A three-dimensional self-adaptive algae inhibition buoy system and a cooperative control method

Abstract

The application provides a three-dimensional self-adaptive algae inhibition buoy system and a cooperative control method, which comprises obtaining historical water quality and meteorological data, combining hydrological flow rate to calculate time cumulative effect characteristics of environmental factors. Then, a random forest model is constructed, and a node split index is modified by introducing a Pearson correlation coefficient to optimize feature selection. In the prediction stage, a preliminary prediction is first made to determine the risk level, based on which high-quality decision trees are screened to form an expert subset, and a weighted average is used to obtain the final algae concentration prediction value. Finally, combined with the prediction value and the prediction confidence based on the dispersion, an "monitoring", "warning" or "emergency" algae inhibition strategy is matched and output from the hierarchical strategy library.

Description

Technical Field

[0001] This application belongs to the field of control, and in particular relates to a three-dimensional adaptive algae-suppressing buoy system and a collaborative control method. Background Technology

[0002] Eutrophication easily triggers algal blooms, which, on a large scale, severely disrupt the balance of aquatic ecosystems, deteriorate water quality, clog pipes, threaten drinking water safety, and cause significant economic losses to fisheries and tourism. Accurate prediction and early warning are crucial for controlling algal blooms. With technological advancements, machine learning models such as random forests and neural networks are widely used in algal bloom prediction. However, most existing methods are based on instantaneous parameters, while environmental factors such as nutrients, water temperature, and light have time-lag and cumulative effects on algal growth. Furthermore, when constructing ensemble learning models like random forests, traditional node splitting criteria, such as information gain, primarily focus on the splitting ability of features on local datasets, failing to utilize the intrinsic correlation between features and target algal concentration over a global time series. In practice, different base learners in the model, such as decision trees, exhibit varying predictive abilities. Averaging all results with equal weight fails to highlight the role of expert base learners or determine the confidence level of the results. All these factors contribute to inaccurate algal bloom predictions, resulting in poor algal suppression effects. Summary of the Invention

[0003] To address the problems mentioned in the background art, in a first aspect of the present invention, a three-dimensional adaptive algae-suppressing buoy system is proposed, comprising a solar photovoltaic panel, an energy storage module, a central control unit, and a vertical guide rail, wherein multiple sensors are arranged on the vertical guide rail, and the system further includes the following modules: The data acquisition module is used to acquire historical water quality and meteorological data of the target water area. Based on the data sequence of each environmental factor within a preset historical length and the attenuation coefficient determined according to the hydrological flow velocity data of the same period, the time cumulative effect characteristics of the environmental factor are obtained by exponential attenuation weighted calculation. The model building module is used to build a random forest prediction model based on the time cumulative effect features. When splitting nodes in each decision tree of the model, the information gain of the candidate split features is multiplied by a preset power of the Pearson correlation coefficient of the feature with the target algae concentration time series to obtain a modified split evaluation index, and the feature with the largest index value is selected for splitting. The prediction module is used to input the data to be predicted, perform preliminary prediction using the random forest prediction model, and determine the algal bloom risk level based on the preliminary prediction results. Based on the algal bloom risk level, decision trees with historical prediction performance better than a preset threshold are selected from the random forest prediction model to form a subset of expert decision trees. The prediction results of each decision tree in the subset of expert decision trees are weighted and averaged according to their respective historical prediction performance to obtain the predicted algal concentration value. The control module is used to determine the prediction confidence level based on the predicted algae concentration value and the dispersion of each predicted value in the subset of the expert decision tree, and to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies by combining the predicted algae concentration value and the prediction confidence level.

[0004] Preferably, the step of selecting decision trees whose historical prediction performance is better than a preset threshold specifically involves: selecting a subset of historical data corresponding to the algal bloom risk level from the validation dataset, using the subset of historical data to evaluate the root mean square error of each decision tree in the random forest model, and selecting decision trees with root mean square errors less than a preset error threshold as the subset of expert decision trees.

[0005] Preferably, in the weighted average, the weight of the j-th decision tree in the subset of the expert decision trees is... The calculation formula is: ; in, Let be the root mean square error of the j-th decision tree in its historical prediction performance, and N be the total number of decision trees in the expert decision tree subset.

