Data-driven ecological system optimization model construction method

Abstract

The invention relates to the technical field of model construction, in particular to a data-driven ecological system optimization model construction method, which comprises the following steps of: performing multi-dimensional cleaning and preprocessing on original monitoring data, filling missing values through a time sequence interpolation algorithm, and removing noise data by adopting a nonlinear filtering technology; key features are extracted based on principal component analysis, and a low-dimensional feature space is constructed in combination with environmental factor correlation; constructing a nitrogen and phosphorus interception performance prediction model by using an integrated learning framework in combination with multiple machine learning algorithms and a deep learning model; preferably selecting a plant type with high adaptability through a dynamic multi-objective optimization algorithm, and updating a plant selection strategy according to real-time environment parameters; the model calculation complexity is reduced through a distributed calculation and model compression technology; according to the method, technologies such as data cleaning, model integration, multi-objective optimization and distributed calculation are introduced, so that the design and optimization efficiency of the ecological ditch-pond system is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of model construction, and particularly relates to a data-driven ecosystem optimization model construction method. BACKGROUND

[0002] As a comprehensive ecological engineering measure, the ecological ditch-pond system is widely used in agricultural non-point source pollution control, urban stormwater management, and river ecological restoration. Through the synergistic effect of ditches and ponds, combined with the purification capacity of plants, soil and microorganisms, the system effectively intercepts and degrades pollutants in water, especially nitrogen (N) and phosphorus (P). Nitrogen and phosphorus are the main driving factors of water eutrophication, and their excessive discharge can cause serious damage to water ecosystems, such as massive algae reproduction, water hypoxia, and other problems. Therefore, the ecological ditch-pond system plays an important role in nitrogen and phosphorus pollution control. Traditional ecological ditch-pond system design mainly relies on empirical formulas and static models. Although these methods have achieved certain results in practice, they often show low prediction accuracy and poor adaptability when faced with complex and variable environmental conditions and pollutant types. In recent years, with the development of big data and artificial intelligence technology, data-driven methods have been gradually introduced into the design and optimization of ecological ditch-pond systems to improve their nitrogen and phosphorus interception performance and operational efficiency.

[0003] In the design of the ecological ditch-pond system, the selection of plant types is a key factor. Different plants have significant differences in nitrogen and phosphorus absorption and degradation capacity, so selecting the appropriate plant type can significantly improve the nitrogen and phosphorus interception performance of the system. Currently, the selection of plant types mainly relies on laboratory research and field trials. Although these methods can provide some reference, they have the following problems: ① Long experimental period: Laboratory research and field trials require a long time to obtain results, making it difficult to meet the rapid decision-making needs of actual engineering. High cost: Large-scale field trials require a large amount of manpower, material resources, and financial resources, increasing the cost of the project. Poor adaptability: The results of laboratory and field trials are often limited to specific environmental conditions and are difficult to generalize to other areas or different environmental conditions.

[0004] In recent years, some researchers have begun to use data-driven methods to predict the nitrogen and phosphorus interception performance of ecological ditch-pond systems and optimize the selection of plant types. These methods are mainly based on machine learning algorithms such as support vector machines (SVM), random forests (RF), and neural networks (NN), which analyze a large amount of historical data to establish prediction models to improve prediction accuracy and adaptability. It is worth noting that although data-driven methods have shown some potential in predicting the nitrogen and phosphorus interception performance of ecological ditch-pond systems and selecting plant types, existing technologies still have the following shortcomings and deficiencies: ① Incomplete, inconsistent, and noisy data: Due to monitoring equipment failure or human factors, there may be missing values in historical data, affecting model training and prediction accuracy. Data collected in different regions or at different times may differ, making it difficult for the model to adapt to different environmental conditions. Environmental monitoring data often contain a large amount of noise, affecting the stability and prediction accuracy of the model. ② Model generalization needs to be improved: Existing data-driven models perform well on training data, but when faced with new environmental conditions or pollutant types, they often show poor generalization ability, which manifests as: model overfitting, performing well on training data but poorly on test data, making it difficult to adapt to new environmental conditions. Or the model has poor adaptability, making it difficult to adapt to different environmental conditions and pollutant types, resulting in decreased prediction accuracy. ③ Inaccurate plant type selection: existing methods have a single selection criterion, mainly based on the plant's ability to absorb nitrogen and phosphorus, ignoring the plant's adaptability to environmental conditions and growth characteristics. Existing methods lack dynamic optimization, mainly based on static data for plant type selection, making it difficult to dynamically optimize according to changes in environmental conditions. In addition, existing models have high computational complexity, requiring a large amount of computing resources during model training and prediction, resulting in a long time for model training and prediction, making it difficult to meet the rapid decision-making needs of actual engineering. SUMMARY

