Water quality prediction system and method based on machine learning

Abstract

The invention discloses a water quality prediction system and method based on machine learning, and the system comprises a data collection and preprocessing module which is used for collecting and preprocessing multi-source monitoring data; the teacher model module is used for constructing a Shenchang differential equation model as a teacher model and outputting a prediction result of the target water quality index; the student model module is used for constructing a lightweight neural network model as a student model to perform structure optimization training; the trajectory evaluation module is used for calculating predicted trajectory similarity of the teacher model and the student model and optimizing structural parameters of the student model; the model reasoning module is used for calling the optimized student model to generate a prediction result of the target water quality index; and the result output module is used for outputting a prediction result and providing a display and storage interface. According to the method, high-precision, light-weight and sustainable self-adaptive prediction of water quality time sequence data is realized by fusing Shenchang differential modeling and a trajectory similar distillation optimization mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and particularly relates to a water quality prediction system and method based on machine learning. Background Art

[0002] At present, water quality prediction, as an important part of water environment monitoring and pollution early warning, has become a key technology in intelligent water conservancy and ecological governance. Traditional water quality prediction methods mostly rely on static models such as linear regression, grey model or support vector regression. These methods have obvious deficiencies in dealing with the non-linear changes and multi-variable coupling relationships in water quality time series data. At the same time, with the popularization of Internet of Things sensor technology, water quality monitoring data presents the characteristics of multi-source heterogeneity and strong time series dependence. Traditional models are increasingly difficult to meet the actual needs in terms of data adaptability, expression ability and prediction accuracy.

[0003] In recent years, deep learning technology has been introduced into the field of water quality prediction. In particular, recurrent neural networks and long short-term memory network models have made certain progress in modeling time series data. However, these models have a large number of parameters and long training time, which is not conducive to deployment on edge devices or real-time operation scenarios. In addition, most current mainstream methods adopt a single model structure, lacking interpretability and model simplification strategies, and it is difficult to balance the contradiction between prediction accuracy and computational efficiency.

[0004] In terms of model optimization, some existing studies have tried to transfer the prediction ability of complex models to simplified models through knowledge distillation technology. However, most methods still rely on the construction of loss functions, such as mean square error or cross entropy, ignoring the dynamic consistency of the overall prediction behavior of the model. At the same time, in the process of structure optimization, the traditional gradient optimization method is still adopted, lacking a behavior-driven parameter regulation mechanism. Especially for time series prediction tasks such as water quality indicators, how to ensure the approximation effect of the simplified model on the dynamic trajectory is the shortcoming of existing methods.

[0005] Therefore, how to provide a water quality prediction system based on machine learning is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] An object of the present invention is to propose a water quality prediction system based on machine learning. The present invention combines a neural ordinary differential equation model, a trajectory similarity-driven optimization mechanism and knowledge distillation technology, and details the realization of efficient modeling and lightweight prediction of multi-source water quality time series data through the behavior consistency guidance between a teacher model and a student model. This method constructs a prediction framework with strong interpretability and flexible structure optimization, and has the advantages of high prediction accuracy, lightweight model structure, low deployment cost and fast response to sudden water quality changes.

[0007] A water quality prediction system based on machine learning according to an embodiment of the present invention includes: A data collection and preprocessing module for collecting multi-source monitoring data from multiple water quality monitoring points and performing preprocessing, where the target water quality index is used as the true label data in the multi-source monitoring data; A teacher model module for constructing a neural ordinary differential equation model as the teacher model and outputting the prediction result of the target water quality index; A student model module for constructing a lightweight neural network model as the student model and performing structure optimization training based on the prediction result and the true label data; A trajectory evaluation module for calculating the similarity of the prediction trajectories of the teacher model and the student model and optimizing the structure parameters of the student model; A model inference module for receiving real-time water quality data and calling the optimized student model to generate the prediction result of the target water quality index; A result output module for outputting the prediction result of the target water quality index and providing display and storage interfaces.

[0008] Optionally, the modules are implemented through the following methods: S1. Collect multi-source monitoring data from multiple water quality monitoring points, where the target water quality index is used as the true label data in the multi-source monitoring data, and perform preprocessing on the multi-source monitoring data to generate a standardized time series training data set; S2. Based on the time series training data set, construct and train a neural ordinary differential equation model as the teacher model and output the prediction results of the target water quality index at each time point; S3. Construct a student model, where the student model is a lightweight neural network with a simplified structure, and input the time series training data set into the student model for training; S4. Use the prediction results of the target water quality index at each time point as soft labels, and at the same time introduce the true label data. By constructing an optimization mechanism based on the similarity of the prediction trajectories of the student model and the teacher model in the time dimension, dynamically perturb and update the structure parameters of the student model; S5. Put the optimized student model into actual water quality prediction operation, receive the latest water quality time series data collected in real time, and output the prediction results of the target water quality index at the corresponding time points; S6. Monitor the prediction error of the student model during the deployment and operation process. When the prediction error exceeds the set threshold, re-execute steps S2 to S4, and perform incremental distillation training on the student model through the teacher model to dynamically improve the adaptability of the student model to sudden water quality changes.

[0009] Optionally, the multi-source monitoring data specifically includes joint time series data of water quality indicators and associated water body physical and chemical parameters, meteorological environment factors, and hydrodynamic information.

[0010] Optionally, the preprocessing of the multi-source monitoring data specifically includes time synchronization, missing value filling, outlier removal, and normalization processing.

