Pollutant prediction method and device based on evolutionary learning strategy, equipment and storage medium

Abstract

The application discloses an evolution learning strategy-based pollutant prediction method and device, equipment and a storage medium, and relates to the technical field of pollutant prediction. The method comprises the following steps: performing data anomaly processing on pollutant data to obtain cleaned data, performing feature conversion and time series data noise reduction on the cleaned data to obtain feature information, identifying the feature information through a target pollutant prediction model to obtain a pollutant prediction result. The evolution learning algorithm is used to dynamically adjust parameters through multi-model fusion optimization, and the target pollutant prediction model is constructed, thereby improving the prediction accuracy of the pollutant.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pollutant prediction, in particular to a pollutant prediction method and device based on an evolutionary learning strategy, equipment and a storage medium. BACKGROUND

[0002] Traditional pollutant prediction methods mainly rely on statistical models and simple machine learning algorithms. These methods perform well when dealing with relatively simple and linear data. However, as the amount of data increases and the complexity of pollutant sources increases, traditional methods gradually show limitations. Statistical models are difficult to effectively process different types of data, especially when data sources are extensive and diverse in form. At the same time, based on the linear assumption, the processing capacity for complex nonlinear relationships is weak, and when dealing with high-dimensional data, overfitting phenomenon is easy to occur, resulting in reduced generalization ability. Therefore, a method is needed to predict pollutants through the collaborative work of multiple models and adaptive optimization mechanism to improve the accuracy of pollutant prediction. SUMMARY

[0003] The main purpose of the present application is to provide a pollutant prediction method and device based on an evolutionary learning strategy, equipment and a storage medium, aiming to solve the technical problem of how to improve the accuracy of pollutant prediction.

[0004] To achieve the above-mentioned purpose, the present application provides a pollutant prediction method based on an evolutionary learning strategy, which comprises:

[0005] Obtaining pollutant data;

[0006] Performing data anomaly processing on the pollutant data to obtain cleaned data;

[0007] Performing feature conversion and time series data denoising on the cleaned data to obtain feature information;

[0008] Identifying the feature information through a target pollutant prediction model to obtain a pollutant prediction result, the prediction result including a pollutant data prediction result, a pollutant category prediction result and a prediction result of nonlinear, high-dimensional feature data of the pollutant;

[0009] Before the step of identifying the feature information through the target pollutant prediction model to obtain the pollutant prediction result, it comprises:

[0010] Optimizing a preset hybrid expert model according to a multi-level knowledge fusion strategy to obtain a preset pollutant prediction model, the preset hybrid expert model including a linear regression expert model, a support vector regression expert model, a random forest expert model and a deep neural network expert model;

[0011] The preset pollutant prediction model is optimized using an evolutionary learning algorithm strategy to obtain the target pollutant prediction model.

[0012] In one embodiment, the step of optimizing the preset hybrid expert model according to a multi-level knowledge fusion strategy to obtain a preset pollutant prediction model includes:

[0013] The task requirement for obtaining pollutant predictions;

[0014] Based on the task requirements, select the corresponding optimized fusion method from the multi-level knowledge fusion strategy. The optimized fusion method includes weighted average fusion, weighted voting fusion, and weighted projection fusion.

[0015] The preset hybrid expert model is fused according to the optimized fusion method to obtain the optimized preset hybrid expert model.

[0016] The optimized preset hybrid expert model is used as the preset pollutant prediction model.

[0017] In one embodiment, the step of selecting the corresponding optimized fusion method in the multi-level knowledge fusion strategy according to the task requirements includes:

[0018] When the task requirement is pollutant data prediction, the optimized fusion method is to prioritize weighted average fusion, as shown in the following formula:

[0019]

[0020] in, The predicted pollutant values ​​after fusion. For the first Pollutant predictions from an expert model. For the first The weights of each expert model, and satisfying , The total number of expert models;

[0021] When the task requirement is pollutant category prediction, the optimized fusion method is to prioritize weighted voting fusion, as shown in the following formula:

[0022]

[0023] in, For pollutant prediction categories, For the first Class category, The total number of pollutant categories. For the first Expert model prediction categories The probability, weights of the first

[0024] When the task requirement is prediction of nonlinear, high-dimensional characteristic data, the optimization fusion mode is to preferentially adopt weighted projection fusion, and the specific expression formula is:

[0025]

[0026] wherein, is the projected pollutant prediction value, is a high-dimensional vector, denotes a projection function, is an optimized output function.

[0027] In an embodiment, the step of optimizing the preset pollutant prediction model according to an evolutionary learning algorithm strategy to obtain a target pollutant prediction model comprises:

[0028] obtaining hyperparameters and model prediction accuracy corresponding to expert models of the preset hybrid expert model;

[0029] encoding the hyperparameters and weight parameters in the multi-level knowledge fusion strategy to obtain individual gene values;

[0030] performing evaluation based on the model prediction accuracy to construct a fitness function and obtain fitness values, the fitness function being in the form of:

[0031]

[0032] wherein, is an actual fitness value, is a predicted fitness value, is a data quantity;

[0033] optimizing the fitness values and the individual gene values according to an evolutionary operation strategy and a multi-objective optimization strategy to obtain optimized hyperparameters and optimized fusion weights;

[0034] updating the preset pollutant prediction model according to the optimized hyperparameters and the optimized fusion weights to obtain a target pollutant prediction model.

[0035] In an embodiment, the step of optimizing the fitness values and the individual gene values according to an evolutionary operation strategy and a multi-objective optimization strategy to obtain optimized hyperparameters and optimized fusion weights comprises:

[0036] selecting an individual corresponding to preset fitness and preset individual gene values by sorting the fitness values and the individual gene values;

[0037] ​The individual is processed by cross operation and mutation operation to obtain an updated individual;

[0038] The fitness value of the updated individual is optimized by a multi-objective optimization strategy to obtain a Pareto optimal solution set;

[0039] A preset solution is selected from the Pareto optimal solution as an optimized hyperparameter and an optimized fusion weight.

[0040] In an embodiment, the step of optimizing the fitness value of the updated individual by a multi-objective optimization strategy to obtain a Pareto optimal solution set comprises:

[0041] The fitness values of the updated individuals are sorted to obtain a plurality of crowded distance levels;

[0042] According to the filtering of the crowded distance levels, the fitness values of a plurality of updated individuals corresponding to a preset level are selected as Pareto optimal solutions;

[0043] The plurality of Pareto optimal solutions are combined to obtain a Pareto optimal solution set.

[0044] In an embodiment, after the step of identifying the feature information by a target pollutant prediction model to obtain a pollutant prediction result, the method further comprises:

[0045] Real-time monitoring and obtaining updated pollutant data;

[0046] The updated pollutant data is input into the target pollutant prediction model to obtain an evaluation performance;

[0047] When the evaluation performance does not meet a preset prediction requirement, adjusting the hyperparameters and fusion weight of the target pollutant prediction model until the evaluation performance meets the preset prediction requirement.

[0048] In addition, to achieve the above-mentioned purpose, the present application also proposes a pollutant prediction device based on an evolutionary learning strategy, which comprises:

[0049] An acquisition module is configured to acquire pollutant data;

[0050] A data cleaning module is configured to obtain cleaned data by filling missing values and detecting and correcting outliers in the pollutant data;

[0051] A data conversion module is configured to obtain feature information by performing feature conversion and time series data noise reduction on the cleaned data;

[0052] A result module is configured to identify the feature information through a target pollutant prediction model to obtain a pollutant prediction result, the prediction result including a pollutant data prediction result, a pollutant category prediction result and a prediction result of nonlinear and high-dimensional feature data of the pollutant.