[0005] The present application introduces data cleaning, model integration, multi-objective optimization, and distributed computing technologies to provide a data-driven ecological system optimization model construction method, effectively solving the problems of low data quality, poor model generalization, inaccurate plant type selection, and high computational complexity in existing technologies, improving the design and optimization efficiency of ecological ditch-pond systems, and having important theoretical and practical significance.

[0006] The object of the present application can be achieved by the following technical solutions:

[0007] A data-driven ecological system optimization model construction method, comprising the following steps:

[0008] Step 1: Multi-dimensional cleaning and preprocessing of original monitoring data, filling missing values through time series interpolation algorithm, and removing noise data using nonlinear filtering technology; extracting key features based on principal component analysis, and constructing low-dimensional feature space combined with environmental factor correlation;

[0009] Step 2: Using an integrated learning framework combined with multiple machine learning algorithms and deep learning models to build a nitrogen and phosphorus interception performance prediction model;

[0010] Step 3: Optimize the plant type with strong adaptability through dynamic multi-objective optimization algorithm, and update the plant selection strategy according to real-time environmental parameters;

[0011] Step 4: Reduce model calculation complexity through distributed computing and model compression technology.

[0012] As a further scheme of the present application: in step 1, data cleaning includes the following:

[0013] Missing value processing: using linear interpolation method to fill in missing values, with an interval of 1 hour.

[0014] Outlier processing: using the 3σ principle to identify outliers, i.e. data points deviating from the mean by more than 3 times the standard deviation are considered outliers and are removed.

[0015] As a further scheme of the present application: in step 1, preprocessing includes the following:

[0016] Normalization range: normalize the nitrogen and phosphorus concentration, flow rate, and temperature parameters to the range [0, 1].

[0017] As a further scheme of the present application: in step 1, constructing a low-dimensional feature space includes:

[0018] Feature extraction: extract key features such as nitrogen and phosphorus interception rate, plant growth characteristics, and environmental factor correlation.

[0019] Feature dimension reduction: use principal component analysis (PCA) to retain 95% of the variance information, reducing the feature dimension from 50 to 10.

[0020] As a further scheme of the present application: in step 2, the deep learning model includes random forest, gradient boosting tree, support vector machine, K-nearest neighbor, multilayer perceptron, long short-term memory network, convolutional neural network, and extreme gradient boosting.

[0021] As a further scheme of the present application: in step 2, constructing a nitrogen and phosphorus interception performance prediction model includes:

[0022] Hyperparameter tuning: Use Bayesian optimization, grid search and random search to search learning rate, range: 0.001-0.1, tree depth, range: 5-20, regularization coefficient, range: 0.01-1;

[0023] Regularization technique: Introduce L2 regularization, regularization coefficient is 0.01;

[0024] Integrated learning optimization: Through Bagging and Boosting technology, improve the generalization ability and prediction accuracy of the model.

[0025] As a further scheme of the present application: in step 3, the multi-objective optimization includes the following steps:

[0026] Optimization goal: maximize nitrogen and phosphorus absorption capacity, minimize environmental adaptability score, maximize growth characteristics score;

[0027] Optimization algorithm: genetic algorithm, population size 100, iteration number 500, crossover probability 0.8, mutation probability 0.1.

[0028] As a further scheme of the present application: in step 3, the process of optimizing plant types with strong adaptability through dynamic multi-objective optimization algorithm is:

[0029] Real-time data update: update environmental parameters every 1 month, including nitrogen and phosphorus concentration, temperature, rainfall, and recalculate plant optimization scheme;

[0030] Feedback mechanism: retrain the model every month to update the model parameters.

[0031] As a further scheme of the present application: in step 4, distributed computing includes the following:

[0032] Framework selection: Apache Spark, cluster size 10 nodes, each node configuration 8 core CPU and 16GB memory;

[0033] Task allocation: divide the data set into 10 subsets and distribute them to 10 nodes for parallel processing. As a further scheme of the present application: in step 4, model compression includes the following:

[0034] Pruning technique: remove neurons with absolute weight less than 0.01 in the model, compression rate about 30%.

