Algae community structure change prediction algorithm and system based on multi-source data fusion

Abstract

The invention relates to the cross technical field of artificial intelligence and environment monitoring, and discloses an algal community structure change prediction algorithm and system based on multi-source data fusion, and the algorithm comprises the steps: obtaining water quality, weather and plankton multi-source time sequence data; performing time alignment and missing value interpolation; eliminating and screening key environment factors through recursive features; performing dynamic weighted fusion on the multi-modal features by using a space-time attention mechanism; inputting a three-layer stacked LSTM network to output future algae dominant species abundance prediction; and model parameters are corrected on line based on measured data. The system comprises a multi-source data acquisition module, a preprocessing module, a key factor extraction module, a space-time attention fusion module, a dynamic prediction module and an adaptive correction module. According to the method, the prediction accuracy and stability are remarkably improved, and algal bloom early warning and ecological regulation are effectively supported.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and environmental monitoring, specifically involving an algorithm and system for predicting changes in algal community structure based on multi-source data fusion. Background Technology

[0002] With the increasing demand for aquatic ecosystem monitoring and eutrophication control, the prediction of algal community structure changes has become a crucial link in water quality early warning and ecological management. As sensitive indicator organisms of aquatic environmental changes, the dynamics of algal community composition and abundance are driven by multiple factors, including water quality physicochemical parameters, meteorological conditions, hydrological characteristics, and biological interactions. Traditional prediction algorithms typically construct statistical regression models or simple machine learning models based on a single data source (such as chlorophyll a concentration or total phosphorus content), which struggle to comprehensively characterize the multi-factor coupling mechanisms in complex aquatic environments. This results in weak generalization ability and poor robustness of prediction results across spatiotemporal scales.

[0003] Dynamic modeling of algal communities based on multi-source heterogeneous data fusion has become an important technological direction for improving prediction accuracy. This direction aims to integrate heterogeneous data from multiple channels such as sensor networks, remote sensing imagery, weather stations, and biological sampling, and achieve a panoramic perception of the aquatic environment through unified representation learning. However, existing technologies generally face challenges in multi-source data fusion, such as inconsistent feature dimensions, large differences in temporal resolution, and spatial coverage mismatch, and lack the ability to effectively model the spatiotemporal dependence between environmental factors and algal responses.

[0004] While deep learning models have been introduced for algae prediction, they often employ static fusion strategies or simple concatenation methods to handle multimodal inputs, failing to fully exploit the correlations between different data sources in terms of temporal evolution and spatial distribution. Furthermore, most mainstream models are trained offline, making them ill-suited to adapting to rapid algal community succession triggered by abrupt changes in aquatic environments (such as torrential rain runoff and heat waves), resulting in a significant decline in online prediction performance. In complex open water scenarios such as lakes and reservoirs, these shortcomings make it difficult for existing prediction systems to support accurate and real-time ecological risk early warning and control decisions. Summary of the Invention

[0005] This invention provides an algorithm and system for predicting algal community structure changes based on multi-source data fusion, aiming to solve the technical problems of insufficient prediction accuracy, poor model adaptability, and difficulty in effectively fusing multi-source heterogeneous data caused by reliance on a single data source in existing technologies. This invention constructs a multi-source data acquisition system covering water quality, meteorological, and biological indicators; employs feature engineering to extract key environmental driving factors; introduces a spatiotemporal attention mechanism for dynamic weighted fusion of multimodal data; and combines a long short-term memory network to establish a dynamic prediction model with adaptive correction capabilities, thereby achieving high-precision, real-time prediction of algal community structure change trends in complex aquatic environments.

[0006] This invention provides an algorithm for predicting changes in algal community structure based on multi-source data fusion, which includes the following steps:

[0007] Acquire historical and real-time multi-source environmental monitoring data of the target water body, including water quality parameter sequences, meteorological observation sequences, and planktonic community structure sequences;

[0008] The multi-source environmental monitoring data are subjected to timestamp alignment and missing value imputation processing to form a multivariate time series dataset under a unified time reference.

[0009] Based on the multivariate time-series dataset, a subset of key environmental factors that have a significant impact on changes in algal community structure were selected using a feature selection algorithm.

[0010] The subset of key environmental factors is input into the spatiotemporal attention fusion module to calculate the attention weight of each environmental factor at different time steps and spatial locations, and a weighted fusion multidimensional feature vector sequence is generated accordingly.

