A cyanobacterial bloom prediction method and system based on non-continuous data

Abstract

The application discloses a cyanobacteria accumulation prediction method and system based on discontinuous data, and belongs to the technical field of cyanobacteria monitoring. The method comprises the following steps: collecting historical discontinuous data of multiple monitoring stations in a target water area and processing the data to obtain pretreated data; establishing a cyanobacteria morphology-risk mapping rule, extracting cyanobacteria condition information from the pretreated data, and converting the information into corresponding risk probability values as cyanobacteria accumulation risk labels; generating a feature variable set to construct an accumulation prediction model, using the accumulation prediction model to output an initial value of cyanobacteria accumulation probability of a target prediction point; constructing a correction function based on spatial position and meteorological factors to correct the initial value of cyanobacteria accumulation probability to obtain a final cyanobacteria accumulation probability; and matching a risk level according to the final cyanobacteria accumulation probability to generate corresponding response measures. The application establishes a cyanobacteria morphology-risk mapping rule, realizes quantitative evaluation of cyanobacteria accumulation risk, and provides clear risk reference basis for subsequent targeted decision-making.

Description

Technical Field

[0001] This invention belongs to the technical field of cyanobacteria monitoring, and in particular relates to a method and system for predicting cyanobacteria accumulation based on discontinuous data. Background Technology

[0002] Blue-green algae accumulation is a common ecological and environmental problem in eutrophic water bodies (such as lakes and reservoirs). Especially in waters like Gonghu Bay, which serve as both drinking water sources and ecological regulators, the massive proliferation and accumulation of blue-green algae not only consumes dissolved oxygen and releases algal toxins, threatening drinking water safety, but also disrupts the ecological balance of the water body.

[0003] However, due to factors such as weather, traffic, and equipment maintenance, the monitoring data suffers from uneven time intervals and incomplete spatial coverage, resulting in a large number of discontinuous data fragments. Traditional models that rely on continuous time series data (such as RNN / LSTM) cannot effectively utilize this type of data, leading to a significant decrease in prediction accuracy.

[0004] The growth and accumulation of cyanobacteria are influenced by multiple factors, including meteorological elements, water quality parameters, spatial characteristics and ecological indicators. There are complex nonlinear coupling relationships among these factors. Traditional linear models (such as linear regression) or single-dimensional modeling methods (such as CNNs that only focus on time features) are unable to capture this synergistic effect of multiple factors, resulting in a large deviation between the prediction results and the actual situation.

[0005] Traditional methods often only output basic indicators such as cyanobacteria concentration, without linking the prediction results to the actual ecological risks of cyanobacteria accumulation or establishing a mapping relationship with emergency measures. This makes it difficult for the prediction results to directly support the decision-making of environmental protection departments and delays the early prevention and control of cyanobacteria accumulation. Summary of the Invention

[0006] To address the technical problems existing in the background art, the present invention provides a method and system for predicting cyanobacterial accumulation based on discontinuous data.

[0007] This invention adopts the following technical solution: a method for predicting cyanobacterial accumulation based on discontinuous data, comprising the following steps:

[0008] Collect historical discontinuous data from multiple monitoring stations within the target water area, and perform data preprocessing on the historical discontinuous data to generate preprocessed data with a unified structure and effective features;

[0009] Based on the ecological threat level corresponding to cyanobacterial morphology, a cyanobacterial morphology-risk mapping rule is established, and cyanobacterial status information is extracted from the preprocessed data. , Indicates a point in time. For monitoring stations;

[0010] Combining the aforementioned cyanobacteria morphology-risk mapping rule to obtain cyanobacteria status information Converted into corresponding risk probability values and the risk probability value As a risk label for cyanobacteria accumulation;

[0011] The feature variables in the preprocessed data are obtained to generate a feature variable set, and a stacking prediction model is constructed. The initial value of the cyanobacteria stacking probability at the target prediction point is output using the stacking prediction model. The stacking prediction model integrates high-dimensional feature dimensionality reduction optimization, nonlinear relationship modeling, and spatiotemporal sequence correlation analysis.

[0012] A correction function is constructed based on spatial location and meteorological factors. The initial value of the cyanobacteria accumulation probability is corrected to obtain the final cyanobacteria accumulation probability. Based on the final probability of cyanobacteria accumulation, a risk level is matched, and corresponding response measures are generated.

[0013] In a further embodiment, the preprocessing process of the preprocessed data includes at least one of: data cleaning, control imputation, feature encoding, redundancy removal, and standardization.