[0053] A construction module is configured to optimize a preset hybrid expert model according to a multi-level knowledge fusion strategy to obtain a preset pollutant prediction model, the preset hybrid expert model including a linear regression expert model, a support vector regression expert model, a random forest expert model and a deep neural network expert model; and optimize the preset pollutant prediction model according to an evolutionary learning algorithm strategy to obtain a target pollutant prediction model.

[0054] In addition, to achieve the above object, the present application also provides a storage medium, which is a computer readable medium, and a computer program is stored on the storage medium, and the computer program is executed by a processor to implement the steps of the pollutant prediction method based on the evolutionary learning strategy.

[0055] In addition, to achieve the above object, the present application also provides a computer program product, which includes a computer program, and the computer program is executed by a processor to implement the steps of the pollutant prediction method based on the evolutionary learning strategy.

[0056] The present application obtains pollutant data, performs data anomaly processing on the pollutant data to obtain cleaned data, performs feature conversion and time series data noise reduction on the cleaned data to obtain feature information, identifies the feature information through a target pollutant prediction model to obtain a pollutant prediction result. The evolutionary learning algorithm is used to dynamically adjust parameters to optimize the multi-model fusion and construct a target pollutant prediction model, thereby improving the prediction accuracy of the pollutant. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced. Obviously, other drawings can also be obtained by those skilled in the art without creative labor.

[0058] Figure 1 Flowchart of the first embodiment of the pollutant prediction method based on the evolutionary learning strategy of the present application;

[0059] Figure 2 Flowchart of the second embodiment of the pollutant prediction method based on the evolutionary learning strategy of the present application;

[0060] Figure 3A flowchart of a third embodiment of the pollutant prediction method based on the evolutionary learning strategy of the present application;

[0061] Figure 4 A module structure diagram of the pollutant prediction device based on the evolutionary learning strategy of the present application;

[0062] Figure 5 A device structure diagram of the hardware operating environment involved in the pollutant prediction method based on the evolutionary learning strategy in the embodiment of the present application.

[0063] The object implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0064] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and do not limit the present application.

[0065] In order to better understand the technical solutions of the present application, the following will be described in detail in combination with the drawings and specific embodiments of the specification.

[0066] Accurate prediction of the concentration of these pollutants and their trend is crucial for formulating effective environmental management strategies, pollution control measures and public health protection measures. However, traditional pollutant prediction methods have defects when faced with data diversity, non-linear relationship and high-dimensional features. These problems mainly manifest in the following aspects: First, the pollutant data has missing values and outliers, which directly affects the accuracy and reliability of subsequent model training. For example, in air quality monitoring, due to equipment failure or external interference, data at certain time points may be lost or have erroneous readings. Second, the relationship between pollutant concentration and its influencing factors is usually complex and non-linear, which poses a challenge to traditional prediction models based on linear assumptions. Third, with the advancement of monitoring technology, the dimension of environmental data is increasing, and how to extract effective information from it has become a major problem. In addition, a single machine learning model cannot cope with all types of pollutant prediction tasks. Each model has its own advantages and limitations, such as linear regression is good at handling simple linear relationships, while deep neural networks are more suitable for capturing complex non-linear patterns. Therefore, how to effectively combine the advantages of multiple models to build a prediction system that can adapt to different pollutant characteristics and handle complex data structures has become an important direction of current research.

[0067] Therefore, the main solution of the embodiment of the present application is to perform data anomaly processing on the pollutant data to obtain cleaned data, perform feature conversion and time series data denoising on the cleaned data to obtain feature information, and identify the feature information through a target pollutant prediction model to obtain a pollutant prediction result.

[0068] Based on this, embodiments of this application provide a pollutant prediction method based on an evolutionary learning strategy, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the pollutant prediction method based on evolutionary learning strategy of this application.

[0069] In this embodiment, the pollutant prediction method based on evolutionary learning strategy includes steps S10 to S40:

[0070] Step S10: Obtain pollutant data.

[0071] It should be noted that in modern society, with increased environmental awareness and technological advancements, the sources of pollutant data have become increasingly diverse and complex. This data typically originates from various monitoring stations, satellite remote sensing, and mobile sensors, encompassing pollutant concentration information in multiple environmental media, including the atmosphere, water bodies, and soil. For example, in air quality monitoring, pollutant data includes PM2.5, PM10, and sulfur dioxide (PM2.5, PM10, and PM2.5). ), nitrogen oxides ( Key indicators include chemical oxygen demand (COD) and ammonia nitrogen; water quality monitoring focuses on pollutant data such as COD and ammonia nitrogen. ), heavy metal content, etc.

[0072] Step S20: By performing data anomaly processing on the pollutant data, cleaned data is obtained.

[0073] It's important to note that missing value issues are a common anomaly in environmental monitoring data, caused by sensor malfunctions, network transmission problems, or human error. Therefore, employing appropriate imputation methods is crucial for addressing this problem. For example, KNN imputation uses the average of the closest neighboring samples to fill in missing values; regression imputation predicts missing values ​​by constructing a regression model based on other intact features; and mean imputation directly uses the average of all known values ​​of that feature. In particular, for time series data, interpolation using data from consecutive time points can be considered to maintain temporal continuity.

[0074] Further, the detection and correction of outliers follows. Outliers are caused by measurement errors or extreme events, which will negatively impact the model training if not handled. Common methods include statistical analysis such as interquartile range (IQR) and Z-score. The IQR method identifies outliers by calculating the difference between the first and third quartiles, any value below the lower quartile 1.5 times IQR or above the upper quartile 1.5 times IQR is considered an outlier. The Z-score method calculates the standard score of each data point, if the absolute value of a point's Z-score exceeds a certain threshold, it is considered an outlier. Once outliers are identified, different strategies can be taken for correction, such as replacing with the mean of neighboring points, deleting abnormal records, or making specific adjustments according to domain knowledge. After completing the above steps, the cleaned dataset is obtained.

[0075] Step S30, feature transformation and time series data denoising are performed on the cleaned data to obtain feature information.

[0076] It should be noted that feature transformation is to transform the original data into a form more suitable for machine learning model processing. This step includes various techniques, for example, by applying logarithmic transformation or Box-Cox transformation, it can effectively handle those data sets with skewed distribution, so that the data is more consistent with the normal distribution, thereby improving the learning efficiency of the model. In addition, appropriate transformation methods can be selected according to specific application scenarios, such as when dealing with pollutants related to concentration, since their concentration changes show an exponential growth or decay trend, appropriate nonlinear transformation can better capture this relationship.

[0077] Secondly, time series data denoising is an important step to optimize the data characteristics in the time dimension. Environmental pollutant monitoring often produces a large amount of time series data, which is easily affected by short-term fluctuations or random noise, and these noises will mask the true trends and periodic changes. In order to solve this problem, signal processing techniques such as wavelet transform or Kalman filter can be used. Wavelet transform can analyze signals at different scales, helping to separate useful information and remove high-frequency noise; Kalman filter is based on state estimation theory, which smoothes the noise in the data through a recursive algorithm, especially suitable for time series prediction of dynamic systems. After the above feature transformation and denoising, we obtain the feature information, which not only retains the core features of the original data, but also enhances the model's understanding of complex nonlinear relationships. This lays the foundation for the subsequent model training and optimization process to more accurately reflect the behavior of pollutants in the actual environment, thereby achieving more accurate prediction results.

[0078] Step S40, the feature information is identified through the target pollutant prediction model to obtain the pollutant prediction result.