[0011] The multidimensional feature vector sequence is input into a long short-term memory network model, which outputs the predicted relative abundance of each dominant species in the algal community within a future preset time window.

[0012] Based on the newly collected measured algal community structure data, the current prediction error is calculated, and this error is used to perform online gradient updates on the internal state parameters of the long short-term memory network model, thereby completing the dynamic correction of the model.

[0013] As one embodiment of the present invention, the water quality parameter sequence includes water temperature, dissolved oxygen concentration, total nitrogen concentration, total phosphorus concentration, chemical oxygen demand, chlorophyll a concentration, turbidity, pH value, and electrical conductivity; the meteorological observation sequence includes daily average temperature, daily maximum temperature, daily minimum temperature, precipitation, sunshine duration, wind speed, and relative humidity; the planktonic community structure sequence includes the cell density or biomass percentage of representative species in cyanobacteria, green algae, diatoms, cryptophytes, and dinoflagellates.

[0014] As one embodiment of the present invention, the timestamp alignment and missing value interpolation process specifically includes: resampling all original monitoring data according to the time axis with 10 minutes as the basic time granularity; for a certain type of parameter value missing at any time point, using cubic spline interpolation to fill it based on the adjacent valid data points before and after it; if more than 30 consecutive time points are missing, the historical average of the parameter in the corresponding season is used as the replacement.

[0015] As one embodiment of the present invention, the feature selection algorithm adopts a recursive feature elimination combined with cross-validation. The initial candidate feature set includes all water quality, meteorological and biological indicator parameters. By gradually eliminating features whose contribution to the prediction target is lower than a preset threshold, the top 15 environmental factors in terms of contribution are finally retained as a subset of key environmental factors.

[0016] As one embodiment of the present invention, the spatiotemporal attention fusion module includes a temporal attention submodule and a spatial attention submodule; the temporal attention submodule is used to calculate the importance weight of the same environmental factor at different historical time steps to the current prediction time; the spatial attention submodule is used to calculate the interaction strength weight between different environmental factors at the same time step; the output of the spatiotemporal attention fusion module is a weighted feature vector corresponding to each time step, the dimension of which is equal to the number of key environmental factor subsets.

[0017] As one embodiment of the present invention, the long short-term memory network model includes three stacked long short-term memory units, with 128 units in each layer. The activation function is the hyperbolic tangent function, and the output layer uses a linear activation function. The model input is a time step sequence of length 720, with each time step corresponding to a weighted and fused multidimensional feature vector. The model output is a relative abundance prediction sequence of each dominant algal species within the next 144 time steps.

[0018] As one embodiment of the present invention, the dynamic correction process specifically includes: when newly collected measured algal community structure data arrives, calculating the mean square error between the model prediction value and the measured value; if the mean square error is greater than a preset tolerance error threshold of 0.05, triggering the online model update mechanism; the online update mechanism adopts the stochastic gradient descent algorithm, adjusting the weight parameters of the forget gate, input gate and output gate of the long short-term memory network along the error backpropagation path with a learning rate of 0.001.

[0019] According to another aspect of the present invention, a system for predicting changes in algal community structure based on multi-source data fusion is provided, comprising:

[0020] The multi-source environmental data acquisition module is used to simultaneously acquire water quality parameter sequences, meteorological observation sequences, and plankton community structure sequences from water quality automatic monitoring stations, meteorological observation stations, and plankton automatic identification equipment deployed in the target water body;

[0021] The data preprocessing module is used to perform timestamp alignment and missing value imputation on the data acquired by the multi-source environmental data acquisition module to form a multivariate time series dataset under a unified time reference.

[0022] The key factor extraction module is used to filter out a subset of key environmental factors that have a significant impact on changes in algal community structure based on the multivariate time-series dataset using a feature selection algorithm.

[0023] The spatiotemporal attention fusion module is used to receive a subset of key environmental factors output by the key factor extraction module, calculate the attention weight of each environmental factor at different time steps and spatial locations, and generate a weighted fusion multidimensional feature vector sequence.

[0024] The dynamic prediction model module is used to receive the multidimensional feature vector sequence output by the spatiotemporal attention fusion module, and output the relative abundance prediction value of each dominant species in the algal community within a future preset time window through a long short-term memory network model.