[0006] Preferably, determining the prediction confidence level based on the dispersion of the predicted algae concentration and each predicted value of the subset of the expert decision tree specifically involves: Calculate the standard deviation of each predicted value in the subset of the expert decision tree, and compare the standard deviation with a preset first confidence threshold and a second confidence threshold to determine whether the prediction confidence is high, medium or low.

[0007] Preferably, the step of combining the predicted algae concentration value with the prediction confidence level to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring," "early warning," and "emergency" strategies specifically involves: The predicted algae concentration is compared with the preset warning concentration threshold and emergency concentration threshold, and combined with the prediction confidence level, the corresponding algae suppression strategy is matched from the hierarchical algae suppression strategy library and output.

[0008] In another aspect, a three-dimensional adaptive algae suppression and synergistic control method is proposed, the method comprising the following steps: Historical water quality and meteorological data of the target water area are obtained. Based on the data sequence of each environmental factor within a preset historical length and the attenuation coefficient determined according to the hydrological flow velocity data of the same period, the time cumulative effect characteristics of the environmental factor are obtained by exponential attenuation weighted calculation. A random forest prediction model is constructed based on the time cumulative effect characteristics. When splitting nodes in each decision tree of the model, the information gain of the candidate splitting feature is multiplied by a preset power of the Pearson correlation coefficient of the feature with the target algae concentration time series to obtain a modified splitting evaluation index, and the feature with the largest index value is selected for splitting. Input the data to be predicted, perform preliminary prediction using the random forest prediction model, and determine the algal bloom risk level based on the preliminary prediction results; based on the algal bloom risk level, select decision trees from the random forest prediction model whose historical prediction performance is better than a preset threshold to form an expert decision tree subset; perform a weighted average of the prediction results of each decision tree in the expert decision tree subset based on their respective historical prediction performance to obtain the algal concentration prediction value. The prediction confidence level is determined based on the predicted algae concentration and the dispersion of each predicted value in the subset of the expert decision tree. Then, the algae suppression strategy is matched and output from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies.

[0009] Preferably, the step of selecting decision trees whose historical prediction performance is better than a preset threshold specifically involves: selecting a subset of historical data corresponding to the algal bloom risk level from the validation dataset, using the subset of historical data to evaluate the root mean square error of each decision tree in the random forest model, and selecting decision trees with root mean square errors less than a preset error threshold as the subset of expert decision trees.

[0010] Preferably, in the weighted average, the weight of the j-th decision tree in the subset of the expert decision trees is... The calculation formula is: ; in, Let be the root mean square error of the j-th decision tree in its historical prediction performance, and N be the total number of decision trees in the expert decision tree subset.

[0011] Preferably, determining the prediction confidence level based on the dispersion of the predicted algae concentration and each predicted value of the subset of the expert decision tree specifically involves: Calculate the standard deviation of each predicted value in the subset of the expert decision tree, and compare the standard deviation with a preset first confidence threshold and a second confidence threshold to determine whether the prediction confidence is high, medium or low.

[0012] Preferably, the step of combining the predicted algae concentration value with the prediction confidence level to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring," "early warning," and "emergency" strategies specifically involves: The predicted algae concentration is compared with the preset warning concentration threshold and emergency concentration threshold, and combined with the prediction confidence level, the corresponding algae suppression strategy is matched from the hierarchical algae suppression strategy library and output.