[0011] Optionally, S2 specifically includes: S21. Define each input sequence in the generated standardized time series training dataset as , where represents the -th sample at the -th time step of the -dimensional input feature vector, being the set of real numbers; S22. Construct a corresponding hidden state function for each type of target water quality index , where represents the -th prediction variable number, and define the basic evolution form as ; S23. Introduce a coupling mechanism between variables in the state evolution and extend it to the stage 1 evolution form ; S24. Introduce a multi-modal input fusion mechanism to encode features of images, text, or external environment data into a latent vector for initializing the state: , where represents the initial hidden state of the -th variable at the starting time S25. Introduce an attention saliency regulation mechanism based on the stage 1 evolution form to form the final evolution expression ; S26. Configure an ODE solver with a dynamic step size and error tolerance adjustment mechanism to adjust the integration granularity according to the degree of input change, and achieve fine-grained solution in high-variation sections and low-computation solution in stable sections; S27. Input the state vector at the prediction time end point into the corresponding prediction function to output the target prediction value , where is the hidden state vector of the -th variable at the prediction time end point being the prediction function parameter; S28. Based on the generated attention weights , construct a saliency heat map for the entire evolution time interval, screen the time segments that satisfy , and extract the corresponding prediction results as candidate soft labels, where is the attention threshold parameter; S29, combining the predicted change trend with the significant distribution, selecting a representative evolution path for each variable to form a soft label subsequence; S210, using the adjoint sensitivity method to calculate the gradient of each parameter in the final teacher model structure, for all evolution function parameter sets , coupling coefficient , attention network parameters and prediction function parameters Joint training is performed to complete the teacher model construction and output the predicted values of the target water quality indicators at each time point as a soft label sequence.

[0012] Optionally, the S3 specifically includes: S31, constructing a student model, wherein the student model is a lightweight neural network with a simplified structure, comprising an input layer, a time series modeling layer, and an output prediction layer, wherein the input layer receives the generated time series training data set; S32. In the time series modeling layer, a gated recurrent unit network structure is used to extract the dynamic representation of the input features of each time step, and a unified hidden state sequence is constructed to represent the state evolution trajectory in each time period; S33, setting a shallow fully connected network module with adjustable parameters in the output prediction layer, receiving the final value of the hidden state and outputting the prediction result of each target water quality indicator; S34, obtain the predicted value of the teacher model at each time point of the target water quality index as a soft label sequence, correspond one-to-one with the predicted result of the student model at each time step, and calculate the error residual sequence at consecutive time points, recorded as ,in, represents the forecast indicator number, represents the time step; S35. Based on the error residual sequence, construct a learnable residual response weight function , characterizes the response sensitivity of the student model to the error intensity in different time periods. The weight function is one of the state update gating parameters, guiding the time series modeling layer to take different learning steps in different time periods. It is constructed based on the prediction error residual of the student model at each time step; S36, introduce the structure distillation gating mechanism, dynamically adjust the state update frequency of the student model according to the residual response weight function, improve the model response accuracy in the high residual time period, and automatically suppress invalid updates in the low residual time period; S37, the generated error residual sequence and the residual response weight function The results are output to the teacher model feedback interface to dynamically control the teacher model evolution path; S38. Iteratively update the parameters of the student model according to the structure distillation gating mechanism and the residual response weight function, gradually enhancing the ability to perceive residual information and the adaptability to feedback control inputs, and generating a student model with residual response characteristics and regulatory feedback capabilities.

[0013] Optionally, the specific steps of S4 are as follows: S41. Use the prediction results of the target water quality indicators output by the teacher model at each time point as soft label inputs to construct a teacher model prediction trajectory sequence, denoted as , where represents the target predicted value of the th water quality indicator by the teacher model at time point , is the prediction time end point, is the starting time; S42. Use the corresponding prediction results output by the student model in step S3 to form a student prediction trajectory, denoted as , where represents the target predicted value of the student model at the same time point; S43. Construct a time trajectory similarity function to measure the evolutionary consistency between the prediction trajectories of the teacher model and the student model; S44. Introduce real label data for supervised comparison and structural screening reference to improve the conservativeness and precision coverage of structural sampling decisions; S45. Construct a student model structure perturbation generator to perform combinatorial sampling on the core structure parameters of the student model to form multiple alternative structure versions; S46. Perform inference runs on each alternative structure version, calculate the predicted output trajectory, and evaluate the structural performance according to the time trajectory similarity function ; S47. Set a trajectory similarity threshold . If the trajectory similarity of the alternative structure version is higher than the trajectory similarity threshold and shows an improvement compared to the current structure, replace the current structure parameters; otherwise, retain the original structure; S48. Repeat the structure perturbation, trajectory evaluation, and screening steps until the similarity converges or the number of iterations reaches the upper limit, and finally determine the structure parameters of the student model.

[0014] Optionally, the specific steps of S5 are as follows: S51. Use the student model with optimized trajectory as the final structure, load it into the operating environment of the water quality prediction system, and complete the initialization and interface binding configuration of the student model. S52. Receive the latest water quality time series data collected in real time from each water quality monitoring point, which includes historical observation values of target water quality indicators and environmental monitoring features consistent with the structure of the time series training dataset; S53. Perform preprocessing operations on the received water quality time series data, including data standardization, missing value repair, and time step reconstruction, to generate standardized input samples that can be directly input into the student model; S54. Input the standardized input samples into the deployed student model, and let the student model perform structured inference operations to output the prediction results of the target water quality indicators corresponding to the current time point; S55. Bind the prediction results to the corresponding timestamps, output them as the online prediction response of the water quality prediction system, and record the prediction results for feedback analysis and anomaly detection; S56. During the operation of the water quality prediction system, continuously receive and process real-time input data, and periodically call the student model for prediction output.

[0015] The beneficial effects of the present invention are as follows: By constructing a water quality prediction method that combines a neural ordinary differential equation model and a knowledge distillation mechanism, the present invention solves the problems of low accuracy and difficult deployment of existing water quality prediction models when dealing with strongly nonlinear and strongly coupled multi-source time series data. By using the neural ordinary differential equation model as the teacher model, giving full play to its ability in continuous dynamic modeling to generate high-quality prediction trajectories; at the same time, designing a student model with a simplified structure, and introducing a trajectory similarity function to measure and compare the output behaviors of the two, so as to complete the optimization of the structural parameters of the student model without the premise of a traditional loss function. This optimization method based on trajectory consistency breaks through the limitation of the existing knowledge distillation technology's dependence on the loss function, and improves the behavioral interpretability and optimization stability of the model training process.