[0014] In a further embodiment, the process of establishing the cyanobacterial morphology-risk mapping rule is as follows:

[0015] Cyanobacteria morphology is classified into the following types: undiscovered, granular, striped, and paint-like, and criteria for judging each type are defined.

[0016] Based on the dissolved oxygen consumption rate, algal toxin release, and the extent of ecological damage to the aquatic ecosystem corresponding to each cyanobacterial morphology, the threat weight coefficients for each morphology were determined: no threat was detected, resulting in a threat weight coefficient of 0; particulate cyanobacteria had a threat weight coefficient of [missing value]. The strip threat weighting coefficient is The threat weighting coefficient for paint-like appearance is ,in, ;

[0017] By mapping cyanobacterial morphology to threat weight coefficients, a cyanobacterial morphology-risk mapping rule is formed.

[0018] In a further embodiment, the risk probability value The conversion formula is:

[0019] Based on information about cyanobacteria Obtain the percentage of cyanobacteria coverage in the target water area. Based on coverage area percentage The coverage area correction factor was calculated. ;

[0020] Based on information about cyanobacteria Determine the density of cyanobacteria in the target water area Based on the density of cyanobacteria aggregation Determine the cluster density correction factor ;

[0021] Based on coverage area correction factor and aggregation density correction factor We perform weighted fusion to obtain the risk probability value. :

[0022] ;

[0023] In the formula, , This refers to the morphological types of cyanobacteria. For granular threat weighting coefficients, For strip-shaped threat weighting coefficients, , and All are weighting coefficients. .

[0024] In a further embodiment, the optimization process for the high-dimensional feature dimensionality reduction optimization is as follows:

[0025] Based on feature dimensions, the set of feature variables is divided into four types of feature sets: time feature sets. Spatial feature set Meteorological feature set and water quality characteristics set Standardize the four feature sets respectively to obtain the corresponding standardized features;

[0026] Weights are assigned based on the degree of impact of cyanobacterial accumulation, with the following weighting coefficients representing time characteristics. Spatial feature weight coefficient Meteorological feature weighting coefficient and water quality characteristic weighting coefficient ;

[0027] Dynamic PCA dimensionality reduction is performed on the four standardized feature sets to obtain the time-low dimensional feature subsets. Low-dimensional feature subsets in space Meteorological low-dimensional feature subset and low-dimensional feature subsets of water quality ;

[0028] Obtain the spatiotemporal fusion low-dimensional feature set :

[0029] .

[0030] In a further embodiment, the specific steps for modeling the nonlinear relationship are as follows:

[0031] A random forest model is constructed to capture the non-linear relationship between features and cyanobacteria accumulation risk labels, with the number of decision trees set to [value missing]. , No. The prediction results of the decision trees are The short-term prediction results of the random forest model for: .

[0032] In a further embodiment, the analysis process of the spatiotemporal sequence correlation analysis is as follows:

[0033] A Transformer model is constructed to capture long-term spatiotemporal correlations, and sine-cosine position encoding is performed on time features;

[0034] The distance and angle between the station and the target in the spatiotemporal features are normalized and then embedded into the encoding vector to generate spatiotemporal fusion features. ;

[0035] Long-term prediction results are obtained based on the Transformer model. :

[0036] In the formula, It is a feedforward neural network. For layer normalization, These are the feature weight coefficients.

[0037] In a further embodiment, the correction function The construction method is as follows:

[0038] Determine monitoring sites With the target prediction point straight-line distance Monitoring stations With the target prediction point The angle between the line connecting the two points and the real-time wind direction Simultaneously acquire real-time wind direction and angle ;

[0039] Based on straight-line distance Angle and real-time wind direction angle Constructing the correction function Its expression is as follows:

[0040] In the formula, This is the attenuation coefficient.

[0041] In a further embodiment, the final cyanobacteria accumulation probability The calculation formula is as follows:

[0042] ,in, The total number of monitoring points. For monitoring sites The output value of the correction function, For monitoring sites The predicted value obtained by the stacking prediction model.

[0043] A system for predicting cyanobacterial accumulation based on discontinuous data, used to implement the cyanobacterial accumulation prediction method described above, includes:

[0044] The first module is set to collect historical discontinuous data from multiple monitoring stations in the target water area, and perform data preprocessing on the historical discontinuous data to generate preprocessed data with uniform structure and effective features.