[0013] This invention improves the effectiveness of model input data by reflecting the time-cumulative effect characteristics of the lagged impact of environmental factors. Furthermore, the split evaluation index enhances the model's ability to identify key influencing factors and optimizes its internal structure. By selecting high-performing decision trees to form an expert subset and weighting them, interference from weak learners within the model is filtered out, ensuring the stability and reliability of the prediction results. Combining the predicted values ​​with their confidence levels achieves matching with stratified algae suppression strategies, providing a more scientific, clear, and targeted decision-making basis for actual algae bloom control work, and improving resource utilization efficiency and algae bloom prevention effectiveness. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the weights; Figure 2 A schematic diagram for selecting a subset of the expert decision tree; Figure 3 A schematic diagram for determining algae suppression strategies. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] First embodiment A three-dimensional adaptive algae-suppressing buoy system is proposed, comprising a solar photovoltaic panel, an energy storage module, a central control unit, and a vertical guide rail. Multiple sensors are arranged on the vertical guide rail. The system also includes the following modules: The data acquisition module is used to acquire historical water quality and meteorological data of the target water area. Based on the data sequence of each environmental factor within a preset historical length and the attenuation coefficient determined according to the hydrological flow velocity data of the same period, the time cumulative effect characteristics of the environmental factor are obtained by exponential attenuation weighted calculation. By deploying online monitoring stations in the target water area or accessing historical databases from environmental protection departments, daily water temperature, pH, dissolved oxygen, total phosphorus, total nitrogen, ammonia nitrogen, and chlorophyll a concentration data for the past three consecutive years are obtained, along with daily temperature, light intensity, wind speed, and rainfall data from meteorological departments for the same period. The preset historical period is set to the past 30 days. For water areas with faster hydrological flow, a larger attenuation coefficient value is set, such as 0.9; for still water areas with slower flow, a smaller attenuation coefficient value is set, such as 0.3. Taking total phosphorus as an example, the total phosphorus concentration on day t is multiplied by 1, the total phosphorus concentration on day t-1 is multiplied by the attenuation coefficient, the total phosphorus concentration on day t-2 is multiplied by the square of the attenuation coefficient, and so on, up to day t-30. Figure 1 As shown, the summation of the 31 product results yields the time-cumulative effect characteristic value of total phosphorus on day t. This calculation is repeated for all environmental factors. A microcurrent electrode module, an ultrasonic transducer module, and a multi-parameter water quality probe, including but not limited to a chlorophyll a fluorescence sensor, dissolved oxygen sensor, pH sensor, and water depth sensor, are mounted on the guide rail. In one embodiment, the attenuation coefficient is calculated as the ratio of the calculated flow velocity to a preset value. The sum of this ratio and the base attenuation coefficient is used as the attenuation coefficient. The preset value is a fixed value or, for example, the monitoring radius around the buoy, or a characteristic scale of the target water area.

[0017] The model building module is used to build a random forest prediction model based on the time cumulative effect features. When splitting nodes in each decision tree of the model, the information gain of the candidate split features is multiplied by a preset power of the Pearson correlation coefficient of the feature with the target algae concentration time series to obtain a modified split evaluation index, and the feature with the largest index value is selected for splitting. The calculated time-cumulative effect features of all environmental factors are used as the input feature set X, and the corresponding historical chlorophyll a concentration is used as the target value Y. A random forest model containing 500 decision trees is trained using this dataset. When a node in a decision tree needs to be split, for each candidate feature, such as the total phosphorus time-cumulative effect feature, the information gain after splitting the dataset is calculated. Furthermore, the Pearson correlation coefficient between the historical data sequence of the total phosphorus time-cumulative effect feature and the historical data sequence of chlorophyll a concentration is calculated, for example, 0.8. Setting the preset power to 2, 0.8 squared is calculated, resulting in 0.64. The information gain is multiplied by 0.64 to obtain the corrected split evaluation index for this feature. This process is repeated for all candidate features, and the feature with the largest corrected split evaluation index is selected as the splitting criterion for the current node. Traditional random forests, when splitting at a specific node, may select a noisy feature that is cleanly split in local data but has no global correlation with algal blooms. This can lead to model overfitting and poor generalization ability to new data. This invention adds a global credibility weight to features by multiplying them by the Pearson coefficient. Even if a feature has a high information gain at the current node, its score will be lowered if it has no global correlation with algae concentration.

[0018] In one embodiment, the maximum mutual information coefficient (MIC) between all candidate environmental features, such as water temperature and light intensity, and the target algae concentration time series is calculated. When constructing each decision tree in the random forest and performing node splitting search, the traditional information gain based on the current node sample is calculated for each candidate feature. This gain value is then weighted with the pre-calculated MIC using a weighted formula, such as information gain × (1 + λ × MIC), to generate a new evaluation index, where λ is the adjustment weight. The feature with the highest comprehensive score and the splitting point are selected for node splitting. This approach utilizes global nonlinear correlation as a priori guidance to accelerate convergence while preserving the local purity optimization mechanism of the decision tree core. In one embodiment, different hierarchical models are constructed for algae concentrations at different depths.