[0016] In addition, the present invention also realizes the adaptive perception and feedback control of the prediction error of the student model through the residual response mechanism and the structure gating strategy, enabling it to adjust the learning structure according to the real-time error changes during operation, and enhancing its adaptability to sudden water quality changes. Finally, the optimized student model has higher generalization ability and operation efficiency, can be stably deployed in the actual water quality prediction environment, and supports continuous and high-precision prediction of target water quality indicators. Compared with the existing technology, the present invention has achieved obvious improvements in terms of model lightweight, prediction accuracy, dynamic adjustability, and actual deployment feasibility, and has good engineering application prospects and promotion value. Description of the Drawings

[0017] The accompanying drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the accompanying drawings: Figure 1 is a flowchart of a water quality prediction method based on machine learning proposed by the present invention; Figure 2 is a schematic diagram of optimizing the structure based on trajectory similarity between the teacher model and the student model of a water quality prediction method based on machine learning proposed by the present invention. Detailed implementation manners

[0018] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0019] A water quality prediction system based on machine learning includes: A data acquisition and preprocessing module, which is used to collect multi-source monitoring data of multiple water quality monitoring points and perform preprocessing. In the multi-source monitoring data, the target water quality index is used as the true label data; A teacher model module, which is used to construct a neural ordinary differential equation model as the teacher model and output the prediction results of the target water quality index; A student model module, which is used to construct a lightweight neural network model as the student model and perform structure optimization training based on the prediction results and the true label data; A trajectory evaluation module, which is used to calculate the similarity of the prediction trajectories of the teacher model and the student model and optimize the structure parameters of the student model; A model inference module, which is used to receive real-time water quality data and call the optimized student model to generate the prediction results of the target water quality index; A result output module, which is used to output the prediction results of the target water quality index and provide display and storage interfaces.

[0020] Reference Figure 1 and Figure 2 In this embodiment, the modules are implemented through the following methods: S1. Collect multi-source monitoring data of multiple water quality monitoring points. In the multi-source monitoring data, the target water quality index is used as the true label data, and preprocess the multi-source monitoring data to generate a standardized time series training data set; S2. Based on the time series training data set, construct and train a neural ordinary differential equation model as the teacher model, and output the prediction results of the target water quality index at each time point; S3. Construct a student model. The student model is a lightweight neural network with a simplified structure, and input the time series training data set into the student model for training; S4. Use the prediction results of the target water quality indicators at each time point as soft label inputs, and at the same time introduce real label data. By constructing an optimization mechanism based on the similarity between the prediction trajectories of the student model and the teacher model in the time dimension, dynamically perturb and update the structural parameters of the student model; S5. Put the optimized student model into actual water quality prediction operation, receive the latest water quality time series data collected in real time, and output the prediction results of the target water quality indicators at the corresponding time points; S6. Monitor the prediction error of the student model during the deployment and operation process. When the prediction error exceeds the set threshold, re-execute steps S2 to S4, and perform incremental distillation training on the student model through the teacher model to dynamically improve the adaptability of the student model to sudden water quality changes.

[0021] In the present invention, by introducing the neural ordinary differential equation model as the teacher model, it is possible to accurately model the dynamic change laws in multi-source water quality time series data and output continuous, smooth, and physically consistent prediction trajectories. By constructing a lightweight student model with a simplified structure, combining the soft labels output by the teacher model and the real label data, and using an optimization mechanism based on trajectory similarity to dynamically perturb and update the structure of the student model, an efficient knowledge transfer process without relying on traditional loss functions is realized, and the prediction ability of the student model under the condition of structural simplification is improved. At the same time, in the present invention, the online prediction output of real-time water quality data is realized by deploying the optimized student model, and the prediction error is continuously monitored during the operation process. When the model responds insufficiently to sudden water quality changes, the incremental distillation training process can be automatically triggered to achieve adaptive correction of the student model, significantly enhancing the robustness and generalization ability of the model in complex environments. The overall solution has the advantages of high prediction accuracy, flexible structure, low deployment cost, and strong self-learning ability, and is suitable for constructing an intelligent water quality prediction application scenario with continuous update ability.

[0022] In this embodiment, the multi-source monitoring data specifically includes the joint time series data of water quality indicators and associated water body physical and chemical parameters, meteorological environment factors, and hydrodynamic information.

[0023] In this embodiment, the preprocessing of the multi-source monitoring data specifically includes time synchronization, missing value filling, outlier removal, and normalization processing. The time synchronization, missing value filling, outlier removal, and normalization processing specifically refer to performing time synchronization on data from different sources to ensure accurate alignment of data at each monitoring point at the same time step; then using linear interpolation or statistical methods to fill in the missing values to ensure data continuity; then removing noise data through outlier detection means to improve the sample quality; finally, performing normalization processing on all features to make the data distributions of all dimensions consistent, so as to provide standardized and structurally unified time series input data for subsequent model training.

[0024] In this embodiment, S2 specifically includes: S21. Define each input sequence in the generated standardized time series training dataset as , where represents the -th sample at the -th time step of the -dimensional input feature vector, and is the set of real numbers; S22. For each type of target water quality index , construct the corresponding hidden state function , where represents the -th prediction variable number, and define the basic evolution form as : ; where is the differential function of the -th variable, and is the set of differential function parameters; The basic evolution form formula defines the dynamic state change law of the neural ordinary differential equation model for each type of target water quality index. Its practical significance lies in constructing a hidden state function that can model the continuous change over time to describe the potential trend of water quality indicators during the evolution process. By mapping water quality monitoring data into hidden states that change over time, the internal driving force of water quality change is characterized. Compared with traditional discrete time series models, it can simulate the state evolution at any time point, with higher time resolution and modeling accuracy. Among them, the hidden state vector represents the internal representation of the m-th prediction target at time t, and the evolution function reflects how this state changes over time and is controlled by trainable parameters for its dynamic behavior. This formula lays the foundation for the time series modeling of the teacher model, and the subsequent introduced coupling mechanism, multi-modal initialization, and attention regulation are all carried out on this basis. Through this continuous modeling method, the mathematical construction of the evolution path of water quality indicators is realized, providing a theoretical support and dynamic expression framework for achieving accurate, smooth, and interpretable prediction outputs.