[0035] Quantization technology: convert floating-point parameters to 8-bit integers, model size reduced to 1 / 4 of the original.

[0036] Knowledge distillation: use teacher model, hidden unit number 128, guide student model, hidden unit number 64, distillation temperature 2.

[0037] Advantages of the present application:

[0038] The present application provides a data-driven ecosystem optimization model construction method, specifically: a data-driven ecological ditch-pond nitrogen and phosphorus interception performance prediction and plant optimization model, which has the following advantages:

[0039] High data quality: the present application improves the quality of historical data by introducing data cleaning and preprocessing techniques, including: data cleaning, which removes missing values and noise data to improve data integrity and consistency. And data standardization, standardizing the data collected in different regions or at different times to improve data comparability.

[0040] High model generalization ability: model integration is performed to improve the stability and prediction accuracy of the model. Regularization terms are introduced to prevent model overfitting and improve model generalization ability.

[0041] Accurate plant type selection: multi-objective optimization algorithm is introduced to accurately select plant types, considering the plant's nitrogen and phosphorus absorption capacity, environmental condition adaptability and growth characteristics, and performing multi-objective optimization. According to the change of environmental conditions, the plant type selection strategy is dynamically adjusted to improve the accuracy of selection.

[0042] Low computational complexity: distributed computing and model compression techniques are introduced to reduce computational complexity, based on distributed computing technology to improve model training and prediction efficiency and reduce computation time. Model compression techniques are used to reduce the computational resource consumption of the model and reduce project costs. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0044] Figure 1 is a flowchart of a data-driven ecosystem optimization model construction method provided by the present application embodiment one;

[0045] Figure 2 is a structural schematic diagram of a data-driven ecosystem optimization model construction system provided by the present application embodiment two. DETAILED DESCRIPTION

[0046] In order to make the person skilled in the art better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the protection scope of the present application.

[0047] Embodiment one

[0048] As Figure 1 shown, the preferred model construction method of the data-driven ecosystem provided by the embodiments of the present application specifically includes the following steps:

[0049] Step 1: Multi-dimensional cleaning and preprocessing of original monitoring data, filling in missing values through time series interpolation algorithm, and removing noise data by using nonlinear filtering technology; extracting key features based on principal component analysis, and constructing low-dimensional feature space combined with environmental factor correlation;

[0050] In step 1, first specifically, data cleaning includes the following contents:

[0051] Missing value processing: using linear interpolation method to fill in missing values, and the interpolation interval is 1 hour.

[0052] Outlier processing: using 3σ principle to identify outliers, that is, when the deviation of data points from the mean value exceeds 3 times the standard deviation, it is considered as an outlier and is removed.

[0053] Secondly, the preprocessing includes the following contents:

[0054] Normalization range: normalizing the nitrogen and phosphorus concentration, flow rate, temperature and other parameters to the range of [0, 1], and the formula is Xnorm=(X-Xmin) / (Xmax-Xmin);

[0055] Standardization processing: the data mean is 0 and the variance is 1, and the formula is Xstd=(X-μ) / σ;

[0056] Feature engineering:

[0057] Feature extraction: extracting key features such as nitrogen and phosphorus interception rate, plant growth characteristics, and environmental factor correlation.

[0058] Feature dimension reduction: using principal component analysis (PCA) to retain 95% of the variance information, and reducing the feature dimension from 50 to 10.

[0059] Step 2: Using an integrated learning framework combined with multiple machine learning algorithms and deep learning models to construct a nitrogen and phosphorus interception performance prediction model;

[0060] In step 2, first specifically, model selection: the following 8 models are selected for calculation:

[0061] Random Forest (RF): number of trees: 100, maximum depth: 10;

[0062] Gradient Boosting Decision Tree (GBDT): learning rate: 0.1, number of trees: 200;

[0063] Support Vector Machine (SVM): kernel function: Radial Basis Function (RBF), regularization parameter C: 1.0;

[0064] K-Nearest Neighbors (KNN): number of neighbors: 5, distance metric: Euclidean distance;

[0065] Multilayer Perceptron (MLP): hidden layer structure: [64, 32], activation function: ReLU;

[0066] Long Short-Term Memory (LSTM): number of hidden layer units: 128, time step: 24 hours, learning rate: 0.001,

[0067] Convolutional Neural Network (CNN): convolutional layer: 2 layers, 32 filters per layer; pooling layer: max pooling, fully connected layer: 64 units;

[0068] Extreme Gradient Boosting (XGBoost): learning rate: 0.1, number of trees: 200, maximum depth: 6;

[0069] Second, specifically, optimization method: the following 3 optimization methods are used for hyperparameter tuning:

[0070] Bayesian Optimization: search space: learning rate (0.001-0.1), tree depth (5-20), regularization coefficient (0.01-1). Number of iterations: 50;

[0071] Grid Search: search range: learning rate (0.001, 0.01, 0.1), tree depth (5, 10, 15, 20), regularization coefficient (0.01, 0.1, 1).