[0045] The second module is configured to establish a cyanobacterial morphology-risk mapping rule based on the ecological threat level corresponding to the cyanobacterial morphology, and extract cyanobacterial status information from the preprocessed data. , Indicates a point in time. For monitoring stations;

[0046] The third module is configured to combine the cyanobacteria morphology-risk mapping rule to obtain cyanobacteria status information. Converted into corresponding risk probability values and the risk probability value As a risk label for cyanobacteria accumulation;

[0047] The fourth module is configured to acquire feature variables from the preprocessed data, generate a set of feature variables, construct a stacking prediction model, and use the stacking prediction model to output the initial value of the cyanobacteria stacking probability at the target prediction point; the stacking prediction model integrates high-dimensional feature dimensionality reduction optimization, nonlinear relationship modeling, and spatiotemporal sequence correlation analysis.

[0048] The fifth module is configured to construct a correction function based on spatial location and meteorological factors. The initial value of the cyanobacteria accumulation probability is corrected to obtain the final cyanobacteria accumulation probability. Based on the final probability of cyanobacteria accumulation, a risk level is matched, and corresponding response measures are generated.

[0049] The beneficial effects of this invention are as follows: When constructing the accumulation prediction model, this invention performs high-dimensional feature reduction optimization on the set of feature variables including time, space, meteorology and water quality, and reasonably allocates weights according to the degree of influence of cyanobacteria accumulation. Then, it uses dynamic PCA dimensionality reduction and other methods to effectively reduce the data dimensionality and computational load, while retaining key feature information, improving the model's ability to process complex data, and thus improving the accuracy of cyanobacteria accumulation probability prediction.

[0050] Based on the ecological threat level corresponding to cyanobacteria morphology, this invention meticulously classifies cyanobacteria morphology into different types such as undiscovered, granular, striped, and paint-like, and determines the threat weight coefficient of each morphology, thereby establishing a cyanobacteria morphology-risk mapping rule. This allows the status of cyanobacteria to be objectively converted into the corresponding risk probability value, realizing a quantitative assessment of the risk of cyanobacteria accumulation.

[0051] In the process of converting risk probability values, multiple factors such as the coverage area and aggregation density of cyanobacteria in the target water area are comprehensively considered, and the risk probability value is obtained through reasonable weighted fusion calculation. This avoids the one-sidedness of assessing risk by a single factor and reflects the actual risk level brought about by cyanobacteria accumulation more comprehensively and scientifically.

[0052] By matching the final probability of cyanobacteria accumulation with the corresponding risk level and generating corresponding response measures, a seamless connection from prediction to decision-making is achieved. This helps relevant management departments or personnel to quickly and intuitively understand the severity of cyanobacteria accumulation and take effective countermeasures, such as timely dredging and adjusting water quality management strategies, thereby enhancing the practical value of the entire prediction method in actual cyanobacteria control work. Attached Figure Description

[0053] Figure 1 This is a flowchart of the cyanobacteria accumulation prediction method based on discontinuous data in Example 1. Detailed Implementation

[0054] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0055] Example 1

[0056] This embodiment takes Gonghu Bay as the research object. In view of the problem of frequent cyanobacterial accumulation and urgent need for ecological risk prevention and control in this water area, Gonghu Bay covers a large area and the distribution of cyanobacteria in the water is affected by multiple factors such as time, space, weather, and water quality. Traditional prediction methods are difficult to accurately capture its accumulation pattern.

[0057] A method for predicting cyanobacterial accumulation based on discontinuous data is proposed, such as... Figure 1 As shown, it includes the following steps:

[0058] Collect historical discontinuous data from multiple monitoring stations within the target water area, and perform data preprocessing on the historical discontinuous data to generate preprocessed data with a unified structure and effective features;

[0059] Based on the ecological threat level corresponding to cyanobacterial morphology, a cyanobacterial morphology-risk mapping rule is established, and cyanobacterial status information is extracted from the preprocessed data. , Indicates a point in time. For monitoring stations;

[0060] Combining the aforementioned cyanobacteria morphology-risk mapping rule to obtain cyanobacteria status information Converted into corresponding risk probability values and the risk probability value As a risk label for cyanobacteria accumulation;

[0061] The feature variables in the preprocessed data are obtained to generate a feature variable set, and a stacking prediction model is constructed. The initial value of the cyanobacteria stacking probability at the target prediction point is output using the stacking prediction model. The stacking prediction model integrates high-dimensional feature dimensionality reduction optimization, nonlinear relationship modeling, and spatiotemporal sequence correlation analysis.

[0062] A correction function is constructed based on spatial location and meteorological factors. The initial value of the cyanobacteria accumulation probability is corrected to obtain the final cyanobacteria accumulation probability. Based on the final probability of cyanobacteria accumulation, a risk level is matched, and corresponding response measures are generated.