[0079] It should be noted that in the present embodiment, the prediction results include pollutant data prediction results, pollutant category prediction results, and prediction results of pollutant nonlinear, high-dimensional feature data. When the feature information is prepared, the feature information is sent to the mixed expert model optimized by the evolutionary learning algorithm. The model is composed of multiple expert models, each of which independently predicts data according to its specific advantages. For example, a linear regression expert model may be responsible for handling simple linear relationships, while a deep neural network expert model focuses on capturing complex nonlinear patterns. In order to maximize the advantages of each model and minimize their limitations, a multi-level knowledge fusion strategy is adopted, such as weighted average fusion, weighted voting fusion, and weighted projection fusion, to integrate the outputs of each expert model. In this process, each fusion strategy is flexibly applied according to specific task requirements. For numerical prediction tasks, weighted average fusion is preferred; for classification problems, weighted voting fusion is used; and for complex nonlinear, high-dimensional feature data, weighted projection fusion is used. Through this comprehensive method, not only can the prediction accuracy be improved, but also the robustness and adaptability of the system can be enhanced. Finally, based on these optimized models and fusion strategies, the prediction results of pollutant concentration or type under specific environmental conditions can be output. This not only helps to understand the current environmental pollution situation in a timely manner, but also provides a scientific basis for formulating effective pollution control measures.

[0080] Specifically, before step S40, it also includes: based on the preset mixed expert model and the multi-level knowledge fusion strategy, performing optimization processing to obtain a preset pollutant prediction model. It should be noted that this process first relies on constructing a mixed expert model containing multiple machine learning models, each of which independently predicts specific types of data features or tasks. The preset mixed expert model includes a linear regression expert model, a support vector regression expert model, a random forest expert model, and a deep neural network expert model. Specifically, the linear regression expert model (LR) is suitable for simple, linear relationship pollutant prediction scenarios. By standardizing the input data, a least squares regression model is constructed, assuming that there is a linear relationship between pollutant concentration and environmental factors, and the formula is:

[0081]

[0082] wherein, is the intercept, is the regression coefficient of each feature, is the current number of feature values, denotes the total number of data. The basic idea of the LR model is to assume that the prediction target is linearly related to the input feature value There is a linear relationship between the predicted and actual values. Then, the sum of squared errors between the predicted and actual values ​​is minimized using the least squares (OLS) method to solve for a set of optimal coefficients. Therefore, the LR model is used to handle the prediction task of the approximately linear relationship between pollutant concentration and environmental factors. It is characterized by strong interpretability and low computational cost.

[0083] Support Vector Regression (SVR) expert models excel at capturing complex, non-linear relationships in pollutant data. SVR uses kernel functions to map low-dimensional data to a high-dimensional space, effectively capturing non-linear patterns. The specific model formula is as follows:

[0084]

[0085] in, For kernel function, For Lagrange multipliers, For bias terms, Indicates the first One data point, This represents the total number of data points. The above formula is derived from the regression form (SVR) of Support Vector Machines (SVM) through a kernel function. The input data is mapped from the original space to a high-dimensional feature space to construct a linear regressor, thereby capturing the nonlinear relationship between input and output. SVR expert models are used to handle complex nonlinear patterns in pollutant data, effectively characterizing nonlinear and non-stationary pollution time-series changes. The SVR expert model, described by the formula, embeds the pollutant prediction problem into a high-dimensional kernel space, realizing the modeling of the relationship between nonlinear features and pollutant concentration. Kernel function. RBF kernel and polynomial kernel are used to adapt to the changing patterns of different pollutant data. Sparsity means that only a subset of training samples (support vectors) influence the prediction, which significantly reduces computational costs when modeling large-scale environmental data. The derivation of the above formula is as follows: In pollutant prediction tasks, using known historical data... The regression function is used to make the regression function respond to new inputs. It can predict values ​​as close as possible to the true value. The numerical value. In the SVR expert model, the regression function can be expressed as:

[0086]

[0087] in, This indicates that the feature vector The mapping function that maps to a high-dimensional feature space. This is the transpose of the weight vector in high-dimensional space. This represents the bias term. Then, by introducing Lagrange multipliers and a weight vector in high-dimensional space... The regression function described above can be expressed as:

[0088]

[0089]

[0090] in, For Lagrange multipliers, Indicates the inner product. Indicates the first One data point, This represents the total number of data points. Then, the inner product is replaced with a kernel function, and this is recorded as... Substitute This will give you the final function representation.

[0091] The Random Forest (RF) expert model reduces overfitting by integrating multiple decision trees. The specific formula is as follows:

[0092]

[0093] in, For the first A tree, For the first Tree samples The formula describes the prediction mechanism of the Random Forest (RF) model, which is conceived by ensemble... Decision trees trained from random feature subsets To reduce variance and improve generalization ability. The prediction function for each tree is... The final output of the forest is the average of the predictions from all trees. The derivation is based on the Bagging principle; ideally, the expected mean squared error of the forest equals the mean error of a single tree minus the covariance of the predictions between trees, indicating that diversity improves accuracy. The RF expert model utilizes a multi-tree averaging mechanism to reduce sensitivity to noise, making it suitable for pollutant prediction scenarios with many feature dimensions and large sample sizes.

[0094] Deep neural network expert models (DNNs) can automatically extract deep features from high-dimensional data and learn deep nonlinear relationships between data through multilayer perceptrons. The specific model formula is as follows:

[0095]

[0096] in, Here are the weight matrices for each layer. The bias matrix, The activation function is used. During the training and evaluation phases of expert models, cross-validation techniques such as K-fold cross-validation are crucial to ensure that each model performs well not only on its training data but also maintains good generalization ability on unseen data. K-fold cross-validation randomly divides the entire dataset into K subsets or "folds," then performs K iterations of training and testing, using one subset as the test set and the remaining K-1 subsets as the training set each time. This method helps evaluate the model's stability and robustness, ensuring consistent performance across different data subsets. When evaluating the performance of each expert model, multiple performance metrics are used to comprehensively measure the model's effectiveness. Mean squared error (MSE) is a commonly used evaluation metric for regression problems; it calculates the average of the sum of squares of the differences between predicted and actual values. A lower MSE indicates higher prediction accuracy. Coefficient of determination (COP) is another important metric. The value of 0 is used to measure the model's ability to explain variability, and its value ranges from 1 to 0. A value closer to 1 indicates a better model fit. Furthermore, other metrics such as Mean Absolute Error (MAE) can be considered to gain a more comprehensive understanding of model performance. In addition, based on hybrid expert model systems, multi-level knowledge fusion strategies are used to integrate the prediction results of various expert models to maximize the advantages of different models and overcome their limitations. These strategies include, but are not limited to, weighted average fusion, weighted voting fusion, and weighted projection fusion. Weighted average fusion assigns weights to each expert model and then averages their predictions, making it particularly suitable for numerical prediction tasks. Weighted voting fusion is mainly used for classification problems; it assigns weights based on the classification ability of each expert model and determines the final category through weighted voting. For complex nonlinear regression tasks, weighted projection fusion maps the outputs of multiple expert models to a low-dimensional space, improving prediction performance by optimizing the fusion weights and mapping function.