[0019] The prediction module is used to input the data to be predicted, perform preliminary prediction using the random forest prediction model, and determine the algal bloom risk level based on the preliminary prediction results. Based on the algal bloom risk level, decision trees with historical prediction performance better than a preset threshold are selected from the random forest prediction model to form a subset of expert decision trees. The prediction results of each decision tree in the subset of expert decision trees are weighted and averaged according to their respective historical prediction performance to obtain the predicted algal concentration value. Acquire weather forecast data for the next day or several days and current water quality monitoring data, calculate the time cumulative effect characteristics of the data to be predicted, and input them into the constructed random forest model. Optionally, acquire real-time monitoring data at the current moment, and combine it with historical sliding window data, calculating the time cumulative effect characteristics in real time according to the same algorithm described in the data acquisition module. Each of the 500 decision trees in the model outputs a predicted value, and their arithmetic mean is taken as the preliminary prediction result, for example, a predicted chlorophyll a concentration of 60 micrograms per liter. Based on a preset concentration threshold, if it is greater than 50 micrograms per liter, it is considered high risk, and the current risk level is determined to be high risk. Subsequently, the root mean square error of historical predictions for all 500 decision trees is evaluated based on the validation set data, and the error threshold for the high risk level is set to 10 micrograms per liter. Decision trees with a historical root mean square error of less than 10 micrograms per liter are selected, for example, 82 trees are selected to form a subset of expert decision trees, see [link to relevant documentation]. Figure 2 For these 82 expert decision trees, the reciprocal of the prediction error of each tree is used as its weight. The predicted value of each tree is multiplied by its corresponding weight, summed, and then divided by the sum of all weights to obtain the predicted value of algae concentration.

[0020] In one embodiment, to avoid feedback bias in model structure selection based on prediction results, the algal bloom risk level is used for expert decision tree selection using a risk level mapping rule based on historical real algal concentration annotations. Specifically, the risk level of each historical sample in the validation dataset is obtained by comparing its corresponding real chlorophyll a concentration with a preset risk threshold. The current sample to be predicted is only used to determine which pre-built subset of expert decision trees should be invoked, and does not participate in the performance evaluation and selection process of that subset, thereby avoiding a circular dependency between model prediction results and model structure selection.

[0021] The control module is used to determine the prediction confidence level based on the predicted algae concentration value and the dispersion of each predicted value in the subset of the expert decision tree, and to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies by combining the predicted algae concentration value and the prediction confidence level.

[0022] The standard deviation of the predicted values ​​from the 82 expert decision trees selected in the previous step is calculated and used as the dispersion. If the standard deviation is less than 5, the prediction confidence is high; if the standard deviation is between 5 and 15, the confidence is medium; and if the standard deviation is greater than 15, the confidence is low. The final algae concentration prediction value and the prediction confidence are combined, and a strategy library is queried for matching. For example, if the final predicted value is 62 micrograms per liter and the confidence is high, an emergency strategy is matched, and the system outputs the instruction: immediately start the aeration and oxygenation equipment in the water area, prepare to add environmentally friendly algaecide, and send an early warning notification to the downstream water plant. As another example, if the final predicted value is 40 micrograms per liter and the confidence is high, an early warning strategy is matched, and the system outputs the instruction: increase the water quality sampling frequency from once a day to once every four hours, and check the readiness status of the physical algae removal vessel. If the final predicted value is 15 micrograms per liter, regardless of the confidence level, a monitoring strategy is matched, and the system outputs the instruction: maintain the routine monitoring plan; no intervention is required. In one embodiment, the determination of the prediction confidence also incorporates a similarity index between the current input feature vector and the training data distribution. When the average distance between the current input feature and the training samples in the feature space is greater than a preset threshold, the prediction confidence will be downgraded by one level, even if the standard deviation of the predicted values ​​of the expert decision tree subset is small, to reduce the risk of overconfidence of the model under out-of-distribution sample conditions.

[0023] In an optional embodiment, the formula for calculating the modified splitting evaluation index M is: Where G is the information gain of the candidate splitting feature. denoted as Pearson correlation coefficient between this feature and the time series of target algae concentration, where p is a preset positive power exponent.

[0024] Information gain G reflects the local discriminative power of features, while the Pearson correlation coefficient... The absolute value of the correlation coefficient represents the strength of the long-term linear correlation between the feature and changes in algae concentration. It also helps the model prioritize features with strong trend consistency with the target variable when the feature dimension is high or noise is high, thereby accelerating convergence and reducing overfitting. By multiplying the two, features that have both good local splitting effects and long-term correlation with the target variable can be prioritized. When constructing a node in the decision tree, assume the candidate splitting feature is total phosphorus concentration in the water. The information gain G of total phosphorus concentration is calculated based on the data of the current node, for example, it is 0.42. Using all historical data in the training set, the Pearson correlation coefficient between the time series of total phosphorus concentration and the time series of algae concentration is calculated. For example, 0.7. The positive power exponent p is used to adjust the weight of the correlation, for example, 2. The modified split evaluation index is calculated to be 0.2058. This process is repeated for all candidate features, and the feature that maximizes the M value is selected as the splitting criterion for that node.