[0063] This embodiment focuses on the Gonghu Bay water area, collecting historical non-continuous monitoring data from 2017 to 2019 from four core monitoring stations: Nanquan Water Source Plant, Renzi Port, Xuxian Port, and Xidong Water Plant. The historical non-continuous data in this embodiment includes: date, time, meteorological elements (such as air pressure (kPa), wind direction, wind speed (m / s), wind level, air temperature (°C)), water temperature (°C), water chemical indicators (transparency, redox potential, pH, dissolved oxygen, algal density, chlorophyll a), cyanobacteria status (not found, granular, streaky, paint-like), weather conditions (sunny, cloudy, light rain, showers, moderate rain, heavy rain), etc.

[0064] Since the data comes from patrol records of different monitoring periods and covers multiple types of indicators such as meteorology, water quality, and ecology, there are differences in data structure (such as only recording core water quality indicators on some dates, and wind direction being presented in text descriptions rather than standardized codes for some periods). It is inevitable that there will be uneven data quality (such as missing dissolved oxygen data for a certain date at some stations, and some wind speed records exceeding reasonable physical ranges). Systematic preprocessing is required to ensure data usability.

[0065] To address the aforementioned issues, the preprocessing process for the preprocessed data described in this embodiment includes at least one of the following: data cleaning, control imputation, feature encoding, redundancy removal, and standardization.

[0066] In a further embodiment, the process of establishing the cyanobacterial morphology-risk mapping rule is as follows:

[0067] Cyanobacteria morphology was classified into the following types: undiscovered, granular, banded, and paint-like, and criteria for judging each type were defined. For ease of understanding, the criteria for judging each type are as follows: "granular" is defined as algal density of 100,000-1,000,000 cells / L and appearing as dispersed granules under a microscope; "banded" is defined as algal density of 1,000,000-5,000,000 cells / L and the appearance of continuous bands on the water surface; "paint-like" is defined as algal density >5,000,000 cells / L and the water surface covered with a uniform, thick layer.

[0068] Based on the dissolved oxygen consumption rate, algal toxin release, and aquatic ecosystem damage extent corresponding to different cyanobacterial morphologies, the analytic hierarchy process (AHP) was used to determine the threat weight coefficients for each morphology: no threat weight coefficient was found (0), and the threat weight coefficient for granular algae was [not specified]. Strip-shaped threat weighting coefficient and paint-like threat weighting coefficient ,in, In this embodiment, , .

[0069] By mapping cyanobacterial morphology to threat weight coefficients, a cyanobacterial morphology-risk mapping rule is formed, as shown in Table 1.

[0070] Table 1. Cyanobacterial morphology-risk mapping table

[0071]

[0072] Traditional risk probability calculations, relying solely on quantitative indicators such as coverage area and aggregation density, can lead to "over-adjustment." Furthermore, traditional models do not explicitly retain the inherent risk weights of cyanobacterial morphologies, making them susceptible to being obscured by fluctuations in quantitative indicators, thus resulting in the loss of inherent risks. To overcome these technical problems, the risk probability values ​​in this embodiment... The conversion formula is:

[0073] Based on information about cyanobacteria Obtain the percentage of cyanobacteria coverage in the target water area. Based on coverage area percentage The coverage area correction factor was calculated. Coverage correction factor The calculation formula is: ,in, The area attenuation coefficient is determined based on historical cyanobacteria coverage area and ecological harm data of the target water area, with a value ranging from 0.03 to 0.05.

[0074] In this embodiment, ,when hour, No overwrite, no correction, and vice versa. ,but Approaching 1, maximum correction.

[0075] Based on information about cyanobacteria Determine the density of cyanobacteria in the target water area Based on the density of cyanobacteria aggregation Determine the cluster density correction factor Furthermore, ,in, The critical density for cyanobacterial blooms in the target water area.

[0076] Based on coverage area correction factor and aggregation density correction factor We perform weighted fusion to obtain the risk probability value. :

[0077] ;

[0078] In the formula, , This refers to the morphological types of cyanobacteria. For granular threat weighting coefficients, For strip-shaped threat weighting coefficients, , and All are weighting coefficients. The weighting coefficients mentioned in this embodiment can be set manually based on past experience.

[0079] It should be noted that, As a basic weighting component, the inherent risk weight of cyanobacterial morphology is retained to ensure that it accounts for no less than 20%, thus avoiding over-correction when quantitative indicators are abnormal. Under normal circumstances... .