[0025] S23. Introduce a coupling mechanism between variables in the state evolution and expand it to the stage 1 evolution form : ; where is the coupling coefficient of the -th variable to the -th variable, is the total number of target prediction variables, is the hidden state vector of the j-th variable at time t; Evolution form of stage 1 The formula introduces a coupling mechanism between variables on the basis of basic neural ordinary differential modeling. Its practical significance lies in simulating the dynamic influence relationship between different water quality indicators during the time series evolution process. In the real environment, water quality indicators often do not change independently, but show complex coupling characteristics among multiple variables. For example, there are significant interaction effects among indicators such as dissolved oxygen, ammonia nitrogen, and chemical oxygen demand. By introducing the linear combination of the states of other variables into the state evolution equation of each predicted variable, a coupling propagation path between multi-dimensional states is constructed, so as to realize the modeling of pollutant diffusion, reaction, and linkage effects in the water body system. Each coupling coefficient represents the influence intensity of one variable on the evolution of another variable. The model automatically learns this coupling weight through training, enabling the causal or synergistic relationship between variables to be dynamically captured during the prediction process. This mechanism significantly enhances the model's ability to express the physical rationality and system linkage of the pollution diffusion process, improves the overall prediction accuracy, and also provides a more structured modeling basis for the subsequent introduction of attention mechanism and trajectory optimization. It expands the neural ordinary differential model from single-variable modeling to a dynamic system model with multi-variable collaborative modeling ability.

[0026] S24. Introduce a multi-modal input fusion mechanism to form a latent vector by feature encoding of images, texts, or external environmental data for initializing the state: , where represents the initial hidden state of the m-th variable at the starting time moment; S25. Introduce an attention significance regulation mechanism on the basis of the evolution form of stage 1 to form the final evolution expression : ; where, is the attention weight of the -th target indicator at the time point ; Final evolution expression The formula introduces an attention significance regulation mechanism based on coupled modeling. Its practical significance lies in realizing the dynamic adjustment of the influence weights in different time periods during the evolution process of water quality indicators, thereby enhancing the model's response ability to key time segments. Specifically, as a time-varying adjustment factor, the attention weight function can automatically determine which time points are more important for the prediction result according to the temporal characteristics of the input state, and give greater modeling attention to these critical moments. This mechanism effectively alleviates the problems of information dilution and noise interference in long-term time series prediction, enabling the model to focus on learning the time segments that are decisive for the change trend of the target variable, thereby improving the prediction accuracy and the interpretability of the model. At the same time, this significance regulation mechanism also has a certain visualization ability. Through the attention distribution, it can assist in analyzing the highly sensitive time periods when pollution events occur, which is helpful for the tracing and intervention of environmental management departments. The finally constructed evolution structure integrates three dynamic factors: the evolution of variables themselves, the coupling between variables, and significance control, making the neural ordinary differential model not only have the continuous modeling characteristics of strong expressive ability, but also take into account time selectivity and feature focus, reflecting the behavior-driven intelligent modeling idea, which is the core basis for achieving high-quality trajectory output and distillation optimization.

[0027] S26. Configure an ODE solver with a dynamic step size and error tolerance adjustment mechanism to adjust the integration granularity according to the degree of input change, and achieve fine-grained solution in high-variation sections and low-computation solution in stable sections; S27. Input the state vector at the prediction time end point into the corresponding prediction function to output the target prediction value , where is the hidden state vector of the m-th variable at the prediction time end point , and are the parameters of the prediction function; S28. Based on the generated attention weights , construct a significance heat map for the entire evolution time interval, screen out the time segments that satisfy , and extract the corresponding prediction results as candidate soft labels, where is the attention threshold parameter; S29. Combine the prediction change trend and significance distribution, select representative evolution paths for each variable to form a soft label subsequence; S210. Use the adjoint sensitivity method to calculate the gradients of the parameters in the final teacher model structure, and for all sets of evolution function parameters , coupling coefficients , attention network parameters, and prediction function parameters Conduct joint training to complete the construction of the teacher model and output the predicted values of the target water quality indicators at each time point as the soft label sequence.

[0028] In the present invention, by constructing a neural ordinary differential equation model with dynamic modeling capabilities, the continuous expression and high-precision prediction of the temporal evolution process of water quality indicators are realized. During the construction of the teacher model, the independent evolution trajectories of each type of water quality indicator are first defined, and then a coupling mechanism between variables is introduced to model the dynamic correlation relationship between multiple factors, effectively improving the model's ability to depict complex pollution diffusion behaviors. By integrating image, text, and external environmental data, the initial state information source of the model is enriched, and the adaptability to multi-modal scenarios is enhanced. At the same time, a time attention mechanism is introduced into the evolution path to realize the significance regulation of key time periods, enabling the model to focus on the time segments that have the greatest impact on the prediction task. The representative prediction results are screened by using the attention heat map to form high-quality soft labels, and the effective paths are further screened through the trajectory change trend, significantly improving the representativeness and compression quality of the soft labels. Finally, the adjoint sensitivity method is used for efficient backpropagation training to realize the joint optimization of the model structure parameters. This solution not only improves the continuity, interpretability, and controllability of the predicted trajectory of the teacher model but also provides a behavior-driven structural guidance for subsequent distillation training, having the beneficial effects of high accuracy, fine modeling, and scientific path screening.