[0025] In some embodiments, the step of selecting decision trees whose historical prediction performance is better than a preset threshold specifically involves: selecting a subset of historical data corresponding to the algal bloom risk level from the validation dataset, using the subset of historical data to evaluate the root mean square error of each decision tree in the random forest model, and selecting decision trees whose root mean square error is less than a preset error threshold as the subset of expert decision trees.

[0026] Each decision tree in a random forest may have varying predictive abilities for different scenarios; some trees may be more accurate in predicting high-risk algal blooms, while others perform better in low-risk scenarios. In this way, the most competent set of decision trees can be combined for prediction tasks at specific risk levels, thereby improving prediction accuracy. Specifically, based on the algal bloom risk level of the current data to be predicted, such as medium risk, all historically labeled medium-risk data are selected from an independent validation dataset to obtain a subset specifically for evaluation. Each decision tree in the random forest model is traversed, tested using the aforementioned medium-risk data subset, and its root mean square error (RMSE) is calculated. For example, if the model has 100 trees, the RMSE value for each tree is calculated, such as tree 1 having an RMSE of 15.3, tree 2 having an RMSE of 8.9, etc. With a preset error threshold, such as 10.0, all decision trees with an RMSE less than 10.0 are selected. These trees, including tree 2 and several others, form a subset of expert decision trees for medium-risk scenarios.

[0027] The random forest model consists of a large number of decision trees, such as 100, with a maximum depth of 15 for each tree. Each decision tree is a tree-like decision structure, where internal nodes represent tests on a specific input feature, such as water temperature or total phosphorus concentration, branches represent test results, and leaf nodes provide a specific predicted algae concentration value. The training set consists of historical monitoring time-series data of the target water area. Each data point contains a set of features and corresponding actual algae concentration values. Features include water temperature, pH, dissolved oxygen, total nitrogen, total phosphorus, and calculated time-cumulative effect features, etc. For example, daily monitoring data from the past three years can be used as the training set. During training, the model randomly selects a subset with replacement from the total training set for each decision tree to train itself. When splitting at each node, only a subset of features is randomly selected from all features to find the optimal split point. The splitting is based on the aforementioned modified split evaluation index M, with the goal of minimizing the mean squared error of the prediction. The input to the model is a feature vector containing current environmental parameters, such as a current water temperature of 25 degrees Celsius and a total phosphorus concentration of 0.05 mg / L. The model outputs a predicted value of algae concentration at a specific future time point. This value is obtained by averaging the predicted values ​​of all trees after passing the input vector to each tree in the forest. For example, it is 55.4 micrograms per liter.

[0028] In some embodiments, the weighted average is the weight of the j-th decision tree in the subset of expert decision trees. The calculation formula is: ; in, Let be the root mean square error of the j-th decision tree in its historical prediction performance, and N be the total number of decision trees in the expert decision tree subset.

[0029] Decision trees with smaller prediction errors should have greater weight. By using the reciprocal of the root mean square error (RMSE) as the basic metric for weights, the predictive performance of each tree can be effectively quantified; the smaller the error, the larger the reciprocal, and the higher the corresponding weight. This approach ensures that decision trees performing better in specific risk scenarios contribute more to the final result, thereby further improving the accuracy and robustness of the ensemble prediction. For example, suppose the subset of expert decision trees selected in the previous step contains three decision trees, and their RMSEs on the corresponding risk level subsets of historical data are respectively... , , Calculate the reciprocal of the error for each tree, resulting in 1 / 5.0 = 0.2, 1 / 8.0 = 0.125, and 1 / 4.0 = 0.25. Sum all the reciprocals, which is 0.575. Divide the reciprocal of each tree's error by this sum to obtain their respective weights: 0.348, 0.217, and 0.435. The predicted algae concentration is the sum of the predicted values ​​of the three trees multiplied by their corresponding weights.