[0097] Further, the preset pollutant prediction model is optimized according to an evolutionary learning algorithm strategy to obtain a target pollutant prediction model. It should be noted that the evolutionary learning algorithm in this embodiment is a genetic algorithm (GA), which is a search heuristic algorithm simulating natural selection and genetic mechanisms, used to solve optimization and search problems, which provides a relatively simple idea for solving complex optimization problems. It can be divided into the following main stages: initialization of population, individual coding, fitness evaluation, selection, crossover, mutation, new population formation, and termination condition judgment. First, a set of solutions is randomly generated as the initial population, and each solution is called an individual. These individuals represent a potential solution to the problem. In the individual coding stage, the hyperparameters of each expert model (such as the kernel function type of SVM, the number of trees of RF, and the number of layers of DNN) and the fusion weights between the expert models are coded to form the individual's gene group. These codes can use binary coding, real number coding, or rule-based coding schemes to facilitate subsequent genetic operations. Then, define the fitness function, which evaluates the performance of individuals based on the accuracy of pollutant prediction (such as mean square error MSE or mean absolute error MAE). The higher the fitness, the closer the individual's model structure and parameter settings are to the optimal solution. During the evolution process, new individual groups are continuously generated through selection, crossover, and mutation operations. The selection operation selects excellent individuals according to the fitness value, and the commonly used methods include roulette wheel selection and tournament selection. The crossover operation generates new individuals through single-point or double-point crossover to explore new combinations of hyperparameters. The mutation operation introduces small perturbations to increase population diversity and avoid premature convergence to local optimal solutions. In addition, in the multi-objective optimization scenario, algorithms such as NSGA-II can be used to balance the trade-off between model complexity and prediction accuracy, ensuring that the model is not only accurate but also not over-fitted. When the maximum number of iterations is reached, a satisfactory solution is found, or the fitness of the population no longer improves, the final best individual is output as the approximate optimal solution to the problem, and if not satisfied, the iteration continues. After multiple iterations, the evolutionary learning algorithm can find a set of optimized hyperparameter configurations and fusion strategies, thereby obtaining a more accurate pollutant prediction model. This pollutant prediction model is used as the target pollutant prediction model. This model not only accurately captures complex patterns and nonlinear relationships in the data, but also dynamically adapts to environmental changes, maintaining long-term stability and efficiency. Finally, the performance of the optimized model is evaluated on the test set, and the indicators such as MSE, determination coefficient and MAE before and after improvement are compared to confirm the optimization effect.

[0098] Further, step S40 is followed by the introduction of a dynamic adaptive and online learning mechanism. Specifically, updated pollutant data is monitored and obtained in real-time, and input into the target pollutant prediction model to obtain an evaluation performance. When the evaluation performance does not meet the preset prediction requirement, the hyperparameters and fusion weights of the target pollutant prediction model are adjusted until the evaluation performance meets the preset prediction requirement. This mechanism enables the system to receive new pollutant data in real-time and dynamically adjust model parameters based on these data, ensuring that it remains efficient and accurate under changing environmental conditions. Specifically, the online learning mechanism allows the model to continuously update its parameters during operation without the need to retrain the entire model. For example, when newly monitored data indicates abnormal fluctuations in the concentration of a certain pollutant, the system can immediately absorb this information and adjust the prediction model accordingly to more accurately reflect the current environmental conditions. Furthermore, by using evolutionary algorithms such as genetic algorithm (GA) or particle swarm optimization (PSO), the weights and fusion strategies of each expert model can be dynamically optimized based on the latest data, further enhancing the adaptive ability of the system. This dynamic adaptive and online learning capability is crucial for addressing complex environmental pollution problems. It not only improves the accuracy and reliability of the prediction model, but also enables the system to quickly respond to sudden pollution events and provide timely warning information.

[0099] The embodiment provides a pollutant prediction method based on an evolutionary learning strategy. The pollutant data is subjected to data anomaly processing to obtain cleaned data, the cleaned data is subjected to feature conversion and time series data denoising to obtain feature information, and the feature information is identified through a target pollutant prediction model to obtain a pollutant prediction result. The evolutionary learning algorithm is used to dynamically adjust parameters through multi-model fusion optimization to construct a target pollutant prediction model, thereby improving the prediction accuracy of pollutants.

[0100] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as the above embodiment one can refer to the above introduction, and the subsequent will not be described in detail. On this basis, please refer to Figure 2 , the pollutant prediction method based on the evolutionary learning strategy further comprises steps S201-S204 before step S40.

[0101] Step S201, obtaining the task requirement of pollutant prediction.

[0102] It should be noted that obtaining the task requirement of pollutant prediction involves multiple aspects. In the present embodiment, according to the different application scenarios, it is divided into three categories: obtaining pollutant data, obtaining pollutant categories, and predicting nonlinear and high-dimensional feature data.

[0103] Specifically, the need for pollutant data acquisition primarily focuses on the accurate prediction of pollutant concentration trends in specific environmental media. For example, in air quality monitoring, it may be necessary to predict the concentrations of pollutants such as PM2.5 and SO2 over the next few days. This requires the system to process large amounts of real-time data and possess high accuracy and rapid response capabilities. The need to acquire pollutant categories focuses on identifying the specific type of pollutant or determining whether it exceeds the standard. For example, in water quality monitoring, it involves determining whether there is heavy metal pollution or excessive organic matter in the water body and further analyzing its source. For the prediction needs of nonlinear, high-dimensional characteristic data, such as the analysis of multi-pollutant interactions in complex atmospheric or aquatic environments, it is usually necessary to process high-dimensional datasets with complex nonlinear relationships.

[0104] Step S202: Based on the task requirements, select the corresponding optimized fusion method in the multi-level knowledge fusion strategy.

[0105] It should be noted that the above-mentioned optimization and fusion methods include weighted average fusion, weighted voting fusion, and weighted projection fusion.

[0106] Step S203: Perform fusion processing on the preset hybrid expert model according to the optimized fusion method to obtain the optimized preset hybrid expert model.

[0107] It should be noted that, further, when the task requirement is pollutant data prediction, the optimized fusion method is to prioritize weighted average fusion, specifically expressed by the following formula:

[0108]

[0109] in, The predicted pollutant values ​​after fusion. For the first Pollutant predictions from an expert model. For the first The weights of each expert model, and satisfying , This represents the total number of expert models. This formula embodies the weighted average fusion concept, that is, it uses the predictions of each expert model... Credibility assignment weight And obtain the final predicted value by linear combination of them. The weights can be normalized by the inverse of the model historical errors or optimized by evolutionary algorithms. First, weight distribution needs to be assigned to each expert model. In this embodiment, the weight distribution method is based on static allocation of model errors, which mainly relies on the prediction error on historical data to determine the weight of each expert model. Specifically, models with lower errors will get higher weights, which can be achieved by calculating the mean square error (MSE) or mean absolute error (MAE) indicators, and the specific formula for static weight allocation is:

[0110]

[0111] wherein, is the mean square error of the th model, represents the error of the th model. However, environmental conditions and data characteristics change over time, which requires the introduction of a dynamic weight adjustment method. By monitoring the prediction error of each expert model in real time and dynamically adjusting its weight according to the latest performance, it can better adapt to changing data patterns. A commonly used technique is to use a sliding window method, which calculates the performance of each expert model in the recent period, giving higher weights to models with better recent performance. This method is particularly suitable for handling non-stationary data, i.e. data distribution changes over time. After weight distribution, first collect the prediction values of each expert model, then assign the corresponding weights according to the historical or real-time errors of these models, and finally calculate the final prediction result using the weighted average formula. In this way, not only can the prediction accuracy be improved, but also the system's adaptability and robustness can be enhanced, ensuring reliable prediction results under different conditions.