[0072] Random Search: Number of searches: 50, search range: learning rate (0.001-0.1), tree depth (5-20), regularization coefficient (0.01-1).

[0073] Thirdly, cross-validation: 5-fold, 7-fold and 10-fold cross-validation are used to evaluate the model performance respectively:

[0074] 5-fold cross-validation: The dataset is divided into 5 subsets, and one subset is used as the validation set and the remaining 4 subsets are used as the training set.

[0075] 7-fold cross-validation: The dataset is divided into 7 subsets, and one subset is used as the validation set and the remaining 6 subsets are used as the training set.

[0076] 10-fold cross-validation: The dataset is divided into 10 subsets, and one subset is used as the validation set and the remaining 9 subsets are used as the training set.

[0077] Fourthly, the impact of parameter changes on the model:

[0078] Through experimental analysis, the impact of different parameter changes on the model performance is analyzed:

[0079] Number of trees (RF, GBDT, XGBoost): Increasing the number of trees can improve the model's expression ability, but may lead to overfitting and increased computational complexity. Experimental results show that when the number of trees is 100-200, the model performance is optimal.

[0080] Maximum depth (RF, GBDT, XGBoost): Increasing the maximum depth can improve the model's fitting ability, but may lead to overfitting. Experimental results show that when the maximum depth is 6-10, the model performance is optimal.

[0081] Learning rate (GBDT, XGBoost): Too high learning rate may lead to model instability, and too low learning rate may lead to slow convergence speed. Experimental results show that when the learning rate is 0.1, the model performance is optimal.

[0082] Regularization coefficient (SVM, MLP, LSTM, CNN): Too small regularization coefficient may lead to overfitting, and too large regularization coefficient may lead to underfitting. Experimental results show that when the regularization coefficient is 0.01-0.1, the model performance is optimal.

[0083] Number of hidden layer units (MLP, LSTM, CNN): Increasing the number of hidden layer units can improve the model's expression ability, but may lead to increased computational complexity. Experimental results show that when the number of hidden layer units is 64-128, the model performance is optimal.

[0084] Time steps (LSTM): Increasing time steps can capture longer temporal dependencies, but may lead to increased computational complexity. Experimental results show that the model performs best with a time step of 24 hours.

[0085] Number of neighbors (KNN): Increasing the number of neighbors can improve the robustness of the model, but may lead to blurred classification boundaries. Experimental results show that the model performs best with a neighbor number of 5.

[0086] Cross-validation folds: Increasing the number of cross-validation folds can improve the stability of the evaluation results, but may lead to increased computational complexity. Experimental results show that the results of 5-fold, 7-fold and 10-fold cross-validation are basically consistent, and the evaluation results of 10-fold cross-validation are the most stable.

[0087] Fifthly, model optimization:

[0088] Hyperparameter tuning: Use Bayesian optimization, grid search and random search to search for learning rate (range: 0.001-0.1), tree depth (range: 5-20) and regularization coefficient (range: 0.01-1).

[0089] Regularization techniques: Introduce L2 regularization with a regularization coefficient of 0.01.

[0090] Integrated learning optimization: Use Bagging and Boosting techniques to improve the generalization ability and prediction accuracy of the model.

[0091] Step 3: Optimize plant types with strong adaptability through dynamic multi-objective optimization algorithm, and update plant selection strategy according to real-time environmental parameters;

[0092] In step 3, first, multi-objective optimization:

[0093] Optimization objectives: Maximize nitrogen and phosphorus absorption capacity (unit: mg / m 2 / day), minimize environmental adaptability score (range: 0-10, 0 is optimal), maximize growth characteristics score (range: 0-10, 10 is optimal).

[0094] Optimization algorithm: Use genetic algorithm (GA) with a population size of 100, iteration number of 500, crossover probability of 0.8 and mutation probability of 0.1.