[0029] In this embodiment, S3 specifically includes: S31. Construct a student model, which is a lightweight neural network with a simplified structure, including an input layer, a temporal modeling layer, and an output prediction layer, where the input layer receives the generated time series training dataset; S32. In the temporal modeling layer, adopt a gated recurrent unit network structure to extract the dynamic representation of the input features at each time step and construct a unified hidden state sequence to represent the state evolution trajectory within each time period; S33. In the output prediction layer, set a shallow fully connected network module with adjustable parameters to receive the final value of the hidden state and output the prediction results of each target water quality indicator; S34. Obtain the predicted values of the teacher model for the target water quality indicators at each time point as the soft label sequence, correspond them one by one with the prediction results of the student model at each time step, and calculate the error residual sequence at consecutive time points, denoted as where, represents the prediction index number, Denote the time step. The error residual sequence calculated at consecutive time points specifically refers to comparing the predicted target water quality index values output by the teacher model at each time point with the prediction results of the student model at the same time point, and calculating the difference value at each time step using methods such as Euclidean distance, absolute error, or mean square error to form an error sequence with respect to the time step t. For each prediction index number k, the error residual sequence characterizes the performance deviation of the student model compared to the teacher model throughout the prediction time window, thereby providing an accurate error information basis for subsequent residual-aware mechanisms and dynamic learning strategies; S35. Based on the error residual sequence, construct a learnable residual response weight function , which characterizes the response sensitivity of the student model to the error intensity in different time periods. The weight function, as one of the state update gating parameters, guides the time series modeling layer to adopt different learning strides in different time periods. It is constructed based on the prediction error residuals of the student model at each time step. The characterization of the response sensitivity of the student model to the error intensity in different time periods specifically refers to analyzing the prediction error residuals of the student model at each time step and using a function mapping method to construct a non-linear relationship between the time step and the response intensity to form a learnable weight vector. This weight vector reflects the degree of attention of the model to high-residual time periods. A higher weight value indicates that the model should respond more actively to the error signal in this time period and enhance the learning intensity, while a lower weight value indicates that the error information in this time period has a lower training value for the model and the update frequency needs to be reduced, thereby dynamically adapting the learning strategy for different time periods; S36. Introduce a structural distillation gating mechanism to dynamically adjust the state update frequency of the student model according to the residual response weight function, improve the model response accuracy in high-residual time periods, and automatically suppress ineffective updates in low-residual time periods. The dynamic adjustment of the state update frequency of the student model according to the residual response weight function specifically refers to, during the model time series modeling process, taking the residual response weight corresponding to each time step as the gating coefficient and introducing it into the state transition function or gating structure to determine whether to perform a state update operation according to the magnitude of this weight value: when the residual response weight is greater than the set threshold, allow the state unit to perform a full update to enhance the fitting of the error information; when the weight is small, skip or partially retain the current state to reduce the sensitivity to redundant data, thereby realizing the dynamic control of the learning frequency and the efficient utilization of resources; S37. The generated error residual sequence and the residual response weight function Output them to the teacher model feedback interface together to dynamically control the evolution path of the teacher model. The dynamic control of the evolution path of the teacher model specifically refers to inputting the error residual sequence output by the student model and the corresponding residual response weight function as feedback signals into the teacher model, and dynamically correcting the state evolution intensity and focus of attention at different time steps by adjusting the coupling coefficient, attention weight or input sensitivity parameter in its state evolution equation, so as to enhance the expression ability of the teacher model in critical time periods, optimize its modeling accuracy for sudden changes or abnormal trends, and improve the overall soft label quality and distillation guidance effect; S38. Iteratively update the parameters of the student model according to the structure distillation gating mechanism and the residual response weight function, gradually enhance the perception ability of residual information and the adaptability to feedback control input, and generate a student model with residual response characteristics and regulation feedback ability. The iterative update of the parameters of the student model specifically refers to using the error residual sequence and the residual response weight function as guiding signals in each round of training, dynamically adjusting the state update frequency and gradient propagation path in the time series modeling layer, and gradually adjusting the weight matrix, bias term and gating parameters in the model, so that the model enhances the feature extraction ability in high-residual sections and suppresses ineffective learning in low-residual sections, thereby continuously optimizing the model's perception ability of error signals and response adaptation ability to teacher feedback in multiple rounds of iteration.

[0030] The present invention solves the problems that traditional distillation models are insensitive to local error responses in time series and the learning process is rigid by constructing a lightweight student model with residual perception and dynamic feedback regulation capabilities. A gated recurrent unit network is introduced into the student model structure as the time series modeling layer to effectively extract dynamic features at each time step, and the prediction output is realized through a shallow fully connected network. By calculating the error residuals between the student model and the teacher model at each time point, a residual response weight function is constructed to clarify the sensitivity of the model to errors in different time periods, and guide the model to dynamically adjust the step size and state update intensity during the learning process. The structure distillation gating mechanism further utilizes this weight function to enhance the model's reaction ability in high-error sections and suppress ineffective updates in low-error sections, thereby achieving precise and energy-saving learning regulation. At the same time, the residual information and the weight function are also fed back to the teacher model for dynamically adjusting the evolution path, and an adaptive bidirectional regulation mechanism between the teacher and the student is constructed. Generally speaking, the present invention improves the local response ability of the student model to error changes and the structure self-adjustment ability, realizes a more stable, flexible and efficient knowledge transfer process, and has beneficial effects such as high prediction accuracy, fast model convergence and friendly deployment.

[0031] In this embodiment, the specific steps of S4 include: S41. Use the prediction results of the target water quality indicators output by the teacher model at each time point as soft label inputs to construct the prediction trajectory sequence of the teacher model, denoted as , where represents the target predicted value of the teacher model for the th water quality indicator at time point , is the end point of the prediction time, is the starting time; S42. Construct the student prediction trajectory from the corresponding prediction results output by the student model in step S3, denoted as , where represents the target predicted value of the student model at the same time point; S43. Construct the time trajectory similarity function to measure the evolutionary consistency between the prediction trajectories of the teacher model and the student model: ; Among them, the closer the similarity value is to 1, the more consistent the two trajectories are; The time trajectory similarity function The actual meaning of the formula is to provide a non-loss function-oriented behavior evaluation index for the structural optimization of the student model. Traditional optimization methods mostly rely on point-to-point prediction errors (such as mean square error), ignoring the overall prediction trend and change law of the model in the time dimension. By constructing the prediction results of the teacher model and the student model into complete time series trajectories respectively, and introducing the normalized squared difference ratio to construct the similarity function, the global matching degree of the two in the time evolution path can be comprehensively reflected. The closer the similarity value is to 1, the more consistent the shapes and change trends of the two trajectories are, and the more effectively the student model can fit the behavior of the teacher model. Therefore, this function not only has interpretability at the behavioral level, but also has the ability to quantitatively compare the results of structural adjustment, and is applicable to multiple processes such as structural sampling optimization, soft label screening, and path-driven distillation. Through this function, the student model can achieve fine-grained regulation based on the trajectory behavior performance, effectively improving the overall performance and response consistency of the model in the time series prediction task.