[0112] Further, when the task requirement is pollutant category prediction, the optimization fusion method is to preferentially adopt weighted voting fusion, which is specifically expressed as:

[0113]

[0114] wherein, is the pollutant prediction category, is the th category, is the total number of pollutant categories, is the probability of the th expert model predicting the category of the pollutant, is the weight of the th expert model. The formula expresses the weighted voting classification fusion process. Each expert model outputs the prediction probability of the th category, and the weight is assigned to the expert model according to the classification accuracy or confidence The method involves summing the weighted probabilities for each category and selecting the category with the highest probability as the final prediction. This derivation stems from Bayesian decision theory: if the model predictions are independent, maximizing the sum of weighted probabilities can approximately minimize the classification error rate. In this patented system, this method is used for pollutant category discrimination (such as pollution level identification), balancing generalization ability and decision reliability. First, weights need to be assigned to each expert model. Specifically, in this embodiment, the weight assignment method is based on classification accuracy, and the specific formula is as follows:

[0115]

[0116] in, For the first While individual expert models can improve classification accuracy, a dynamic weight allocation strategy based on prediction confidence is more effective when dealing with rapidly changing data or non-stationary distributions. This method adjusts weights in real time based on the expert model's prediction confidence (e.g., classification probability) for the current sample. When an expert model has a higher confidence in predicting a particular category, it is given a higher weight. For example, in dealing with multi-class pollutant identification problems, if an expert model's prediction probability for a specific pollution type is significantly higher than other categories, then that expert model's judgment for that category is considered more reliable. After weight allocation, the classification results or category probabilities output by each expert model are collected; then, the corresponding weights are calculated based on the expert model's classification accuracy or the prediction confidence of the current sample; finally, these weights and category probabilities are combined, and the final predicted category is determined through weighted voting or other fusion mechanisms. This dynamic weight allocation method not only improves the system's flexibility and adaptability but also more accurately captures subtle changes in the data, thus providing more reliable support for environmental monitoring and management.

[0117] Furthermore, when the task requires prediction of nonlinear, high-dimensional feature data, the optimized fusion method is to prioritize weighted projection fusion, specifically expressed by the following formula:

[0118]

[0119] in, These are the projected pollutant predictions. It is a high-dimensional vector. Represents the projection function. For the optimized output function, Indicates the first The weights of each expert model. The formula is used for fusion of nonlinear, high-dimensional prediction tasks. The high-dimensional output of each model. First through projection function Map to a latent space (could be a dimensionality reduction map or a kernel space map), then weighted Linear combination, finally through the output function Generate a prediction result. The derivation idea refers to the combination of feature transformation and regression, and the projection step before fusion can avoid the problem of information loss or noise amplification caused by directly weighting on inconsistent high-dimensional outputs. This method is used to integrate the modeling advantages of different models on high-dimensional spatio-temporal features. Weighted projection fusion includes linear projection and nonlinear projection. Directly using linear weighted projection is a simple and effective strategy, especially suitable for scenarios where the prediction value is a scalar. By assigning weights to each expert model and weighting the sum of their prediction values, the final prediction result can be obtained. This method is computationally efficient and suitable for handling relatively direct and linear data relationships. However, when faced with complex nonlinear and high-dimensional features, nonlinear projection is more appropriate. A common approach is to use kernel methods (such as RBF kernels) to map the prediction values of each expert model to a high-dimensional space, perform fusion in this space, and then project back to the original space. In addition, deep autoencoders can also be used to learn the optimal fusion representation of prediction values, automatically extracting features to capture deep patterns in the data. This method not only can handle complex nonlinear relationships, but also can effectively reduce the impact of the curse of dimensionality. After obtaining the weights, define the projection function and the output function, calculate the prediction result of each expert model and project it to the target space, and then optimize the final prediction value based on the weights according to the projected result.

[0120] Through the above series of optimization and fusion processing, we can get an optimized preset hybrid expert model. This model not only inherits the advantages of each expert model, but also improves the prediction ability and adaptability of the overall system through multi-level knowledge fusion strategies.

[0121] Step S204, the optimized preset hybrid expert model is used as a preset pollutant prediction model.

[0122] This embodiment obtains the task requirements for pollutant prediction, selects the corresponding optimization and fusion mode in the multi-level knowledge fusion strategy according to the task requirements, performs fusion processing on the preset hybrid expert model according to the optimization and fusion mode, obtains the optimized preset hybrid expert model, and uses the optimized preset hybrid expert model as a preset pollutant prediction model. According to the task requirements, select the appropriate multi-level knowledge fusion strategy to optimize the preset hybrid expert model, obtain a high-precision pollutant prediction model, and improve the accuracy and adaptability of the pollutant prediction.

[0123] Based on the first embodiment of the present application, in the third embodiment of the present application, the same or similar contents as the above embodiment one can refer to the above introduction, and the subsequent will not be repeated. On this basis, please refer to Figure 3The pollution prediction method based on the evolutionary learning strategy further comprises steps S301-S305 before step S40.

[0124] In step S301, the hyperparameters corresponding to the expert models of the preset mixed expert model and the model prediction accuracy are obtained.

[0125] It should be noted that before model prediction, the hyperparameters of each expert model need to be determined. For example, in the linear regression (LR) model, no special hyperparameters need to be set, but in the support vector regression (SVR) model, the kernel function type, the regularization parameter, and the kernel coefficient need to be determined; for the random forest (RF), the number of trees and the maximum depth need to be set; and in the deep neural network (DNN), the number of layers, the number of neurons in each layer, the activation function type, and the learning rate are important hyperparameters. These hyperparameters directly affect the performance of the model, so it is crucial to select appropriate values. The optimal hyperparameter combination can be explored by methods such as grid search, random search, or more advanced Bayesian optimization. In order to evaluate the performance of each expert model, cross-validation techniques such as K-fold cross-validation are usually used. In this process, the data is divided into K subsets, and the model is trained on K-1 subsets and tested on the remaining one subset. This process is repeated K times, with a different subset used as the test set each time. In this way, a series of performance indicators can be calculated, including mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R2), to quantify the prediction accuracy of the model.

[0126] In step S302, the hyperparameters and the weight parameters in the multi-level knowledge fusion strategy are encoded to obtain individual gene values.

[0127] It should be noted that for hyperparameters, real number encoding can be used to represent continuous variables (such as learning rate, regularization coefficient, etc.), while integer encoding can be used for discrete variables (such as the number of trees, the number of layers, etc.). For example, in support vector regression (SVR), the kernel function type can be encoded by an integer (1 represents linear kernel, 2 represents RBF kernel, etc.), while the regularization parameter and the kernel coefficient are represented by real number encoding. Similarly, the number of trees and the maximum depth in the random forest can also be represented by the corresponding encoding method. Next, for the weight parameters in the multi-level knowledge fusion strategy, real number encoding can be used to ensure that the sum of all weights is equal to 1. For example, in weighted average fusion, the weight of each expert model can be represented by a real number vector that satisfies , where N is the total number of expert models. In this way, the entire individual genome is composed of the encoding of these hyperparameters and weight parameters, forming a complete solution, and based on the individual genome, the corresponding individual gene values can be obtained. ​

[0128] Through this encoding mode, the evolutionary algorithm can search for the optimal configuration in a vast solution space. The individual genome not only contains the best hyperparameter settings for each expert model, but also reflects the best way of cooperation between different models, thereby improving the prediction accuracy and adaptability of the hybrid expert model.

[0129] Step S303, based on the model prediction accuracy, evaluation is carried out, the fitness function is constructed, and the fitness value is obtained.

[0130] It should be noted that the design of the fitness function needs to be closely related to the specific task requirements and performance indicators. The fitness function is constructed according to the true value and the predicted value, which is used for evolutionary algorithm optimization. The fitness function is in the form of:

[0131]

[0132] Among them, is the actual value of the fitness, is the predicted value of the fitness, is the number of data. The formula defines the calculation method of individual fitness, which uses mean square error (MSE) to measure the performance of individuals (expert model configuration and fusion weight combination) in the prediction task. The derivation is derived from the error measurement commonly used in regression tasks, which is suitable for optimization target of genetic algorithm or other evolutionary learning strategies, because it is continuous, comparable and minimizing MSE is equivalent to maximizing prediction accuracy. This fitness function drives the evolutionary learning process, constantly filters and optimizes model parameters and fusion strategies, so that the pollutant prediction model is iteratively improved under multi-objective constraints. For a single sample:

[0133]

[0134] Among them, represents the square of the prediction error, which is a positive number and punishes large deviations. For a sample, take the average:

[0135]

[0136] Therefore, directly use MSE as Fitness (the smaller the better).