[0080] The original high-dimensional feature data (multi-dimensional, including time, space, meteorology, and water quality) is massive, resulting in a heavy workload and low efficiency for backend data processing. The optimization process for high-dimensional feature dimensionality reduction in this embodiment is as follows:

[0081] Based on feature dimensions, the set of feature variables is divided into four types of feature sets: time feature sets. Spatial feature set Meteorological feature set and water quality characteristics set Standardize the four feature sets separately to obtain the corresponding standardized features, such as... standardization;

[0082] The analytic hierarchy process (AHP) was used to assign weights based on the degree of impact of cyanobacterial accumulation, with the following weights being time-specific weight coefficients. Spatial feature weight coefficient Meteorological feature weighting coefficient and water quality characteristic weighting coefficient In other embodiments, corresponding coefficient values ​​can be manually assigned according to the season and time to increase the applicability of the scenario.

[0083] Dynamic PCA dimensionality reduction is performed on the four standardized feature sets to obtain the time-low dimensional feature subsets. Low-dimensional feature subsets in space Meteorological low-dimensional feature subset and low-dimensional feature subsets of water quality ;

[0084] Obtain the spatiotemporal fusion low-dimensional feature set :

[0085] .

[0086] Dynamic PCA dimensionality reduction and feature classification significantly reduce the amount of data to be processed, alleviate the background workload, and improve the efficiency of data processing and model training.

[0087] By employing the analytic hierarchy process (AHP) to scientifically allocate feature weight coefficients and combining this with dynamic PCA to retain key features, a spatiotemporal fusion of low-dimensional feature sets is achieved. It can accurately reflect the impact of various factors on cyanobacteria accumulation, ensuring the effectiveness of the input features for subsequent prediction models.

[0088] Additionally, it should be noted that the execution process of dynamic PCA dimensionality reduction described in this embodiment is as follows:

[0089] For time feature set (e.g., year, month, day, season): For the standardized time features, first count the frequency of cyanobacterial accumulation in different time periods (e.g., high accumulation frequency in summer, low in winter), and dynamically adjust the PCA principal component screening threshold: prioritize retaining principal components with high correlation to accumulation frequency (e.g., the "season-month" comprehensive factor), and finally retain 2-3 principal components. Redundant time features (such as "hours" and other dimensions that have a weak impact on stacking) are removed.

[0090] Spatial feature set (e.g., station latitude and longitude, distance from the target point): Combining the cyanobacteria diffusion pattern in Gonghu Bay (e.g., diffusion along wind direction), the correlation between various spatial features and cyanobacteria accumulation locations is dynamically calculated. During PCA dimensionality reduction, the principal components related to "distance from the target point" and "diffusion area to which the station belongs" are retained, generating two principal components. This reduces spatial dimensional redundancy.

[0091] Meteorological feature set (e.g., wind direction, wind speed, temperature): Based on the driving effect of meteorological factors on cyanobacterial accumulation in historical data (e.g., high wind speed promotes diffusion, suitable temperature accelerates growth), high-impact meteorological features (e.g., wind speed, temperature) are dynamically assigned higher weights. After PCA dimensionality reduction, two core principal components (e.g., "wind speed-wind direction" and "temperature-air pressure" factors) are retained. This reduces the amount of data on low-impact meteorological features (such as wind speed).

[0092] Water quality characteristics set (e.g., water temperature, algae density, dissolved oxygen): Combining the water quality requirements for cyanobacteria growth (e.g., suitable water temperature, high algae density), dynamically analyze the correlation between various water quality indicators and cyanobacteria accumulation. During PCA dimensionality reduction, prioritize retaining strongly correlated principal components such as "algae density-chlorophyll a" and "water temperature-dissolved oxygen," generating principal components... Weakly correlated water quality characteristics such as conductivity were eliminated.

[0093] In another embodiment, a spatiotemporal fusion low-dimensional feature set is used. Inputting the data into a random forest model yields short-term prediction results. Specifically, a random forest model is constructed to capture the non-linear relationship between features and cyanobacteria accumulation risk labels, with the number of decision trees set to [value missing]. , No. The prediction results of the decision trees are The short-term prediction results of the random forest model for: .

[0094] Furthermore, to predict the short-term (12-24 hours) risk probability of cyanobacteria accumulation, a Bagging algorithm is used to integrate multiple decision trees to reduce the overfitting risk caused by noise in discontinuous data. The number of decision trees is [not specified]. Optimized using 5-fold cross-validation, the value is set between 150 and 200.