[0032] S44. Introduce real label data for supervised comparison and structural screening reference to improve the conservativeness and precision coverage of the structural sampling decision; S45. Construct a student model structure perturbation generator to perform combinatorial sampling on the core structural parameters of the student model to form multiple alternative structure versions. The combinatorial sampling of the core structural parameters of the student model specifically refers to, on the premise of keeping the overall structural framework of the student model unchanged, selecting its key structural parameters such as the number of hidden layer units, the depth of the time series modeling layer, the type of activation function, and the gating mechanism parameters, and using strategies such as Bayesian optimization or random search to generate multiple groups of parameter combinations, thereby constructing multiple candidate versions of the student model with different structural characteristics; S46. Perform inference runs on each alternative structure version, calculate the predicted output trajectory, and evaluate the structural performance according to the time trajectory similarity function The performing inference runs on each alternative structure version and calculating the predicted output trajectory means inputting the standardized time series training data into the alternative student model structure in sequence. The model gradually completes the forward propagation according to the set network layer parameters and activation functions, and generates a corresponding sequence of predicted values of the target water quality indicators. Subsequently, a complete output trajectory is constructed based on the prediction results at each time step for subsequent one-to-one comparison and similarity evaluation with the teacher model prediction trajectory in the time dimension; S47. Set a trajectory similarity threshold , and if the trajectory similarity of the alternative structure version is higher than the trajectory similarity threshold and shows an improvement compared to the current structure, replace the current structural parameters; otherwise, retain the original structure; S48. Repeat the steps of structure perturbation, trajectory evaluation, and screening until the similarity converges or the number of iterations reaches the upper limit, and finally determine the structural parameters of the student model.

[0033] The present invention constructs a structure optimization mechanism guided by trajectory similarity, which solves the problem of lacking behavioral-level evaluation in the adjustment of the student model structure during the traditional distillation process. By constructing a complete trajectory based on the prediction results of the teacher model and the student model at consecutive time points, and calculating the trajectory similarity between the two over the entire time period, it replaces the traditional optimization method relying on the loss function and improves the consistency expression ability at the model behavioral level. At the same time, the system introduces real label data as an auxiliary reference for structure screening, enhancing the supervision and robustness of the structure perturbation screening process. During the optimization process of the student model, a structure perturbation generator is used to perform combinatorial sampling on the model structure parameters to form multiple alternative versions, and each version is evaluated for its behavior through a trajectory similarity function. By setting a trajectory similarity threshold for structure selection, it is ensured that the retained structure version has better behavior fitting performance. Finally, through repeated iteration, the structure of the student model continuously approaches the behavior performance of the teacher model without relying on the loss function. This mechanism significantly improves the behavior consistency, structural flexibility, and optimization efficiency of the student model, and has the advantages of strong interpretability, no gradient dependence, and being suitable for multi-structure space exploration. It is suitable for deployment in water quality prediction tasks that emphasize both the model response ability and structural compactness.

[0034] In this embodiment, step S5 specifically includes: S51. Use the student model with optimized trajectory as the final structure, load it into the operating environment of the water quality prediction system, and complete the initialization and interface binding configuration of the student model; S52. Receive the latest water quality time series data collected in real time from each water quality monitoring point, which includes historical observation values of target water quality indicators and environmental monitoring features consistent with the structure of the time series training data set; S53. Perform preprocessing operations on the received water quality time series data, including data standardization, missing value repair, and time step reconstruction, to generate standardized input samples that can be directly input into the student model. The data standardization, missing value repair, and time step reconstruction refer to performing normalization processing on the received water quality time series data using Z-score standardization, then using the K-nearest neighbor interpolation method to repair missing values, and finally reconstructing the input sample sequence with a unified time step through the sliding window technique; S54. Input the standardized input samples into the deployed student model, and let the student model perform structured inference operations to output the prediction results of the target water quality indicators corresponding to the current time point; S55. Bind the prediction results to the corresponding timestamps, output them as the online prediction response of the water quality prediction system, and record the prediction results for feedback analysis and anomaly detection; S56. During the operation of the water quality prediction system, continuously receive and process real-time input data, and periodically call the student model for prediction output.

[0035] By putting the student model with optimized trajectory into the actual operating environment, the present invention realizes the efficient and continuous prediction of water quality indicators, and solves the problems of difficult deployment and slow response of traditional models. After structural optimization, the student model has the characteristics of lightweight, which is convenient for rapid deployment and low-cost operation in water quality prediction scenarios. During actual operation, the system can receive the latest time series data from each water quality monitoring point in real time, and combine preprocessing mechanisms such as standardization, missing value repair, and time step reconstruction to ensure the consistency and integrity of the input data quality, thus ensuring the stability of the model prediction accuracy. After receiving the standardized input, the student model can quickly output the prediction results of the target water quality indicators at the current time point, meeting the requirements of online prediction for response speed and real-time performance. The prediction results generate a continuous response stream in combination with timestamps, and are recorded for subsequent error analysis and anomaly identification, providing a data basis for the subsequent adaptive update mechanism. The entire operation process has high automation and scalability, and can support water quality prediction tasks with multiple monitoring points and multiple variables. The present invention significantly improves the real-time performance, stability, and deployment flexibility of the prediction process, and is applicable to various actual application scenarios such as dynamic water quality supervision, pollution warning, and smart water services.

[0036] In this embodiment, the step S6 specifically includes: The system continuously monitors the prediction output of the student model during actual operation and compares it with the real observed water quality data. When it is detected that the prediction errors at consecutive time steps exceed the preset threshold, it is determined as a sign of sudden water quality change or model performance degradation. At this time, the model retraining mechanism will be automatically triggered, and the processes of teacher model construction and training, student model construction and initial training, and trajectory similarity optimization will be executed again. By introducing the latest data and performing incremental learning on the student model based on the distillation optimization strategy, it is ensured that the student model can quickly adapt to environmental changes, improve prediction accuracy and robustness while maintaining a lightweight structure, thus ensuring the stable operation and efficient response of the system under complex water quality dynamic conditions.