[0095] Secondly, dynamic optimization:

[0096] Real-time data update: Update environmental parameters (such as nitrogen and phosphorus concentration, temperature, rainfall) every 1 month and recalculate the plant optimization plan.

[0097] Feedback mechanism: Re-train the model every month to update the model parameters.

[0098] Step 4: Reducing model computational complexity through distributed computing and model compression techniques;

[0099] In Step 4, first specifically, distributed computing includes the following:

[0100] Framework selection: Apache Spark is adopted, with a cluster size of 10 nodes, each node configured with 8-core CPU and 16 GB memory.

[0101] Task allocation: The dataset is divided into 10 subsets and allocated to 10 nodes for parallel processing.

[0102] First specifically, model compression includes the following:

[0103] Pruning technique: Remove neurons with weights less than 0.01 in absolute value, compression rate about 30%.

[0104] Quantization technique: Convert floating-point parameters to 8-bit integers, model size reduced to 1 / 4 of the original.

[0105] Knowledge distillation: Use the teacher model (LSTM, hidden unit number 128) to guide the student model (LSTM, hidden unit number 64), distillation temperature is 2.

[0106] Example Two

[0107] As Figure 2 shown, the data-driven ecosystem preferred model construction system provided by the embodiment of the application specifically includes the following modules:

[0108] Data cleaning and preprocessing module: Multi-dimensional cleaning and preprocessing of original monitoring data, missing values are filled by time series interpolation algorithm, and nonlinear filtering technology is used to remove noise data; based on principal component analysis to extract key features, combined with environmental factor correlation to construct low-dimensional feature space;

[0109] In the data cleaning and preprocessing module, first specifically, data cleaning includes the following:

[0110] Missing value processing: Linear interpolation method is used to fill in missing values, with an interpolation interval of 1 hour.

[0111] Outlier processing: 3σ principle is used to identify outliers, i.e. data points deviating from the mean by more than 3 times the standard deviation are considered outliers and removed.

[0112] Secondly, preprocessing includes the following:

[0113] Normalization: Normalize nitrogen and phosphorus concentrations, flow rates, temperatures, etc. to the range [0, 1] using the formula Xnorm = (X - Xmin) / (Xmax - Xmin);

[0114] Standardization: Standardize data with mean 0 and variance 1 using the formula Xstd = (X - μ) / σ;

[0115] Feature Engineering:

[0116] Feature Extraction: Extract key features such as nitrogen and phosphorus interception rates, plant growth characteristics, and environmental factor correlations.

[0117] Feature Dimension Reduction: Use principal component analysis (PCA) to retain 95% of the variance information and reduce the feature dimension from 50 to 10.

[0118] Model Training and Optimization Module: Use an integrated learning framework combining multiple machine learning algorithms and deep learning models to build a nitrogen and phosphorus interception performance prediction model.

[0119] In the model training and optimization module, first, model selection: select the following 8 models for calculation:

[0120] Random Forest (RF): Number of trees: 100, maximum depth: 10;

[0121] Gradient Boosting Decision Tree (GBDT): Learning rate: 0.1, number of trees: 200;

[0122] Support Vector Machine (SVM): Kernel function: Radial Basis Function (RBF), regularization parameter C: 1.0;

[0123] K-Nearest Neighbors (KNN): Number of neighbors: 5, distance metric: Euclidean distance;

[0124] Multilayer Perceptron (MLP): Hidden layer structure: [64, 32], activation function: ReLU;

[0125] Long Short-Term Memory (LSTM): Number of hidden layer units: 128, time step: 24 hours, learning rate: 0.001,

[0126] Convolutional Neural Network (CNN): Convolutional layers: 2 layers, each with 32 filters; Pooling layers: Max pooling; Fully connected layers: 64 units;

[0127] Extreme Gradient Boosting (XGBoost): Learning rate: 0.1, Number of trees: 200, Maximum depth: 6;

[0128] Secondly, optimization methods: The following 3 optimization methods are used for hyperparameter tuning:

[0129] Bayesian Optimization: Search space: Learning rate (0.001-0.1), Tree depth (5-20), Regularization coefficient (0.01-1). Number of iterations: 50;

[0130] Grid Search: Search range: Learning rate (0.001, 0.01, 0.1), Tree depth (5, 10, 15, 20), Regularization coefficient (0.01, 0.1, 1).