[0025] The model adaptive correction module is used to calculate the current prediction error based on newly acquired measured algal community structure data, and to use this error to perform online gradient updates on the internal state parameters of the dynamic prediction model module.

[0026] In one embodiment of the present invention, the automatic water quality monitoring station in the multi-source environmental data acquisition module is equipped with a water temperature sensor, a dissolved oxygen sensor, a total nitrogen analyzer, a total phosphorus analyzer, a chemical oxygen demand analyzer, a chlorophyll a fluorescence probe, a turbidity meter, a pH electrode, and a conductivity electrode; the meteorological observation station is equipped with a thermometer, a rain gauge, a sunshine meter, an anemometer, and a humidity sensor; and the automatic plankton identification device is equipped with a flow cytometry imaging microscopy system and an image recognition algorithm unit for automatically classifying and counting planktonic algae in water samples.

[0027] As one embodiment of the present invention, the data preprocessing module resamples all the original monitoring data with a basic time granularity of 10 minutes, and fills the missing data with cubic spline interpolation. For parameters that are missing for more than 30 consecutive time points, the historical average of the parameter in the corresponding season is used as the substitute.

[0028] As one embodiment of the present invention, the key factor extraction module adopts a recursive feature elimination combined with five-fold cross-validation to gradually eliminate features whose contribution to the prediction target is lower than a preset threshold from the initial candidate feature set, and finally retains the top 15 environmental factors in terms of contribution.

[0029] As one embodiment of the present invention, the spatiotemporal attention fusion module includes a temporal attention submodule and a spatial attention submodule. The temporal attention submodule calculates the importance weight of the same environmental factor at different historical time steps, and the spatial attention submodule calculates the interaction strength weight between different environmental factors at the same time step, and finally outputs the weighted feature vector corresponding to each time step.

[0030] As one embodiment of the present invention, the dynamic prediction model module includes three stacked long short-term memory units, with 128 units in each layer. The model input sequence has a length of 720 time steps, and the output is a prediction sequence of the relative abundance of each dominant algal species within the next 144 time steps.

[0031] In one embodiment of the present invention, after receiving new measured data, the model adaptive correction module calculates the mean square error between the predicted value and the measured value. If the error is greater than 0.05, the stochastic gradient descent algorithm is started to update the gating weight parameters of the dynamic prediction model module online with a learning rate of 0.001.

[0032] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0033] 1. This invention solves the problem of incomplete environmental representation caused by the single data source in traditional algorithms by constructing a multi-source heterogeneous data acquisition system covering water quality, meteorological and biological indicators; it adopts a recursive feature elimination algorithm to accurately screen key environmental driving factors, effectively reducing the input dimensionality and improving the interpretability of the model.

[0034] 2. By introducing a spatiotemporal attention mechanism, dynamic weighted fusion of different environmental factors in the dimensions of temporal evolution and spatial interaction is achieved, overcoming the shortcomings of simple splicing or multilayer perceptron fusion methods in modeling complex nonlinear relationships.

[0035] 3. The constructed long short-term memory network model has the ability to remember long-term dependencies and can capture the lag effect and cumulative effect in the process of algal community succession.

[0036] 4. By using the model adaptive correction mechanism, the model parameters are updated online using newly arrived measured data, which significantly improves the model's generalization ability and prediction stability in dynamically changing water environments.

[0037] 5. In the field application of this invention in typical lakes and reservoirs, the accuracy of 7-day prediction of the relative abundance of major algal phyla is improved and the standard deviation of prediction bias is reduced, effectively supporting the decision-making of algal bloom early warning and ecological regulation. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0039] Figure 2 This is a schematic diagram of the core principle framework of the spatiotemporal attention fusion module in this invention;

[0040] Figure 3 This is a logical flowchart of the multi-source environmental data acquisition and preprocessing stage in this invention;

[0041] Figure 4 This is a flowchart illustrating the logical process of the key environmental factor extraction and feature selection stages in this invention.

[0042] Figure 5 This is a logical flowchart of the dynamic prediction model and adaptive correction mechanism in this invention;

[0043] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the multi-source data acquisition device and the various functional modules of the prediction system in this invention. Detailed Implementation

[0044] Please refer to Figures 1 to 6 This invention provides an algorithm and system for predicting algal community structure changes based on multi-source data fusion. It aims to address the technical problems in existing technologies, such as insufficient prediction accuracy, poor model adaptability, and difficulty in effectively fusing heterogeneous multi-source data due to reliance on a single data source. The algorithm constructs a multi-source data acquisition system covering water quality, meteorological, and biological indicators; employs feature engineering to extract key environmental driving factors; introduces a spatiotemporal attention mechanism for dynamic weighted fusion of multimodal data; and combines a long short-term memory network to establish a dynamic prediction model with adaptive correction capabilities. This enables high-precision, real-time prediction of algal community structure change trends in complex aquatic environments.

