indicates that the observation vector used for updating the assimilation is in the same order as the observation vector components. Then, using the set of assimilation state vectors from the previous time step as a priori, an ensemble Kalman filter is applied to correct the assimilation state sequence to obtain the updated assimilation state sequence and its uncertainty. The set size is fixed. Process noise covariance matrix Covariance matrix of observation noise Take the value set in step S5 and observe the operator. Take the extraction from the state vector set in step S5. and The constant matrix of the four components is used, and the iterative correction is performed on each time stamp using the same Kalman gain calculation and perturbation observation set update method as in step S5, thereby obtaining the updated assimilation state vector set. The updated assimilation state sequence is formed by using its set mean, and the updated uncertainty is formed by using its set covariance;

[0170] Based on the performance monitoring data, updated lake flooding labels corresponding to preset time slices are determined, where the preset time slice length is 1 hour and the preset prediction time domain is [not specified]. The occurrence of algal blooms was determined using the black and odorous criterion defined in step S1, with a duration of 72 hours. And search for the earliest satisfying condition within the interval after the end of each time slice. The timing is used to determine the expected outbreak time window label, and at the same time, the lake flood risk level label is determined according to the mapping rules in step S1;

[0171] Update the feature vector for each time slice. It is combined with the updated lake flood risk level label and the updated expected outbreak time window label to form an updated training sample and aggregated into an updated training sample set. ,in This represents the dataset used for incremental training.

[0172] Based on the updated training sample set, incremental training is performed on the trained aura pattern recognition model while maintaining the model structure consistent with step S2. The physically constrained neurodynamic model still uses the right-hand side function of the neuronormal differential equation. And the parameter is denoted as The multi-task output network parameters are denoted as Incremental training uses the same training objective function as step S2. And maintain the weights of physical constraint terms. Unchanged, the stated The data fitting terms for the lake flood risk level, the data fitting terms for the predicted outbreak time window, and the physical constraint terms for the state evolution relationship of the neural network's ordinary differential equation are combined. Incremental training uses the Adam optimizer and sets the learning rate to [value missing]. The first-order moment coefficients are set to 0.9, the second-order moment coefficients are set to 0.999, and the numerical stability term is set to... The batch size was set to 64, the number of training epochs was set to 50, and the gradient norm was pruned to a threshold of 1.0 to ensure numerical stability. Parameter updates were performed accordingly. It means that, among them This represents the parameters of the physically constrained neurodynamic model after incremental training. This represents the multi-task output network parameters after incremental training. The operator represents the set of parameters that minimizes the objective function. This indicates that the training sample set is being updated. The training objective function calculated above;

[0173] After obtaining the updated precursor pattern recognition model, the updated assimilation state sequence and its uncertainty are used as the initial conditions for prediction, and the updated model parameters are used. Generate an updated predicted state sequence, and then input the updated predicted state sequence into the updated multi-task output network parameters. The updated warning results are generated and used as input for the next step S8 to generate prevention and control instructions and implement the whole-process linkage prevention and control.

[0174] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0175] This invention transforms lake flood warning from "single-point threshold discrimination" to "state evolution-driven precursor pattern recognition" by synergistically combining time alignment and spatial matching of multi-source online monitoring data, ensemble Kalman filtering state assimilation, physically constrained neurodynamic prediction, and multi-task early warning inference. Specifically, the ensemble Kalman filtering iteratively corrects for biases introduced by observation noise, missing data, and asynchronous sampling, outputting an assimilated state sequence and its uncertainties. This assimilation result is used as the initial state and driving force of the physically constrained neurodynamic model, thereby generating a continuous predicted state sequence within a preset prediction time domain. The multi-task output network simultaneously provides the lake flood risk level and expected outbreak time window based on the same predicted state, enabling the early warning result to be directly converted into prevention and control instructions and trigger coordinated execution, thus achieving the technical effect of "earlier early warning and more accurate decision-making" for algal-originating lake floods.