[0030] In some embodiments, determining the prediction confidence based on the dispersion of the final algae concentration prediction value and each prediction value of the expert decision tree subset specifically involves: Calculate the standard deviation of each predicted value in the subset of the expert decision tree, and compare the standard deviation with a preset first confidence threshold and a second confidence threshold to determine whether the prediction confidence is high, medium or low.

[0031] In ensemble learning, the consistency of predictions from various base learners is a crucial indicator of model confidence. If most trees in a subset of expert decision trees provide very similar predictions, their standard deviations will be small, indicating high consensus and confidence in the predictions. Conversely, if the predictions are widely distributed and have large standard deviations, it suggests significant internal disagreement within the model, resulting in lower reliability of the predictions. This process involves obtaining the predictions from each tree in the expert decision tree subset for the current input sample, resulting in a set of predictions. For example, a subset containing five expert trees might provide predictions of 52.5, 53.1, 51.9, 53.5, and 52.8 micrograms per liter. Calculating the standard deviation of this set yields approximately 0.6 micrograms per liter. Two pre-set thresholds are used, such as a first confidence threshold of 1.0 and a second confidence threshold of 3.0. Since the calculated standard deviation of 0.6 is less than the first threshold of 1.0, the confidence level of this prediction is considered high. If the standard deviation is between 1.0 and 3.0, the confidence level is medium; if it is greater than 3.0, the confidence level is low.

[0032] In some embodiments, the step of combining the predicted algae concentration value with the prediction confidence level to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring," "early warning," and "emergency" strategies specifically involves: The predicted algae concentration is compared with the preset warning concentration threshold and emergency concentration threshold, and combined with the prediction confidence level, the corresponding algae suppression strategy is matched from the hierarchical algae suppression strategy library and output.

[0033] See Figure 3By establishing a hierarchical algae suppression strategy library, this system can provide differentiated responses based on the predicted severity of algal blooms and the reliability of the prediction results. The dual judgment mechanism combining predicted values ​​and confidence levels avoids the risks associated with decisions based solely on a single predicted value. For example, the system's preset warning concentration threshold is 70 micrograms per liter, and the emergency concentration threshold is 100 micrograms per liter. The hierarchical algae suppression strategy library stores rules similar to the following: if the predicted value is below 70, execute the regular monitoring strategy; if the predicted value is between 70 and 100 and the confidence level is high, initiate preventative measures such as aeration; if the predicted value is between 70 and 100 but the confidence level is low, recommend intensive manual sampling for verification; if the predicted value is above 100 and the confidence level is high or medium, activate the emergency plan and prepare to release algaecides. Assume the model's final output algae concentration prediction is 85 micrograms per liter with a high confidence level. The results are matched with the strategy library. If the predicted value is between the warning and emergency thresholds and has a high confidence level, the corresponding algae suppression strategy is matched and output: start oxygenation and aeration and increase ecological diversion.

[0034] In some embodiments, the step of combining the predicted algae concentration value with the prediction confidence level to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies specifically involves: if the predicted value is lower than the early warning concentration threshold, the "monitoring" strategy is directly matched; if the predicted value is between the early warning concentration threshold and the emergency concentration threshold, or if the predicted value is higher than the emergency concentration threshold but is judged to have low confidence due to the large dispersion of the expert decision tree subset prediction, the "early warning" strategy is matched to reduce the system's false alarm rate; the "emergency" strategy is matched only when the predicted value is higher than the emergency concentration threshold and the prediction confidence level is medium or high.

[0035] If the strategy is determined to be a monitoring strategy, groups of interconnected solar photovoltaic panels are used to block sunlight and suppress algae, while surface aeration is intermittently activated to prevent oxygen depletion in the water. If the strategy is determined to be an early warning strategy, such as for medium or high concentrations with low confidence, a time-series coordinated mode is activated, executing "ultrasonic pulse disturbance for 30 seconds each time, with a 5-minute interval + continuous microcurrent algae suppression" to disperse the algae population and inhibit its recovery. If the strategy is determined to be an emergency strategy, with high concentrations and high confidence, stratified algae suppression is implemented based on vertical monitoring data. For surface algae, a 20-40kHz ultrasonic transducer is activated and aeration is enhanced. For algae in the middle and lower layers, the equipment is lowered to the target depth via a vertical guide rail, and deep coordinated removal is carried out using a 5-20mA microcurrent and a 10-20kHz low-frequency ultrasonic wave. At the same time, it supports forced switching between microcurrent or ultrasonic single mode based on manually input specific algae species classification results.