[0045] In one embodiment of the present invention, the acquisition of historical and real-time multi-source environmental monitoring data of the target water body includes water quality parameter sequences, meteorological observation sequences, and plankton community structure sequences. Specifically, the water quality parameter sequences are acquired through automatic water quality monitoring stations deployed in the target water body, and include water temperature, dissolved oxygen concentration, total nitrogen concentration, total phosphorus concentration, chemical oxygen demand, chlorophyll a concentration, turbidity, pH value, and conductivity; the meteorological observation sequences are acquired through meteorological observation stations in adjacent water areas, and include daily average temperature, daily maximum temperature, daily minimum temperature, precipitation, sunshine duration, wind speed, and relative humidity; the plankton community structure sequences are acquired through automatic plankton identification equipment, and include the cell density or biomass percentage of representative species in cyanobacteria, green algae, diatoms, cryptophytes, and dinoflagellates. All data are continuously collected at the original sampling frequency and transmitted in real time to the central processing unit via wired or wireless communication links.

[0046] In one embodiment of the present invention, the multi-source environmental monitoring data are subjected to timestamp alignment and missing value imputation processing to form a multivariate time-series dataset under a unified time reference. This step uses a 10-minute basic time granularity to resample all original monitoring data along the time axis. For any missing parameter value at any given time point, cubic spline interpolation is used to fill the missing value based on adjacent valid data points. If more than 30 consecutive time points are missing, the historical mean of the parameter in the corresponding season is used as the replacement. The historical mean is calculated based on statistical data from the same season over the past three years (March to May in spring, June to August in summer, September to November in autumn, and December to February of the following year in winter). The timestamp alignment process ensures that all parameters are comparable under the same time coordinate, avoiding data misalignment caused by differences in sampling frequency. After missing value imputation, all parameters are organized into a two-dimensional tensor, where the row index represents the time step, the column index represents different environmental factors, and each element represents the value of a specific factor at a specific time point.

[0047] In one embodiment of the present invention, a subset of key environmental factors that significantly influence changes in algal community structure is selected based on the multivariate time-series dataset using a feature selection algorithm. This feature selection algorithm employs a combination of recursive feature elimination and five-fold cross-validation. The initial candidate feature set includes all water quality, meteorological, and biological indicators, totaling 19 items. The recursive feature elimination process starts from the entire set, training a support vector regression model each time. Features are sorted according to their absolute weights in the model, and the feature with the smallest weight is removed. The model is then retrained using the remaining features, iterating until 15 features remain. Five-fold cross-validation is used to evaluate the model's generalization performance after each removal. If the mean squared error on the validation set increases by more than 5% after removal, the feature is retained. The final 15 retained key environmental factors constitute a subset of key environmental factors, serving as input for subsequent fusion modules. This subset may be dynamically adjusted for different water bodies or seasons, but a complete feature selection process must be completed before each prediction task begins.

[0048] In one embodiment of the present invention, the subset of key environmental factors is input into a spatiotemporal attention fusion module to calculate the attention weight of each environmental factor at different time steps and spatial locations, and thereby generate a weighted fused multidimensional feature vector sequence. The spatiotemporal attention fusion module is composed of a cascaded temporal attention submodule and a spatial attention submodule. The temporal attention submodule and the spatial attention submodule perform weighted fusion on the subset of key environmental factors to obtain a multidimensional feature vector corresponding to each time step. The temporal attention submodule receives a sequence of key factors with a length of 720 time steps and independently calculates the importance weight of each factor at each time step within the historical window to the current prediction time. This weight is generated through a learnable query-key-value mechanism, and the specific formula is as follows:

[0049]

[0050] in, This represents the attention weight of the i-th environmental factor at time step t. For learnable query vectors, Let i be the key vector of the i-th factor at time step t. The time window length is 720. After time attention weighting, each factor generates a compressed time series representation.