[0176] Furthermore, this invention improves the algorithm structure to address the characteristics of algal blooms, which are characterized by "strong mechanistic constraints, strong temporal evolution, and strong intervention feedback." First, physical constraints are introduced into the neurodynamic training objective to constrain the evolutionary relationships of key states such as dissolved oxygen, redox state at the mud-water interface, and ferrous ions and sulfides, reducing the risk of non-physical predictions from purely data-driven models under extreme conditions. Second, the uncertainty obtained through assimilation is incorporated into the prediction stage and the prediction uncertainty is output, providing a reliable basis for risk level and time window judgment and enhancing the robustness of the early warning. Third, the monitoring data of the effect of the whole-process linkage prevention and control is used for assimilation updates and incremental training, enabling the model to adapt to the distribution drift caused by seasonal changes and prevention and control interventions, forming a rolling iterative closed-loop optimization mechanism, thereby continuously improving the accuracy of early warning and the effectiveness of prevention and control.

Claims

1. A method for early warning and prevention of algal blooms in lacustrine environments based on precursor pattern recognition, characterized in that, include: S1. Collect historical monitoring data of the target water area and form a training sample set according to the preset time slices. Each training sample includes observation data of precursor indicators and corresponding lake flooding labels. S2. Train the precursor pattern recognition model using the training sample set, including a physically constrained neurodynamic model and a multi-task output network; S3. Conduct online monitoring of the target water area to obtain observation data sequences; S4. Perform time alignment and spatial matching on the observation data sequences, and calculate feature sequences, including the numerical features and rate of change features of each precursor index; S5. Establish a state vector based on the feature sequences, and use ensemble Kalman filtering to iteratively correct the state vector to obtain an assimilated state sequence and its uncertainty; S6. Use the assimilated state sequence and its uncertainty to drive the physically constrained neurodynamic model in the precursor pattern recognition model to generate a predicted state sequence; S7. Input the predicted state sequence into the precursor pattern recognition model for multi-task inference to generate early warning results; S8. Convert the early warning results into prevention and control instructions and execute the entire process of coordinated prevention and control to obtain effect monitoring data; S9. Update the assimilation state sequence using effect monitoring data, and update the parameters of the precursor pattern recognition model using effect monitoring data to generate updated early warning results for the next round of full-process linkage prevention and control execution.

2. The method for early warning and prevention of algal blooms in lacustrine environments based on precursor pattern recognition according to claim 1, characterized in that, S1 includes: Set preset time slots and collect historical monitoring data for each preset time slot in the target water area to form a training sample set; Each training sample includes the meteorological conditions of the target water area, the amount of algae accumulation in the surface water, the redox potential of the mud-water interface, the ferrous ion concentration in the bottom and overlying water, and the sulfide concentration in the bottom and overlying water within the same preset time slice, and a lake flooding label is configured for the preset time slice. The lake flooding label includes a lake flooding risk level label and an expected outbreak time window label. The lake flooding risk level label and the expected outbreak time window label are determined based on whether algal-originating lake flooding occurs within a preset prediction time domain after the preset time slice and the timing of algal-originating lake flooding. The occurrence of algal blooms is determined based on at least one of the following preset black and odorous criteria: either the on-site water black and odorous phenomenon record or the concentration of ferrous ions in the bottom and overlying water and the concentration of sulfides in the bottom and overlying water meet the record.

3. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S2 include: The training objective function of the precursor pattern recognition model is constructed using the training sample set, and the trained precursor pattern recognition model is obtained by iteratively updating the model parameters. In the precursor pattern recognition model, the physical constraint neurodynamic model establishes at least the state evolution relationship of the dissolved oxygen state, the redox state of the mud-water interface, the ferrous ion state of the bottom layer and the overlying water, the sulfide state of the bottom layer and the overlying water, and the state of the surface algae accumulation, which is used to generate state prediction results corresponding to the training samples. The multi-task output network outputs the lake flooding risk level prediction result and the expected outbreak time window prediction result corresponding to the training sample based on the state prediction result. The training objective function includes a data fitting term and a physical constraint term. The data fitting term is used to characterize the error between the predicted lake flood risk level and the lake flood risk level label in the training sample set, as well as the error between the predicted outbreak time window and the predicted outbreak time window label in the training sample set. The physical constraint term is used to characterize the degree of deviation of the state prediction result from the state evolution relationship of the neural ordinary differential equation. The model parameters of the physically constrained neurodynamic model and the multi-task output network are jointly trained by minimizing the training objective function, and the trained aura pattern recognition model is output.

4. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S3 includes: Monitoring points are set up in areas prone to lacustrine flooding in the target waters, and observation data of precursor indicators corresponding to the same timestamp are acquired at preset time intervals during the early warning and prevention operation period, and the data are collected in chronological order to form an observation data sequence. The meteorological conditions of the target water area include the temperature, wind direction, and wind speed near the water surface; The amount of algae accumulation in the surface water was obtained by collecting cyanobacteria from the surface water using a plankton net, filtering out the water, and then calculating the amount of algae accumulation per unit area. The pore size of the plankton net was less than 64 μm. The redox potential of the mud-water interface was obtained by measuring within 2 cm above and below the mud-water interface using a high-precision, high-sensitivity microelectrode system. The high-precision, high-sensitivity microelectrode system has a vertical resolution of more than 200 μm for the mud-water interface. The concentrations of ferrous ions and sulfides in the overlying water were obtained by collecting samples of the overlying water within 20 cm above the mud-water interface and performing chemical analysis. The concentration of ferrous ions in the overlying water was analyzed using the o-phenanthroline spectrophotometric method or the phenanthroline spectrophotometric method, and the concentration of sulfides in the overlying water was analyzed using the methylene blue method.

5. The method for early warning and prevention of algal blooms in lacustrine ecosystems based on precursor pattern recognition according to claim 1, characterized in that, S4 include: A unified time axis is set for the observation data sequence, and the meteorological conditions of the target water area, the amount of algae in the surface water, the redox potential of the mud-water interface, the ferrous ion concentration in the bottom and overlying water, and the sulfide concentration in the bottom and overlying water are mapped to the unified time axis to complete the time alignment. When the sampling times of various precursor indicators are inconsistent, interpolation or resampling is used to convert each precursor indicator into data consistent with the unified time axis. When multiple monitoring points are set up in the target water area, each precursor indicator is limited to the data corresponding to the same monitoring point, or the data corresponding to multiple monitoring points are aggregated according to a preset spatial range to complete spatial matching. After completing time alignment and spatial matching, a feature sequence is constructed for each precursor indicator. The feature sequence includes numerical features and rate of change features, wherein the rate of change feature is obtained by dividing the numerical difference between adjacent time points by the time interval between adjacent time points.

6. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S5 include: Based on the feature sequence, an observation vector is constructed at each time point. The observation vector includes numerical features of the redox potential of the mud-water interface, the ferrous ion concentration in the bottom and overlying water, the sulfide concentration in the bottom and overlying water, and the amount of algae accumulation in the surface water. For each time point, initialize a set of state vectors containing multiple set members. Each set member contains dissolved oxygen state, mud-water interface redox state, bottom and overlying water ferrous ion state, bottom and overlying water sulfide state, and surface algae accumulation state. The mud-water interface redox state is characterized by the mud-water interface redox potential, the bottom and overlying water ferrous ion state is characterized by the bottom and overlying water ferrous ion concentration, the bottom and overlying water sulfide state is characterized by the bottom and overlying water sulfide concentration, and the surface algae accumulation state is characterized by the amount of algae in the surface water. For adjacent time points, the state transition relationship is used to predict each set member to obtain a predicted state vector set, and the predicted state vector set is mapped to a predicted observation vector set through the observation operator; The Kalman gain is calculated based on the difference between the observed vector and the predicted observed vector set, and the predicted state vector set is corrected using the Kalman gain to obtain the assimilated state vector set. The assimilation state sequence is determined by the set mean of the assimilation state vector set, and the uncertainty of the assimilation state sequence is determined by the set covariance of the assimilation state vector set. The assimilation state sequence and its uncertainty are then output.

7. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S6 include: The current state of the assimilation state sequence is taken as the initial state, and the meteorological conditions of the target water area and the amount of algae accumulation in the surface water corresponding to the current time in the feature sequence are taken as external driving quantities and input into the physical constraint neural dynamics model in the trained precursor pattern recognition model, so as to obtain the predicted state sequence by iteratively solving in the preset prediction time domain according to the preset prediction step size. The predicted state sequence includes the redox potential of the mud-water interface, the ferrous ion concentration in the overlying water, and the sulfide concentration in the overlying water at each predicted time within the preset prediction time domain. The predicted state sequence also provides the prediction uncertainty corresponding to the assimilation state sequence and its uncertainty.

8. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S7 includes: The predicted state sequence is input into the multi-task output network of the trained precursor pattern recognition model to obtain the lake flooding risk level prediction result and the expected outbreak time window prediction result corresponding to the predicted state sequence. The prediction results of the lake flood risk level include the probability of occurrence corresponding to each risk level, and the probability of occurrence is mapped to the lake flood risk level according to the preset risk level classification rules. The predicted outbreak time window is determined by the predicted remaining outbreak time output by the multi-task output network, or by the predicted moment when the lake flood risk level first reaches the medium risk level in the predicted state sequence.

9. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S8 includes: The prevention and control level is determined based on the early warning results, and prevention and control instructions are generated and issued in accordance with the prevention and control level to implement the whole-process linkage prevention and control. When the lake flood risk level characterized by the early warning result is low risk, the air-floating physical enclosure is kept in a non-outflow state, and the intermittent solar aeration device is operated. When the lake flooding risk level indicated by the warning results reaches medium risk or above, cyanobacteria interception, capture, and emergency removal in the lakeshore zone will be implemented. This includes: controlling the air-floating physical enclosure to be in an outflow state within 1000 m of the shore for interception and capture; setting up cyanobacteria capture trenches outside the air-floating physical enclosures; when the algal sludge thickness in the capture trenches reaches more than 50% of the trench's thickness, using a sludge removal device to remove the algal sludge; and removing accumulated cyanobacteria in the surface water to control the algal accumulation in the surface water to 0.5 kg / m³. 2 the following; When the lake flooding risk level indicated by the warning results reaches medium risk or above, aeration and oxygenation regulation is implemented. This regulation includes: starting high-intensity aeration on top of the operation of the intermittent solar-powered aeration device, ensuring an aeration flow rate greater than 0.18 m³ / s. 3 air / (min m 3 (water), and the emergency elimination standard is defined as the reduction of both the ferrous ion concentration and the sulfide concentration in the bottom and overlying water to below the preset low-risk threshold. When the content of volatile acid sulfides in the sediment is found to be higher than 100 mg / kg and the content of ferrous iron in the sediment is higher than 5000 mg / kg, the polluted sediment should be removed. The dredging depth should be controlled to be greater than both the depth of volatile acid sulfide pollution and the depth of ferrous iron pollution. After the removal of the polluted sediment is completed, a multi-series aquatic plant system should be constructed. The execution process of the whole-process linkage prevention and control is recorded as effect monitoring data. The effect monitoring data includes meteorological conditions of the target water area, algae accumulation in the surface water, redox potential of the mud-water interface, ferrous ion concentration in the bottom and overlying water, sulfide concentration in the bottom and overlying water, as well as the enclosure status, floating mud removal amount and aeration flow rate corresponding to the whole-process linkage prevention and control.

10. The method for early warning and prevention of algal blooms based on precursor pattern recognition according to claim 1, characterized in that, S9 includes: The effect monitoring data is time-aligned and spatially matched according to a preset time interval to generate an updated feature sequence. The updated feature sequence includes the numerical characteristics and rate of change characteristics of the meteorological conditions of the target water area, the amount of algae accumulation in the surface water, the redox potential of the mud-water interface, the concentration of ferrous ions in the bottom and overlying water, and the concentration of sulfides in the bottom and overlying water. An updated observation vector is constructed based on the updated feature sequence, and an ensemble Kalman filter is used to correct the assimilation state sequence to obtain the updated assimilation state sequence and its uncertainty. Based on the effect monitoring data, the updated lake flooding labels corresponding to the preset time slices are determined, and the updated feature sequences and the updated lake flooding labels are aggregated into updated training samples. Based on the updated training samples, incremental training is performed on the trained precursor pattern recognition model to minimize both the data fitting term and the physical constraint term in the training objective function, thus obtaining the updated precursor pattern recognition model. The updated early warning results are generated using the updated precursor pattern recognition model, and the updated early warning results are used in the next round of full-process joint prevention and control execution.

Images