[0095] Correspondingly, the analysis process of spatiotemporal sequence correlation analysis is as follows:

[0096] A Transformer model is constructed to capture long-term spatiotemporal correlations, and sine-cosine position encoding is performed on time features; ;

[0097] In the formula, For time steps, To guide the dimension, For feature dimension, This is the encoded output value.

[0098] The distance and angle between the station and the target in the spatiotemporal features are normalized and then embedded into the encoding vector to generate spatiotemporal fusion features. It solves the problem of "separate processing of time and space features" in traditional models and adapts to the actual laws of cyanobacteria "growing with time and spreading with space".

[0099] Long-term prediction results are obtained based on the Transformer model. : In the formula, It is a feedforward neural network. For layer normalization, These are the feature weight coefficients.

[0100] Therefore, the predicted value obtained by the stacking prediction model described in this embodiment is... The prediction model is:

[0101] In the formula, and These are all weighting coefficients, which can be assigned by human experience or by using the analytic hierarchy process.

[0102] Traditional forecasting often treats the contribution of data from each monitoring station to the target point as equal, ignoring the actual law that "cyanobacteria diffusion decreases with distance." Ignoring distance differences can easily lead to data from distant stations interfering with the target point forecast (e.g., high cyanobacteria density at distant stations that cannot diffuse to the target point are still included in the forecast), reducing forecast accuracy.

[0103] Similarly, ignoring the relationship between wind direction and the angle of the line connecting the station and the target point will lead to a mismatch of "insufficient weight of upstream station data and excessive weight of downstream station data", which cannot fit the actual spatial correlation logic of cyanobacteria spreading with wind direction.

[0104] Therefore, this embodiment introduces the following method for constructing the correction function:

[0105] Determine monitoring sites With the target prediction point straight-line distance Monitoring stations With the target prediction point The angle between the line connecting the two points and the real-time wind direction Simultaneously acquire real-time wind direction and angle ;

[0106] Based on straight-line distance Angle and real-time wind direction angle Constructing the correction function Its expression is as follows:

[0107] In the formula, This is the attenuation coefficient, which can be determined based on factors such as water area and cyanobacteria diffusion rate.

[0108] It should be noted that an exponential function is used to represent the effect of distance on weight. The greater the distance, the smaller the weight. For wind direction matching weights, a cosine function is used to represent the degree of matching between the wind direction and the station's direction relative to the target point. The weight is 1 when the wind direction and the target direction are the same; the weight is 0 when the wind direction and the target direction are opposite.

[0109] Finally, the probability of cyanobacteria accumulation. The calculation formula is as follows:

[0110] ,in, The total number of monitoring points. For monitoring sites The output value of the correction function, For monitoring sites The predicted value obtained by the stacking prediction model.

[0111] The predicted values ​​of each station are weighted and summed, and then divided by the sum of the weights of all stations to obtain the prediction result of the target point. This method comprehensively considers the information of different stations and their spatial relationship with the target point, and more accurately predicts the situation of the target point.

[0112] Based on the above predictive analysis, according to the final probability of cyanobacteria accumulation Classify risk levels and match corresponding response measures:

[0113] If the cyanobacteria morphology is predominantly "not detected," routine monitoring should be conducted, with monthly sampling and analysis, focusing on changes in basic water quality indicators; The value is less than 0.2.

[0114] If the cyanobacteria are mostly granular, intensify patrols, conduct sampling and monitoring every 3 days, and record changes in the cyanobacteria aggregation range simultaneously; The value ranges from 0.2 to 0.4.

[0115] If the cyanobacteria morphology includes "band-like" patterns, an early warning mechanism should be activated, with daily sampling and monitoring, and small-scale retrieval equipment deployed for localized physical intervention to control the spread of the cyanobacteria; The value ranges from 0.4 to 0.7.

[0116] If the cyanobacteria exhibit a "paint-like" appearance, an emergency response will be activated, with monitoring every 6 hours. A large-scale cleanup fleet will be mobilized, and water quality regulators (such as clay) will be added to inhibit algae growth. Coordination with nearby water plants will be undertaken to adjust their water intake strategies to ensure water supply safety. The value is greater than or equal to 0.7.

[0117] This embodiment was verified in the Gonghu Bay waters during the summer of 2020 (June-August). Ten verification sites that were not involved in the modeling were selected, and the predicted results were compared with the actual cyanobacteria accumulation.

[0118] Prediction accuracy: During the validation period, the correlation coefficients between the final cyanobacteria accumulation probability P and the actual cyanobacteria accumulation degree (expressed by a comprehensive score of algal density and coverage area) were R=0.89, R²=0.7921, MSE=0.0032, RMSE=0.0566, MAE=0.0245, and the prediction error was controlled within a low range.