[0037] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the water quality monitoring of a river in a certain area, and a two-week water quality monitoring application test is carried out. 10 representative monitoring points are selected, and the goal is to predict the changes in the concentration of key indicators in the water body in real time, including chemical oxygen demand, ammonia nitrogen, and total phosphorus. The data collection frequency of the monitoring points is set to once per hour, and each data point records parameters such as historical water quality indicator values, environmental temperature, rainfall, and water flow velocity at the same time, constituting a complete multi-source time series input sample.

[0038] To evaluate the performance by comparison, a traditional LSTM model and the lightweight student model optimized by trajectory distillation in the present invention were respectively constructed and deployed. Under the same conditions, the two types of models were trained based on historical monitoring data for the same time period and received the same real-time data for prediction during subsequent actual operation.

[0039] During the test period, the traditional model generally showed the defects of slow response speed and poor adaptability to drastic water quality changes. Especially under the condition of multi-variable coupled fluctuations, the prediction error increased significantly. The model of the present invention optimizes the structure through the output trajectory of the teacher model and combines the attention mechanism and residual gating adjustment, enabling the student model to accurately fit the changing trend of the target indicators and showing higher prediction sensitivity at critical times.

[0040] In addition, under mutation conditions such as continuous rainfall and sudden pollution input, the model of the present invention can react in time and predict water quality fluctuations, successfully capturing multiple abnormal change trends, while the traditional model generally has problems of delayed response or misjudgment. This test verifies the accuracy of the present invention in time series modeling, the sensitivity of dynamic change response, and its practicability in the edge deployment environment, and is applicable to the accurate identification and real-time early warning of high-frequency change trends in actual water quality monitoring scenarios.

[0041] Table 1 Performance comparison table of water quality prediction models

[0042] According to the results shown in the data of Table 1, it can be clearly seen that the model of the present invention has significant advantages over the traditional model in key performance indicators. First, in terms of the average prediction error, the prediction errors of the traditional model at all 10 monitoring points are between 0.78 and 0.91 mg / L, showing a certain upward trend. Especially at points such as M2, M4, and M5, the errors exceed 0.85 mg / L, indicating that the model lacks accuracy in the face of complex fluctuation conditions. Under the same data input and scenario conditions, the average prediction error of the model of the present invention generally remains in the lower range of 0.36 to 0.43 mg / L, with stronger stability and smaller error fluctuation amplitude, proving its precision advantage in fitting the time evolution trend of the target water quality indicators.

[0043] Secondly, the difference is particularly obvious in terms of response time. The response time of the traditional model is generally above 250 milliseconds, reaching a maximum of 265 milliseconds, resulting in possible problems of delayed output or response accumulation when processing high-frequency data streams. The model of the present invention, with the help of the lightweight network structure and structure distillation optimization strategy, greatly shortens the inference time. The response time at all monitoring points is controlled within 100 milliseconds, and at multiple points it is below 95 milliseconds, significantly improving the real-time performance and computing efficiency, and is especially suitable for edge computing or water quality monitoring terminals deployed with limited resources.

[0044] Generally speaking, the model proposed by the present invention is superior to the traditional water quality prediction method in terms of two key dimensions of accuracy and real-time performance. Moreover, its performance at each monitoring point is consistent, with strong error control ability, greatly enhancing the prediction ability for sudden water quality situations and the adaptability to practical applications. This performance advantage verifies the practical value of the trajectory similarity optimization mechanism, residual regulation, and dynamic gating mechanism in water quality prediction tasks, providing technical support for constructing an intelligent water quality prediction model with high responsiveness, high accuracy, and low computational load.

[0045] The above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered by the protection scope of the present invention.

Claims

1. A water quality prediction system based on machine learning, characterized in that, Including: A data acquisition and preprocessing module, which is used to acquire multi-source monitoring data of multiple water quality monitoring points and perform preprocessing. In the multi-source monitoring data, the target water quality index is used as the true label data; A teacher model module, which is used to construct a neural ordinary differential equation model as the teacher model and output the prediction results of the target water quality index; A student model module, which is used to construct a lightweight neural network model as the student model and perform structure optimization training based on the prediction results and the true label data; A trajectory evaluation module, which is used to calculate the similarity of the prediction trajectories between the teacher model and the student model and optimize the structural parameters of the student model; A model inference module, which is used to receive real-time water quality data and call the optimized student model to generate the prediction results of the target water quality index; A result output module, which is used to output the prediction results of the target water quality index and provide display and storage interfaces.

2. A water quality prediction method based on machine learning, applied to the water quality prediction system based on machine learning described in claim 1, characterized in that, Including the following steps: S1. Acquire multi-source monitoring data of multiple water quality monitoring points. In the multi-source monitoring data, the target water quality index is used as the true label data, and preprocess the multi-source monitoring data to generate a standardized time series training data set; S2. Based on the time series training data set, construct and train a neural ordinary differential equation model as the teacher model and output the prediction results of the target water quality index at each time point; S3. Construct a student model. The student model is a lightweight neural network with a simplified structure, and input the time series training data set into the student model for training; S4. Use the prediction results of the target water quality index at each time point as soft labels, and at the same time introduce the true label data. By constructing an optimization mechanism based on the similarity between the prediction trajectories of the student model and the teacher model in the time dimension, dynamically perturb and update the structural parameters of the student model; S5. Put the optimized student model into actual water quality prediction operation, receive the latest water quality time series data collected in real time, and output the prediction results of the target water quality index at the corresponding time points; S6. Monitor the prediction error of the student model during the deployment and operation process. When the prediction error exceeds the set threshold, re-execute steps S2 to S4, and perform incremental distillation training on the student model through the teacher model to dynamically improve the adaptability of the student model to sudden water quality changes.

3. The water quality prediction method based on machine learning according to claim 2, characterized in that The multi-source monitoring data specifically includes the joint time series data of water quality indicators and associated water body physical and chemical parameters, meteorological environment factors, and hydrodynamic information.