[0137] Step S304, according to the evolutionary operation strategy and multi-objective optimization strategy, the fitness value and individual gene value are optimized, and the optimized hyperparameters and optimized fusion weights are obtained.

[0138] It should be noted that by selecting evolutionary operation strategies, crossover, mutation and introducing multi-objective optimization strategies, the hyperparameters and fusion weights are dynamically adjusted to improve the overall performance of the model.

[0139] Further, the step S304 further comprises: selecting individuals corresponding to preset fitness and preset individual gene values by sorting the fitness values and the individual gene values. Specifically, the selecting operation selects individuals with excellent performance based on the fitness values to participate in the reproduction process of the next generation. Common selection methods include roulette selection and tournament selection. For example, in the roulette selection, individuals with higher fitness and higher individual gene values are more likely to be selected, ensuring that excellent characteristics are inherited. Then, the individuals are processed through crossover operation and mutation operation to obtain updated individuals. Specifically, the crossover operation generates new offspring by combining gene fragments of two parent individuals. Single-point or double-point crossover method can effectively explore new combinations of hyperparameter space. For example, combining the regularization parameter of one parent individual of the support vector regression (SVR) model with the kernel coefficient parameter of another parent individual produces a better configuration. Then, the mutation operation increases population diversity by randomly disturbing individual genes, avoiding falling into local optimal solution. For example, small amplitude mutation on the depth neural network (DNN) layer encoding of a certain individual produces a new excellent population. The fitness values of the updated individuals are optimized through a multi-objective optimization strategy to obtain a Pareto optimal solution set, and a preset solution is selected from the Pareto optimal solution set as the optimized hyperparameters and the optimized fusion weights. Specifically, in order to balance the relationship between model complexity and prediction accuracy, a multi-objective optimization strategy is adopted, and in this embodiment, the NSGA-II method is adopted. The fitness values of the updated individuals are sorted to obtain multiple crowded distance levels, and the fitness values of multiple updated individuals corresponding to a preset level are selected as the Pareto optimal solution according to the selection of the crowded distance levels. The multiple Pareto optimal solutions are combined to obtain a Pareto optimal solution set. Specifically, the fitness values are sorted, and all updated individuals are non-dominantly sorted according to their performance on different objectives. This means that the individuals are divided into different levels, and the first level contains all non-dominated solutions (i.e. the Pareto front), the second level contains individuals that are only dominated by individuals in the first level, and so on. Next, the crowded distance is calculated, which is a density indicator that measures the density of an individual in its own level. Specifically, for each individual, the sum of the distances between adjacent individuals along each objective dimension is calculated. Higher crowded distance indicates that the individual is sparser in its neighborhood, so it is more likely to be retained to maintain population diversity. Then, multiple updated individuals in a preset level (usually the first few levels) are selected as candidate Pareto optimal solutions according to the selection of the crowded distance levels. These individuals not only perform well in their respective non-dominated layers, but also have good diversity due to their larger crowded distance. Based on the combination of multiple Pareto optimal solutions, a comprehensive Pareto optimal solution set is formed.The set contains solutions that achieve the best balance between different objectives, and then finds the Pareto optimal solution between multiple objectives (such as MSE and model complexity), so that the final optimized hyperparameters and fusion weights can ensure high precision and will not cause overfitting.

[0140] Through the above series of evolutionary operations and multi-objective optimization strategies, the hyperparameters and fusion weights are continuously iteratively optimized, and finally the optimized preset hybrid expert model is obtained.

[0141] Step S305, update the preset pollutant prediction model according to the optimized hyperparameters and the optimized fusion weights to obtain the target pollutant prediction model.

[0142] It should be noted that by applying the optimal solution obtained by the evolutionary algorithm and the multi-objective optimization strategy to each expert model and the fusion strategy, a target pollutant prediction model with better performance is constructed.

[0143] Specifically, the optimal hyperparameters obtained by optimization are used to update each expert model. For example, in the SVR model, the kernel function type, regularization parameter, and kernel coefficient are adjusted; in the RF model, the number of trees and maximum depth are updated; for DNN, the number of layers, the number of neurons in each layer, the type of activation function, and the learning rate are adjusted. These optimized hyperparameters can improve the performance of each expert model in a specific task. Then, the multi-level knowledge fusion strategy is reconfigured using the optimized fusion weights. If the task requirement is numerical prediction, the weighted average fusion method is used to ensure that the model with lower error has higher weight; for classification tasks, weighted voting fusion is used to make the model with high classification accuracy play a greater role; and when dealing with complex nonlinear relationships, weighted projection fusion is applied, and deep autoencoder or kernel method is further used to optimize the fusion result. Through the above steps, an optimized pollutant prediction model is obtained.

[0144] Further, after obtaining the optimized pollutant prediction model, it further includes: obtaining pollutant verification data, which should be independent of the test data set to ensure the fairness and objectivity of the evaluation. The verification data is usually derived from actual monitoring stations, satellite remote sensing or other reliable environmental monitoring means, covering various types of pollutants and their corresponding environmental conditions. Then the target pollutant prediction model is tested according to the pollutant verification data, and the test result is obtained, and the model parameters are iteratively updated based on the optimized hyperparameters and optimized fusion weights until the maximum number of iterations is reached or the test result meets the preset detection requirements, and the target pollutant prediction model is obtained. Specifically, the verification data is input into the model, and the prediction result is generated, and by comparing the predicted value with the true value, the performance of the model is comprehensively evaluated, and the test result is obtained. Based on the test result, if the model does not meet the preset detection requirements, it enters the iterative update stage. In this stage, evolutionary algorithms or particle swarm optimization are used to dynamically adjust the model parameters combined with the optimized hyperparameters and fusion weights. After each iteration, the fitness function is recalculated and the best individual is selected as the basis for the next generation. The iterative update process continues until the maximum number of iterations is reached or the model's performance on the validation set meets the preset detection criteria. Each iteration aims to further improve the model's prediction accuracy and robustness, ensuring its efficiency and stability in complex real-world environments. When the model finally meets all conditions, the target pollutant prediction model is obtained. This model not only has higher prediction accuracy and stability, but also can handle various complex real-world application scenarios, providing strong technical support for environmental protection.

[0145] The embodiment obtains the hyperparameters and prediction accuracy of the preset hybrid expert model, optimizes the optimized hyperparameters and fusion weights using evolutionary operations and multi-objective optimization strategies, and finally updates the preset model to obtain the target pollutant prediction model. The optimized target pollutant prediction model is used to predict pollutants, improving the accuracy and adaptability of pollutant prediction.

[0146] The application also provides a pollutant prediction device based on an evolutionary learning strategy, which is described in detail in Figure 4 , and the device comprises:

[0147] The obtaining module 10 is configured to obtain pollutant data.

[0148] The data cleaning module 20 is configured to obtain cleaned data by performing data anomaly processing on the pollutant data.

[0149] The data conversion module 30 is configured to perform feature conversion and time series data denoising on the cleaned data to obtain feature information.

[0150] The result module 40 is configured to identify the feature information through the target pollutant prediction model to obtain a pollutant prediction result, and the prediction result includes a pollutant data prediction result, a pollutant category prediction result and a prediction result of nonlinear and high-dimensional feature data of the pollutant.

[0151] The construction module 50 is configured to optimize a preset hybrid expert model according to a multi-level knowledge fusion strategy to obtain a preset pollutant prediction model, and the preset hybrid expert model includes a linear regression expert model, a support vector regression expert model, a random forest expert model and a deep neural network expert model; and the preset pollutant prediction model is optimized according to an evolutionary learning algorithm strategy to obtain a target pollutant prediction model.