[0119] Response effectiveness: The early warning and emergency measures initiated based on the forecast results reduced the number of large-scale cyanobacterial outbreaks in Gonghu Bay in 2020 by 40% compared to 2019, reduced the area of ​​water with excessive algal toxins by 35%, and increased the average dissolved oxygen level in the water by 0.8 mg / L, effectively protecting the ecological environment and drinking water source safety of Gonghu Bay.

[0120] Example 2

[0121] This embodiment discloses a cyanobacteria accumulation prediction method system based on discontinuous data, used to implement the prediction method described above, including:

[0122] The first module is set to collect historical discontinuous data from multiple monitoring stations in the target water area, and perform data preprocessing on the historical discontinuous data to generate preprocessed data with uniform structure and effective features.

[0123] The second module is configured to establish a cyanobacterial morphology-risk mapping rule based on the ecological threat level corresponding to the cyanobacterial morphology, and extract cyanobacterial status information from the preprocessed data. , Indicates a point in time. For monitoring stations;

[0124] The third module is configured to combine the cyanobacteria morphology-risk mapping rule to obtain cyanobacteria status information. Converted into corresponding risk probability values and the risk probability value As a risk label for cyanobacteria accumulation;

[0125] The fourth module is configured to acquire feature variables from the preprocessed data, generate a set of feature variables, construct a stacking prediction model, and use the stacking prediction model to output the initial value of the cyanobacteria stacking probability at the target prediction point; the stacking prediction model integrates high-dimensional feature dimensionality reduction optimization, nonlinear relationship modeling, and spatiotemporal sequence correlation analysis.

[0126] The fifth module is configured to construct a correction function based on spatial location and meteorological factors. The initial value of the cyanobacteria accumulation probability is corrected to obtain the final cyanobacteria accumulation probability. Based on the final probability of cyanobacteria accumulation, a risk level is matched, and corresponding response measures are generated.

Claims

1. A method for predicting cyanobacterial accumulation based on discontinuous data, characterized in that, Includes the following steps: Collect historical discontinuous data from multiple monitoring stations within the target water area, and perform data preprocessing on the historical discontinuous data to generate preprocessed data with a unified structure and effective features; Based on the ecological threat level corresponding to cyanobacterial morphology, a cyanobacterial morphology-risk mapping rule is established, and cyanobacterial status information is extracted from the preprocessed data. , Indicates a point in time. For monitoring stations; Combining the aforementioned cyanobacteria morphology-risk mapping rule to obtain cyanobacteria status information Converted into corresponding risk probability values and the risk probability value As a risk label for cyanobacterial accumulation; the risk probability value The formula is as follows: ； In the formula, , This refers to the morphological types of cyanobacteria. For granular threat weighting coefficients, This represents the threat weighting coefficient in a strip format. , and All are weighting coefficients. , This is the coverage area correction factor. This is a cluster density correction factor; Feature variables are obtained from the preprocessed data to generate a feature variable set, and a cyanobacteria accumulation prediction model is constructed. The cyanobacteria accumulation probability initial value for the target prediction point is then output using this model. The cyanobacteria accumulation prediction model integrates high-dimensional feature reduction optimization, nonlinear relationship modeling, and spatiotemporal sequence correlation analysis to obtain the predicted values. The prediction model is: In the formula, and All are weighting coefficients. The short-term prediction results are obtained by modeling nonlinear relationships. This is the long-term prediction result of spatiotemporal sequence correlation analysis; the specific steps of nonlinear relationship modeling are as follows: A random forest model is constructed to capture the non-linear relationship between features and cyanobacteria accumulation risk labels, with the number of decision trees set to [value missing]. , No. The prediction results of the decision trees are The short-term prediction results of the random forest model for: , A spatiotemporal fusion set of low-dimensional features; The analysis process of spatiotemporal series correlation analysis is as follows: A Transformer model is constructed to capture long-term spatiotemporal correlations, and sine-cosine position encoding is performed on time features; The distance and angle between the station and the target in the spatiotemporal features are normalized and then embedded into the encoding vector to generate spatiotemporal fusion features. ; Long-term prediction results are obtained based on the Transformer model. : In the formula, It is a feedforward neural network. For layer normalization, These are the feature weight coefficients; A correction function is constructed based on spatial location and meteorological factors. The initial value of the cyanobacteria accumulation probability is corrected to obtain the final cyanobacteria accumulation probability. Based on the final probability of cyanobacteria accumulation, a risk level is matched, and corresponding response measures are generated. The final probability of cyanobacteria accumulation The calculation formula is as follows: ,in, The total number of monitoring points. For monitoring sites The output value of the correction function, For monitoring sites Predicted values ​​obtained from the stacking prediction model; Among them, the correction function The construction method is as follows: Determine monitoring sites With the target prediction point straight-line distance Monitoring stations With the target prediction point The angle between the line connecting the two points and the real-time wind direction Simultaneously acquire real-time wind direction and angle ; Based on straight-line distance Angle and real-time wind direction angle Constructing the correction function Its expression is as follows: In the formula, This is the attenuation coefficient.