4. A water quality prediction method based on machine learning according to claim 2, characterized in that, The preprocessing of the multi-source monitoring data specifically includes time synchronization, missing value filling, outlier removal, and normalization processing.

5. The water quality prediction method based on machine learning according to claim 2, wherein, The specific content of S2 includes: S21. Define each input sequence in the generated standardized time series training dataset as , where represents the -th sample at the -th time step of the -dimensional input feature vector, being the set of real numbers; S22. For each type of target water quality index Construct the corresponding hidden state function , where represents the th predictor variable number, and the basic evolution form is defined as ; S23. Introduce the coupling mechanism between variables in the state evolution and extend it to the evolution form of Stage 1 ; S24. Introduce a multimodal input fusion mechanism to form latent vectors by feature encoding of images, texts, or external environmental data for initializing the state: , where: where represents the initial hidden state of the m-th variable at the starting time moment; S25. Introduce an attention saliency regulation mechanism based on the evolved form in Phase 1 to form the final evolved expression ; S26. Configure an ODE solver with a dynamic step size and error tolerance adjustment mechanism, adjust the integration granularity according to the degree of input change, and achieve fine-grained solution in high-variation sections and low-computation solution in stable sections; S27. Input the state vector at the predicted time endpoint into the corresponding prediction function to output the target predicted value , where is the hidden state vector of the m-th variable at the predicted time endpoint , and are the prediction function parameters; S28. Based on the generated attention weights , construct a saliency heat map for the entire evolution time interval, and screen out the time segments that satisfy . Extract the corresponding prediction results as candidate soft labels, where is the attention threshold parameter; S29. Combine the prediction change trend and significance distribution, select representative evolution paths for each variable, and form a soft label subsequence; S210. Calculate the gradients of the parameters in the final teacher model structure using the adjoint sensitivity method for all sets of evolutionary function parameters and coupling coefficients jointly train the attention network parameters and prediction function parameters complete the construction of the teacher model, and output the predicted values of the target water quality indicators at each time point as the soft label sequence.

6. The water quality prediction method based on machine learning according to claim 5, wherein, The specific content of S3 includes: S31. Construct a student model. The student model is a lightweight neural network with a simplified structure, including an input layer, a time series modeling layer, and an output prediction layer, where the input layer receives the generated time series training data set; S32. Adopt a gated recurrent unit network structure in the temporal modeling layer to extract the dynamic representation of the input features at each time step and construct a unified hidden state sequence to represent the state evolution trajectory within each time period; S33. Set a shallow fully connected network module with adjustable parameters in the output prediction layer to receive the final value of the hidden state and output the prediction results of each target water quality index; S34. Obtain the predicted values of the teacher model for the target water quality indicators at each time point as the soft label sequence, and correspond them one by one with the prediction results of the student model at each time step. Calculate the error residual sequence at consecutive time points, denoted as , where represents the prediction index number, represents the time step; S35. Construct a learnable residual response weight function based on the error residual sequence , to characterize the response sensitivity of the student model to the error intensity in different time periods. The weight function, as one of the state update gating parameters, guides the time series modeling layer to adopt different learning strides in different time periods and is constructed according to the prediction error residuals of the student model at each time step; S36. Introduce a structure distillation gating mechanism to dynamically adjust the state update frequency of the student model according to the residual response weight function, improve the model response accuracy in high-residual time periods, and automatically suppress invalid updates in low-residual time periods; S37. Output the generated error residual sequence together with the residual response weight function to the teacher model feedback interface to dynamically regulate the evolution path of the teacher model; S38. According to the structure distillation gating mechanism and the residual response weight function, iteratively update the parameters of the student model, gradually enhance the perception ability of residual information and the adaptation ability to feedback control inputs, and generate a student model with residual response characteristics and regulatory feedback capabilities.

7. A water quality prediction method based on machine learning according to claim 6, characterized in that, The specific content of S4 includes: S41. Use the prediction results of the target water quality indicators output by the teacher model at each time point as soft label inputs to construct a teacher model prediction trajectory sequence, denoted as , where represents the target predicted value of the th water quality indicator by the teacher model at time point , is the prediction time end point, is the start time; S42. Construct the corresponding prediction results output by the student model in step S3 into a student prediction trajectory, denoted as , where represents the target prediction value of the student model at the same time point; S43. Construct a time trajectory similarity function , and measure the evolutionary consistency between the prediction trajectories of the teacher model and the student model; S44. Introduce real label data for supervised comparison and structure screening reference to improve the conservativeness and accuracy coverage of structure sampling decisions; S45. Construct a student model structure perturbation generator to perform combined sampling on the core structure parameters of the student model to form multiple alternative structure versions; S46. Perform an inference run on each alternative structural version, calculate the predicted output trajectory, and evaluate the structural performance according to the temporal trajectory similarity function Evaluate the structural performance; S47. Set the trajectory similarity threshold , if the trajectory similarity of the alternative structure version is higher than the trajectory similarity threshold and shows an improvement compared to the current structure, then replace the current structure parameters; otherwise, retain the original structure; S48. Repeat the structure perturbation, trajectory evaluation, and screening steps until the similarity converges or the number of iterations reaches the upper limit, and finally determine the structure parameters of the student model.

8. A water quality prediction method based on machine learning according to claim 7, characterized in that, The specific content of S5 includes: S51. Use the student model with optimized trajectory as the final structure, load it into the water quality prediction system operating environment, and complete the initialization and interface binding configuration of the student model; S52. Receive the latest water quality time series data collected in real time from each water quality monitoring point, which includes historical observed values of target water quality indicators and environmental monitoring features consistent with the structure of the time series training data set; S53. Perform preprocessing operations on the received water quality time series data, including data standardization, missing value repair, and time step reconstruction, to generate standardized input samples that can be directly input into the student model; S54. Input the standardized input samples into the deployed student model, and the student model performs structured inference operations to output the prediction results of the target water quality indicators corresponding to the current time point; S55. Bind the prediction results to the corresponding timestamps, output them as the online prediction response of the water quality prediction system, and record the prediction results for feedback analysis and anomaly detection; S56. During the operation of the water quality prediction system, continuously receive and process real-time input data, and periodically call the student model for prediction output.

Images