[0152] The pollutant prediction device based on the evolutionary learning strategy provided in the application can solve the technical problem of how to improve the accuracy of pollutant prediction by using the pollutant prediction method based on the evolutionary learning strategy in the above embodiment. Compared with the prior art, the pollutant prediction device based on the evolutionary learning strategy provided in the application has the same beneficial effects as the pollutant prediction method based on the evolutionary learning strategy provided in the above embodiment, and other technical features in the pollutant prediction device based on the evolutionary learning strategy are the same as the features disclosed in the above embodiment method, which will not be repeated here.

[0153] In an embodiment, the construction module 50 is further configured to obtain a task requirement of pollutant prediction; select a corresponding optimization fusion mode in the multi-level knowledge fusion strategy according to the task requirement, and the optimization fusion mode includes weighted average fusion, weighted voting fusion and weighted projection fusion; fuse the preset hybrid expert model according to the optimization fusion mode to obtain an optimized preset hybrid expert model; and use the optimized preset hybrid expert model as the preset pollutant prediction model.

[0154] In an embodiment, the construction module 50 is further configured to, when the task requirement is pollutant data prediction, the optimization fusion mode is to preferentially adopt weighted average fusion; when the task requirement is pollutant category prediction, the optimization fusion mode is to preferentially adopt weighted voting fusion; and when the task requirement is prediction of nonlinear and high-dimensional feature data, the optimization fusion mode is to preferentially adopt weighted projection fusion.

[0155] In an embodiment, the constructing module 50 is further configured to obtain hyperparameters corresponding to the expert models of the preset hybrid expert model and model prediction accuracy; encode the hyperparameters and weight parameters in the multi-level knowledge fusion strategy to obtain individual gene values; evaluate based on the model prediction accuracy to construct a fitness function and obtain fitness values; optimize the fitness values and the individual gene values according to an evolutionary operation strategy and a multi-objective optimization strategy to obtain optimized hyperparameters and optimized fusion weights; and update the preset pollutant prediction model according to the optimized hyperparameters and the optimized fusion weights to obtain the target pollutant prediction model.

[0156] In an embodiment, the constructing module 50 is further configured to select an individual corresponding to a preset fitness and a preset individual gene value by sorting the fitness values and the individual gene values; process the individual by performing cross operation and mutation operation to obtain an updated individual; optimize the fitness values of the updated individual by a multi-objective optimization strategy to obtain a Pareto optimal solution set; and select a preset solution from the Pareto optimal solution set as the optimized hyperparameters and the optimized fusion weights.

[0157] In an embodiment, the constructing module 50 is further configured to sort the fitness values of the updated individual to obtain a plurality of crowding distance levels; select fitness values of a plurality of updated individuals corresponding to a preset level as Pareto optimal solutions by screening the crowding distance levels; and combine a plurality of Pareto optimal solutions to obtain a Pareto optimal solution set.

[0158] In an embodiment, the result module 40 is further configured to monitor and obtain updated pollutant data in real time; input the updated pollutant data into the target pollutant prediction model to obtain evaluation performance; and adjust the hyperparameters and the fusion weights of the target pollutant prediction model until the evaluation performance meets a preset prediction requirement when the evaluation performance does not meet the preset prediction requirement.

[0159] The present application provides a pollutant prediction device based on an evolutionary learning strategy, which comprises at least one processor and a memory in communication connection with the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the pollutant prediction method based on the evolutionary learning strategy in the above-mentioned embodiment one.

[0160] The following will be described with reference to the accompanying drawings Figure 5This document illustrates a structural schematic diagram of a pollutant prediction device based on an evolutionary learning strategy suitable for implementing embodiments of this application. The pollutant prediction device based on an evolutionary learning strategy in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 5 The pollutant prediction device based on the evolutionary learning strategy shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0161] like Figure 5 The pollutant prediction device based on the evolutionary learning strategy may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in ROM 1002 (Read Only Memory) or a program loaded from storage device 1003 into RAM 1004 (Random Access Memory). RAM 1004 also stores various programs and data required for the operation of the pollutant prediction device based on the evolutionary learning strategy. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the pollutant prediction device based on an evolutionary learning strategy to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a pollutant prediction device based on an evolutionary learning strategy with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0162] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program codes for executing the method described in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network through a communication device, or installed from the storage device 1003, or installed from the ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiments disclosed in the present application are executed.

[0163] The pollution prediction device based on the evolutionary learning strategy provided by the present application adopts the pollution prediction method based on the evolutionary learning strategy in the above-mentioned embodiments, and can solve the technical problem of how to improve the accuracy of pollution prediction. Compared with the prior art, the pollution prediction device based on the evolutionary learning strategy provided by the present application has the same beneficial effects as the pollution prediction method based on the evolutionary learning strategy provided by the above-mentioned embodiments, and other technical features in the pollution prediction device based on the evolutionary learning strategy are the same as the features disclosed in the previous embodiment method, which will not be repeated here.

[0164] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above-mentioned embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0165] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0166] The present application provides a computer readable medium having computer readable program instructions (i.e. computer program) stored thereon for performing calculations, the computer readable program instructions for executing the pollution prediction method based on the evolutionary learning strategy in the above-mentioned embodiments.

[0167] The computer readable medium provided in the present application may, for example, be a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination of the above. More specific examples of the computer readable medium can include, but are not limited to, an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer readable medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to an electric wire, an optical cable, an RF (Radio Frequency), and the like, or any suitable combination of the above.

[0168] The above computer readable medium can be included in the pollutant prediction device based on the evolutionary learning strategy, or can exist separately without being assembled into the pollutant prediction device based on the evolutionary learning strategy.

[0169] The above computer readable medium carries one or more programs, which, when executed by the pollutant prediction device based on the evolutionary learning strategy, enable the pollutant prediction device based on the evolutionary learning strategy to be written in one or more programming languages or combinations thereof for computer program code to perform the operations of the present application, the programming languages including object-oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as the "C" language or similar programming languages. The program code can be executed entirely on a user computer, partially on a user computer, as a separate software package, partially on a user computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, through the Internet using an Internet service provider).

[0170] The flowcharts and block diagrams in the drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or combinations of hardware and software.

[0171] The modules involved in the embodiments of the present application can be implemented in the form of software or in the form of hardware. In some cases, the name of the module does not constitute a limitation on the unit itself.

[0172] The readable medium provided by the present application is a computer readable medium, which stores computer readable program instructions (i.e. computer programs) for executing the above-mentioned pollutant prediction method based on the evolutionary learning strategy, and can solve the technical problem of how to improve the accuracy of pollutant prediction. Compared with the prior art, the computer readable medium provided by the present application has the same beneficial effects as the pollutant prediction method based on the evolutionary learning strategy provided by the above-mentioned embodiments, which will not be repeated here.

[0173] The present application also provides a computer program product comprising a computer program, which, when executed by a processor, implements the steps of the above-mentioned pollutant prediction method based on the evolutionary learning strategy.

[0174] The computer program product provided by the present application can solve the technical problem of how to improve the accuracy of pollutant prediction. Compared with the prior art, the computer program product provided by the present application has the same beneficial effects as the pollutant prediction method based on the evolutionary learning strategy provided by the above-mentioned embodiments, which will not be repeated here.