[0131] Random Search: Number of searches: 50, Search range: Learning rate (0.001-0.1), Tree depth (5-20), Regularization coefficient (0.01-1).

[0132] Thirdly, cross-validation: 5-fold, 7-fold and 10-fold cross-validation are used to evaluate model performance:

[0133] 5-fold cross-validation: The dataset is divided into 5 subsets, and one subset is used as the validation set and the remaining 4 subsets are used as the training set.

[0134] 7-fold cross-validation: The dataset is divided into 7 subsets, and one subset is used as the validation set and the remaining 6 subsets are used as the training set.

[0135] 10-fold cross-validation: The dataset is divided into 10 subsets, and one subset is used as the validation set and the remaining 9 subsets are used as the training set.

[0136] Fourthly, the impact of parameter changes on the model:

[0137] Through experimental analysis of the impact of different parameter changes on model performance:

[0138] Number of trees (RF, GBDT, XGBoost): Increasing the number of trees can improve the expression ability of the model, but may lead to overfitting and increased computational complexity. Experimental results show that when the number of trees is 100-200, the model performance is optimal.

[0139] Maximum depth (RF, GBDT, XGBoost): Increasing the maximum depth can improve the fitting ability of the model, but may cause overfitting. Experimental results show that when the maximum depth is 6-10, the model performance is optimal.

[0140] Learning rate (GBDT, XGBoost): Too high learning rate may cause the model to be unstable, and too low learning rate may cause slow convergence. Experimental results show that when the learning rate is 0.1, the model performance is optimal.

[0141] Regularization coefficient (SVM, MLP, LSTM, CNN): Too small regularization coefficient may cause overfitting, and too large regularization coefficient may cause underfitting. Experimental results show that when the regularization coefficient is 0.01-0.1, the model performance is optimal.

[0142] Number of hidden layer units (MLP, LSTM, CNN): Increasing the number of hidden layer units can improve the expression ability of the model, but may cause the increase of computational complexity. Experimental results show that when the number of hidden layer units is 64-128, the model performance is optimal.

[0143] Time step (LSTM): Increasing the time step can capture longer time dependence, but may cause increased computational complexity. Experimental results show that when the time step is 24 hours, the model performance is optimal.

[0144] Number of neighbors (KNN): Increasing the number of neighbors can improve the robustness of the model, but may cause the classification boundary to be blurred. Experimental results show that when the number of neighbors is 5, the model performance is optimal.

[0145] Cross-validation folds: Increasing the number of cross-validation folds can improve the stability of the evaluation results, but may cause increased computational complexity. Experimental results show that the results of 5-fold, 7-fold and 10-fold cross-validation are basically consistent, and the evaluation results of 10-fold cross-validation are the most stable.

[0146] Fifthly, model optimization:

[0147] Hyperparameter tuning: Bayesian optimization, grid search and random search are used to search for learning rate (range: 0.001-0.1), tree depth (range: 5-20) and regularization coefficient (range: 0.01-1).

[0148] Regularization technique: L2 regularization is introduced, with a regularization coefficient of 0.01.

[0149] Ensemble learning optimization: Bagging and Boosting techniques are used to improve the generalization ability and prediction accuracy of the model.

[0150] Plant type optimization module: dynamically optimize plant types with strong adaptability and update plant selection strategy based on real-time environmental parameters;

[0151] In the plant type optimization module, first specifically, multi-objective optimization:

[0152] Optimization objectives: maximize nitrogen and phosphorus absorption capacity (unit: mg / m 2 / day), minimize environmental adaptability score (range: 0-10, 0 is optimal), maximize growth characteristics score (range: 0-10, 10 is optimal).

[0153] Optimization algorithm: genetic algorithm (GA) is adopted, population size is 100, iteration number is 500, crossover probability is 0.8, and mutation probability is 0.1.

[0154] Secondly, dynamic optimization:

[0155] Real-time data update: update environmental parameters (such as nitrogen and phosphorus concentration, temperature, rainfall) every 1 month and recalculate plant optimization scheme.

[0156] Feedback mechanism: retrain the model once a month and update the model parameters.

[0157] Distributed computing and model compression module: reduce model calculation complexity through distributed computing and model compression technology;

[0158] In the distributed computing and model compression module, first specifically, distributed computing includes the following:

[0159] Framework selection: Apache Spark is adopted, cluster size is 10 nodes, each node is configured with 8-core CPU and 16GB memory.

[0160] Task allocation: divide the data set into 10 subsets and distribute them to 10 nodes for parallel processing.