[0036] Second embodiment A three-dimensional adaptive synergistic control method for algae suppression is proposed, the method comprising the following steps: Historical water quality and meteorological data of the target water area are obtained. Based on the data sequence of each environmental factor within a preset historical length and the attenuation coefficient determined according to the hydrological flow velocity data of the same period, the time cumulative effect characteristics of the environmental factor are obtained by exponential attenuation weighted calculation. A random forest prediction model is constructed based on the time cumulative effect characteristics. When splitting nodes in each decision tree of the model, the information gain of the candidate splitting feature is multiplied by a preset power of the Pearson correlation coefficient of the feature with the target algae concentration time series to obtain a modified splitting evaluation index, and the feature with the largest index value is selected for splitting. Input the data to be predicted, perform preliminary prediction using the random forest prediction model, and determine the algal bloom risk level based on the preliminary prediction results; based on the algal bloom risk level, select decision trees from the random forest prediction model whose historical prediction performance is better than a preset threshold to form an expert decision tree subset; perform a weighted average of the prediction results of each decision tree in the expert decision tree subset based on their respective historical prediction performance to obtain the algal concentration prediction value. The prediction confidence level is determined based on the predicted algae concentration and the dispersion of each predicted value in the subset of the expert decision tree. Then, the algae suppression strategy is matched and output from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies.

[0037] In some embodiments, the step of selecting decision trees whose historical prediction performance is better than a preset threshold specifically involves: selecting a subset of historical data corresponding to the algal bloom risk level from the validation dataset, using the subset of historical data to evaluate the root mean square error of each decision tree in the random forest model, and selecting decision trees whose root mean square error is less than a preset error threshold as the subset of expert decision trees.

[0038] In some embodiments, the weight of the j-th decision tree in the expert decision tree subset is used in the weighted average. The calculation formula is: ; in, Let be the root mean square error of the j-th decision tree in its historical prediction performance, and N be the total number of decision trees in the expert decision tree subset.

[0039] In some embodiments, determining the prediction confidence based on the dispersion of the final algae concentration prediction value and each prediction value of the expert decision tree subset specifically involves: Calculate the standard deviation of each predicted value in the subset of the expert decision tree, and compare the standard deviation with a preset first confidence threshold and a second confidence threshold to determine whether the prediction confidence is high, medium or low.

[0040] In some embodiments, the step of combining the predicted algae concentration value with the prediction confidence level to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring," "early warning," and "emergency" strategies specifically involves: The predicted algae concentration is compared with the preset warning concentration threshold and emergency concentration threshold, and combined with the prediction confidence level, the corresponding algae suppression strategy is matched from the hierarchical algae suppression strategy library and output.

[0041] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0042] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A three-dimensional adaptive algae-suppressing buoy system, comprising a solar photovoltaic panel, an energy storage module, a central control unit, and a vertical guide rail, wherein multiple sensors are arranged on the vertical guide rail, characterized in that, The system also includes the following modules: The data acquisition module is used to acquire historical water quality and meteorological data of the target water area. Based on the data sequence of each environmental factor within a preset historical length and the attenuation coefficient determined according to the hydrological flow velocity data of the same period, the time cumulative effect characteristics of the environmental factor are obtained by exponential attenuation weighted calculation. The model building module is used to build a random forest prediction model based on the time cumulative effect features. When splitting nodes in each decision tree of the model, the information gain of the candidate split features is multiplied by a preset power of the Pearson correlation coefficient of the feature with the target algae concentration time series to obtain a modified split evaluation index, and the feature with the largest index value is selected for splitting. The prediction module is used to input the data to be predicted, perform preliminary prediction using the random forest prediction model, and determine the algal bloom risk level based on the preliminary prediction results. Based on the algal bloom risk level, decision trees with historical prediction performance better than a preset threshold are selected from the random forest prediction model to form a subset of expert decision trees. The prediction results of each decision tree in the subset of expert decision trees are weighted and averaged according to their respective historical prediction performance to obtain the predicted algal concentration value. The control module is used to determine the prediction confidence level based on the predicted algae concentration value and the dispersion of each predicted value in the subset of the expert decision tree, and to match and output an algae suppression strategy from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies by combining the predicted algae concentration value and the prediction confidence level.