[0051] Subsequently, the spatial attention submodule performs interactive modeling of the representations of 15 factors at the same time step. This module calculates the interaction strength weights between any two factors, forming a 15×15 attention matrix. This matrix is ​​generated through a bilinear transformation:

[0052]

[0053] in, This represents the spatial attention weights between factor i and factor j. and Factor representation of time-attention output, For learnable weight matrix, The activation function is Sigmoid. Finally, the output feature vector at each time step is a weighted sum of the factor representations and their corresponding spatial attention weights, with a dimension of 15. This multidimensional feature vector sequence serves as the input to the Long Short-Term Memory (LSTM) network.

[0054] In one embodiment of the present invention, the multidimensional feature vector sequence is input into a Long Short-Term Memory (LSTM) network model, which outputs predicted relative abundance values ​​of dominant species in algal communities within a future preset time window. The LTM network model comprises three stacked LTM units, with 128 units per layer. The input sequence has a length of 720 time steps, with each time step corresponding to a 15-dimensional feature vector. The network's internal state is regulated through forgetting gates, input gates, and output gates. The activation function for the gating mechanism is the Sigmoid function, and the cell state update uses the hyperbolic tangent function. The output layer uses a linear activation function, directly mapping to the prediction target space. The prediction target is the relative abundance of dominant species in Cyanobacteria, Chlorophyta, Diatoms, Cryptophyta, and Dinophyta within the next 144 time steps (i.e., 24 hours), totaling 10 output variables. Each output variable ranges from 0 to 1, and the sum of abundance values ​​for all phyla at the same time step is constrained to 1. This constraint is achieved by introducing Lagrange multipliers into the loss function, ensuring that the prediction results are biologically plausible.

[0055] In one embodiment of the present invention, the current prediction error is calculated based on newly collected measured algal community structure data, and this error is used to perform online gradient updates on the internal state parameters of the Long Short-Term Memory (LSTM) network model, thus completing the dynamic correction of the model. The dynamic correction process is triggered each time new measured data arrives. The measured data is provided by an automatic plankton identification device and includes the actual relative abundance of each dominant species at the current time point. The prediction error is defined as the mean square error between the model's predicted value and the measured value at that time point. If the mean square error is greater than a preset tolerance threshold of 0.05, an online update mechanism is initiated. This mechanism uses a stochastic gradient descent algorithm to adjust the weight parameters of the forget gate, input gate, and output gate of the LTM network along the error backpropagation path with a learning rate of 0.001. The update process only performs gradient backpropagation on the hidden states of the most recent 144 steps to reduce computational overhead. After the parameters are updated, the model is immediately used for the next round of prediction, achieving closed-loop adaptive optimization. This mechanism ensures that the model can continuously track the dynamic changes in the aquatic environment and avoids performance degradation caused by long-term static training.

[0056] At the system level, the algal community structure change prediction system based on multi-source data fusion includes a multi-source environmental data acquisition module, a data preprocessing module, a key factor extraction module, a spatiotemporal attention fusion module, a dynamic prediction model module, and a model adaptive correction module. The multi-source environmental data acquisition module is equipped with an automatic water quality monitoring station, a meteorological observation station, and an automatic plankton identification device. The automatic water quality monitoring station integrates a water temperature sensor, a dissolved oxygen sensor, a total nitrogen analyzer, a total phosphorus analyzer, a chemical oxygen demand (COD) meter, a chlorophyll a fluorescence probe, a turbidimeter, a pH electrode, and a conductivity electrode; all sensors sample synchronously at 10-minute intervals. The meteorological observation station is equipped with a thermometer, a rain gauge, a sunshine meter, an anemometer, and a humidity sensor; data is also uploaded at 10-minute granularity. The automatic plankton identification device uses a flow cytometry imaging microscopy system in conjunction with a deep convolutional neural network image recognition algorithm unit to perform real-time imaging and classification of water samples flowing through the detection chamber, outputting the cell density and biomass percentage of each algal phylum. The sampling frequency is once per hour, and it is extended to a 10-minute granularity through linear interpolation to match other data sources.

[0057] The data preprocessing module receives the raw data stream and performs timestamp alignment and missing value imputation. This module incorporates a time resampling engine, forcing all data to align to a uniform 10-minute grid. The interpolation engine prioritizes cubic spline interpolation, automatically switching to a seasonal historical mean imputation strategy when more than 30 consecutive time points are detected as missing. The preprocessed data is written to a time-series database for use by subsequent modules.