[0161] Firstly, model compression includes the following:

[0162] Pruning technology: remove neurons with absolute weight less than 0.01 in the model, compression rate is about 30%.

[0163] Quantization technology: convert floating-point parameters to 8-bit integers, model size is reduced to 1 / 4 of the original.

[0164] Knowledge distillation: use teacher model (LSTM, hidden unit number 128) to guide student model (LSTM, hidden unit number 64), distillation temperature is 2.

[0165] The above has been described in detail one embodiment of the present application, but the content is only the preferred embodiment of the present application, cannot be considered for limiting the scope of the present application. Any equivalent changes and improvements made in the scope of the present application, should still belong to the scope of the present application.

Claims

1. A data-driven method for constructing an ecosystem optimization model, characterized in that, Includes the following steps: Step 1: Perform multi-dimensional cleaning and preprocessing on the raw monitoring data, fill in missing values ​​using time series interpolation algorithms, and remove noise data using nonlinear filtering techniques; extract key features based on principal component analysis, and construct a low-dimensional feature space by combining the correlation of environmental factors; Step 2: Use an ensemble learning framework to combine multiple machine learning algorithms and deep learning models to build a nitrogen and phosphorus interception performance prediction model; Step 3: Select the most adaptable plant types using a dynamic multi-objective optimization algorithm, and update the plant selection strategy based on real-time environmental parameters; Step 4: Reduce the computational complexity of the model through distributed computing and model compression techniques.

2. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 1, data cleaning includes the following: Missing value handling: Missing values ​​are filled using linear interpolation with an interpolation interval of 1 hour. Outlier handling: Outliers are identified using the 3σ principle, which means that data points that deviate from the mean by more than 3 times the standard deviation are considered outliers and removed.

3. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 1, the preprocessing includes the following: Normalization range: Normalize nitrogen and phosphorus concentrations, flow rates, and temperature parameters to the range [0,1].

4. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 1, constructing the low-dimensional feature space includes: Feature extraction: Extract key features, such as nitrogen and phosphorus interception rate, plant growth characteristics, and correlation with environmental factors. Feature dimensionality reduction: Using principal component analysis (PCA), 95% of the variance information is retained, reducing the feature dimension from 50 dimensions to 10 dimensions.

5. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 2, the deep learning models include random forest, gradient boosting tree, support vector machine, K-nearest neighbors, multilayer perceptron, long short-term memory network, convolutional neural network, and extreme gradient boosting.

6. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 2, a nitrogen and phosphorus interception performance prediction model is constructed, including: Hyperparameter tuning: Bayesian optimization, grid search and random search were used. The search learning rate ranged from 0.001 to 0.1, the tree depth ranged from 5 to 20, and the regularization coefficient ranged from 0.01 to 1. Regularization technique: L2 regularization is introduced, with a regularization coefficient of 0.01; Ensemble learning optimization: Improve the model's generalization ability and prediction accuracy through Bagging and Boosting techniques.

7. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 3, multi-objective optimization includes the following steps: Optimization objectives: Maximize nitrogen and phosphorus uptake capacity, minimize environmental adaptability score, and maximize growth characteristic score; Optimization algorithm: Genetic algorithm is used, with a population size of 100, 500 iterations, a crossover probability of 0.8, and a mutation probability of 0.

1.

8. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 3, the process of selecting the most adaptable plant type using a dynamic multi-objective optimization algorithm is as follows: Real-time data updates: Environmental parameters, including nitrogen and phosphorus concentrations, temperature, and rainfall, are updated every month, and the optimal plant selection scheme is recalculated. Feedback mechanism: The model is retrained monthly to update the model parameters.

9. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 4, distributed computing includes the following: Framework selection: Apache Spark is used, with a cluster size of 10 nodes, each node configured with 8 CPU cores and 16GB of memory; Task allocation: Divide the dataset into 10 subsets and distribute them to 10 nodes for parallel processing.

10. The data-driven ecosystem optimization model construction method according to claim 1, characterized in that, In step 4, model compression includes the following: Pruning technique: Remove neurons in the model whose absolute weight is less than 0.01, with a compression rate of approximately 30%. Quantization technique: Convert floating-point parameters into 8-bit integers, reducing the model size to 1 / 4 of the original. Knowledge distillation: Using the teacher model with 128 hidden units, and the student model with 64 hidden units, the distillation temperature is 2.

Images