2. The system according to claim 1, characterized in that, The process of selecting decision trees whose historical prediction performance is better than a preset threshold is as follows: based on the algal bloom risk level, a subset of historical data corresponding to the risk level is selected from the validation dataset; the root mean square error of each decision tree in the random forest model is evaluated using the historical data subset; and decision trees whose root mean square error is less than a preset error threshold are selected as the subset of expert decision trees.

3. The system according to claim 1, characterized in that, In the weighted average, the weight of the j-th decision tree in the subset of the expert decision trees. The calculation formula is: ； in, Let be the root mean square error of the j-th decision tree in its historical prediction performance, and N be the total number of decision trees in the expert decision tree subset.

4. The system according to claim 1, characterized in that, The determination of prediction confidence based on the predicted algae concentration and the dispersion of each predicted value in the subset of the expert decision tree is specifically as follows: Calculate the standard deviation of each predicted value in the subset of the expert decision tree, and compare the standard deviation with a preset first confidence threshold and a second confidence threshold to determine whether the prediction confidence is high, medium or low.

5. The system according to claim 1, characterized in that, The process involves combining the predicted algae concentration with the prediction confidence level, matching and outputting an algae suppression strategy from a pre-defined tiered algae suppression strategy library that includes "monitoring," "early warning," and "emergency" strategies. Specifically: The predicted algae concentration is compared with the preset warning concentration threshold and emergency concentration threshold, and combined with the prediction confidence level, the corresponding algae suppression strategy is matched from the hierarchical algae suppression strategy library and output.

6. A three-dimensional adaptive algae suppression and synergistic control method, characterized in that, The method includes the following steps: Historical water quality and meteorological data of the target water area are obtained. Based on the data sequence of each environmental factor within a preset historical length and the attenuation coefficient determined according to the hydrological flow velocity data of the same period, the time cumulative effect characteristics of the environmental factor are obtained by exponential attenuation weighted calculation. A random forest prediction model is constructed based on the time cumulative effect characteristics. When splitting nodes in each decision tree of the model, the information gain of the candidate splitting feature is multiplied by a preset power of the Pearson correlation coefficient of the feature with the target algae concentration time series to obtain a modified splitting evaluation index, and the feature with the largest index value is selected for splitting. Input the data to be predicted, perform preliminary prediction using the random forest prediction model, and determine the algal bloom risk level based on the preliminary prediction results; based on the algal bloom risk level, select decision trees from the random forest prediction model whose historical prediction performance is better than a preset threshold to form an expert decision tree subset; perform a weighted average of the prediction results of each decision tree in the expert decision tree subset based on their respective historical prediction performance to obtain the algal concentration prediction value. The prediction confidence level is determined based on the predicted algae concentration and the dispersion of each predicted value in the subset of the expert decision tree. The algae suppression strategy is matched and output from a preset hierarchical algae suppression strategy library containing "monitoring", "early warning" and "emergency" strategies by combining the predicted algae concentration and the prediction confidence level.

7. The method according to claim 6, characterized in that, The process of selecting decision trees whose historical prediction performance is better than a preset threshold is as follows: based on the algal bloom risk level, a subset of historical data corresponding to the risk level is selected from the validation dataset; the root mean square error of each decision tree in the random forest model is evaluated using the historical data subset; and decision trees whose root mean square error is less than a preset error threshold are selected as the subset of expert decision trees.

8. The method according to claim 7, characterized in that, In the weighted average, the weight of the j-th decision tree in the subset of the expert decision trees. The calculation formula is: ； in, Let be the root mean square error of the j-th decision tree in its historical prediction performance, and N be the total number of decision trees in the expert decision tree subset.

9. The method according to claim 6, characterized in that, The determination of prediction confidence based on the predicted algae concentration and the dispersion of each predicted value in the subset of the expert decision tree is specifically as follows: Calculate the standard deviation of each predicted value in the subset of the expert decision tree, and compare the standard deviation with a preset first confidence threshold and a second confidence threshold to determine whether the prediction confidence is high, medium or low.

10. The method according to claim 6, characterized in that, The process involves combining the predicted algae concentration with the prediction confidence level, matching and outputting an algae suppression strategy from a pre-defined tiered algae suppression strategy library that includes "monitoring," "early warning," and "emergency" strategies. Specifically: The predicted algae concentration is compared with the preset warning concentration threshold and emergency concentration threshold, and combined with the prediction confidence level, the corresponding algae suppression strategy is matched from the hierarchical algae suppression strategy library and output.

Images