[0058] The key factor extraction module periodically runs the feature selection algorithm. This module maintains a feature contribution ranking list and initiates a recursive feature elimination process after receiving complete data for a new week. Five-fold cross-validation is executed in parallel on the local computing cluster, ensuring that the feature selection process is completed within two hours. The list of 15 selected key factors is broadcast to the spatiotemporal attention fusion module as the basis for its input channel configuration.

[0059] The spatiotemporal attention fusion module is deployed on high-performance computing nodes and its implementation comprises two independent neural sub-networks: a temporal attention network and a spatial attention network. The temporal attention network processes each factor channel independently and outputs compressed temporal features; the spatial attention network receives compressed features from all channels and calculates cross-factor interaction weights. The fused feature sequences are then batch-processed and fed into the dynamic prediction model module.

[0060] The dynamic prediction model module is built on the TensorFlow framework, and model parameters are stored on a distributed parameter server. Inference requests use a sliding window approach, with each prediction covering the next 24 hours. Prediction results are published to the water environment management platform via an API for use by the decision support system.

[0061] The model adaptive calibration module monitors the measured data stream and immediately initiates error calculation and parameter update procedures upon detecting new algal community structure data. This module shares memory space with the dynamic prediction model module to ensure the atomicity of gradient update operations. Update logs are recorded in an audit database for model performance backtesting analysis.

[0062] The entire system runs on a containerized microservice architecture, with modules decoupled through message queues to ensure high availability and scalability. The end-to-end latency of data flow from the acquisition end to the prediction output end is controlled within 5 minutes, meeting the requirements for real-time alerts. The system performs a full model retraining once every morning at midnight, updating the basic parameters based on data from the past 30 days, while the online calibration mechanism is responsible for intraday fine-tuning. The two work together to ensure the long-term stability and short-term adaptability of the model.

[0063] In summary, this embodiment constructs a complete, robust, and high-precision system for predicting algal community structure changes through rigorous multi-source data acquisition specifications, a refined preprocessing workflow, a scientific feature selection mechanism, an innovative spatiotemporal attention fusion architecture, a deep long short-term memory prediction model, and a closed-loop adaptive correction strategy. This system not only solves the challenge of multi-source heterogeneous data fusion but also significantly improves the prediction accuracy and practicality in complex aquatic environments through a dynamic modeling mechanism.

Claims

1. An algorithm for predicting changes in algal community structure based on multi-source data fusion, characterized in that, include: Acquire historical and real-time multi-source environmental monitoring data of the target water body, including water quality parameter sequences, meteorological observation sequences, and planktonic community structure sequences; The multi-source environmental monitoring data are subjected to timestamp alignment and missing value imputation processing to form a multivariate time series dataset under a unified time reference. Based on the multivariate time-series dataset, a subset of key environmental factors that have a significant impact on changes in algal community structure were selected using a feature selection algorithm. The subset of key environmental factors is input into the spatiotemporal attention fusion module to calculate the attention weight of each environmental factor at different time steps and spatial locations, and a weighted fusion multidimensional feature vector sequence is generated accordingly. The multidimensional feature vector sequence is input into a long short-term memory network model, which outputs the predicted relative abundance of each dominant species in the algal community within a future preset time window. Based on the newly collected measured algal community structure data, the current prediction error is calculated, and this error is used to perform online gradient updates on the internal state parameters of the long short-term memory network model, thereby completing the dynamic correction of the model.

2. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 1, characterized in that, The multi-source environmental monitoring data are subjected to timestamp alignment and missing value imputation processing to form a multivariate time series dataset under a unified time reference, including: Using 10-minute intervals as the basic time granularity, all raw monitoring data were resampled along the time axis. For any missing parameter value at any given time point, cubic spline interpolation is used to fill the missing value based on the adjacent valid data points before and after it. If more than 30 consecutive time points are missing, the historical average of the parameter in the corresponding season will be used instead.

3. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 2, characterized in that, Based on the aforementioned multivariate time-series dataset, a subset of key environmental factors that significantly influence changes in algal community structure was selected using a feature selection algorithm, including: A recursive feature elimination method combined with five-fold cross-validation is used to gradually remove features whose contribution to the prediction target is lower than a preset threshold from the initial candidate feature set; The initial candidate feature set includes all water quality, meteorological, and biological indicator parameters; The top 15 environmental factors in terms of contribution were retained as a subset of the key environmental factors.

4. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 3, characterized in that, The spatiotemporal attention fusion module is composed of a cascaded temporal attention submodule and a spatial attention submodule. It inputs the subset of key environmental factors into the spatiotemporal attention fusion module, calculates the attention weight of each environmental factor at different time steps and spatial locations, and generates a weighted fused multidimensional feature vector sequence accordingly, including: The time attention submodule calculates the importance weight of the same environmental factor at different historical time steps to the current prediction time. The spatial attention submodule is used to calculate the interaction strength weights between different environmental factors at the same time step. Based on the aforementioned temporal attention submodule and spatial attention submodule, a weighted fusion of a subset of key environmental factors is performed to obtain a multidimensional feature vector corresponding to each time step.

5. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 4, characterized in that, The time attention submodule calculates the importance weight of the same environmental factor at different historical time steps to the current prediction time, including: For each environmental factor, a query-key-value mechanism is independently constructed to generate its attention coefficient at each time step within the historical time window; The factor values ​​at each time step are weighted and summed based on the attention coefficients to obtain the compressed time series representation.

6. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 5, characterized in that, The spatial attention submodule calculates the interaction strength weights between different environmental factors at the same time step, including: The temporal representations of each environmental factor at the current time step are input into a bilinear transformation function to generate inter-factor interaction scores. Apply the Sigmoid activation function to the interaction score to obtain the spatial attention weight matrix; The spatial attention weight matrix is ​​used to weight and aggregate the representations of each factor to generate the final multidimensional feature vector.

7. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 6, characterized in that, The multidimensional feature vector sequence is input into a long short-term memory network model, which outputs the predicted relative abundance of each dominant species in the algal community within a preset time window, including: A multidimensional feature vector sequence with a length of 720 time steps is input into a three-layer stacked long short-term memory unit, with each layer containing 128 memory units; Cell state updates are regulated by forgetting gates, input gates, and output gates, with the Sigmoid function used as the gating activation function and the hyperbolic tangent function used for cell state updates. The linear output layer is used to map the relative abundance prediction sequences of each dominant algal species over the next 144 time steps.

8. The algal community structure change prediction algorithm based on multi-source data fusion according to claim 7, characterized in that, Based on newly acquired measured algal community structure data, the current prediction error is calculated, and this error is used to perform online gradient updates on the internal state parameters of the long short-term memory network model, completing the dynamic correction of the model, including: Calculate the mean square error between the model's predicted value and the measured value at the current time point; If the mean square error is greater than the preset tolerance error threshold of 0.05, then the stochastic gradient descent algorithm is started. The weight parameters of the forget gate, input gate, and output gate of the Long Short-Term Memory Network are adjusted along the error backpropagation path.

9. A system for predicting changes in algal community structure based on multi-source data fusion, characterized in that, include: The multi-source environmental data acquisition module is used to simultaneously acquire water quality parameter sequences, meteorological observation sequences, and plankton community structure sequences from water quality automatic monitoring stations, meteorological observation stations, and plankton automatic identification equipment deployed in the target water body; The data preprocessing module is used to perform timestamp alignment and missing value imputation on the data acquired by the multi-source environmental data acquisition module to form a multivariate time series dataset under a unified time reference. The key factor extraction module is used to filter out a subset of key environmental factors that have a significant impact on changes in algal community structure based on the multivariate time-series dataset using a feature selection algorithm. The spatiotemporal attention fusion module is used to receive a subset of key environmental factors output by the key factor extraction module, calculate the attention weight of each environmental factor at different time steps and spatial locations, and generate a weighted fusion multidimensional feature vector sequence. The dynamic prediction model module is used to receive the multidimensional feature vector sequence output by the spatiotemporal attention fusion module, and output the relative abundance prediction value of each dominant species in the algal community within a future preset time window through a long short-term memory network model. The model adaptive correction module is used to calculate the current prediction error based on newly acquired measured algal community structure data, and to use this error to perform online gradient updates on the internal state parameters of the dynamic prediction model module.

10. The algal community structure change prediction system based on multi-source data fusion according to claim 9, characterized in that, The data preprocessing module is used for: All original monitoring data were resampled using a 10-minute time granularity. The missing data was filled using cubic spline interpolation. For parameters that are missing for more than 30 consecutive time points, the historical mean of the parameter in the corresponding season is used as a substitute.

Images