2. The method for predicting cyanobacterial accumulation based on discontinuous data according to claim 1, characterized in that, The preprocessing process of the preprocessed data includes at least one of the following: data cleaning, control imputation, feature encoding, redundancy removal, and standardization.

3. The method for predicting cyanobacterial accumulation based on discontinuous data according to claim 1, characterized in that, The process of establishing the cyanobacterial morphology-risk mapping rule is as follows: Cyanobacteria morphology is classified into the following types: undiscovered, granular, striped, and paint-like, and criteria for judging each type are defined. Based on the dissolved oxygen consumption rate, algal toxin release, and the extent of ecological damage to the aquatic ecosystem corresponding to each cyanobacterial morphology, the threat weight coefficients for each morphology were determined: no threat was detected, resulting in a threat weight coefficient of 0; particulate cyanobacteria had a threat weight coefficient of [missing value]. The strip threat weighting coefficient is The threat weighting coefficient for paint-like appearance is ,in, ; By mapping cyanobacterial morphology to threat weight coefficients, a cyanobacterial morphology-risk mapping rule is formed.

4. The method for predicting cyanobacterial accumulation based on discontinuous data according to claim 1, characterized in that, The coverage correction factor It is based on information about cyanobacteria status. Obtain the percentage of cyanobacteria coverage in the target water area. Based on coverage area percentage The calculation yielded: ,in, This is the area attenuation coefficient; The aggregation density correction factor Based on cyanobacteria status information Determine the density of cyanobacteria in the target water area Based on the density of cyanobacteria aggregation Sure: ,in, The critical density for cyanobacterial blooms in the target water area.

5. The method for predicting cyanobacterial accumulation based on discontinuous data according to claim 1, characterized in that, The optimization process for the high-dimensional feature dimensionality reduction optimization is as follows: Based on feature dimensions, the set of feature variables is divided into four types of feature sets: time feature sets. Spatial feature set Meteorological feature set and water quality characteristics set Standardize the four feature sets respectively to obtain the corresponding standardized features; Weights are assigned based on the degree of impact of cyanobacterial accumulation, with the following weighting coefficients representing time characteristics. Spatial feature weight coefficient Meteorological feature weighting coefficient and water quality characteristic weighting coefficient ; Dynamic PCA dimensionality reduction is performed on the four standardized feature sets to obtain the time-low dimensional feature subsets. Low-dimensional feature subsets in space Meteorological low-dimensional feature subset and low-dimensional feature subsets of water quality ; Obtain the spatiotemporal fusion low-dimensional feature set : 。 6. A method system for predicting cyanobacterial accumulation based on discontinuous data, used to implement the cyanobacterial accumulation prediction method as described in any one of claims 1 to 5, characterized in that, include: The first module is set to collect historical discontinuous data from multiple monitoring stations in the target water area, and perform data preprocessing on the historical discontinuous data to generate preprocessed data with uniform structure and effective features. The second module is configured to establish a cyanobacterial morphology-risk mapping rule based on the ecological threat level corresponding to the cyanobacterial morphology, and extract cyanobacterial status information from the preprocessed data. , Indicates a point in time. For monitoring stations; The third module is configured to combine the cyanobacteria morphology-risk mapping rule to obtain cyanobacteria status information. Converted into corresponding risk probability values and the risk probability value As a risk label for cyanobacteria accumulation; The fourth module is configured to acquire feature variables from the preprocessed data, generate a set of feature variables, construct a stacking prediction model, and use the stacking prediction model to output the initial value of the cyanobacteria stacking probability at the target prediction point; the stacking prediction model integrates high-dimensional feature dimensionality reduction optimization, nonlinear relationship modeling, and spatiotemporal sequence correlation analysis. The fifth module is configured to construct a correction function based on spatial location and meteorological factors. The initial value of the cyanobacteria accumulation probability is corrected to obtain the final cyanobacteria accumulation probability. Based on the final probability of cyanobacteria accumulation, a risk level is matched, and corresponding response measures are generated.

Images