[0175] The above only describes some embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent structural transformation made by using the content of the present application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A method for predicting pollutants based on an evolutionary learning strategy, characterized in that, The method comprises: acquiring pollutant data; performing data anomaly processing on the pollutant data to obtain cleaned data; performing feature conversion and time series data denoising on the cleaned data to obtain feature information; identifying the feature information through a target pollutant prediction model to obtain a pollutant prediction result, wherein the prediction result comprises a pollutant data prediction result, a pollutant category prediction result and a prediction result of pollutant nonlinear and high-dimensional feature data, the pollutant data prediction result is focused on accurate prediction of the concentration change trend of pollutants in the environmental medium, the pollutant category prediction result is used to identify the specific type of pollutants or judge whether the pollutants exceed the standard, and the prediction result of the pollutant nonlinear and high-dimensional feature data is used to process a high-dimensional data set with complex nonlinear relationship; Before the step of identifying the feature information through the target pollutant prediction model to obtain the pollutant prediction result, the method comprises: optimizing a preset mixed expert model according to a multi-level knowledge fusion strategy to obtain a preset pollutant prediction model, wherein the preset mixed expert model comprises a linear regression expert model, a support vector regression expert model, a random forest expert model and a deep neural network expert model; optimizing the preset pollutant prediction model according to an evolutionary learning algorithm strategy to obtain a target pollutant prediction model; The step of optimizing the preset mixed expert model according to the multi-level knowledge fusion strategy to obtain the preset pollutant prediction model comprises: acquiring a task requirement for pollutant prediction; selecting a corresponding optimization fusion mode in the multi-level knowledge fusion strategy according to the task requirement, wherein the optimization fusion mode comprises weighted average fusion, weighted voting fusion and weighted projection fusion; fusing the preset mixed expert model according to the optimization fusion mode to obtain an optimized preset mixed expert model; using the optimized preset mixed expert model as the preset pollutant prediction model; The step of optimizing the preset pollutant prediction model according to the evolutionary learning algorithm strategy to obtain the target pollutant prediction model comprises: acquiring hyperparameters and model prediction accuracy of the expert model of the preset mixed expert model; encoding the hyperparameters and weight parameters in the multi-level knowledge fusion strategy to obtain individual gene values; evaluating based on the model prediction accuracy to construct a fitness function and obtain a fitness value, wherein the fitness function is in the form of: wherein, is the fitness actual value, is the fitness predicted value, is the number of data; optimizing the fitness value and the individual gene values according to an evolutionary operation strategy and a multi-objective optimization strategy to obtain optimized hyperparameters and optimized fusion weights; updating the preset pollutant prediction model according to the optimized hyperparameters and the optimized fusion weights to obtain the target pollutant prediction model.

2. The method of claim 1, wherein, The step of selecting the corresponding optimization fusion mode in the multi-level knowledge fusion strategy according to the task requirement comprises: when the task requirement is pollutant data prediction, the optimization fusion mode is weighted average fusion, and the specific expression formula is: wherein, is the predicted value of the pollutant after fusion, is the predicted value of the pollutant of the expert model, is the weight of the expert model, and satisfies , is the total number of expert models. When the task requirement is pollutant category prediction, the optimization fusion mode is to preferentially adopt weighted voting fusion, and the specific performance formula is: in, For pollutant prediction categories, For the first Class category, The total number of pollutant categories. For the first Expert model prediction categories The probability, For the first The weights of each expert model; When the task requirement is prediction of nonlinear and high-dimensional feature data, the optimization fusion mode is to preferentially adopt weighted projection fusion, and the specific performance formula is: wherein, is the predicted value of the pollutant after projection, is a high-dimensional vector, denotes a projection function, is the optimized output function.

3. The method of claim 1, wherein, The step of optimizing the fitness value and the individual gene value according to the evolution operation strategy and the multi-objective optimization strategy to obtain the optimized hyperparameters and the optimized fusion weight includes: By sorting the fitness value and the individual gene value, an individual corresponding to a preset fitness and a preset individual gene value is selected; The individual is processed through crossover operation and mutation operation to obtain an updated individual; The fitness value of the updated individual is optimized through a multi-objective optimization strategy to obtain a Pareto optimal solution set; A preset solution is selected from the Pareto optimal solution set and used as the optimized hyperparameters and the optimized fusion weight.

4. The method of claim 3, wherein, The step of optimizing the fitness value of the updated individual through the multi-objective optimization strategy to obtain the Pareto optimal solution set includes: The fitness values of the updated individuals are sorted to obtain multiple crowded distance levels; According to the filtering of the crowded distance levels, the fitness values of multiple updated individuals corresponding to a preset level are selected as Pareto optimal solutions; The Pareto optimal solution set is obtained by combining multiple Pareto optimal solutions.

5. The method of claim 1, wherein, After the step of identifying the feature information through the target pollutant prediction model to obtain a pollutant prediction result, the method further includes: Real-time monitoring and obtaining updated pollutant data; Inputting the updated pollutant data into the target pollutant prediction model to obtain an evaluation performance; When the evaluation performance does not meet a preset prediction requirement, adjusting the hyperparameters and the fusion weight of the target pollutant prediction model until the evaluation performance meets the preset prediction requirement.

6. A pollution prediction device based on an evolutionary learning strategy, characterized by, The device includes: An acquisition module configured to acquire pollutant data; A data cleaning module configured to obtain cleaned data by performing data anomaly processing on the pollutant data; A data conversion module configured to perform feature conversion and time series data denoising on the cleaned data to obtain feature information; A result module configured to identify the feature information through a target pollutant prediction model to obtain a pollutant prediction result, the prediction result including a pollutant data prediction result, a pollutant category prediction result, and a prediction result of nonlinear and high-dimensional feature data of a pollutant, wherein the pollutant data prediction result is focused on accurate prediction of the concentration change trend of pollutants in an environmental medium, the pollutant category prediction result is used to identify the specific type of pollutants or determine whether the pollutants exceed the standard, and the prediction result of nonlinear and high-dimensional feature data of a pollutant is used to process a high-dimensional data set with complex nonlinear relationships. The construction module is configured to optimize a preset hybrid expert model according to a multi-level knowledge fusion strategy to obtain a preset pollutant prediction model, the preset hybrid expert model including a linear regression expert model, a support vector regression expert model, a random forest expert model, and a deep neural network expert model; optimize the preset pollutant prediction model according to an evolutionary learning algorithm strategy to obtain a target pollutant prediction model; and obtain a task requirement for pollutant prediction; select a corresponding optimization fusion mode in the multi-level knowledge fusion strategy according to the task requirement, the optimization fusion mode including weighted average fusion, weighted voting fusion, and weighted projection fusion; fuse the preset hybrid expert model according to the optimization fusion mode to obtain an optimized preset hybrid expert model; use the optimized preset hybrid expert model as the preset pollutant prediction model; obtain an expert model corresponding hyperparameter and model prediction accuracy of the preset hybrid expert model; encode the hyperparameter and a weight parameter in the multi-level knowledge fusion strategy to obtain an individual gene value; evaluate based on the model prediction accuracy to construct a fitness function and obtain a fitness value; optimize the fitness value and the individual gene value according to an evolutionary operation strategy and a multi-objective optimization strategy to obtain an optimized hyperparameter and an optimized fusion weight; and update the preset pollutant prediction model according to the optimized hyperparameter and the optimized fusion weight to obtain a target pollutant prediction model.

7. A pollution prediction device based on an evolutionary learning strategy, characterized by, The device comprises a memory, a processor, and an evolutionary learning strategy-based pollutant prediction program stored on the memory and running on the processor, the evolutionary learning strategy-based pollutant prediction program being configured to implement the steps of the evolutionary learning strategy-based pollutant prediction method according to any one of claims 1-5.

8. A storage medium, characterized by The storage medium stores an evolutionary learning strategy-based pollutant prediction program, the evolutionary learning strategy-based pollutant prediction program being executed by the processor to implement the steps of the evolutionary learning strategy-based pollutant prediction method according to any one of claims 1-5.

Images